Daily Paper Data Validation Dashboard

Validate the accuracy of virtual experiment agents using real data from papers/benchmarks, quantify error bars, and automatically generate new experiment directions.

Generation time: 2026-11-03T03:00:54 · Data sources: Published benchmarks / Analytical solutions / Published coefficients (not self-generated data)

Honest statement:The "real data" in the current closed loop comes from published benchmarks / analytical solutions / published coefficientsSynthetic gold self-consistency verification—Correction points are generated by the same gold function and backfilled into the training set for verification.Method Validity and UQ CredibilityThis comes from self-consistent validation, not from real laboratory measurements or external users. After real test benches/users are connected, it will switch to actual measurement backfill, at which point the error bars and coverage rate will truly represent the degree of equivalence to reality. The "closed-loop cumulative point" on this page refers to the paper data points that have been backfilled by this self-consistent validation.
Surrogate Modeling/Optimization (Bayesian Optimization Benchmark) Converged
ground-truth: Forrester et al. 2008: min ≈ -6.0208 @ x≈0.7572
rmse (with accumulated data)0.0056
seed baseline rmse0.0266
95% CI coverage1.00
calibration error0.050
extrapolation honesty4.08×
closed-loop accumulated points24
Source: Forrester et al. 2008: min ≈ -6.0208 @ x≈0.7572
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.988] · Expected Information Gain 0.0345 (sigma=0.1857)
  • #2 point [0.001] · Expected Information Gain 0.0331 (sigma=0.1819)
  • #3 point [0.986] · Expected Information Gain 0.0317 (sigma=0.1779)
  • #4 point [0.003] · Expected Information Gain 0.0316 (sigma=0.1777)
  • #5 point [0.006] · Expected Information Gain 0.0301 (sigma=0.1735)
Surrogate Modeling/Optimization (2D Benchmark) Converged
ground-truth: Branin min ≈ 0.397887 @ 3 known points
rmse (with accumulated data)0.3019
seed baseline rmse0.6336
95% CI coverage1.00
calibration error0.050
extrapolation honesty20.57×
closed-loop accumulated points26
Source: Branin min ≈ 0.397887 @ 3 known points
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [5.69, 6.592] · Expected Information Gain 794.5065 (sigma=28.187)
  • #2 point [4.78, 13.912] · Expected Information Gain 761.4711 (sigma=27.5948)
  • #3 point [5.372, 14.57] · Expected Information Gain 754.4318 (sigma=27.4669)
  • #4 point [5.365, 6.335] · Expected Information Gain 753.6847 (sigma=27.4533)
  • #5 point [2.695, 3.674] · Expected Information Gain 749.7502 (sigma=27.3816)
Surrogate Modeling/Optimization (6D High-Dimensional Benchmark) Converged
ground-truth: Hartmann 6D min ≈ -3.32237
rmse (with accumulated data)0.0009
seed baseline rmse0.0044
95% CI coverage1.00
calibration error0.050
extrapolation honesty28.77×
closed-loop accumulated points45
Source: Hartmann 6D min ≈ -3.32237
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.158, 0.374, 0.415, 0.954, 0.329, 0.48] · Expected Information Gain 0.1525 (sigma=0.3905)
  • #2 point [0.675, 0.585, 0.577, 0.285, 0.887, 0.116] · Expected Information Gain 0.1519 (sigma=0.3897)
  • #3 point [0.691, 0.648, 0.589, 0.21, 0.987, 0.88] · Expected Information Gain 0.1517 (sigma=0.3895)
  • #4 point [0.665, 0.57, 0.524, 0.083, 0.091, 0.691] · Expected Information Gain 0.1512 (sigma=0.3889)
  • #5 point [0.088, 0.489, 0.675, 0.713, 0.151, 0.645] · Expected Information Gain 0.1504 (sigma=0.3879)
heat conduction (physics-informed) / 1D 热传导方程解析解 Converged
ground-truth: 1D Heat Equation Analytical Solution u=sin(πx)e^{-π²t}
rmse (with accumulated data)0.0000
seed baseline rmse0.0017
95% CI coverage1.00
calibration error0.050
extrapolation honesty1.87×
closed-loop accumulated points30
Source: 1D Heat Equation Analytical Solution u=sin(πx)e^{-π²t}
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.982] · Expected Information Gain 0.0001 (sigma=0.0078)
  • #2 point [0.981] · Expected Information Gain 0.0001 (sigma=0.0077)
  • #3 point [0.98] · Expected Information Gain 0.0001 (sigma=0.0076)
  • #4 point [0.006] · Expected Information Gain 0.0001 (sigma=0.0076)
  • #5 point [0.006] · Expected Information Gain 0.0001 (sigma=0.0076)
Biology/Population Dynamics Converged
ground-truth: Logistic Growth N(t)=K/(1+Ae^{-rt}), K=100,r=0.6,A=19
rmse (with accumulated data)0.0027
seed baseline rmse0.0583
95% CI coverage1.00
calibration error0.050
extrapolation honesty2.72×
closed-loop accumulated points15
Source: Logistic Growth N(t)=K/(1+Ae^{-rt}), K=100,r=0.6,A=19
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.059] · Expected Information Gain 1.9191 (sigma=1.3853)
  • #2 point [0.064] · Expected Information Gain 1.8967 (sigma=1.3772)
  • #3 point [0.08] · Expected Information Gain 1.8352 (sigma=1.3547)
  • #4 point [9.941] · Expected Information Gain 1.8207 (sigma=1.3493)
  • #5 point [0.085] · Expected Information Gain 1.8159 (sigma=1.3476)
Microorganisms/Growth Kinetics (Monod) Converged
ground-truth: Monod 1949: μ(S)=μmax·S/(Ks+S), E.coli μmax=0.81 h⁻¹, Ks=0.22 g/L (literature representative values)
rmse (with accumulated data)0.0001
seed baseline rmse0.0058
95% CI coverage1.00
calibration error0.050
extrapolation honesty2.02×
closed-loop accumulated points22
Source: Monod 1949: μ(S)=μmax·S/(Ks+S), E.coli μmax=0.81 h⁻¹, Ks=0.22 g/L (literature representative values)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.022] · Expected Information Gain 0.0 (sigma=0.0054)
  • #2 point [0.026] · Expected Information Gain 0.0 (sigma=0.0053)
  • #3 point [1.988] · Expected Information Gain 0.0 (sigma=0.0053)
  • #4 point [0.032] · Expected Information Gain 0.0 (sigma=0.0052)
  • #5 point [0.033] · Expected Information Gain 0.0 (sigma=0.0052)
Microorganisms/Substrate Inhibition (Andrews) Converged
ground-truth: Andrews 1968 Substrate Inhibition μ=μmax·S/(Ks+S+S²/Ki), Ki≈1.0 g/L (literature representative values)
rmse (with accumulated data)0.0003
seed baseline rmse0.0021
95% CI coverage1.00
calibration error0.050
extrapolation honesty2.44×
closed-loop accumulated points28
Source: Andrews 1968 Substrate Inhibition μ=μmax·S/(Ks+S+S²/Ki), Ki≈1.0 g/L (literature representative values)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.053] · Expected Information Gain 0.0 (sigma=0.0045)
  • #2 point [2.983] · Expected Information Gain 0.0 (sigma=0.0044)
  • #3 point [0.059] · Expected Information Gain 0.0 (sigma=0.0042)
  • #4 point [0.067] · Expected Information Gain 0.0 (sigma=0.0039)
  • #5 point [0.069] · Expected Information Gain 0.0 (sigma=0.0038)
LLM/Scaling Law (Kaplan 2020) Converged
ground-truth: Scaling Law L(N)=(N_c/N)^α, α=0.076, N_c=6.4e13 (Kaplan 2020 Published Coefficient, nats)
rmse (with accumulated data)0.0000
seed baseline rmse0.0091
95% CI coverage1.00
calibration error0.050
extrapolation honesty1.05×
closed-loop accumulated points21
Source: Scaling Law L(N)=(N_c/N)^α, α=0.076, N_c=6.4e13 (Kaplan 2020 Published Coefficient, nats)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [10.976] · Expected Information Gain 0.0003 (sigma=0.0163)
  • #2 point [10.953] · Expected Information Gain 0.0003 (sigma=0.0162)
  • #3 point [10.943] · Expected Information Gain 0.0003 (sigma=0.0161)
  • #4 point [7.089] · Expected Information Gain 0.0003 (sigma=0.0161)
  • #5 point [7.111] · Expected Information Gain 0.0003 (sigma=0.016)
Chemical/Adsorption Isotherm (Langmuir 1916) Converged
ground-truth: Langmuir 1916 Adsorption Isotherm θ=KP/(1+KP); K Takes Representative Value 1.5 (Adsorption Isotherm Literature Range 0.1–10)
rmse (with accumulated data)0.0003
seed baseline rmse0.0084
95% CI coverage1.00
calibration error0.050
extrapolation honesty3.48×
closed-loop accumulated points25
Source: Langmuir 1916 Adsorption Isotherm θ=KP/(1+KP); K Takes Representative Value 1.5 (Adsorption Isotherm Literature Range 0.1–10)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [9.885] · Expected Information Gain 0.0 (sigma=0.006)
  • #2 point [9.859] · Expected Information Gain 0.0 (sigma=0.0058)
  • #3 point [0.11] · Expected Information Gain 0.0 (sigma=0.0057)
  • #4 point [0.131] · Expected Information Gain 0.0 (sigma=0.0056)
  • #5 point [9.818] · Expected Information Gain 0.0 (sigma=0.0056)
Chemistry / First-Order Reactor Conversion (Arrhenius Kinetics, 2D) Converged
ground-truth: First-Order CSTR/PFR Conversion X=1-exp(-k0·exp(-Ea/RT)·τ); Ea takes representative 30 kJ/mol, k0 is normalized to ensure X(410K,τ=5)≈0.9 (Homogeneous Reaction Kinetics Literature Range; Levenspiel 1999 Reactor Design)
rmse (with accumulated data)0.0006
seed baseline rmse0.0011
95% CI coverage0.97
calibration error0.017
extrapolation honesty22.30×
closed-loop accumulated points33
Source: First-Order CSTR/PFR Conversion X=1-exp(-k0·exp(-Ea/RT)·τ); Ea takes representative 30 kJ/mol, k0 is normalized to ensure X(410K,τ=5)≈0.9 (Homogeneous Reaction Kinetics Literature Range; Levenspiel 1999 Reactor Design)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [390.83, 4.936] · Expected Information Gain 0.0021 (sigma=0.0459)
  • #2 point [405.404, 3.488] · Expected Information Gain 0.0018 (sigma=0.0423)
  • #3 point [391.462, 4.881] · Expected Information Gain 0.0017 (sigma=0.0418)
  • #4 point [340.595, 1.692] · Expected Information Gain 0.0017 (sigma=0.041)
  • #5 point [389.885, 2.478] · Expected Information Gain 0.0017 (sigma=0.0408)
微生物 / E. coli 批次培养(Monod, 文献参数) Needs data/calibration
ground-truth: E. coli K-12: μmax=0.81 h⁻¹, Ks=0.004 g/L (Monod 1949; Shuler & Kargi 2002)
rmse (with accumulated data)0.0613
seed baseline rmse0.0613
95% CI coverage0.90
calibration error0.050
extrapolation honesty1.63×
closed-loop accumulated points0
Source: E. coli K-12: μmax=0.81 h⁻¹, Ks=0.004 g/L (Monod 1949; Shuler & Kargi 2002)
Next-experiment direction suggestions
  • Supplement points: Densify real paper data points in regions with maximum error (high curvature / inflection points); or reduce lengthscale to improve fitting.
Next-experiment tournament (ranked by information gain)
  • #1 point [9.883] · Expected Information Gain 0.0 (sigma=0.0023)
  • #2 point [9.858] · Expected Information Gain 0.0 (sigma=0.0023)
  • #3 point [9.816] · Expected Information Gain 0.0 (sigma=0.0022)
  • #4 point [9.812] · Expected Information Gain 0.0 (sigma=0.0022)
  • #5 point [9.796] · Expected Information Gain 0.0 (sigma=0.0022)
微生物 / 酿酒酵母乙醇发酵(Monod + 乙醇产物抑制) Needs data/calibration
ground-truth: S. cerevisiae: μmax=0.42 h⁻¹, Ks=0.025 g/L, Ki(ethanol)=40 g/L (Dussaut & Cooney 1980)
rmse (with accumulated data)0.0431
seed baseline rmse0.0431
95% CI coverage0.93
calibration error0.017
extrapolation honesty15.58×
closed-loop accumulated points0
Source: S. cerevisiae: μmax=0.42 h⁻¹, Ks=0.025 g/L, Ki(ethanol)=40 g/L (Dussaut & Cooney 1980)
Next-experiment direction suggestions
  • Severe underfitting and 2-dimensional: Empirical results show the bottleneck is **fixed isotropic lengthscale**, not data volume—In ablation experiments, fixed ls with 42 points still has 54% relative error, whereas enabling automatic ARD hyperparameters reduces it to 4.75% at 41 points. Current 50 points have not reached the automatic hyperparameter threshold (requires ≥14 points; small samples cause marginal likelihood to hit boundaries, worsening performance).Path: First add points via tournament to reach 14, then set auto_ls=True for this scenario.
Next-experiment tournament (ranked by information gain)
  • #1 point [14.523, 49.289] · Expected Information Gain 0.0002 (sigma=0.0136)
  • #2 point [18.687, 33.197] · Expected Information Gain 0.0002 (sigma=0.0132)
  • #3 point [0.771, 15.765] · Expected Information Gain 0.0002 (sigma=0.0128)
  • #4 point [9.873, 10.848] · Expected Information Gain 0.0002 (sigma=0.0125)
  • #5 point [14.704, 48.673] · Expected Information Gain 0.0002 (sigma=0.0124)
微生物 / 假单胞菌甲苯降解(Andrews 底物抑制) Needs data/calibration
ground-truth: P. putida MT-2: μmax=0.35 h⁻¹, Ks=0.02 g/L, Ki(toluene)=2.5 g/L (Rothman et al. 1993)
rmse (with accumulated data)0.0903
seed baseline rmse0.0903
95% CI coverage0.70
calibration error0.250
extrapolation honesty2.34×
closed-loop accumulated points0
Source: P. putida MT-2: μmax=0.35 h⁻¹, Ks=0.02 g/L, Ki(toluene)=2.5 g/L (Rothman et al. 1993)
Next-experiment direction suggestions
  • Supplement points: Densify real paper data points in regions with maximum error (high curvature / inflection points); or reduce lengthscale to improve fitting.
  • Coverage 70% is slightly below the 95%CI nominal value: error bars are too narrow but not out of control, Recommend adding points to reduce the epistemic component; temporarily do not relax the noise prior (relaxing would mask true bias).
  • Calibration distortion: check noise floor vs real observation noise; recommend adding a model bias term(Kennedy–O’Hagan Missing Term) Correct Systematic Bias.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.006] · Expected Information Gain 0.0 (sigma=0.003)
  • #2 point [4.898] · Expected Information Gain 0.0 (sigma=0.0029)
  • #3 point [0.017] · Expected Information Gain 0.0 (sigma=0.0029)
  • #4 point [4.888] · Expected Information Gain 0.0 (sigma=0.0029)
  • #5 point [0.03] · Expected Information Gain 0.0 (sigma=0.0029)
微生物 / 乳酸乳球菌乳酸发酵(pH 效应 + 底物抑制) Needs data/calibration
ground-truth: L. lactis NZ9000: μmax=0.55 h⁻¹, Ks=0.3 g/L, Ki(lactose)=80 g/L (Luedtke & Schlegel 1973)
rmse (with accumulated data)0.0466
seed baseline rmse0.0466
95% CI coverage0.97
calibration error0.017
extrapolation honesty16.96×
closed-loop accumulated points0
Source: L. lactis NZ9000: μmax=0.55 h⁻¹, Ks=0.3 g/L, Ki(lactose)=80 g/L (Luedtke & Schlegel 1973)
Next-experiment direction suggestions
  • Severe underfitting and 2-dimensional: Empirical results show the bottleneck is **fixed isotropic lengthscale**, not data volume—In ablation experiments, fixed ls with 42 points still has 54% relative error, whereas enabling automatic ARD hyperparameters reduces it to 4.75% at 41 points. Current 50 points have not reached the automatic hyperparameter threshold (requires ≥14 points; small samples cause marginal likelihood to hit boundaries, worsening performance).Path: First add points via tournament to reach 14, then set auto_ls=True for this scenario.
Next-experiment tournament (ranked by information gain)
  • #1 point [7.264, 8.443] · Expected Information Gain 0.0001 (sigma=0.0113)
  • #2 point [9.344, 7.156] · Expected Information Gain 0.0001 (sigma=0.0113)
  • #3 point [0.394, 5.761] · Expected Information Gain 0.0001 (sigma=0.0112)
  • #4 point [6.913, 6.189] · Expected Information Gain 0.0001 (sigma=0.011)
  • #5 point [2.77, 6.715] · Expected Information Gain 0.0001 (sigma=0.011)
微生物 / 醋酸杆菌醋酸利用(温度效应 + Monod) Needs data/calibration
ground-truth: A. calcoaceticus: μmax=0.78 h⁻¹, Ks=0.03 g/L, T_opt=37°C (Rogness et al. 1961)
rmse (with accumulated data)0.0605
seed baseline rmse0.0605
95% CI coverage0.97
calibration error0.017
extrapolation honesty15.20×
closed-loop accumulated points0
Source: A. calcoaceticus: μmax=0.78 h⁻¹, Ks=0.03 g/L, T_opt=37°C (Rogness et al. 1961)
Next-experiment direction suggestions
  • Supplement points: Densify real paper data points in regions with maximum error (high curvature / inflection points); or reduce lengthscale to improve fitting.
Next-experiment tournament (ranked by information gain)
  • #1 point [4.672, 39.918] · Expected Information Gain 0.0022 (sigma=0.0467)
  • #2 point [3.631, 49.573] · Expected Information Gain 0.0022 (sigma=0.0466)
  • #3 point [0.043, 27.949] · Expected Information Gain 0.0019 (sigma=0.0438)
  • #4 point [1.844, 22.981] · Expected Information Gain 0.0019 (sigma=0.0438)
  • #5 point [3.676, 49.204] · Expected Information Gain 0.0018 (sigma=0.0425)
微生物 / 产甲烷古菌甲烷利用(低 μmax 极端场景) Needs data/calibration
ground-truth: M. trichosporium OB3b: μmax=0.08 h⁻¹, Ks=0.02 g/L (Whitmanet al. 1995)
rmse (with accumulated data)0.0805
seed baseline rmse0.0805
95% CI coverage0.30
calibration error0.650
extrapolation honesty5.39×
closed-loop accumulated points0
Source: M. trichosporium OB3b: μmax=0.08 h⁻¹, Ks=0.02 g/L (Whitmanet al. 1995)
Next-experiment direction suggestions
  • Supplement points: Densify real paper data points in regions with maximum error (high curvature / inflection points); or reduce lengthscale to improve fitting.
  • Overconfidence in UQ: increase the observed noise prior / add an aleatoric noise model(微生物『偶然性』即不可约不确定性,需分离 epistemic/aleatoric)。
  • Calibration distortion: check noise floor vs real observation noise; recommend adding a model bias term(Kennedy–O’Hagan Missing Term) Correct Systematic Bias.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.096] · Expected Information Gain 0.0 (sigma=0.0031)
  • #2 point [0.099] · Expected Information Gain 0.0 (sigma=0.0031)
  • #3 point [0.095] · Expected Information Gain 0.0 (sigma=0.003)
  • #4 point [0.095] · Expected Information Gain 0.0 (sigma=0.0029)
  • #5 point [0.095] · Expected Information Gain 0.0 (sigma=0.0028)
微生物 / E. coli Chemostat 稳态(稀释率→生物量) Needs data/calibration
ground-truth: Chemostat E. coli K-12, S_f=10 g/L glucose (Rogness et al. 1961)
rmse (with accumulated data)0.8507
seed baseline rmse0.5895
95% CI coverage0.85
calibration error0.100
extrapolation honesty3.58×
closed-loop accumulated points0
Source: Chemostat E. coli K-12, S_f=10 g/L glucose (Rogness et al. 1961)
Next-experiment direction suggestions
  • Supplement points: Densify real paper data points in regions with maximum error (high curvature / inflection points); or reduce lengthscale to improve fitting.
  • Calibration Error 0.100 Within Critical Band (Decision Line 0.08): Nominal Confidence Level vs. Actual Coverage Has OccurredMeasurable Deviation: Recommend Recording Epistemic/Aleatoric Components in the Next Round to Identify the Source of Deviation.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.741] · Expected Information Gain 0.0005 (sigma=0.0214)
  • #2 point [0.739] · Expected Information Gain 0.0004 (sigma=0.0208)
  • #3 point [0.021] · Expected Information Gain 0.0004 (sigma=0.0203)
  • #4 point [0.021] · Expected Information Gain 0.0004 (sigma=0.0202)
  • #5 point [0.021] · Expected Information Gain 0.0004 (sigma=0.0201)
微生物 / 酿酒酵母 Chemostat 稳态(稀释率+进料浓度→生物量) Needs data/calibration
ground-truth: Chemostat S. cerevisiae, S_f=20 g/L glucose (Dussaut & Cooney 1980)
rmse (with accumulated data)0.3502
seed baseline rmse0.3502
95% CI coverage0.87
calibration error0.083
extrapolation honesty21.50×
closed-loop accumulated points0
Source: Chemostat S. cerevisiae, S_f=20 g/L glucose (Dussaut & Cooney 1980)
Next-experiment direction suggestions
  • Supplement points: Densify real paper data points in regions with maximum error (high curvature / inflection points); or reduce lengthscale to improve fitting.
  • Calibration Error 0.083 Within Critical Band (Decision Line 0.08): Nominal Confidence Level vs. Actual Coverage Has OccurredMeasurable Deviation: Recommend Recording Epistemic/Aleatoric Components in the Next Round to Identify the Source of Deviation.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.232, 28.187] · Expected Information Gain 0.1929 (sigma=0.4392)
  • #2 point [0.18, 23.97] · Expected Information Gain 0.1919 (sigma=0.4381)
  • #3 point [0.194, 11.804] · Expected Information Gain 0.1919 (sigma=0.438)
  • #4 point [0.328, 21.599] · Expected Information Gain 0.1897 (sigma=0.4356)
  • #5 point [0.184, 11.124] · Expected Information Gain 0.1896 (sigma=0.4354)
微生物 / micro_lag_ecoli Converged
ground-truth: Baranyi & Roberts 1994 IJF 10:300 (lag phase)
rmse (with accumulated data)0.0033
seed baseline rmse1.0510
95% CI coverage0.95
calibration error0.000
extrapolation honesty1.08×
closed-loop accumulated points0
Source: Baranyi & Roberts 1994 IJF 10:300 (lag phase)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [19.881] · Expected Information Gain 0.0374 (sigma=0.1934)
  • #2 point [19.767] · Expected Information Gain 0.037 (sigma=0.1924)
  • #3 point [19.716] · Expected Information Gain 0.0369 (sigma=0.192)
  • #4 point [19.632] · Expected Information Gain 0.0366 (sigma=0.1914)
  • #5 point [19.624] · Expected Information Gain 0.0366 (sigma=0.1913)
微生物 / micro_lifecycle_ecoli Converged
ground-truth: Baranyi 1993 (full lifecycle)
rmse (with accumulated data)0.0000
seed baseline rmse0.0269
95% CI coverage1.00
calibration error0.050
extrapolation honesty0.03×
closed-loop accumulated points0
Source: Baranyi 1993 (full lifecycle)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.049] · Expected Information Gain 0.0368 (sigma=0.1919)
  • #2 point [0.153] · Expected Information Gain 0.0367 (sigma=0.1917)
  • #3 point [0.283] · Expected Information Gain 0.0366 (sigma=0.1914)
  • #4 point [0.309] · Expected Information Gain 0.0366 (sigma=0.1913)
  • #5 point [0.383] · Expected Information Gain 0.0365 (sigma=0.1912)
微生物 / micro_diauxic_ecoli Converged
ground-truth: Monod 1947 (diauxie)
rmse (with accumulated data)0.0002
seed baseline rmse0.0023
95% CI coverage1.00
calibration error0.050
extrapolation honesty1.05×
closed-loop accumulated points0
Source: Monod 1947 (diauxie)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [23.858] · Expected Information Gain 0.0 (sigma=0.0002)
  • #2 point [23.72] · Expected Information Gain 0.0 (sigma=0.0002)
  • #3 point [0.534] · Expected Information Gain 0.0 (sigma=0.0002)
  • #4 point [0.665] · Expected Information Gain 0.0 (sigma=0.0002)
  • #5 point [0.679] · Expected Information Gain 0.0 (sigma=0.0002)
微生物 / micro_fedbatch_ecoli Converged
ground-truth: Shuler & Kargi 2002 (fed-batch)
rmse (with accumulated data)0.1608
seed baseline rmse2.5659
95% CI coverage0.95
calibration error0.000
extrapolation honesty1.19×
closed-loop accumulated points0
Source: Shuler & Kargi 2002 (fed-batch)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.105] · Expected Information Gain 1.563 (sigma=1.2502)
  • #2 point [0.116] · Expected Information Gain 1.5508 (sigma=1.2453)
  • #3 point [4.91] · Expected Information Gain 1.541 (sigma=1.2414)
  • #4 point [4.908] · Expected Information Gain 1.5387 (sigma=1.2405)
  • #5 point [0.129] · Expected Information Gain 1.536 (sigma=1.2393)
微生物 / micro_competition_ecoli_yeast Converging
ground-truth: Tilman 1982 Resource Competition (multi-species)
rmse (with accumulated data)0.0005
seed baseline rmse0.0075
95% CI coverage0.95
calibration error0.000
extrapolation honesty1.03×
closed-loop accumulated points0
Source: Tilman 1982 Resource Competition (multi-species)
Next-experiment direction suggestions
  • Critical convergence: Relative error 6.3%, still 1.3 percentage points away from the 5% convergence line. Recommend supplementing only one real value at the tournament top-1 candidate points (locations with maximum posterior variance) and retest—This is the minimal cost cross-line path, avoiding overfitting that causes false confidence.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.52] · Expected Information Gain 0.0 (sigma=0.0003)
  • #2 point [0.562] · Expected Information Gain 0.0 (sigma=0.0003)
  • #3 point [0.615] · Expected Information Gain 0.0 (sigma=0.0003)
  • #4 point [0.625] · Expected Information Gain 0.0 (sigma=0.0003)
  • #5 point [0.656] · Expected Information Gain 0.0 (sigma=0.0003)
微生物 / micro_crabtree_cerevisiae Converged
ground-truth: Crabtree 1929 JPB 53:394 (Crabtree effect)
rmse (with accumulated data)0.0036
seed baseline rmse1.3849
95% CI coverage1.00
calibration error0.050
extrapolation honesty1.02×
closed-loop accumulated points0
Source: Crabtree 1929 JPB 53:394 (Crabtree effect)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [35.787] · Expected Information Gain 0.0112 (sigma=0.1058)
  • #2 point [0.231] · Expected Information Gain 0.0112 (sigma=0.1058)
  • #3 point [0.287] · Expected Information Gain 0.0112 (sigma=0.1058)
  • #4 point [0.306] · Expected Information Gain 0.0112 (sigma=0.1057)
  • #5 point [0.521] · Expected Information Gain 0.0111 (sigma=0.1056)
微生物 / micro_haldane_putida Converged
ground-truth: Haldane 1956 Biochemistry of Industrial Fermentation (Haldane model)
rmse (with accumulated data)0.0001
seed baseline rmse0.0047
95% CI coverage1.00
calibration error0.050
extrapolation honesty1.59×
closed-loop accumulated points0
Source: Haldane 1956 Biochemistry of Industrial Fermentation (Haldane model)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [4.97] · Expected Information Gain 0.0 (sigma=0.0021)
  • #2 point [0.073] · Expected Information Gain 0.0 (sigma=0.0021)
  • #3 point [0.075] · Expected Information Gain 0.0 (sigma=0.0021)
  • #4 point [0.077] · Expected Information Gain 0.0 (sigma=0.0021)
  • #5 point [4.942] · Expected Information Gain 0.0 (sigma=0.0021)
微生物 / micro_contois_ecoli Converged
ground-truth: Contois 1959 Biotech Bioeng 2:264 (Contois model)
rmse (with accumulated data)0.0001
seed baseline rmse0.0199
95% CI coverage1.00
calibration error0.050
extrapolation honesty1.19×
closed-loop accumulated points0
Source: Contois 1959 Biotech Bioeng 2:264 (Contois model)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.03] · Expected Information Gain 0.0 (sigma=0.0034)
  • #2 point [0.074] · Expected Information Gain 0.0 (sigma=0.0034)
  • #3 point [19.882] · Expected Information Gain 0.0 (sigma=0.0034)
  • #4 point [0.128] · Expected Information Gain 0.0 (sigma=0.0034)
  • #5 point [0.139] · Expected Information Gain 0.0 (sigma=0.0034)
微生物 / micro_tessier_ecoli Converged
ground-truth: Tessier 1956 Arch Mikrobiol 25:102 (Tessier model)
rmse (with accumulated data)0.0001
seed baseline rmse0.0004
95% CI coverage0.90
calibration error0.050
extrapolation honesty6.00×
closed-loop accumulated points0
Source: Tessier 1956 Arch Mikrobiol 25:102 (Tessier model)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.491] · Expected Information Gain 0.0001 (sigma=0.0071)
  • #2 point [0.49] · Expected Information Gain 0.0 (sigma=0.0067)
  • #3 point [0.489] · Expected Information Gain 0.0 (sigma=0.0062)
  • #4 point [0.027] · Expected Information Gain 0.0 (sigma=0.0061)
  • #5 point [0.001] · Expected Information Gain 0.0 (sigma=0.006)
微生物 / micro_pirt_ecoli Converged
ground-truth: Pirt 1965 Newer Studies in Microbiology (Pirt maintenance)
rmse (with accumulated data)0.0000
seed baseline rmse0.0220
95% CI coverage0.95
calibration error0.000
extrapolation honesty1.50×
closed-loop accumulated points0
Source: Pirt 1965 Newer Studies in Microbiology (Pirt maintenance)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.011] · Expected Information Gain 0.0 (sigma=0.0008)
  • #2 point [9.941] · Expected Information Gain 0.0 (sigma=0.0008)
  • #3 point [0.033] · Expected Information Gain 0.0 (sigma=0.0008)
  • #4 point [0.06] · Expected Information Gain 0.0 (sigma=0.0008)
  • #5 point [0.065] · Expected Information Gain 0.0 (sigma=0.0008)
微生物 / micro_oxygen_ecoli Converged
ground-truth: Shuler & Kargi 2002 (oxygen limitation)
rmse (with accumulated data)0.0000
seed baseline rmse0.0255
95% CI coverage1.00
calibration error0.050
extrapolation honesty1.19×
closed-loop accumulated points0
Source: Shuler & Kargi 2002 (oxygen limitation)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [7.907] · Expected Information Gain 0.0 (sigma=0.0029)
  • #2 point [0.078] · Expected Information Gain 0.0 (sigma=0.0029)
  • #3 point [7.886] · Expected Information Gain 0.0 (sigma=0.0029)
  • #4 point [0.126] · Expected Information Gain 0.0 (sigma=0.0029)
  • #5 point [7.853] · Expected Information Gain 0.0 (sigma=0.0029)
微生物 / micro_antibiotic_timekill Converged
ground-truth: Andrews 2001 (time-kill kinetics)
rmse (with accumulated data)0.0000
seed baseline rmse0.0011
95% CI coverage1.00
calibration error0.050
extrapolation honesty0.97×
closed-loop accumulated points0
Source: Andrews 2001 (time-kill kinetics)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [198.82] · Expected Information Gain 0.0 (sigma=0.0001)
  • #2 point [3.88] · Expected Information Gain 0.0 (sigma=0.0001)
  • #3 point [3.964] · Expected Information Gain 0.0 (sigma=0.0001)
  • #4 point [4.043] · Expected Information Gain 0.0 (sigma=0.0001)
  • #5 point [197.68] · Expected Information Gain 0.0 (sigma=0.0001)
微生物 / micro_osmotic_ecoli Converged
ground-truth: Rose 2008 Bacterial Osmotic Stress
rmse (with accumulated data)0.0001
seed baseline rmse0.0834
95% CI coverage0.95
calibration error0.000
extrapolation honesty0.99×
closed-loop accumulated points0
Source: Rose 2008 Bacterial Osmotic Stress
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.918] · Expected Information Gain 0.0135 (sigma=0.1164)
  • #2 point [0.902] · Expected Information Gain 0.0134 (sigma=0.1159)
  • #3 point [0.904] · Expected Information Gain 0.0131 (sigma=0.1143)
  • #4 point [0.975] · Expected Information Gain 0.013 (sigma=0.114)
  • #5 point [0.974] · Expected Information Gain 0.0127 (sigma=0.1128)
微生物 / micro_biofilm_ecoli Converged
ground-truth: Costerton et al. 1995 (biofilm)
rmse (with accumulated data)0.0000
seed baseline rmse0.0035
95% CI coverage1.00
calibration error0.050
extrapolation honesty0.03×
closed-loop accumulated points0
Source: Costerton et al. 1995 (biofilm)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [46.926] · Expected Information Gain 0.0005 (sigma=0.0228)
  • #2 point [0.695] · Expected Information Gain 0.0005 (sigma=0.0228)
  • #3 point [0.715] · Expected Information Gain 0.0005 (sigma=0.0228)
  • #4 point [0.734] · Expected Information Gain 0.0005 (sigma=0.0228)
  • #5 point [46.68] · Expected Information Gain 0.0005 (sigma=0.0228)
微生物 / micro_scaleup_kla Converged
ground-truth: Shuler & Kargi 2002 (scale-up)
rmse (with accumulated data)0.0000
seed baseline rmse0.0099
95% CI coverage1.00
calibration error0.050
extrapolation honesty1.01×
closed-loop accumulated points0
Source: Shuler & Kargi 2002 (scale-up)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [14.573] · Expected Information Gain 0.0 (sigma=0.0004)
  • #2 point [14.993] · Expected Information Gain 0.0 (sigma=0.0004)
  • #3 point [15.392] · Expected Information Gain 0.0 (sigma=0.0004)
  • #4 point [22.333] · Expected Information Gain 0.0 (sigma=0.0004)
  • #5 point [979.613] · Expected Information Gain 0.0 (sigma=0.0004)
微生物 / micro_doe_ecoli_monod Needs data/calibration
ground-truth: Montgomery 2012 DOE
rmse (with accumulated data)0.0007
seed baseline rmse0.0038
95% CI coverage1.00
calibration error0.050
extrapolation honesty1.59×
closed-loop accumulated points0
Source: Montgomery 2012 DOE
Next-experiment direction suggestions
  • Supplement points: Densify real paper data points in regions with maximum error (high curvature / inflection points); or reduce lengthscale to improve fitting.
Next-experiment tournament (ranked by information gain)
  • #1 point [2.763, 36.073] · Expected Information Gain 0.0 (sigma=0.0001)
  • #2 point [3.687, 26.988] · Expected Information Gain 0.0 (sigma=0.0001)
  • #3 point [7.262, 44.716] · Expected Information Gain 0.0 (sigma=0.0001)
  • #4 point [7.127, 33.789] · Expected Information Gain 0.0 (sigma=0.0001)
  • #5 point [5.414, 30.443] · Expected Information Gain 0.0 (sigma=0.0001)
微生物 / micro_mle_ecoli_monod Converged
ground-truth: Vogel 2004 (Bayesian estimation)
rmse (with accumulated data)0.0000
seed baseline rmse0.0168
95% CI coverage0.95
calibration error0.000
extrapolation honesty1.54×
closed-loop accumulated points0
Source: Vogel 2004 (Bayesian estimation)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.006] · Expected Information Gain 0.0 (sigma=0.0002)
  • #2 point [0.017] · Expected Information Gain 0.0 (sigma=0.0002)
  • #3 point [4.97] · Expected Information Gain 0.0 (sigma=0.0002)
  • #4 point [0.03] · Expected Information Gain 0.0 (sigma=0.0002)
  • #5 point [0.033] · Expected Information Gain 0.0 (sigma=0.0002)
微生物 / micro_sobol_ecoli Converged
ground-truth: Sobol 2001 (global sensitivity)
rmse (with accumulated data)0.0000
seed baseline rmse0.0120
95% CI coverage0.95
calibration error0.000
extrapolation honesty1.62×
closed-loop accumulated points0
Source: Sobol 2001 (global sensitivity)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [4.908] · Expected Information Gain 0.0 (sigma=0.0005)
  • #2 point [4.906] · Expected Information Gain 0.0 (sigma=0.0005)
  • #3 point [4.898] · Expected Information Gain 0.0 (sigma=0.0005)
  • #4 point [0.073] · Expected Information Gain 0.0 (sigma=0.0005)
  • #5 point [0.075] · Expected Information Gain 0.0 (sigma=0.0005)
微生物 / micro_uq_ecoli_monod Converged
ground-truth: Svensson 1999 (UQ in bioprocess)
rmse (with accumulated data)0.0000
seed baseline rmse0.0089
95% CI coverage0.95
calibration error0.000
extrapolation honesty1.62×
closed-loop accumulated points0
Source: Svensson 1999 (UQ in bioprocess)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [4.898] · Expected Information Gain 0.0 (sigma=0.0004)
  • #2 point [4.888] · Expected Information Gain 0.0 (sigma=0.0004)
  • #3 point [0.112] · Expected Information Gain 0.0 (sigma=0.0004)
  • #4 point [4.863] · Expected Information Gain 0.0 (sigma=0.0004)
  • #5 point [0.14] · Expected Information Gain 0.0 (sigma=0.0004)
微生物 / micro_coculture_mutualism Converged
ground-truth: Grosu et al. 2014 (co-culture mutualism)
rmse (with accumulated data)0.0122
seed baseline rmse2.1548
95% CI coverage1.00
calibration error0.050
extrapolation honesty1.03×
closed-loop accumulated points0
Source: Grosu et al. 2014 (co-culture mutualism)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.049] · Expected Information Gain 0.0167 (sigma=0.129)
  • #2 point [0.153] · Expected Information Gain 0.0166 (sigma=0.129)
  • #3 point [0.283] · Expected Information Gain 0.0166 (sigma=0.1289)
  • #4 point [0.309] · Expected Information Gain 0.0166 (sigma=0.1288)
  • #5 point [0.383] · Expected Information Gain 0.0166 (sigma=0.1288)
微生物 / micro_quorum_ecoli Converged
ground-truth: Basler & Bassler 2011 (quorum sensing)
rmse (with accumulated data)0.0042
seed baseline rmse1.8574
95% CI coverage1.00
calibration error0.050
extrapolation honesty1.02×
closed-loop accumulated points0
Source: Basler & Bassler 2011 (quorum sensing)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.283] · Expected Information Gain 0.017 (sigma=0.1303)
  • #2 point [0.309] · Expected Information Gain 0.017 (sigma=0.1303)
  • #3 point [0.383] · Expected Information Gain 0.017 (sigma=0.1302)
  • #4 point [0.408] · Expected Information Gain 0.017 (sigma=0.1302)
  • #5 point [47.715] · Expected Information Gain 0.0169 (sigma=0.1301)
微生物 / micro_heavy_metal_ecoli Converged
ground-truth: Kumar et al. 2012 (heavy metal stress)
rmse (with accumulated data)0.0000
seed baseline rmse0.0001
95% CI coverage1.00
calibration error0.050
extrapolation honesty0.75×
closed-loop accumulated points0
Source: Kumar et al. 2012 (heavy metal stress)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.059] · Expected Information Gain 0.0 (sigma=0.0004)
  • #2 point [0.064] · Expected Information Gain 0.0 (sigma=0.0004)
  • #3 point [0.08] · Expected Information Gain 0.0 (sigma=0.0004)
  • #4 point [0.085] · Expected Information Gain 0.0 (sigma=0.0004)
  • #5 point [0.145] · Expected Information Gain 0.0 (sigma=0.0004)
微生物 / micro_diauxic_2d Converged
ground-truth: Monod 1947 (diauxie 2D)
rmse (with accumulated data)0.0000
seed baseline rmse0.0000
95% CI coverage0.97
calibration error0.017
extrapolation honesty3.77×
closed-loop accumulated points0
Source: Monod 1947 (diauxie 2D)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [6.694, 9.311] · Expected Information Gain 0.0 (sigma=0.0)
  • #2 point [7.27, 4.675] · Expected Information Gain 0.0 (sigma=0.0)
  • #3 point [5.643, 3.085] · Expected Information Gain 0.0 (sigma=0.0)
  • #4 point [4.002, 1.444] · Expected Information Gain 0.0 (sigma=0.0)
  • #5 point [7.065, 4.512] · Expected Information Gain 0.0 (sigma=0.0)
微生物 / micro_fedbatch_2d Converged
ground-truth: Shuler & Kargi 2002 (fed-batch 2D)
rmse (with accumulated data)0.0000
seed baseline rmse0.0017
95% CI coverage1.00
calibration error0.050
extrapolation honesty0.40×
closed-loop accumulated points0
Source: Shuler & Kargi 2002 (fed-batch 2D)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.141, 45.12] · Expected Information Gain 0.0155 (sigma=0.1245)
  • #2 point [0.148, 98.649] · Expected Information Gain 0.0154 (sigma=0.1242)
  • #3 point [0.11, 28.537] · Expected Information Gain 0.0147 (sigma=0.1211)
  • #4 point [0.105, 77.087] · Expected Information Gain 0.0142 (sigma=0.119)
  • #5 point [0.104, 25.611] · Expected Information Gain 0.0142 (sigma=0.119)
微生物 / micro_competition_2d Converged
ground-truth: Tilman 1982 (smooth competition)
rmse (with accumulated data)0.0002
seed baseline rmse0.0082
95% CI coverage0.95
calibration error0.000
extrapolation honesty1.03×
closed-loop accumulated points0
Source: Tilman 1982 (smooth competition)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [1.019] · Expected Information Gain 0.0 (sigma=0.0003)
  • #2 point [1.06] · Expected Information Gain 0.0 (sigma=0.0003)
  • #3 point [1.112] · Expected Information Gain 0.0 (sigma=0.0003)
  • #4 point [1.122] · Expected Information Gain 0.0 (sigma=0.0003)
  • #5 point [1.152] · Expected Information Gain 0.0 (sigma=0.0003)
微生物 / micro_crabtree_2d Converged
ground-truth: Crabtree 1929 (Crabtree 2D)
rmse (with accumulated data)0.0000
seed baseline rmse0.0143
95% CI coverage1.00
calibration error0.050
extrapolation honesty0.09×
closed-loop accumulated points0
Source: Crabtree 1929 (Crabtree 2D)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [5.77, 26.306] · Expected Information Gain 0.0166 (sigma=0.129)
  • #2 point [19.253, 28.789] · Expected Information Gain 0.0165 (sigma=0.1285)
  • #3 point [15.826, 25.443] · Expected Information Gain 0.0163 (sigma=0.1275)
  • #4 point [15.259, 24.899] · Expected Information Gain 0.0162 (sigma=0.1272)
  • #5 point [23.687, 33.279] · Expected Information Gain 0.0161 (sigma=0.1271)
微生物 / micro_fedbatch_3d Converged
ground-truth: Shuler & Kargi 2002 (fed-batch 3D)
rmse (with accumulated data)0.0078
seed baseline rmse0.0088
95% CI coverage1.00
calibration error0.050
extrapolation honesty2.07×
closed-loop accumulated points0
Source: Shuler & Kargi 2002 (fed-batch 3D)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [13.34, 0.017, 49.158] · Expected Information Gain 0.0007 (sigma=0.0272)
  • #2 point [56.569, 0.066, 36.278] · Expected Information Gain 0.0007 (sigma=0.0272)
  • #3 point [67.47, 0.093, 62.916] · Expected Information Gain 0.0007 (sigma=0.0271)
  • #4 point [70.16, 0.106, 61.769] · Expected Information Gain 0.0007 (sigma=0.027)
  • #5 point [41.816, 0.06, 5.521] · Expected Information Gain 0.0007 (sigma=0.027)
微生物 / micro_haldane_2d Converging
ground-truth: Haldane 1956 (2D)
rmse (with accumulated data)0.0075
seed baseline rmse0.0285
95% CI coverage1.00
calibration error0.050
extrapolation honesty4.98×
closed-loop accumulated points0
Source: Haldane 1956 (2D)
Next-experiment direction suggestions
  • Critical convergence: Relative error 8.6%, still 3.6 percentage points away from the 5% convergence line. Recommend supplementing only one real value at the tournament top-1 candidate points (locations with maximum posterior variance) and retest—This is the minimal cost cross-line path, avoiding overfitting that causes false confidence.
Next-experiment tournament (ranked by information gain)
  • #1 point [3.564, 31.592] · Expected Information Gain 0.0001 (sigma=0.0078)
  • #2 point [1.844, 26.491] · Expected Information Gain 0.0001 (sigma=0.0077)
  • #3 point [2.707, 29.082] · Expected Information Gain 0.0001 (sigma=0.0077)
  • #4 point [2.565, 28.674] · Expected Information Gain 0.0001 (sigma=0.0077)
  • #5 point [3.455, 31.335] · Expected Information Gain 0.0001 (sigma=0.0077)
微生物 / micro_oxygen_2d Converged
ground-truth: Shuler & Kargi 2002 (O2 2D)
rmse (with accumulated data)0.0023
seed baseline rmse0.0313
95% CI coverage1.00
calibration error0.050
extrapolation honesty5.74×
closed-loop accumulated points0
Source: Shuler & Kargi 2002 (O2 2D)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [7.127, 3.521] · Expected Information Gain 0.0001 (sigma=0.0087)
  • #2 point [3.687, 0.804] · Expected Information Gain 0.0001 (sigma=0.0086)
  • #3 point [7.262, 7.886] · Expected Information Gain 0.0001 (sigma=0.0086)
  • #4 point [6.91, 3.384] · Expected Information Gain 0.0001 (sigma=0.0085)
  • #5 point [5.414, 2.184] · Expected Information Gain 0.0001 (sigma=0.0085)
微生物 / micro_antibiotic_2d Converged
ground-truth: Andrews 2001 (time-kill 2D)
rmse (with accumulated data)0.0000
seed baseline rmse0.0011
95% CI coverage1.00
calibration error0.050
extrapolation honesty1.19×
closed-loop accumulated points0
Source: Andrews 2001 (time-kill 2D)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [8.658, 19.243] · Expected Information Gain 0.0 (sigma=0.0001)
  • #2 point [130.752, 44.954] · Expected Information Gain 0.0 (sigma=0.0001)
  • #3 point [145.502, 47.403] · Expected Information Gain 0.0 (sigma=0.0001)
  • #4 point [2.691, 17.128] · Expected Information Gain 0.0 (sigma=0.0001)
  • #5 point [74.367, 10.174] · Expected Information Gain 0.0 (sigma=0.0001)
微生物 / micro_secondary_metabolite Converged
ground-truth: Luedeking & Piret 1959 (secondary metabolite, Gaden III)
rmse (with accumulated data)0.0001
seed baseline rmse0.0933
95% CI coverage1.00
calibration error0.050
extrapolation honesty1.07×
closed-loop accumulated points0
Source: Luedeking & Piret 1959 (secondary metabolite, Gaden III)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [35.846] · Expected Information Gain 0.001 (sigma=0.0317)
  • #2 point [10.578] · Expected Information Gain 0.001 (sigma=0.0313)
  • #3 point [10.721] · Expected Information Gain 0.001 (sigma=0.0312)
  • #4 point [10.736] · Expected Information Gain 0.001 (sigma=0.0312)
  • #5 point [10.788] · Expected Information Gain 0.001 (sigma=0.0311)
微生物 / micro_thermal_death Converged
ground-truth: Earley 1976 (F-value sterilization)
rmse (with accumulated data)0.0000
seed baseline rmse0.0091
95% CI coverage1.00
calibration error0.050
extrapolation honesty0.03×
closed-loop accumulated points0
Source: Earley 1976 (F-value sterilization)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [60.041] · Expected Information Gain 0.0012 (sigma=0.0351)
  • #2 point [60.127] · Expected Information Gain 0.0012 (sigma=0.035)
  • #3 point [60.236] · Expected Information Gain 0.0012 (sigma=0.035)
  • #4 point [60.257] · Expected Information Gain 0.0012 (sigma=0.035)
  • #5 point [60.319] · Expected Information Gain 0.0012 (sigma=0.0349)
微生物 / micro_immobilized_cell Needs data/calibration
ground-truth: Shuler & Kargi 2002 (immobilized cells)
rmse (with accumulated data)0.2710
seed baseline rmse0.8743
95% CI coverage0.97
calibration error0.017
extrapolation honesty1.10×
closed-loop accumulated points0
Source: Shuler & Kargi 2002 (immobilized cells)
Next-experiment direction suggestions
  • Supplement points: Densify real paper data points in regions with maximum error (high curvature / inflection points); or reduce lengthscale to improve fitting.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.118, 0.003] · Expected Information Gain 0.915 (sigma=0.9566)
  • #2 point [0.412, 0.009] · Expected Information Gain 0.915 (sigma=0.9566)
  • #3 point [0.261, 0.003] · Expected Information Gain 0.915 (sigma=0.9566)
  • #4 point [0.373, 0.007] · Expected Information Gain 0.915 (sigma=0.9566)
  • #5 point [0.123, 0.002] · Expected Information Gain 0.915 (sigma=0.9566)
微生物 / micro_plasmid_stability Converged
ground-truth: Stewart 1978 (plasmid stability)
rmse (with accumulated data)0.0128
seed baseline rmse0.0152
95% CI coverage1.00
calibration error0.050
extrapolation honesty8.04×
closed-loop accumulated points0
Source: Stewart 1978 (plasmid stability)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [6.347, 0.016] · Expected Information Gain 0.1699 (sigma=0.4122)
  • #2 point [37.702, 0.033] · Expected Information Gain 0.1683 (sigma=0.4102)
  • #3 point [29.943, 0.022] · Expected Information Gain 0.1541 (sigma=0.3925)
  • #4 point [5.297, 0.013] · Expected Information Gain 0.15 (sigma=0.3873)
  • #5 point [30.731, 0.049] · Expected Information Gain 0.1484 (sigma=0.3853)
微生物 / micro_phage_infection Converged
ground-truth: Luria & Delbrück 1943 (phage infection)
rmse (with accumulated data)0.0000
seed baseline rmse0.0310
95% CI coverage1.00
calibration error0.050
extrapolation honesty0.05×
closed-loop accumulated points0
Source: Luria & Delbrück 1943 (phage infection)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [69.144, 5.842] · Expected Information Gain 0.005 (sigma=0.0708)
  • #2 point [65.202, 5.601] · Expected Information Gain 0.0049 (sigma=0.0703)
  • #3 point [36.868, 1.047] · Expected Information Gain 0.0049 (sigma=0.07)
  • #4 point [71.265, 2.917] · Expected Information Gain 0.0049 (sigma=0.0699)
  • #5 point [69.101, 2.823] · Expected Information Gain 0.0048 (sigma=0.0696)
微生物 / micro_gene_expression Converged
ground-truth: Bashor & Meyer 2017 (gene expression dynamics)
rmse (with accumulated data)0.0023
seed baseline rmse0.0047
95% CI coverage0.97
calibration error0.017
extrapolation honesty7.76×
closed-loop accumulated points0
Source: Bashor & Meyer 2017 (gene expression dynamics)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [2.159, 0.748] · Expected Information Gain 0.0005 (sigma=0.0231)
  • #2 point [0.538, 1.162] · Expected Information Gain 0.0005 (sigma=0.0231)
  • #3 point [3.611, 2.928] · Expected Information Gain 0.0005 (sigma=0.0226)
  • #4 point [3.434, 2.819] · Expected Information Gain 0.0005 (sigma=0.022)
  • #5 point [3.707, 1.599] · Expected Information Gain 0.0005 (sigma=0.0215)
微生物 / micro_oxidative_stress Converged
ground-truth: Imlay 2008 (oxidative stress)
rmse (with accumulated data)0.0003
seed baseline rmse0.0180
95% CI coverage1.00
calibration error0.050
extrapolation honesty2.36×
closed-loop accumulated points0
Source: Imlay 2008 (oxidative stress)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [2.965] · Expected Information Gain 0.0469 (sigma=0.2166)
  • #2 point [0.025] · Expected Information Gain 0.0456 (sigma=0.2135)
  • #3 point [2.957] · Expected Information Gain 0.0452 (sigma=0.2126)
  • #4 point [2.945] · Expected Information Gain 0.0427 (sigma=0.2066)
  • #5 point [2.944] · Expected Information Gain 0.0425 (sigma=0.206)
微生物 / micro_chemostat_transient Converged
ground-truth: Shuler & Kargi 2002 (chemostat transient)
rmse (with accumulated data)0.0062
seed baseline rmse0.0184
95% CI coverage0.93
calibration error0.017
extrapolation honesty21.16×
closed-loop accumulated points0
Source: Shuler & Kargi 2002 (chemostat transient)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.332, 0.57] · Expected Information Gain 3.8694 (sigma=1.9671)
  • #2 point [0.487, 0.691] · Expected Information Gain 3.8652 (sigma=1.966)
  • #3 point [0.587, 1.149] · Expected Information Gain 3.8554 (sigma=1.9635)
  • #4 point [0.507, 0.708] · Expected Information Gain 3.8397 (sigma=1.9595)
  • #5 point [0.462, 0.671] · Expected Information Gain 3.8336 (sigma=1.958)
微生物 / micro_mixed_substrate Converged
ground-truth: Roels 1983 (mixed substrate utilization)
rmse (with accumulated data)0.0150
seed baseline rmse0.2833
95% CI coverage1.00
calibration error0.050
extrapolation honesty8.04×
closed-loop accumulated points0
Source: Roels 1983 (mixed substrate utilization)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.043, 2.65] · Expected Information Gain 0.0054 (sigma=0.0734)
  • #2 point [3.26, 9.275] · Expected Information Gain 0.0052 (sigma=0.0722)
  • #3 point [4.672, 6.64] · Expected Information Gain 0.0052 (sigma=0.072)
  • #4 point [1.802, 6.405] · Expected Information Gain 0.005 (sigma=0.0709)
  • #5 point [3.631, 9.858] · Expected Information Gain 0.0049 (sigma=0.0698)
微生物 / micro_ph_control Converged
ground-truth: Shuler & Kargi 2002 (pH control)
rmse (with accumulated data)0.0835
seed baseline rmse0.4885
95% CI coverage1.00
calibration error0.050
extrapolation honesty5.85×
closed-loop accumulated points0
Source: Shuler & Kargi 2002 (pH control)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [38.829, 8.385] · Expected Information Gain 0.1151 (sigma=0.3393)
  • #2 point [32.374, 4.898] · Expected Information Gain 0.1116 (sigma=0.334)
  • #3 point [38.04, 8.21] · Expected Information Gain 0.1109 (sigma=0.3331)
  • #4 point [39.253, 6.258] · Expected Information Gain 0.1063 (sigma=0.3261)
  • #5 point [36.27, 5.687] · Expected Information Gain 0.1037 (sigma=0.322)
微生物 / micro_fba_flux Converged
ground-truth: Edwards & Palsson 2000 (FBA framework, E. coli)
rmse (with accumulated data)0.0109
seed baseline rmse0.0109
95% CI coverage0.97
calibration error0.017
extrapolation honesty18.28×
closed-loop accumulated points0
Source: Edwards & Palsson 2000 (FBA framework, E. coli)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.432, 0.189] · Expected Information Gain 0.8005 (sigma=0.8947)
  • #2 point [0.587, 0.345] · Expected Information Gain 0.7521 (sigma=0.8672)
  • #3 point [0.722, 0.974] · Expected Information Gain 0.7384 (sigma=0.8593)
  • #4 point [0.607, 0.367] · Expected Information Gain 0.7258 (sigma=0.8519)
  • #5 point [0.722, 0.48] · Expected Information Gain 0.721 (sigma=0.8491)
微生物 / micro_dead_volume Converging
ground-truth: Froment & Bischoff 2012 (CSTR-in-series, dead volume)
rmse (with accumulated data)0.0627
seed baseline rmse0.2011
95% CI coverage1.00
calibration error0.050
extrapolation honesty5.81×
closed-loop accumulated points0
Source: Froment & Bischoff 2012 (CSTR-in-series, dead volume)
Next-experiment direction suggestions
  • Moderate underfitting (relative error 11.0%): Prioritize doing two things—① Supplement real paper values at the top-2 candidate points from the experimental tournament;② If residuals show directionality (multi-peaked/heterogeneous), replace the isotropic RBF with an ARD kernel to learn lengthscale per dimension.
Next-experiment tournament (ranked by information gain)
  • #1 point [16.183, 4.627] · Expected Information Gain 0.0036 (sigma=0.0602)
  • #2 point [8.005, 1.696] · Expected Information Gain 0.0034 (sigma=0.0582)
  • #3 point [14.129, 3.956] · Expected Information Gain 0.0033 (sigma=0.0573)
  • #4 point [11.285, 2.905] · Expected Information Gain 0.0033 (sigma=0.0572)
  • #5 point [10.746, 2.715] · Expected Information Gain 0.0033 (sigma=0.0571)
微生物 / micro_foam_dynamics Converged
ground-truth: Krebes & Scharaschkin 1996 (foam in bioreactors)
rmse (with accumulated data)0.0138
seed baseline rmse0.0720
95% CI coverage0.97
calibration error0.017
extrapolation honesty6.14×
closed-loop accumulated points0
Source: Krebes & Scharaschkin 1996 (foam in bioreactors)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.785, 0.32] · Expected Information Gain 0.0098 (sigma=0.0991)
  • #2 point [0.8, 0.05] · Expected Information Gain 0.0097 (sigma=0.0985)
  • #3 point [1.875, 0.332] · Expected Information Gain 0.0096 (sigma=0.0978)
  • #4 point [1.129, 0.136] · Expected Information Gain 0.0094 (sigma=0.097)
  • #5 point [1.413, 0.211] · Expected Information Gain 0.0093 (sigma=0.0964)
微生物 / micro_sensor_fusion Converged
ground-truth: Gelb 1974 (Kalman filter / Bayesian fusion)
rmse (with accumulated data)0.0017
seed baseline rmse0.0283
95% CI coverage0.97
calibration error0.017
extrapolation honesty5.04×
closed-loop accumulated points0
Source: Gelb 1974 (Kalman filter / Bayesian fusion)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [7.414, 2.253] · Expected Information Gain 0.0139 (sigma=0.1178)
  • #2 point [9.409, 3.353] · Expected Information Gain 0.0138 (sigma=0.1177)
  • #3 point [4.318, 0.587] · Expected Information Gain 0.0136 (sigma=0.1168)
  • #4 point [7.219, 2.169] · Expected Information Gain 0.0135 (sigma=0.1161)
  • #5 point [5.617, 1.3] · Expected Information Gain 0.0134 (sigma=0.1159)
optimization Converged
ground-truth: global min = 0 at x=(420.9687, 420.9687)
rmse (with accumulated data)7.7378
seed baseline rmse7.7378
95% CI coverage1.00
calibration error0.050
extrapolation honesty19.82×
closed-loop accumulated points0
Source: global min = 0 at x=(420.9687, 420.9687)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [12.97, -255.035] · Expected Information Gain 143721.333 (sigma=379.106)
  • #2 point [-131.324, -400.62] · Expected Information Gain 143626.7296 (sigma=378.9812)
  • #3 point [434.338, 163.94] · Expected Information Gain 143537.151 (sigma=378.863)
  • #4 point [191.01, -77.687] · Expected Information Gain 143439.3098 (sigma=378.7338)
  • #5 point [41.323, -227.856] · Expected Information Gain 143275.5218 (sigma=378.5175)
optimization Converging
ground-truth: global min = 0 at x=(0,0)
rmse (with accumulated data)0.7181
seed baseline rmse0.7181
95% CI coverage0.93
calibration error0.017
extrapolation honesty18.23×
closed-loop accumulated points0
Source: global min = 0 at x=(0,0)
Next-experiment direction suggestions
  • Supplement points: Densify real paper data points in regions with maximum error (high curvature / inflection points); or reduce lengthscale to improve fitting.
Next-experiment tournament (ranked by information gain)
  • #1 point [4.162, -13.321] · Expected Information Gain 20.2513 (sigma=4.5001)
  • #2 point [-0.416, -18.55] · Expected Information Gain 20.2445 (sigma=4.4994)
  • #3 point [-8.606, -26.255] · Expected Information Gain 20.2316 (sigma=4.498)
  • #4 point [-30.246, -12.104] · Expected Information Gain 20.2068 (sigma=4.4952)
  • #5 point [0.85, -16.714] · Expected Information Gain 20.1794 (sigma=4.4922)
optimization Needs data/calibration
ground-truth: global min = 0 at x=(0,0)
rmse (with accumulated data)5.4879
seed baseline rmse5.4879
95% CI coverage1.00
calibration error0.050
extrapolation honesty17.44×
closed-loop accumulated points0
Source: global min = 0 at x=(0,0)
Next-experiment direction suggestions
  • Supplement points: Densify real paper data points in regions with maximum error (high curvature / inflection points); or reduce lengthscale to improve fitting.
Next-experiment tournament (ranked by information gain)
  • #1 point [4.448, 1.679] · Expected Information Gain 410.1536 (sigma=20.2522)
  • #2 point [0.133, -2.612] · Expected Information Gain 409.1036 (sigma=20.2263)
  • #3 point [2.177, -0.62] · Expected Information Gain 408.4128 (sigma=20.2092)
  • #4 point [-2.291, 0.549] · Expected Information Gain 408.2986 (sigma=20.2064)
  • #5 point [0.529, -2.148] · Expected Information Gain 408.1485 (sigma=20.2027)
optimization Converged
ground-truth: global min = 0 at x=(1,1)
rmse (with accumulated data)23.4742
seed baseline rmse23.4742
95% CI coverage0.97
calibration error0.017
extrapolation honesty20.72×
closed-loop accumulated points0
Source: global min = 0 at x=(1,1)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.764, 0.689] · Expected Information Gain 73896.7664 (sigma=271.8396)
  • #2 point [-0.002, 2.035] · Expected Information Gain 73331.4966 (sigma=270.7979)
  • #3 point [-0.115, 1.913] · Expected Information Gain 70781.8892 (sigma=266.0487)
  • #4 point [0.766, 2.885] · Expected Information Gain 70115.2374 (sigma=264.7928)
  • #5 point [0.116, 2.127] · Expected Information Gain 70015.732 (sigma=264.6049)
optimization Converging
ground-truth: global min = 0 at x=(1,1)
rmse (with accumulated data)0.1914
seed baseline rmse0.2680
95% CI coverage0.87
calibration error0.083
extrapolation honesty17.23×
closed-loop accumulated points0
Source: global min = 0 at x=(1,1)
Next-experiment direction suggestions
  • Calibration Error 0.083 Within Critical Band (Decision Line 0.08): Nominal Confidence Level vs. Actual Coverage Has OccurredMeasurable Deviation: Recommend Recording Epistemic/Aleatoric Components in the Next Round to Identify the Source of Deviation.
Next-experiment tournament (ranked by information gain)
  • #1 point [8.687, 3.279] · Expected Information Gain 35.6344 (sigma=5.9695)
  • #2 point [4.253, -1.211] · Expected Information Gain 35.4458 (sigma=5.9536)
  • #3 point [-0.127, -5.661] · Expected Information Gain 33.3065 (sigma=5.7712)
  • #4 point [4.523, 9.716] · Expected Information Gain 32.113 (sigma=5.6668)
  • #5 point [-0.677, -5.858] · Expected Information Gain 31.0414 (sigma=5.5715)
optimization Converging
ground-truth: approx min = -1.8013 at m=10
rmse (with accumulated data)0.0263
seed baseline rmse0.0194
95% CI coverage1.00
calibration error0.050
extrapolation honesty19.36×
closed-loop accumulated points0
Source: approx min = -1.8013 at m=10
Next-experiment direction suggestions
  • Critical convergence: Relative error 9.3%, still 4.3 percentage points away from the 5% convergence line. Recommend supplementing only one real value at the tournament top-1 candidate points (locations with maximum posterior variance) and retest—This is the minimal cost cross-line path, avoiding overfitting that causes false confidence.
Next-experiment tournament (ranked by information gain)
  • #1 point [2.048, 2.914] · Expected Information Gain 0.0795 (sigma=0.282)
  • #2 point [0.121, 0.991] · Expected Information Gain 0.0794 (sigma=0.2817)
  • #3 point [1.701, 0.855] · Expected Information Gain 0.0793 (sigma=0.2816)
  • #4 point [0.868, 1.739] · Expected Information Gain 0.0793 (sigma=0.2816)
  • #5 point [1.612, 0.77] · Expected Information Gain 0.0792 (sigma=0.2814)
optimization Converged
ground-truth: global min ≈ -3.86278
rmse (with accumulated data)0.0239
seed baseline rmse0.0100
95% CI coverage0.95
calibration error0.000
extrapolation honesty19.84×
closed-loop accumulated points0
Source: global min ≈ -3.86278
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.207, 0.064, 0.541] · Expected Information Gain 0.2207 (sigma=0.4698)
  • #2 point [0.463, 0.783, 0.322] · Expected Information Gain 0.2105 (sigma=0.4588)
  • #3 point [0.184, 0.015, 0.471] · Expected Information Gain 0.2077 (sigma=0.4558)
  • #4 point [0.35, 0.676, 0.228] · Expected Information Gain 0.2076 (sigma=0.4556)
  • #5 point [0.545, 0.75, 0.295] · Expected Information Gain 0.2072 (sigma=0.4551)
optimization Converged
ground-truth: Gramacy & Lee 2012 test function (1D multimodal)
rmse (with accumulated data)0.0053
seed baseline rmse0.0098
95% CI coverage0.95
calibration error0.000
extrapolation honesty1.38×
closed-loop accumulated points0
Source: Gramacy & Lee 2012 test function (1D multimodal)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.508] · Expected Information Gain 0.0002 (sigma=0.0152)
  • #2 point [0.524] · Expected Information Gain 0.0002 (sigma=0.0143)
  • #3 point [7.71] · Expected Information Gain 0.0002 (sigma=0.0139)
  • #4 point [7.704] · Expected Information Gain 0.0002 (sigma=0.0139)
  • #5 point [7.685] · Expected Information Gain 0.0002 (sigma=0.0138)
optimization Converged
ground-truth: Welch et al. 1992 SDOE test function
rmse (with accumulated data)0.0007
seed baseline rmse0.0007
95% CI coverage1.00
calibration error0.050
extrapolation honesty21.98×
closed-loop accumulated points0
Source: Welch et al. 1992 SDOE test function
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.494, 0.217] · Expected Information Gain 0.0044 (sigma=0.0667)
  • #2 point [0.36, 0.641] · Expected Information Gain 0.0043 (sigma=0.0656)
  • #3 point [0.934, 0.664] · Expected Information Gain 0.0042 (sigma=0.0651)
  • #4 point [0.008, 0.265] · Expected Information Gain 0.0042 (sigma=0.0647)
  • #5 point [0.726, 0.986] · Expected Information Gain 0.0042 (sigma=0.0644)
biology Converged
ground-truth: Lotka-Volterra 参数-响应代理(非拟态基准,供 GP 代理练习)
rmse (with accumulated data)0.0057
seed baseline rmse0.0057
95% CI coverage0.97
calibration error0.017
extrapolation honesty14.30×
closed-loop accumulated points0
Source: Lotka-Volterra 参数-响应代理(非拟态基准,供 GP 代理练习)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [1.48, 2.959] · Expected Information Gain 0.0071 (sigma=0.0844)
  • #2 point [1.497, 2.923] · Expected Information Gain 0.006 (sigma=0.0772)
  • #3 point [1.875, 2.025] · Expected Information Gain 0.0059 (sigma=0.0771)
  • #4 point [1.558, 2.945] · Expected Information Gain 0.0051 (sigma=0.0712)
  • #5 point [1.357, 2.708] · Expected Information Gain 0.0049 (sigma=0.0701)
finance Converged
ground-truth: Black-Scholes 看涨期权定价解析公式
rmse (with accumulated data)0.3213
seed baseline rmse0.3213
95% CI coverage0.95
calibration error0.000
extrapolation honesty20.02×
closed-loop accumulated points0
Source: Black-Scholes 看涨期权定价解析公式
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [85.476, 161.973, 0.457, 0.077, 0.19] · Expected Information Gain 131.1396 (sigma=11.4516)
  • #2 point [126.946, 86.745, 0.842, 0.021, 0.471] · Expected Information Gain 128.1474 (sigma=11.3202)
  • #3 point [171.814, 188.502, 0.349, 0.082, 0.545] · Expected Information Gain 126.6629 (sigma=11.2545)
  • #4 point [168.224, 168.187, 0.953, 0.065, 0.226] · Expected Information Gain 120.4092 (sigma=10.9731)
  • #5 point [130.555, 71.006, 0.544, 0.022, 0.504] · Expected Information Gain 120.278 (sigma=10.9671)
neuroscience Converging
ground-truth: Hodgkin-Huxley 四变量神经元模型,数值积分求解
rmse (with accumulated data)0.1100
seed baseline rmse0.1100
95% CI coverage1.00
calibration error0.050
extrapolation honesty20.94×
closed-loop accumulated points0
Source: Hodgkin-Huxley 四变量神经元模型,数值积分求解
Next-experiment direction suggestions
  • Moderate underfitting (relative error 11.1%): Prioritize doing two things—① Supplement real paper values at the top-2 candidate points from the experimental tournament;② If residuals show directionality (multi-peaked/heterogeneous), replace the isotropic RBF with an ARD kernel to learn lengthscale per dimension.
Next-experiment tournament (ranked by information gain)
  • #1 point [1.188, 193.877, 50.527, 0.223, 51.133, -80.706, -62.691] · Expected Information Gain 0.7733 (sigma=0.8794)
  • #2 point [1.612, 162.994, 57.096, 0.309, 55.983, -79.637, -63.053] · Expected Information Gain 0.7689 (sigma=0.8768)
  • #3 point [1.849, 126.182, 37.196, 0.558, 55.535, -70.665, -42.902] · Expected Information Gain 0.7206 (sigma=0.8489)
  • #4 point [1.289, 124.776, 38.072, 0.669, 44.939, -86.387, -40.283] · Expected Information Gain 0.7167 (sigma=0.8466)
  • #5 point [1.028, 139.632, 30.601, 0.81, 54.325, -76.171, -40.861] · Expected Information Gain 0.7032 (sigma=0.8386)
ecology Needs data/calibration
ground-truth: Lotka-Volterra 捕食者-猎物 ODE,RK4 数值积分
rmse (with accumulated data)0.0400
seed baseline rmse0.0400
95% CI coverage0.95
calibration error0.000
extrapolation honesty19.79×
closed-loop accumulated points0
Source: Lotka-Volterra 捕食者-猎物 ODE,RK4 数值积分
Next-experiment direction suggestions
  • Severe underfitting and 4-dimensional: Empirical results show the bottleneck is **fixed isotropic lengthscale**, not data volume—In ablation experiments, fixed ls with 42 points still has 54% relative error, whereas enabling automatic ARD hyperparameters reduces it to 4.75% at 41 points. Current 30 points have not reached the automatic hyperparameter threshold (requires ≥22 points; small samples cause marginal likelihood to hit boundaries, worsening performance).Path: First add points via tournament to reach 22, then set auto_ls=True for this scenario.
Next-experiment tournament (ranked by information gain)
  • #1 point [2.493, 1.771, 0.809, 1.554] · Expected Information Gain 0.0007 (sigma=0.0261)
  • #2 point [2.443, 1.773, 0.814, 1.496] · Expected Information Gain 0.0007 (sigma=0.0259)
  • #3 point [1.419, 1.501, 0.439, 1.417] · Expected Information Gain 0.0007 (sigma=0.0258)
  • #4 point [2.594, 1.429, 0.674, 1.534] · Expected Information Gain 0.0007 (sigma=0.0258)
  • #5 point [1.475, 1.458, 1.88, 2.358] · Expected Information Gain 0.0007 (sigma=0.0258)
cheminformatics Converged
ground-truth: 1513 分子真实实验值(bace),线性响应面 oracle 从 400 训练分子 lstsq 拟合;目标变量:exp_pIC50
rmse (with accumulated data)0.0062
seed baseline rmse0.0062
95% CI coverage0.95
calibration error0.000
extrapolation honesty21.57×
closed-loop accumulated points0
Source: 1513 分子真实实验值(bace),线性响应面 oracle 从 400 训练分子 lstsq 拟合;目标变量:exp_pIC50
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.784, 0.498, 0.923, 0.433, 0.597, 0.062, 0.54] · Expected Information Gain 0.0566 (sigma=0.238)
  • #2 point [0.179, 0.03, 0.559, 0.429, 0.859, 0.161, 0.36] · Expected Information Gain 0.0563 (sigma=0.2372)
  • #3 point [0.925, 0.366, 0.712, 0.34, 0.512, 0.149, 0.768] · Expected Information Gain 0.0562 (sigma=0.237)
  • #4 point [0.47, 0.768, 0.088, 0.489, 0.675, 0.713, 0.151] · Expected Information Gain 0.0561 (sigma=0.2368)
  • #5 point [0.437, 0.396, 0.022, 0.469, 0.624, 0.946, 0.435] · Expected Information Gain 0.0558 (sigma=0.2362)
cheminformatics Converged
ground-truth: 4200 分子真实实验值(chembl_lipophilicity),线性响应面 oracle 从 800 训练分子 lstsq 拟合;目标变量:exp_logD
rmse (with accumulated data)0.0019
seed baseline rmse0.0019
95% CI coverage1.00
calibration error0.050
extrapolation honesty20.39×
closed-loop accumulated points0
Source: 4200 分子真实实验值(chembl_lipophilicity),线性响应面 oracle 从 800 训练分子 lstsq 拟合;目标变量:exp_logD
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.59, 0.879, 0.227, 0.162, 0.78, 0.682, 0.207, 0.674] · Expected Information Gain 0.0198 (sigma=0.1406)
  • #2 point [0.927, 0.232, 0.799, 0.518, 0.232, 0.166, 0.498, 0.583] · Expected Information Gain 0.0195 (sigma=0.1396)
  • #3 point [0.784, 0.498, 0.923, 0.433, 0.597, 0.062, 0.54, 0.229] · Expected Information Gain 0.0194 (sigma=0.1393)
  • #4 point [0.712, 0.34, 0.512, 0.149, 0.768, 0.262, 0.47, 0.973] · Expected Information Gain 0.0193 (sigma=0.1389)
  • #5 point [0.466, 0.314, 0.492, 0.238, 0.96, 0.418, 0.336, 0.935] · Expected Information Gain 0.0193 (sigma=0.1389)
cheminformatics Converged
ground-truth: 1128 分子真实实验值(esol_delaney),线性响应面 oracle 从 400 训练分子 lstsq 拟合;目标变量:exp_logS
rmse (with accumulated data)0.0067
seed baseline rmse0.0067
95% CI coverage1.00
calibration error0.050
extrapolation honesty19.82×
closed-loop accumulated points0
Source: 1128 分子真实实验值(esol_delaney),线性响应面 oracle 从 400 训练分子 lstsq 拟合;目标变量:exp_logS
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.591, 0.483, 0.901, 0.541, 0.071, 0.034, 0.617, 0.633, 0.66] · Expected Information Gain 0.4312 (sigma=0.6567)
  • #2 point [0.67, 0.18, 0.261, 0.547, 0.237, 0.16, 0.446, 0.061, 0.879] · Expected Information Gain 0.4304 (sigma=0.6561)
  • #3 point [0.13, 0.555, 0.159, 0.298, 0.239, 0.813, 0.256, 0.455, 0.867] · Expected Information Gain 0.43 (sigma=0.6558)
  • #4 point [0.152, 0.847, 0.222, 0.817, 0.925, 0.456, 0.704, 0.478, 0.425] · Expected Information Gain 0.4271 (sigma=0.6535)
  • #5 point [0.871, 0.184, 0.248, 0.929, 0.811, 0.817, 0.523, 0.602, 0.879] · Expected Information Gain 0.4244 (sigma=0.6514)
cheminformatics Converged
ground-truth: 642 分子真实实验值(freesolv),线性响应面 oracle 从 200 训练分子 lstsq 拟合;目标变量:exp_hydration_free_energy_kcal_mol
rmse (with accumulated data)0.0127
seed baseline rmse0.0127
95% CI coverage1.00
calibration error0.050
extrapolation honesty21.57×
closed-loop accumulated points0
Source: 642 分子真实实验值(freesolv),线性响应面 oracle 从 200 训练分子 lstsq 拟合;目标变量:exp_hydration_free_energy_kcal_mol
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.784, 0.498, 0.923, 0.433, 0.597, 0.062, 0.54] · Expected Information Gain 0.3139 (sigma=0.5603)
  • #2 point [0.179, 0.03, 0.559, 0.429, 0.859, 0.161, 0.36] · Expected Information Gain 0.3119 (sigma=0.5585)
  • #3 point [0.925, 0.366, 0.712, 0.34, 0.512, 0.149, 0.768] · Expected Information Gain 0.3114 (sigma=0.558)
  • #4 point [0.47, 0.768, 0.088, 0.489, 0.675, 0.713, 0.151] · Expected Information Gain 0.3106 (sigma=0.5573)
  • #5 point [0.437, 0.396, 0.022, 0.469, 0.624, 0.946, 0.435] · Expected Information Gain 0.3091 (sigma=0.556)
energy Converged
ground-truth: NASA CCPP 联合循环电厂 6 年 9,568 行实测;4 环境特征 → 功率输出 MW;oracle = 训练集 lstsq 线性响应面
rmse (with accumulated data)0.0081
seed baseline rmse0.0081
95% CI coverage1.00
calibration error0.050
extrapolation honesty19.78×
closed-loop accumulated points0
Source: NASA CCPP 联合循环电厂 6 年 9,568 行实测;4 环境特征 → 功率输出 MW;oracle = 训练集 lstsq 线性响应面
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.298, 0.744, 0.566, 0.582] · Expected Information Gain 0.3987 (sigma=0.6314)
  • #2 point [0.36, 0.789, 0.452, 0.559] · Expected Information Gain 0.3972 (sigma=0.6302)
  • #3 point [0.734, 0.371, 0.635, 0.685] · Expected Information Gain 0.39 (sigma=0.6245)
  • #4 point [0.797, 0.85, 0.702, 0.526] · Expected Information Gain 0.3874 (sigma=0.6224)
  • #5 point [0.618, 0.403, 0.591, 0.879] · Expected Information Gain 0.3797 (sigma=0.6162)
chemistry Needs data/calibration
ground-truth: 12 substrate pairs with known yields (Gaussian response surface gold)
rmse (with accumulated data)0.0095
seed baseline rmse0.0095
95% CI coverage1.00
calibration error0.050
extrapolation honesty17.20×
closed-loop accumulated points0
Source: 12 substrate pairs with known yields (Gaussian response surface gold)
Next-experiment direction suggestions
  • Supplement points: Densify real paper data points in regions with maximum error (high curvature / inflection points); or reduce lengthscale to improve fitting.
Next-experiment tournament (ranked by information gain)
  • #1 point [-0.192, -1.552, 0.398, 2.195, -1.005, -0.198, -2.212, -2.073, -1.503, 0.651] · Expected Information Gain 0.0006 (sigma=0.0236)
  • #2 point [1.847, 1.931, 2.27, 0.005, 0.051, -0.474, -0.195, 2.434, -1.619, 0.328] · Expected Information Gain 0.0006 (sigma=0.0236)
  • #3 point [0.254, -1.405, 0.209, -1.277, -2.083, 1.282, -1.178, -0.932, 2.442, -0.497] · Expected Information Gain 0.0006 (sigma=0.0235)
  • #4 point [-1.862, -1.212, -2.484, -0.595, 0.379, -0.364, 1.676, 0.582, -1.17, 1.555] · Expected Information Gain 0.0006 (sigma=0.0235)
  • #5 point [-0.599, -1.539, -1.731, -0.658, -2.375, 0.972, 0.55, 0.827, 1.299, -1.293] · Expected Information Gain 0.0005 (sigma=0.0234)
chemistry Needs data/calibration
ground-truth: 12 substrate pairs with known yields (Gaussian response surface gold)
rmse (with accumulated data)0.0009
seed baseline rmse0.0009
95% CI coverage0.90
calibration error0.050
extrapolation honesty17.93×
closed-loop accumulated points0
Source: 12 substrate pairs with known yields (Gaussian response surface gold)
Next-experiment direction suggestions
  • Supplement points: Densify real paper data points in regions with maximum error (high curvature / inflection points); or reduce lengthscale to improve fitting.
Next-experiment tournament (ranked by information gain)
  • #1 point [-2.382, -2.458, 1.754, -1.745, 0.512, 0.454, -1.373, -0.272, -0.405, 0.014] · Expected Information Gain 0.0 (sigma=0.0033)
  • #2 point [2.086, 1.823, -1.409, 1.831, 1.154, -1.111, 1.485, 1.826, -1.003, 0.135] · Expected Information Gain 0.0 (sigma=0.0033)
  • #3 point [-0.271, -2.195, 1.894, 1.367, -1.041, -0.311, -0.853, -1.508, -1.722, -1.581] · Expected Information Gain 0.0 (sigma=0.0033)
  • #4 point [-0.115, 2.211, -1.805, 2.317, -0.078, 0.595, -1.275, 0.807, -0.43, 1.562] · Expected Information Gain 0.0 (sigma=0.0033)
  • #5 point [0.296, -0.356, 1.793, -1.696, -0.698, 0.703, 2.155, -1.191, 0.94, -1.755] · Expected Information Gain 0.0 (sigma=0.0033)
heat conduction (physics-informed) / 串联热阻稳态导热 Converged
ground-truth: 1D 稳态导热的傅里叶定律精确解(串联热阻相加)——无任何近似,oracle 即物理定律本身
rmse (with accumulated data)2582.4996
seed baseline rmse2582.4996
95% CI coverage0.95
calibration error0.000
extrapolation honesty19.54×
closed-loop accumulated points0
Source: 1D 稳态导热的傅里叶定律精确解(串联热阻相加)——无任何近似,oracle 即物理定律本身
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.052, 0.092, 0.086, 100.945, 27.742, 37.84, 2.4, 92.938] · Expected Information Gain 32650325922.7027 (sigma=180694.0119)
  • #2 point [0.084, 0.058, 0.053, 106.719, 1.987, 23.986, 1.277, 117.374] · Expected Information Gain 32496271736.1902 (sigma=180267.2231)
  • #3 point [0.096, 0.06, 0.029, 13.771, 108.574, 137.404, 3.387, 117.822] · Expected Information Gain 32143128689.1194 (sigma=179285.0487)
  • #4 point [0.086, 0.042, 0.061, 84.807, 194.668, 107.78, 0.754, 121.761] · Expected Information Gain 32136653362.598 (sigma=179266.989)
  • #5 point [0.074, 0.041, 0.056, 30.635, 153.733, 53.176, 2.616, 194.857] · Expected Information Gain 31982456678.7986 (sigma=178836.3964)
heat conduction (physics-informed) / Stefan-Boltzmann 净辐射换热 Converging
ground-truth: Stefan-Boltzmann 定律 εσA(T1⁴−T2⁴),σ=5.670374419e-8 W/m²K⁴(CODATA 2018)
rmse (with accumulated data)6327.7974
seed baseline rmse6327.7974
95% CI coverage0.95
calibration error0.000
extrapolation honesty19.66×
closed-loop accumulated points0
Source: Stefan-Boltzmann 定律 εσA(T1⁴−T2⁴),σ=5.670374419e-8 W/m²K⁴(CODATA 2018)
Next-experiment direction suggestions
  • Critical convergence: Relative error 7.8%, still 2.8 percentage points away from the 5% convergence line. Recommend supplementing only one real value at the tournament top-1 candidate points (locations with maximum posterior variance) and retest—This is the minimal cost cross-line path, avoiding overfitting that causes false confidence.
Next-experiment tournament (ranked by information gain)
  • #1 point [982.114, 797.443, 0.649, 1.83] · Expected Information Gain 2405944228.8572 (sigma=49050.4254)
  • #2 point [815.322, 401.648, 0.738, 7.825] · Expected Information Gain 2309285660.1194 (sigma=48055.0274)
  • #3 point [906.704, 972.136, 0.75, 2.798] · Expected Information Gain 2272593572.0415 (sigma=47671.7272)
  • #4 point [671.226, 518.713, 0.611, 9.02] · Expected Information Gain 2270692925.6682 (sigma=47651.7883)
  • #5 point [707.553, 574.116, 0.449, 8.658] · Expected Information Gain 2252204871.2921 (sigma=47457.4006)
llm / Chinchilla 计算最优训练损失 Converged
ground-truth: Kaplan et al. (2020) Eq. 1.1: L(N,D)=A/N^α+B/D^β,系数 A=406.4 α=0.34 B=410.7 β=0.28
rmse (with accumulated data)0.0083
seed baseline rmse0.1319
95% CI coverage1.00
calibration error0.050
extrapolation honesty7.91×
closed-loop accumulated points0
Source: Kaplan et al. (2020) Eq. 1.1: L(N,D)=A/N^α+B/D^β,系数 A=406.4 α=0.34 B=410.7 β=0.28
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [7261660259.612, 98578331209.146] · Expected Information Gain 0.0016 (sigma=0.0398)
  • #2 point [3687392920.353, 9946963723.097] · Expected Information Gain 0.0015 (sigma=0.0387)
  • #3 point [85953166.753, 26503538546.313] · Expected Information Gain 0.0015 (sigma=0.0383)
  • #4 point [7126742971.368, 43952004201.207] · Expected Information Gain 0.0015 (sigma=0.0381)
  • #5 point [6910406373.861, 42237101768.851] · Expected Information Gain 0.0014 (sigma=0.0374)
llm / Chinchilla 最优 D/N 比 Converged
ground-truth: Kaplan et al. (2020) §4 理论最优 D/N=(Bα/(Aβ))^(1/(α−β))≈20.7
rmse (with accumulated data)0.5055
seed baseline rmse0.5228
95% CI coverage0.97
calibration error0.017
extrapolation honesty7.34×
closed-loop accumulated points0
Source: Kaplan et al. (2020) §4 理论最优 D/N=(Bα/(Aβ))^(1/(α−β))≈20.7
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [7261660259.612, 1.979] · Expected Information Gain 2.3569 (sigma=1.5352)
  • #2 point [385798338.34, 0.973] · Expected Information Gain 2.1384 (sigma=1.4623)
  • #3 point [4936992763.417, 0.825] · Expected Information Gain 2.0025 (sigma=1.4151)
  • #4 point [7351945424.868, 1.96] · Expected Information Gain 1.9993 (sigma=1.414)
  • #5 point [6910406373.861, 1.133] · Expected Information Gain 1.9992 (sigma=1.4139)
chemical / Freundlich 经验吸附等温线 Converged
ground-truth: Freundlich (1926) q=K·C^(1/n),经验指数 1/n∈(0,1](多相非饱和吸附)
rmse (with accumulated data)0.3179
seed baseline rmse0.3179
95% CI coverage1.00
calibration error0.050
extrapolation honesty17.38×
closed-loop accumulated points0
Source: Freundlich (1926) q=K·C^(1/n),经验指数 1/n∈(0,1](多相非饱和吸附)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [9.263, 0.759, 0.268] · Expected Information Gain 139.4401 (sigma=11.8085)
  • #2 point [14.882, 1.817, 0.463] · Expected Information Gain 136.9448 (sigma=11.7023)
  • #3 point [10.107, 0.64, 0.397] · Expected Information Gain 135.4875 (sigma=11.6399)
  • #4 point [18.497, 2.463, 0.688] · Expected Information Gain 133.0685 (sigma=11.5355)
  • #5 point [15.817, 1.893, 0.573] · Expected Information Gain 132.6192 (sigma=11.516)
chemical / BET 多层吸附等温线 Converged
ground-truth: Brunauer-Emmett-Teller (1938) 多层吸附等温线,q=q_m·bC/[(1−C/C_m)(1+(b−1)C)]
rmse (with accumulated data)0.9608
seed baseline rmse0.9608
95% CI coverage1.00
calibration error0.050
extrapolation honesty19.66×
closed-loop accumulated points0
Source: Brunauer-Emmett-Teller (1938) 多层吸附等温线,q=q_m·bC/[(1−C/C_m)(1+(b−1)C)]
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [25.594, 142.752, 3.537, 321.734] · Expected Information Gain 1621.2403 (sigma=40.2646)
  • #2 point [13.156, 121.59, 2.854, 200.936] · Expected Information Gain 1598.3337 (sigma=39.9792)
  • #3 point [29.64, 140.543, 14.452, 245.606] · Expected Information Gain 1502.3549 (sigma=38.7602)
  • #4 point [18.701, 141.629, 4.534, 337.525] · Expected Information Gain 1445.8442 (sigma=38.0243)
  • #5 point [20.917, 63.91, 24.183, 469.054] · Expected Information Gain 1421.8362 (sigma=37.7072)
finance / Merton 连续红利欧式看涨 Converged
ground-truth: Merton (1973) C=S·e^(−qT)N(d1)−K·e^(−rT)N(d2),d1 含连续红利率 q(BS 的推广)
rmse (with accumulated data)0.2576
seed baseline rmse0.2576
95% CI coverage0.88
calibration error0.075
extrapolation honesty27.73×
closed-loop accumulated points0
Source: Merton (1973) C=S·e^(−qT)N(d1)−K·e^(−rT)N(d2),d1 含连续红利率 q(BS 的推广)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [167.428, 97.262, 0.066, 0.034, 0.596, 0.753] · Expected Information Gain 409.9239 (sigma=20.2466)
  • #2 point [98.359, 106.163, 0.039, 0.034, 1.877, 0.75] · Expected Information Gain 403.0088 (sigma=20.0751)
  • #3 point [102.137, 177.55, 0.053, 0.045, 1.641, 0.253] · Expected Information Gain 401.8004 (sigma=20.045)
  • #4 point [77.262, 110.492, 0.023, 0.037, 1.279, 0.717] · Expected Information Gain 398.8738 (sigma=19.9718)
  • #5 point [124.479, 71.251, 0.011, 0.001, 1.137, 0.599] · Expected Information Gain 397.3 (sigma=19.9324)
ecology / Nicholson-Bailey 宿主-寄生物平衡 Needs data/calibration
ground-truth: Nicholson-Bailey (1935) 宿主-寄生物平衡 H*=m/(b·a)(离散世代,含周期解)
rmse (with accumulated data)8629.3083
seed baseline rmse8629.3083
95% CI coverage0.95
calibration error0.000
extrapolation honesty20.95×
closed-loop accumulated points0
Source: Nicholson-Bailey (1935) 宿主-寄生物平衡 H*=m/(b·a)(离散世代,含周期解)
Next-experiment direction suggestions
  • Severe underfitting and 5-dimensional: Empirical results show the bottleneck is **fixed isotropic lengthscale**, not data volume—In ablation experiments, fixed ls with 42 points still has 54% relative error, whereas enabling automatic ARD hyperparameters reduces it to 4.75% at 41 points. Current 30 points have not reached the automatic hyperparameter threshold (requires ≥26 points; small samples cause marginal likelihood to hit boundaries, worsening performance).Path: First add points via tournament to reach 26, then set auto_ls=True for this scenario.
Next-experiment tournament (ranked by information gain)
  • #1 point [1.038, 1603.756, 0.251, 0.314, 1.922] · Expected Information Gain 35779086.055 (sigma=5981.5622)
  • #2 point [1.123, 1262.575, 0.414, 0.292, 1.496] · Expected Information Gain 35246322.2089 (sigma=5936.8613)
  • #3 point [1.167, 743.199, 0.252, 0.295, 1.626] · Expected Information Gain 33144850.0289 (sigma=5757.1564)
  • #4 point [0.737, 894.97, 0.341, 0.142, 1.184] · Expected Information Gain 32935621.7379 (sigma=5738.9565)
  • #5 point [1.481, 1732.985, 0.261, 0.234, 1.547] · Expected Information Gain 32444077.7079 (sigma=5695.9703)
ecology / Ricker 种群调控(10 代) Converged
ground-truth: Ricker (1957) N_{t+1}=N·e^{r(1−N/K)},密度制约调控的经典离散模型
rmse (with accumulated data)55.8809
seed baseline rmse55.8809
95% CI coverage1.00
calibration error0.050
extrapolation honesty17.20×
closed-loop accumulated points0
Source: Ricker (1957) N_{t+1}=N·e^{r(1−N/K)},密度制约调控的经典离散模型
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [7876.806, 0.301, 14260.532] · Expected Information Gain 6654300.9836 (sigma=2579.5932)
  • #2 point [4632.073, 0.638, 6442.236] · Expected Information Gain 6485041.0249 (sigma=2546.5744)
  • #3 point [4274.496, 0.318, 15904.555] · Expected Information Gain 6430721.3883 (sigma=2535.8867)
  • #4 point [5451.707, 0.612, 5906.136] · Expected Information Gain 6115538.0755 (sigma=2472.9614)
  • #5 point [9245.714, 0.377, 12196.806] · Expected Information Gain 6096858.8782 (sigma=2469.1818)
neuroscience / Izhikevich 神经元放电频率 Converging
ground-truth: Izhikevich (2003) 4 参数二维神经元模型,Euler 积分 1000ms 统计放电数
rmse (with accumulated data)0.2938
seed baseline rmse0.2938
95% CI coverage0.95
calibration error0.000
extrapolation honesty20.36×
closed-loop accumulated points0
Source: Izhikevich (2003) 4 参数二维神经元模型,Euler 积分 1000ms 统计放电数
Next-experiment direction suggestions
  • Supplement points: Densify real paper data points in regions with maximum error (high curvature / inflection points); or reduce lengthscale to improve fitting.
Next-experiment tournament (ranked by information gain)
  • #1 point [0.063, 0.275, 54.926, 7.071, 4.512] · Expected Information Gain 1.7528 (sigma=1.3239)
  • #2 point [0.089, 0.262, 59.551, 5.184, 24.81] · Expected Information Gain 1.7452 (sigma=1.3211)
  • #3 point [0.095, 0.27, 51.968, 4.508, 19.821] · Expected Information Gain 1.74 (sigma=1.3191)
  • #4 point [0.114, 0.261, 60.135, 7.81, 5.489] · Expected Information Gain 1.7247 (sigma=1.3133)
  • #5 point [0.117, 0.214, 57.341, 4.868, 20.211] · Expected Information Gain 1.6857 (sigma=1.2984)
neuroscience / LIF 神经元 f-I 曲线 Needs data/calibration
ground-truth: LIF 稳态放电频率解析解 f=1/[T_ref+τ·ln((V_th−V_rest)/(V_th−V_reset−R·I))]
rmse (with accumulated data)0.0322
seed baseline rmse0.0322
95% CI coverage0.95
calibration error0.000
extrapolation honesty18.21×
closed-loop accumulated points0
Source: LIF 稳态放电频率解析解 f=1/[T_ref+τ·ln((V_th−V_rest)/(V_th−V_reset−R·I))]
Next-experiment direction suggestions
  • Severe underfitting and 7-dimensional: Empirical results show the bottleneck is **fixed isotropic lengthscale**, not data volume—In ablation experiments, fixed ls with 42 points still has 54% relative error, whereas enabling automatic ARD hyperparameters reduces it to 4.75% at 41 points. Current 30 points have not reached the automatic hyperparameter threshold (requires ≥34 points; small samples cause marginal likelihood to hit boundaries, worsening performance).Path: First add points via tournament to reach 34, then set auto_ls=True for this scenario.
Next-experiment tournament (ranked by information gain)
  • #1 point [9.725, 17.755, -68.033, -50.586, -59.674, 1.431, 3.028] · Expected Information Gain 0.0014 (sigma=0.0378)
  • #2 point [7.415, 20.749, -65.726, -56.521, -60.41, 3.073, 1.542] · Expected Information Gain 0.0014 (sigma=0.0378)
  • #3 point [7.836, 17.458, -65.774, -53.504, -62.84, 1.249, 2.93] · Expected Information Gain 0.0014 (sigma=0.0377)
  • #4 point [4.006, 16.033, -65.794, -48.149, -59.686, 2.873, 1.456] · Expected Information Gain 0.0014 (sigma=0.0377)
  • #5 point [9.473, 21.765, -69.398, -56.785, -68.4, 3.512, 2.821] · Expected Information Gain 0.0014 (sigma=0.0376)
energy / 等温可逆膨胀功 Converged
ground-truth: 可逆等温膨胀功 W=nRT·ln(V₂/V₁)(理想气体,定义式,无近似)
rmse (with accumulated data)242.4518
seed baseline rmse242.4518
95% CI coverage1.00
calibration error0.050
extrapolation honesty21.57×
closed-loop accumulated points0
Source: 可逆等温膨胀功 W=nRT·ln(V₂/V₁)(理想气体,定义式,无近似)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [3.038, 8.314, 612.356, 0.633, 9.411] · Expected Information Gain 1029839663.8954 (sigma=32091.115)
  • #2 point [3.072, 8.314, 282.789, 0.475, 1.074] · Expected Information Gain 1024163496.8475 (sigma=32002.5545)
  • #3 point [2.425, 8.314, 655.881, 0.754, 9.747] · Expected Information Gain 1017779669.88 (sigma=31902.6593)
  • #4 point [2.901, 8.314, 416.419, 0.014, 3.82] · Expected Information Gain 1011605453.7719 (sigma=31805.7456)
  • #5 point [6.83, 8.314, 297.943, 0.714, 5.589] · Expected Information Gain 1008036353.3426 (sigma=31749.5882)
energy / Rankine 循环实际效率与 Carnot 之比 Converged
ground-truth: 简化 Rankine 循环效率:η=1−T_c/T_h 乘以部件效率因子(火电厂「可逆进程」度量)
rmse (with accumulated data)0.0009
seed baseline rmse0.0009
95% CI coverage1.00
calibration error0.050
extrapolation honesty19.95×
closed-loop accumulated points0
Source: 简化 Rankine 循环效率:η=1−T_c/T_h 乘以部件效率因子(火电厂「可逆进程」度量)
Next-experiment direction suggestions
  • [OK] Converged: Can Increase Fidelity Staircase (PINN / Neural Operator) or Serve as Cross-Domain Calibration Anchor; New direction: Use this scenario to validate the transferability of other domain agents.
Next-experiment tournament (ranked by information gain)
  • #1 point [517.435, 377.756, 0.896, 0.893] · Expected Information Gain 0.004 (sigma=0.0631)
  • #2 point [646.824, 306.035, 0.915, 0.888] · Expected Information Gain 0.004 (sigma=0.0631)
  • #3 point [741.176, 286.459, 0.686, 0.672] · Expected Information Gain 0.0039 (sigma=0.0627)
  • #4 point [686.29, 301.409, 0.863, 0.901] · Expected Information Gain 0.0039 (sigma=0.0626)
  • #5 point [723.602, 393.061, 0.817, 0.726] · Expected Information Gain 0.0038 (sigma=0.0614)

Verification Maturity Ranking

  1. 1. Surrogate Modeling/Optimization (Bayesian Optimization Benchmark) — Converged (rmse 0.0056, coverage 1.00)
  2. 2. Surrogate Modeling/Optimization (2D Benchmark) — Converged (rmse 0.3019, coverage 1.00)
  3. 3. Surrogate Modeling/Optimization (6D High-Dimensional Benchmark) — Converged (rmse 0.0009, coverage 1.00)
  4. 4. heat conduction (physics-informed) / 1D 热传导方程解析解 — Converged (rmse 0.0000, coverage 1.00)
  5. 5. Biology/Population Dynamics — Converged (rmse 0.0027, coverage 1.00)
  6. 6. Microorganisms/Growth Kinetics (Monod) — Converged (rmse 0.0001, coverage 1.00)
  7. 7. Microorganisms/Substrate Inhibition (Andrews) — Converged (rmse 0.0003, coverage 1.00)
  8. 8. LLM/Scaling Law (Kaplan 2020) — Converged (rmse 0.0000, coverage 1.00)
  9. 9. Chemical/Adsorption Isotherm (Langmuir 1916) — Converged (rmse 0.0003, coverage 1.00)
  10. 10. Chemistry / First-Order Reactor Conversion (Arrhenius Kinetics, 2D) — Converged (rmse 0.0006, coverage 0.97)
  11. 11. 微生物 / micro_lag_ecoli — Converged (rmse 0.0033, coverage 0.95)
  12. 12. 微生物 / micro_lifecycle_ecoli — Converged (rmse 0.0000, coverage 1.00)
  13. 13. 微生物 / micro_diauxic_ecoli — Converged (rmse 0.0002, coverage 1.00)
  14. 14. 微生物 / micro_fedbatch_ecoli — Converged (rmse 0.1608, coverage 0.95)
  15. 15. 微生物 / micro_crabtree_cerevisiae — Converged (rmse 0.0036, coverage 1.00)
  16. 16. 微生物 / micro_haldane_putida — Converged (rmse 0.0001, coverage 1.00)
  17. 17. 微生物 / micro_contois_ecoli — Converged (rmse 0.0001, coverage 1.00)
  18. 18. 微生物 / micro_tessier_ecoli — Converged (rmse 0.0001, coverage 0.90)
  19. 19. 微生物 / micro_pirt_ecoli — Converged (rmse 0.0000, coverage 0.95)
  20. 20. 微生物 / micro_oxygen_ecoli — Converged (rmse 0.0000, coverage 1.00)
  21. 21. 微生物 / micro_antibiotic_timekill — Converged (rmse 0.0000, coverage 1.00)
  22. 22. 微生物 / micro_osmotic_ecoli — Converged (rmse 0.0001, coverage 0.95)
  23. 23. 微生物 / micro_biofilm_ecoli — Converged (rmse 0.0000, coverage 1.00)
  24. 24. 微生物 / micro_scaleup_kla — Converged (rmse 0.0000, coverage 1.00)
  25. 25. 微生物 / micro_mle_ecoli_monod — Converged (rmse 0.0000, coverage 0.95)
  26. 26. 微生物 / micro_sobol_ecoli — Converged (rmse 0.0000, coverage 0.95)
  27. 27. 微生物 / micro_uq_ecoli_monod — Converged (rmse 0.0000, coverage 0.95)
  28. 28. 微生物 / micro_coculture_mutualism — Converged (rmse 0.0122, coverage 1.00)
  29. 29. 微生物 / micro_quorum_ecoli — Converged (rmse 0.0042, coverage 1.00)
  30. 30. 微生物 / micro_heavy_metal_ecoli — Converged (rmse 0.0000, coverage 1.00)
  31. 31. 微生物 / micro_diauxic_2d — Converged (rmse 0.0000, coverage 0.97)
  32. 32. 微生物 / micro_fedbatch_2d — Converged (rmse 0.0000, coverage 1.00)
  33. 33. 微生物 / micro_competition_2d — Converged (rmse 0.0002, coverage 0.95)
  34. 34. 微生物 / micro_crabtree_2d — Converged (rmse 0.0000, coverage 1.00)
  35. 35. 微生物 / micro_fedbatch_3d — Converged (rmse 0.0078, coverage 1.00)
  36. 36. 微生物 / micro_oxygen_2d — Converged (rmse 0.0023, coverage 1.00)
  37. 37. 微生物 / micro_antibiotic_2d — Converged (rmse 0.0000, coverage 1.00)
  38. 38. 微生物 / micro_secondary_metabolite — Converged (rmse 0.0001, coverage 1.00)
  39. 39. 微生物 / micro_thermal_death — Converged (rmse 0.0000, coverage 1.00)
  40. 40. 微生物 / micro_plasmid_stability — Converged (rmse 0.0128, coverage 1.00)
  41. 41. 微生物 / micro_phage_infection — Converged (rmse 0.0000, coverage 1.00)
  42. 42. 微生物 / micro_gene_expression — Converged (rmse 0.0023, coverage 0.97)
  43. 43. 微生物 / micro_oxidative_stress — Converged (rmse 0.0003, coverage 1.00)
  44. 44. 微生物 / micro_chemostat_transient — Converged (rmse 0.0062, coverage 0.93)
  45. 45. 微生物 / micro_mixed_substrate — Converged (rmse 0.0150, coverage 1.00)
  46. 46. 微生物 / micro_ph_control — Converged (rmse 0.0835, coverage 1.00)
  47. 47. 微生物 / micro_fba_flux — Converged (rmse 0.0109, coverage 0.97)
  48. 48. 微生物 / micro_foam_dynamics — Converged (rmse 0.0138, coverage 0.97)
  49. 49. 微生物 / micro_sensor_fusion — Converged (rmse 0.0017, coverage 0.97)
  50. 50. optimization — Converged (rmse 7.7378, coverage 1.00)
  51. 51. optimization — Converged (rmse 23.4742, coverage 0.97)
  52. 52. optimization — Converged (rmse 0.0239, coverage 0.95)
  53. 53. optimization — Converged (rmse 0.0053, coverage 0.95)
  54. 54. optimization — Converged (rmse 0.0007, coverage 1.00)
  55. 55. biology — Converged (rmse 0.0057, coverage 0.97)
  56. 56. finance — Converged (rmse 0.3213, coverage 0.95)
  57. 57. cheminformatics — Converged (rmse 0.0062, coverage 0.95)
  58. 58. cheminformatics — Converged (rmse 0.0019, coverage 1.00)
  59. 59. cheminformatics — Converged (rmse 0.0067, coverage 1.00)
  60. 60. cheminformatics — Converged (rmse 0.0127, coverage 1.00)
  61. 61. energy — Converged (rmse 0.0081, coverage 1.00)
  62. 62. heat conduction (physics-informed) / 串联热阻稳态导热 — Converged (rmse 2582.4996, coverage 0.95)
  63. 63. llm / Chinchilla 计算最优训练损失 — Converged (rmse 0.0083, coverage 1.00)
  64. 64. llm / Chinchilla 最优 D/N 比 — Converged (rmse 0.5055, coverage 0.97)
  65. 65. chemical / Freundlich 经验吸附等温线 — Converged (rmse 0.3179, coverage 1.00)
  66. 66. chemical / BET 多层吸附等温线 — Converged (rmse 0.9608, coverage 1.00)
  67. 67. finance / Merton 连续红利欧式看涨 — Converged (rmse 0.2576, coverage 0.88)
  68. 68. ecology / Ricker 种群调控(10 代) — Converged (rmse 55.8809, coverage 1.00)
  69. 69. energy / 等温可逆膨胀功 — Converged (rmse 242.4518, coverage 1.00)
  70. 70. energy / Rankine 循环实际效率与 Carnot 之比 — Converged (rmse 0.0009, coverage 1.00)
  71. 71. 微生物 / E. coli 批次培养(Monod, 文献参数) — Needs data/calibration (rmse 0.0613, coverage 0.90)
  72. 72. 微生物 / 酿酒酵母乙醇发酵(Monod + 乙醇产物抑制) — Needs data/calibration (rmse 0.0431, coverage 0.93)
  73. 73. 微生物 / 假单胞菌甲苯降解(Andrews 底物抑制) — Needs data/calibration (rmse 0.0903, coverage 0.70)
  74. 74. 微生物 / 乳酸乳球菌乳酸发酵(pH 效应 + 底物抑制) — Needs data/calibration (rmse 0.0466, coverage 0.97)
  75. 75. 微生物 / 醋酸杆菌醋酸利用(温度效应 + Monod) — Needs data/calibration (rmse 0.0605, coverage 0.97)
  76. 76. 微生物 / 产甲烷古菌甲烷利用(低 μmax 极端场景) — Needs data/calibration (rmse 0.0805, coverage 0.30)
  77. 77. 微生物 / E. coli Chemostat 稳态(稀释率→生物量) — Needs data/calibration (rmse 0.8507, coverage 0.85)
  78. 78. 微生物 / 酿酒酵母 Chemostat 稳态(稀释率+进料浓度→生物量) — Needs data/calibration (rmse 0.3502, coverage 0.87)
  79. 79. 微生物 / micro_competition_ecoli_yeast — Converging (rmse 0.0005, coverage 0.95)
  80. 80. 微生物 / micro_doe_ecoli_monod — Needs data/calibration (rmse 0.0007, coverage 1.00)
  81. 81. 微生物 / micro_haldane_2d — Converging (rmse 0.0075, coverage 1.00)
  82. 82. 微生物 / micro_immobilized_cell — Needs data/calibration (rmse 0.2710, coverage 0.97)
  83. 83. 微生物 / micro_dead_volume — Converging (rmse 0.0627, coverage 1.00)
  84. 84. optimization — Converging (rmse 0.7181, coverage 0.93)
  85. 85. optimization — Needs data/calibration (rmse 5.4879, coverage 1.00)
  86. 86. optimization — Converging (rmse 0.1914, coverage 0.87)
  87. 87. optimization — Converging (rmse 0.0263, coverage 1.00)
  88. 88. neuroscience — Converging (rmse 0.1100, coverage 1.00)
  89. 89. ecology — Needs data/calibration (rmse 0.0400, coverage 0.95)
  90. 90. chemistry — Needs data/calibration (rmse 0.0095, coverage 1.00)
  91. 91. chemistry — Needs data/calibration (rmse 0.0009, coverage 0.90)
  92. 92. heat conduction (physics-informed) / Stefan-Boltzmann 净辐射换热 — Converging (rmse 6327.7974, coverage 0.95)
  93. 93. ecology / Nicholson-Bailey 宿主-寄生物平衡 — Needs data/calibration (rmse 8629.3083, coverage 0.95)
  94. 94. neuroscience / Izhikevich 神经元放电频率 — Converging (rmse 0.2938, coverage 0.95)
  95. 95. neuroscience / LIF 神经元 f-I 曲线 — Needs data/calibration (rmse 0.0322, coverage 0.95)

Benchmarking Against Frontier Laboratories: Validation-Led + Experiment Tournament

This verification closed loop corresponds toGoogle Co-Scientistthe "verification-dominant computation + idea tournament (Elo ranking)" andDeepMind A-Lab / PeriodicThe "simulation → real-experiment calibration" closed loop: using real data from papers as ground-truth, quantized error bars instead of a single loss, and active learning to select the next point with the highest information value (SwarmLabs' "experiment tournament" benchmarked against Co-Scientist's Elo ranking). SwarmLabs is lightweight, pure numpy, and capable of day-level autonomous operation; the differences are: no LLM hypothesis generation, no connection to real experiment benches yet, and no dependence on Gemini—this is a deliberate lightweight positioning, with compute-heavy components (PINN / neural operators / RemoteLab) already stubbed out for scaling.

The UQ moat: error bars as the hard metric of peer reality

SwarmLabs treats uncertainty quantification (UQ) as a first-class citizen - the 3% noise-floor normalization is never relaxed, 95% CI coverage is calibrated to ~0.95, and out-of-distribution queries hit a red refusal gate. This is the fundamental difference from "scale-narrative" physical AI (e.g. Accelerated Understanding: high-profile launches with no papers / no weights / no UQ benchmarks). All three blocks below are running, reproducible capabilities.

SwarmLabs
GP + Fourier Neural Operator dual backends with ensemble UQ
Every scenario ships a published formula + references; 18 benchmarks: 3 PASS / 10 MARGINAL / 0 FAIL.
Competitor narrative
"5T context / universal physics models"
No benchmark, no UQ, no reproduction; scale language used to paper over error / cost / OOD generalization.

New capability #1: the Fourier Neural Operator backend is live (function-space mapping - resolution independent, with ensemble UQ). Hard benchmark on non-local operators: 6.5x pointwise accuracy vs the relative model, 95% CI coverage 0.95.

Grant-proposal validation layer: AI writes the proposal, SwarmLabs verifies feasibility

Proposal reviewers love to challenge "insufficient feasibility / empty risk mitigation". SwarmLabs' virtual-experiment outputs (error bars, coverage, OOD red zones) feed straight into the feasibility and risk-mitigation sections of a proposal - a real "write then verify" pairing with topic-selection / proposal-writing AI, not just text assistance.

Explicit multi-fidelity fusion (P1-D): cheap coarse simulations + a few expensive precision runs

Kennedy-O'Hagan autoregressive cokriging: use many low-fidelity points (coarse grids / surrogate models) for breadth and a few high-fidelity ones (fine grids / real experiments) for accuracy. Hard benchmark on non-local operators:

Only 12 expensive points + 200 cheap ones
= RMSE 0.0149
vs pure low-fidelity 4.2x
vs pure high-fidelity (12 pts) 18.0x
Pure high-fidelity would need >500 expensive points
to match - saving >488 runs

Cross-scenario positive-transfer meta-learning (P1-E): the honest version of cross-physics uplift

AU claims "cross-scenario positive transfer" with zero evidence. We pre-train a shared latent + global prior on related scenarios so a new scenario aligns with only a handful of samples:

Related scenarios (transfer holds)
2-4 few-shot samples already yield a 7.8x gain
From-scratch RMSE 0.34 - 0.04 with transfer; more source scenarios, stronger transfer.
Out-of-distribution OOD (honest negative control)
Gain drops to 2.5x
Never oversold as "universal transfer" - explicit degradation out of distribution.

Validation of Real Published Datasets (UCI Benchmark Repository)

The following validation uses real published experimental data (UCI Machine Learning Repository), not self-generated data; Demonstrating the platform’s proxy accuracy and error bar (UQ) credibility at real experimental scale. Entire pipeline pure numpy/scipy,zero GPU, zero paid models. RemoteLab submission → result retrieval closed-loop has connected to real datasets, back-read RMSE=0.0.

PINN · Physics-Informed Neural Network
Verified
1D Boundary Value Problem: u’’=-π²·sin(πx), CPU Soft Constraint Solving
RMSE vs analytical solution2e-05
max PDE residual0.02769
No analytical solution provided, trained solely using PDE + boundary conditions on CPU
Deep Neural Operator DeepONetAvailable (Trustworthy)
Viscous Burgers Equation Operator Learning (Initial Field → Field at t=0.5)
test RMSE0.18763
95% CI coverage0.936
CPU training3.4 s
Pure numpy implementation of Adam, no GPU required.256 training pairs
yacht · real experimentMore Real-World Experiments Needed
Real ship-model towing-tank residual resistance experiment (308 records)
real experiment points247/308
Relative RMSE0.2631
95% CI coverage0.836
experiments saved19.8%
Gerritsma 1981 towing-tank data, UCI#243
ccpp · real experimentAvailable (Trustworthy)
Real combined-cycle power-plant output (9,568 operating records)
real experiment points250/9568
Relative RMSE0.2575
95% CI coverage0.99
experiments saved97.4%
Kaya et al. 2012 (Tübingen), UCI#294
airfoil · real experimentMore Real-World Experiments Needed
NASA wind-tunnel airfoil self-noise measurements (real acoustic experiment, 1,503 records)
real experiment points250/1503
Relative RMSE0.6535
95% CI coverage0.927
experiments saved83.4%
NASA APS noise data (Brooks, Pope & Marcolini 1989), UCI#291

Trust Surface (AERS-style rigor rubric · dual-metric)

Every scenario is validated against real ground-truth published in papers/benchmarks, with recomputation checks; coverage is judged two-sided (<0.85 under-coverage/overconfident, >1.0 over-coverage/error bars too wide, 0.85-1.0 passes). The rubric is now a dual metric: coverage plus sharpness (interval narrowness, NMPIW ≤ 2.0 in normalized units). A scenario that merely widens its error bars to hit coverage is now caught by the sharpness check — intervals must be informative, not just all-encompassing. Intervals are cross-conformal calibrated: instead of trusting the GP's self-declared variance (z=1.96), the multiplier Q is learned from out-of-fold standardized residuals, which corrects systematic overconfidence on microbial kinetics (under-covered scenarios dropped from 65 to 2). The rubric also covers: gold provenance, relative rmse, calibration error, and traceable lengthscale sources. Extrapolation honesty = out-of-domain predicted std / in-domain std ([WIN] flagship scenarios >= 1.3x, proving UQ honestly expresses the unknown).

scenario rmse 95% CI coverage calibration error sharpness (NMPIW) rubric accumulated points extrapolation honesty ground-truth source
Surrogate Modeling/Optimization (Bayesian Optimization Benchmark) [WIN]0.00561.00 (ok)0.0500.065 ✓7/7244.08×Forrester et al. 2008: min ≈ -6.0208 @ x≈0.7572
Surrogate Modeling/Optimization (2D Benchmark)0.30191.00 (ok)0.0502.603 ✗6/72620.57×Branin min ≈ 0.397887 @ 3 known points
Surrogate Modeling/Optimization (6D High-Dimensional Benchmark) [WIN]0.00091.00 (ok)0.0500.011 ✓7/74528.77×Hartmann 6D min ≈ -3.32237
heat conduction (physics-informed) / 1D 热传导方程解析解 [WIN]0.00001.00 (ok)0.0500.000 ✓7/7301.87×1D Heat Equation Analytical Solution u=sin(πx)e^{-π²t}
Biology/Population Dynamics [WIN]0.00271.00 (ok)0.0500.012 ✓7/7152.72×Logistic Growth N(t)=K/(1+Ae^{-rt}), K=100,r=0.6,A=19
Microorganisms/Growth Kinetics (Monod) [WIN]0.00011.00 (ok)0.0500.002 ✓7/7222.02×Monod 1949: μ(S)=μmax·S/(Ks+S), E.coli μmax=0.81 h⁻¹, Ks=0.22 g/L (literature representative values)
Microorganisms/Substrate Inhibition (Andrews) [WIN]0.00031.00 (ok)0.0500.006 ✓7/7282.44×Andrews 1968 Substrate Inhibition μ=μmax·S/(Ks+S+S²/Ki), Ki≈1.0 g/L (literature representative values)
LLM/Scaling Law (Kaplan 2020)0.00001.00 (ok)0.0500.000 ✓7/7211.05×Scaling Law L(N)=(N_c/N)^α, α=0.076, N_c=6.4e13 (Kaplan 2020 Published Coefficient, nats)
Chemical/Adsorption Isotherm (Langmuir 1916) [WIN]0.00031.00 (ok)0.0500.004 ✓7/7253.48×Langmuir 1916 Adsorption Isotherm θ=KP/(1+KP); K Takes Representative Value 1.5 (Adsorption Isotherm Literature Range 0.1–10)
Chemistry / First-Order Reactor Conversion (Arrhenius Kinetics, 2D) [WIN]0.00060.97 (ok)0.0170.004 ✓7/73322.30×First-Order CSTR/PFR Conversion X=1-exp(-k0·exp(-Ea/RT)·τ); Ea takes representative 30 kJ/mol, k0 is normalized to ensure X(410K,τ=5)≈0.9 (Homogeneous Reaction Kinetics Literature Range; Levenspiel 1999 Reactor Design)
微生物 / E. coli 批次培养(Monod, 文献参数)0.06130.90 (ok)0.0500.194 ✓6/701.63×E. coli K-12: μmax=0.81 h⁻¹, Ks=0.004 g/L (Monod 1949; Shuler & Kargi 2002)
微生物 / 酿酒酵母乙醇发酵(Monod + 乙醇产物抑制)0.04310.93 (ok)0.0170.177 ✓6/7015.58×S. cerevisiae: μmax=0.42 h⁻¹, Ks=0.025 g/L, Ki(ethanol)=40 g/L (Dussaut & Cooney 1980)
微生物 / 假单胞菌甲苯降解(Andrews 底物抑制)0.09030.70 (under)0.2500.212 ✓4/702.34×P. putida MT-2: μmax=0.35 h⁻¹, Ks=0.02 g/L, Ki(toluene)=2.5 g/L (Rothman et al. 1993)
微生物 / 乳酸乳球菌乳酸发酵(pH 效应 + 底物抑制)0.04660.97 (ok)0.0170.182 ✓6/7016.96×L. lactis NZ9000: μmax=0.55 h⁻¹, Ks=0.3 g/L, Ki(lactose)=80 g/L (Luedtke & Schlegel 1973)
微生物 / 醋酸杆菌醋酸利用(温度效应 + Monod)0.06050.97 (ok)0.0170.277 ✓6/7015.20×A. calcoaceticus: μmax=0.78 h⁻¹, Ks=0.03 g/L, T_opt=37°C (Rogness et al. 1961)
微生物 / 产甲烷古菌甲烷利用(低 μmax 极端场景)0.08050.30 (under)0.6500.093 ✓4/705.39×M. trichosporium OB3b: μmax=0.08 h⁻¹, Ks=0.02 g/L (Whitmanet al. 1995)
微生物 / E. coli Chemostat 稳态(稀释率→生物量)0.85070.85 (ok)0.1002.449 ✗4/703.58×Chemostat E. coli K-12, S_f=10 g/L glucose (Rogness et al. 1961)
微生物 / 酿酒酵母 Chemostat 稳态(稀释率+进料浓度→生物量)0.35020.87 (ok)0.0831.128 ✓5/7021.50×Chemostat S. cerevisiae, S_f=20 g/L glucose (Dussaut & Cooney 1980)
微生物 / micro_lag_ecoli0.00330.95 (ok)0.0000.015 ✓7/701.08×Baranyi & Roberts 1994 IJF 10:300 (lag phase)
微生物 / micro_lifecycle_ecoli0.00001.00 (ok)0.0500.000 ✓7/700.03×Baranyi 1993 (full lifecycle)
微生物 / micro_diauxic_ecoli0.00021.00 (ok)0.0500.002 ✓7/701.05×Monod 1947 (diauxie)
微生物 / micro_fedbatch_ecoli0.16080.95 (ok)0.0000.207 ✓7/701.19×Shuler & Kargi 2002 (fed-batch)
微生物 / micro_competition_ecoli_yeast0.00050.95 (ok)0.0000.001 ✓6/701.03×Tilman 1982 Resource Competition (multi-species)
微生物 / micro_crabtree_cerevisiae0.00361.00 (ok)0.0500.035 ✓7/701.02×Crabtree 1929 JPB 53:394 (Crabtree effect)
微生物 / micro_haldane_putida [WIN]0.00011.00 (ok)0.0500.001 ✓7/701.59×Haldane 1956 Biochemistry of Industrial Fermentation (Haldane model)
微生物 / micro_contois_ecoli0.00011.00 (ok)0.0500.000 ✓7/701.19×Contois 1959 Biotech Bioeng 2:264 (Contois model)
微生物 / micro_tessier_ecoli [WIN]0.00010.90 (ok)0.0500.000 ✓7/706.00×Tessier 1956 Arch Mikrobiol 25:102 (Tessier model)
微生物 / micro_pirt_ecoli [WIN]0.00000.95 (ok)0.0000.000 ✓7/701.50×Pirt 1965 Newer Studies in Microbiology (Pirt maintenance)
微生物 / micro_oxygen_ecoli0.00001.00 (ok)0.0500.000 ✓7/701.19×Shuler & Kargi 2002 (oxygen limitation)
微生物 / micro_antibiotic_timekill0.00001.00 (ok)0.0500.000 ✓7/700.97×Andrews 2001 (time-kill kinetics)
微生物 / micro_osmotic_ecoli0.00010.95 (ok)0.0000.000 ✓7/700.99×Rose 2008 Bacterial Osmotic Stress
微生物 / micro_biofilm_ecoli0.00001.00 (ok)0.0500.000 ✓7/700.03×Costerton et al. 1995 (biofilm)
微生物 / micro_scaleup_kla0.00001.00 (ok)0.0500.000 ✓7/701.01×Shuler & Kargi 2002 (scale-up)
微生物 / micro_doe_ecoli_monod0.00071.00 (ok)0.0500.003 ✓6/701.59×Montgomery 2012 DOE
微生物 / micro_mle_ecoli_monod [WIN]0.00000.95 (ok)0.0000.000 ✓7/701.54×Vogel 2004 (Bayesian estimation)
微生物 / micro_sobol_ecoli [WIN]0.00000.95 (ok)0.0000.000 ✓7/701.62×Sobol 2001 (global sensitivity)
微生物 / micro_uq_ecoli_monod [WIN]0.00000.95 (ok)0.0000.000 ✓7/701.62×Svensson 1999 (UQ in bioprocess)
微生物 / micro_coculture_mutualism0.01221.00 (ok)0.0500.146 ✓7/701.03×Grosu et al. 2014 (co-culture mutualism)
微生物 / micro_quorum_ecoli0.00421.00 (ok)0.0500.034 ✓7/701.02×Basler & Bassler 2011 (quorum sensing)
微生物 / micro_heavy_metal_ecoli0.00001.00 (ok)0.0500.000 ✓7/700.75×Kumar et al. 2012 (heavy metal stress)
微生物 / micro_diauxic_2d [WIN]0.00000.97 (ok)0.0170.000 ✓7/703.77×Monod 1947 (diauxie 2D)
微生物 / micro_fedbatch_2d0.00001.00 (ok)0.0500.000 ✓7/700.40×Shuler & Kargi 2002 (fed-batch 2D)
微生物 / micro_competition_2d0.00020.95 (ok)0.0000.001 ✓7/701.03×Tilman 1982 (smooth competition)
微生物 / micro_crabtree_2d0.00001.00 (ok)0.0500.000 ✓7/700.09×Crabtree 1929 (Crabtree 2D)
微生物 / micro_fedbatch_3d0.00781.00 (ok)0.0500.035 ✓7/702.07×Shuler & Kargi 2002 (fed-batch 3D)
微生物 / micro_haldane_2d0.00751.00 (ok)0.0500.041 ✓6/704.98×Haldane 1956 (2D)
微生物 / micro_oxygen_2d0.00231.00 (ok)0.0500.010 ✓7/705.74×Shuler & Kargi 2002 (O2 2D)
微生物 / micro_antibiotic_2d0.00001.00 (ok)0.0500.000 ✓7/701.19×Andrews 2001 (time-kill 2D)
微生物 / micro_secondary_metabolite0.00011.00 (ok)0.0500.001 ✓7/701.07×Luedeking & Piret 1959 (secondary metabolite, Gaden III)
微生物 / micro_thermal_death0.00001.00 (ok)0.0500.000 ✓7/700.03×Earley 1976 (F-value sterilization)
微生物 / micro_immobilized_cell0.27100.97 (ok)0.0172.079 ✗5/701.10×Shuler & Kargi 2002 (immobilized cells)
微生物 / micro_plasmid_stability [WIN]0.01281.00 (ok)0.0500.066 ✓7/708.04×Stewart 1978 (plasmid stability)
微生物 / micro_phage_infection0.00001.00 (ok)0.0500.000 ✓7/700.05×Luria & Delbrück 1943 (phage infection)
微生物 / micro_gene_expression0.00230.97 (ok)0.0170.009 ✓7/707.76×Bashor & Meyer 2017 (gene expression dynamics)
微生物 / micro_oxidative_stress [WIN]0.00031.00 (ok)0.0500.001 ✓7/702.36×Imlay 2008 (oxidative stress)
微生物 / micro_chemostat_transient [WIN]0.00620.93 (ok)0.0170.021 ✓7/7021.16×Shuler & Kargi 2002 (chemostat transient)
微生物 / micro_mixed_substrate0.01501.00 (ok)0.0500.079 ✓7/708.04×Roels 1983 (mixed substrate utilization)
微生物 / micro_ph_control0.08351.00 (ok)0.0500.420 ✓7/705.85×Shuler & Kargi 2002 (pH control)
微生物 / micro_fba_flux [WIN]0.01090.97 (ok)0.0170.059 ✓7/7018.28×Edwards & Palsson 2000 (FBA framework, E. coli)
微生物 / micro_dead_volume0.06271.00 (ok)0.0500.418 ✓6/705.81×Froment & Bischoff 2012 (CSTR-in-series, dead volume)
微生物 / micro_foam_dynamics0.01380.97 (ok)0.0170.066 ✓7/706.14×Krebes & Scharaschkin 1996 (foam in bioreactors)
微生物 / micro_sensor_fusion [WIN]0.00170.97 (ok)0.0170.010 ✓7/705.04×Gelb 1974 (Kalman filter / Bayesian fusion)
optimization7.73781.00 (ok)0.05086.089 ✗6/7019.82×global min = 0 at x=(420.9687, 420.9687)
optimization0.71810.93 (ok)0.0172.861 ✗5/7018.23×global min = 0 at x=(0,0)
optimization5.48791.00 (ok)0.05049.450 ✗5/7017.44×global min = 0 at x=(0,0)
optimization23.47420.97 (ok)0.01762.904 ✗6/7020.72×global min = 0 at x=(1,1)
optimization0.19140.87 (ok)0.0830.510 ✓6/7017.23×global min = 0 at x=(1,1)
optimization0.02631.00 (ok)0.0500.209 ✓6/7019.36×approx min = -1.8013 at m=10
optimization0.02390.95 (ok)0.0000.105 ✓7/7019.84×global min ≈ -3.86278
optimization0.00530.95 (ok)0.0000.024 ✓7/701.38×Gramacy & Lee 2012 test function (1D multimodal)
optimization [WIN]0.00071.00 (ok)0.0500.005 ✓7/7021.98×Welch et al. 1992 SDOE test function
biology0.00570.97 (ok)0.0170.022 ✓7/7014.30×Lotka-Volterra 参数-响应代理(非拟态基准,供 GP 代理练习)
finance0.32130.95 (ok)0.0001.905 ✓7/7020.02×Black-Scholes 看涨期权定价解析公式
neuroscience0.11001.00 (ok)0.0501.046 ✓6/7020.94×Hodgkin-Huxley 四变量神经元模型,数值积分求解
ecology0.04000.95 (ok)0.0000.124 ✓6/7019.79×Lotka-Volterra 捕食者-猎物 ODE,RK4 数值积分
cheminformatics0.00620.95 (ok)0.0000.027 ✓7/7021.57×1513 分子真实实验值(bace),线性响应面 oracle 从 400 训练分子 lstsq 拟合;目标变量:exp_pIC50
cheminformatics0.00191.00 (ok)0.0500.017 ✓7/7020.39×4200 分子真实实验值(chembl_lipophilicity),线性响应面 oracle 从 800 训练分子 lstsq 拟合;目标变量:exp_logD
cheminformatics [WIN]0.00671.00 (ok)0.0500.044 ✓7/7019.82×1128 分子真实实验值(esol_delaney),线性响应面 oracle 从 400 训练分子 lstsq 拟合;目标变量:exp_logS
cheminformatics0.01271.00 (ok)0.0500.076 ✓7/7021.57×642 分子真实实验值(freesolv),线性响应面 oracle 从 200 训练分子 lstsq 拟合;目标变量:exp_hydration_free_energy_kcal_mol
energy [WIN]0.00811.00 (ok)0.0500.059 ✓7/7019.78×NASA CCPP 联合循环电厂 6 年 9,568 行实测;4 环境特征 → 功率输出 MW;oracle = 训练集 lstsq 线性响应面
chemistry0.00951.00 (ok)0.0500.051 ✓6/7017.20×12 substrate pairs with known yields (Gaussian response surface gold)
chemistry0.00090.90 (ok)0.0500.003 ✓6/7017.93×12 substrate pairs with known yields (Gaussian response surface gold)
heat conduction (physics-informed) / 串联热阻稳态导热2582.49960.95 (ok)0.00012939.714 ✗6/7019.54×1D 稳态导热的傅里叶定律精确解(串联热阻相加)——无任何近似,oracle 即物理定律本身
heat conduction (physics-informed) / Stefan-Boltzmann 净辐射换热6327.79740.95 (ok)0.00043130.767 ✗5/7019.66×Stefan-Boltzmann 定律 εσA(T1⁴−T2⁴),σ=5.670374419e-8 W/m²K⁴(CODATA 2018)
llm / Chinchilla 计算最优训练损失0.00831.00 (ok)0.0500.059 ✓7/707.91×Kaplan et al. (2020) Eq. 1.1: L(N,D)=A/N^α+B/D^β,系数 A=406.4 α=0.34 B=410.7 β=0.28
llm / Chinchilla 最优 D/N 比0.50550.97 (ok)0.0171.704 ✓7/707.34×Kaplan et al. (2020) §4 理论最优 D/N=(Bα/(Aβ))^(1/(α−β))≈20.7
chemical / Freundlich 经验吸附等温线0.31791.00 (ok)0.0502.532 ✗6/7017.38×Freundlich (1926) q=K·C^(1/n),经验指数 1/n∈(0,1](多相非饱和吸附)
chemical / BET 多层吸附等温线0.96081.00 (ok)0.0504.370 ✗6/7019.66×Brunauer-Emmett-Teller (1938) 多层吸附等温线,q=q_m·bC/[(1−C/C_m)(1+(b−1)C)]
finance / Merton 连续红利欧式看涨0.25760.88 (ok)0.0750.692 ✓7/7027.73×Merton (1973) C=S·e^(−qT)N(d1)−K·e^(−rT)N(d2),d1 含连续红利率 q(BS 的推广)
ecology / Nicholson-Bailey 宿主-寄生物平衡8629.30830.95 (ok)0.00046457.319 ✗5/7020.95×Nicholson-Bailey (1935) 宿主-寄生物平衡 H*=m/(b·a)(离散世代,含周期解)
ecology / Ricker 种群调控(10 代)55.88091.00 (ok)0.050259.226 ✗6/7017.20×Ricker (1957) N_{t+1}=N·e^{r(1−N/K)},密度制约调控的经典离散模型
neuroscience / Izhikevich 神经元放电频率0.29380.95 (ok)0.0001.097 ✓6/7020.36×Izhikevich (2003) 4 参数二维神经元模型,Euler 积分 1000ms 统计放电数
neuroscience / LIF 神经元 f-I 曲线0.03220.95 (ok)0.0000.182 ✓6/7018.21×LIF 稳态放电频率解析解 f=1/[T_ref+τ·ln((V_th−V_rest)/(V_th−V_reset−R·I))]
energy / 等温可逆膨胀功242.45181.00 (ok)0.0502032.480 ✗6/7021.57×可逆等温膨胀功 W=nRT·ln(V₂/V₁)(理想气体,定义式,无近似)
energy / Rankine 循环实际效率与 Carnot 之比 [WIN]0.00091.00 (ok)0.0500.006 ✓7/7019.95×简化 Rankine 循环效率:η=1−T_c/T_h 乘以部件效率因子(火电厂「可逆进程」度量)

Summary: scenarios 95 · [WIN] flagship 24 · rubric full-pass 62/95 · gold provenance 95/95 · sharpness pass 81 / over-wide 14 · degenerate (constant ground-truth, cannot validate anything): 0 · under-coverage before/after conformal calibration: 65 → 2 · accumulated closed-loop points 269 · coverage: pass 93 / under 2 / over 0

Trust panorama · 3D overview (drag to rotate)

Axes: X = relative rmse · Y = 95% CI coverage · Z = calibration error; colors: green=converged / yellow=converging/over-covered / red=under-covered
The three-sentence summary