{"id":"bayesian-active-learning","title":"Bayesian Active Learning for Potential Selection","subtitle":"Gaussian Process surrogates across 23 potentials × 12,000 materials.","category":"methods","tags":["bayesian","active-learning"],"source":"articles/docs/bayesian_active_learning_report.md","lang":"en","words":958,"readMinutes":4,"toc":[{"depth":2,"text":"Kimi Deep Research Report — Comprehensive Literature Review & Practical Guide","id":"kimi-deep-research-report-comprehensive-literature-review-practical-guide"},{"depth":2,"text":"Report Structure (10 Major Sections, 22+ Tables)","id":"report-structure-10-major-sections-22-tables"},{"depth":3,"text":"1. Problem Formulation and Theoretical Foundations","id":"1-problem-formulation-and-theoretical-foundations"},{"depth":3,"text":"2. Bayesian Model Averaging (BMA) for Multi-Potential Ensembles","id":"2-bayesian-model-averaging-bma-for-multi-potential-ensembles"},{"depth":3,"text":"3. Gaussian Process Surrogates for Error Prediction","id":"3-gaussian-process-surrogates-for-error-prediction"},{"depth":3,"text":"4. Active Learning and Query-by-Committee for DFT Prioritization","id":"4-active-learning-and-query-by-committee-for-dft-prioritization"},{"depth":3,"text":"5. Multi-Fidelity Bayesian Optimization","id":"5-multi-fidelity-bayesian-optimization"},{"depth":3,"text":"6. Domain-Specific Considerations for Crystalline Materials","id":"6-domain-specific-considerations-for-crystalline-materials"},{"depth":3,"text":"7. Gaussian Process Architecture Design for GLIM","id":"7-gaussian-process-architecture-design-for-glim"},{"depth":3,"text":"8. Literature Synthesis: Key References and Methods","id":"8-literature-synthesis-key-references-and-methods"},{"depth":3,"text":"9. Comparative Analysis of Acquisition Strategies (Tables 18-21)","id":"9-comparative-analysis-of-acquisition-strategies-tables-18-21"},{"depth":3,"text":"10. Implementation Roadmap","id":"10-implementation-roadmap"},{"depth":2,"text":"Key Findings for GLIM","id":"key-findings-for-glim"}],"html":"

Bayesian Active Learning for Interatomic Potential Selection

Kimi Deep Research Report — Comprehensive Literature Review & Practical Guide

Tags: Bayesian Optimization | Active Learning | Materials Science | GLIM

\n
\n

Report Structure (10 Major Sections, 22+ Tables)

1. Problem Formulation and Theoretical Foundations

1.1 The GLIM Benchmark Challenge

1.1.1 Dataset Characteristics: 23 Potentials x 12,000 Materials

The GLIM benchmark represents a paradigm shift in computational materials science, providing systematic evaluation of 23 distinct interatomic potentials across approximately 12,000 materials from the JARVIS-FF database. This massive-scale comparison creates a performance matrix of 276,000 potential-material combinations for each target property. The 23 potentials encompass a diverse methodological spectrum: classical empirical potentials (EAM, MEAM, ReaxFF), machine learning interatomic potentials (SNAP, GAP, neural network potentials), and recent universal pre-trained models (M3GNet, CHGNet, MACE-MP0).

\n

Table 1: Taxonomy of interatomic potentials in the GLIM benchmark, illustrating the cost-accuracy spectrum that motivates multi-fidelity optimization.

\n
1.1.2 Target Properties: Bulk Modulus (K), Shear Moduli (C', C44)

Table 2: Target elastic constants in the GLIM benchmark, showing their complementary information content and varying sensitivity to potential quality.

\n
1.1.3 Dimensionality Reduction: PCA Analysis with 55% First-Component Variance

The PCA decomposition operates on the centered prediction error matrix, where entries represent deviations from DFT reference values. The 55% variance capture is remarkably high for such heterogeneous data, indicating that potential performance is not randomly distributed but follows predictable patterns exploitable for selection.

\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Principal ComponentVariance ExplainedInterpretationImplications for Selection
PC155%Global "potential quality" / "material difficulty" axisCoarse ranking; identifies broadly good/bad potentials
PC2~15%Ionic vs. covalent bonding discriminationSeparates potential families by bonding type
PC3~10%Crystal structure effects (fcc vs. bcc vs. hcp)Structure-specific performance patterns
PC4-5~5% eachProperty-specific effectsFine-grained, property-dependent selection
1.1.4 Core Decision Problem: Material-Specific Potential Trustworthiness

Table 4: Taxonomy of uncertainty sources in potential selection, distinguishing reducible (epistemic) from irreducible (aleatoric) components.

\n

1.2 Statistical Framework for Model Selection

\n
\n

2. Bayesian Model Averaging (BMA) for Multi-Potential Ensembles

2.1 Classical BMA Theory Adapted to Interatomic Potentials

\n

Specialist behavior is valuable for application-targeted selection — the same material may warrant different potentials for bulk modulus screening versus dislocation modeling.

\n

2.2 Implementation Strategies for Interatomic Potential Ensembles

2.2.1 Training Set Likelihood Evaluation Across JARVIS-FF Materials

Practical BMA implementation requires efficient likelihood computation at GLIM scale. The heteroscedastic error model:

\n

log p(y_i | m, M_k) ~ -(y_i - y_ki)^2 / (2 sigma_k^2(m_i)) - log sigma_k(m_i)

\n

The material-dependent variance sigma_k^2(m) is learned via GP regression on descriptors, enabling location-adaptive uncertainty.

\n
\n

3. Gaussian Process Surrogates for Error Prediction

\n
\n

4. Active Learning and Query-by-Committee for DFT Prioritization

4.1 Ensemble Disagreement as Uncertainty Measure

4.1.2 Jensen-Shannon Divergence for Multi-Potential Comparison

JSD(m) = (1/K) sum_k KL(p_k(.|m) || p_ens(.|m))

\n

Captures both mean and variance disagreement, identifying materials where potentials differ in confidence as well as point estimates.

\n
4.1.3 Committee Variance as Proxy for Epistemic Uncertainty

sigma^2_committee(m) = (1/(K-1)) sum_k (y_k(m) - y_bar(m))^2

\n

Decomposition: Total error^2 = Bias^2 + sigma^2_committee + sigma^2_aleatoric. The committee variance term is reducible epistemic uncertainty — the target of active learning.

\n

4.2 Uncertainty-Driven DFT Prioritization

4.2.1 Maximum Expected Prediction Error (MEPE) Criteria

alpha_MEPE(m) = mu^2_bias(m) + sigma^2_committee(m)

\n\n
\n

5. Multi-Fidelity Bayesian Optimization

\n
\n

6. Domain-Specific Considerations for Crystalline Materials

\n
\n

7. Gaussian Process Architecture Design for GLIM

\n

Table 17: Kernel selection for GP surrogates in materials applications.

\n
\n

8. Literature Synthesis: Key References and Methods

\n
\n

9. Comparative Analysis of Acquisition Strategies (Tables 18-21)

\n

10. Implementation Roadmap

10.1 Software Ecosystem

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
ToolRoleKey Features
BoTorchBayesian optimizationGPU-accelerated, modular acquisition functions
GPyTorchGP modelingScalable exact/approximate inference
ASEAtomistic simulationCalculator interface for all potentials
pymatgenStructure analysisComposition/structure featurization
MatminerFeature extractionMagpie, SOAP, structure featurizers
AiiDA/FireWorksWorkflow orchestrationProvenance, distributed execution

10.2 End-to-End Workflow Design

Input: Material m with descriptors phi(m)\n1. Predict: mu_k(m), sigma_k(m) for all k potentials via GP surrogate\n2. Select: (m, k*) = argmax acquisition(mu, sigma, cost)\n3. Evaluate: Run potential k* on material m, observe y\n4. Update: Incorporate (m, k*, y) into GP posterior\n5. Recommend: Return best predicted potential or BMA prediction\n
\n

10.3 Validation and Benchmarking Protocols

\n

10.4 Future Directions

\n
\n

Key Findings for GLIM

    \n
  1. Ensemble variance from 23 potentials is the most promising acquisition criterion for active learning
    2. \n
  2. Heteroscedastic GP surrogates with Matern-5/2 kernels provide optimal balance for error prediction
    3. \n
  3. BMA with property-dependent weights captures the specialist nature of potentials
    4. \n
  4. Multi-fidelity optimization exploits the cost hierarchy via alpha-UA acquisition functions
    5. \n
  5. Committee variance decomposes into bias + epistemic + aleatoric components
    6. \n
  6. Leave-one-material-out cross-validation is the primary validation protocol
    7. \n
\n
\n

Report extracted from Kimi Deep Research. March 28, 2026.\nNote: Partial extraction (~40% of 55K original) due to content filtering. Full report accessible in browser tab 1899282409.

\n"}