Skip to content

API Reference

The API reference is generated from the package docstrings with mkdocstrings. Use this section when you need exact function signatures, parameters, return types, and class methods.

Public Workflow Modules

  • Estimators: scikit-learn-style single-task, physics-informed, and multitask GPR estimator classes.
  • Featurizers: scikit-learn-style materials transformers.
  • Kernels: Tanimoto, element-fraction, structure-feature, and feature-subset kernel helpers.
  • Target Transforms: positive, bounded, standardized, physics-residual, and materials-property preset helpers.
  • Physics Constraints: soft virtual observations for known limits and monotonic trends.
  • Derivative-Constrained GPR: exact RBF GPR with function-value and derivative observations.
  • Noise Models: source-dependent, replicate-aware, and feature-dependent observation-noise profiles.
  • Heteroscedastic GPR: learned input-dependent observation noise with a residual-noise GP.
  • Multitask GPR: correlated multi-output GPR for complete materials-property target matrices.
  • Multi-Fidelity GPR: delta correction models for low-/high-fidelity materials-property data.
  • Sparse Multitask GPR: correlated multi-output GPR for incomplete target matrices with missing task observations.
  • Physics Equation Templates: reusable Arrhenius, power-law, Hall-Petch, free-volume, and mixture mean equations.
  • Safe Custom Equations: JSON-safe custom-equation schema records, safe expression validation, NumPy previews, torch evaluation, and trainable PI-GPR mean adapters for user-defined safe equation specs.
  • Physics-Informed GPR: GPyTorch training, prediction, and mean-function utilities.
  • Validation: train/test evaluation, cross-validation, and configurable learning-curve utilities.
  • Candidate Generation: finite chemistry, composition, formulation, and processing-condition candidate pools.
  • Bayesian Optimization: BoTorch finite-pool candidate ranking for next-experiment selection.
  • BO Benchmarking: offline finite-pool strategy replay against known outcomes.
  • Multi-Objective Selection: Pareto-front and weighted finite-pool ranking helpers.
  • Experiment Logging: closed-loop recommendation, selection, observation, and campaign-summary logs.
  • Uncertainty: coverage, calibration, NLPD, and uncertainty diagnostics.
  • Fingerprints: lower-level composition, SMILES, polymer, and cache helpers.
  • Data And Metrics: cleaning, splitting, preprocessing, metrics, PCA, reporting, artifact utilities, and optional-dependency helpers.
  • Visualization: plotting helpers for model analysis.

Prefer the high-level estimator and featurizer APIs for reusable scripts. Use the lower-level functions when building custom research workflows.