Changelog

All notable changes to PyPIELM are documented here. This project adheres to Semantic Versioning.

[0.1.0] — 2026-05-15

Added

  • Repository scaffold and pyproject.toml packaging (Hatchling).

  • core/: BasePIELM, RandomFeatureMap, FourierFeatureMap, AutogradFeatureMap; ridge_solve, rrqr_solve, bayesian_solve, tikhonov_solve; seed_everything, get_device.

  • data/: PIELMDataset, adapters for CSV, NPZ, PINNacle .dat, PDEBench HDF5, and torch.utils.data.Dataset; Normalizer, FeatureExpander; auto_load dispatcher.

  • pde/: autograd gradient, laplacian, divergence, advection_term; AnalyticLaplacian; UniformSampler, LHSSampler, GridSampler, AdaptiveSampler; BoxDomain, UnionDomain; DirichletBC, NeumannBC, InitialCondition, PeriodicBC.

  • 26 PIELM variants: VanillaPIELM, CorePIELM, GFFPIELM, BayesianPIELM, DPIELM, LocELM, DDELMCoarse, CurriculumPIELM, NullSpacePIELM, EigPIELM, LSE_ELM, StefanPIELM, FPIELM, SGE_PIELM, RINN, RaNN, XPIELM, PIELM_RVDS, TSPIELM, KAPIELM, SoftPartitionKAPIELM, NormalEquationELM, ParameterRetentionELM, PiecewiseELM, DELM, PinnacleELM; model registry with @register decorator and get_model.

  • PINN baselines: VanillaPINN, AdaptivePINN, FourierPINN, MuonPINN.

  • metrics/: rmse, mae, relative_l2, max_error, r2_score, MetricsBundle; io/checkpoint.py (save_model, load_model); io/export.py (to_onnx, to_torchscript).

  • visualization/plots.py: plot_solution_1d, plot_solution_2d, plot_training_history, plot_pareto, plot_leaderboard_heatmap, save_figure.

  • benchmarks/: perf_profile.py, sweep_hidden_dim.py, sweep_solver.py, compare_numpy_torch.py, stats_analysis.py, compare_devices.py; all support --platform argument.

  • utils/config.py (ExperimentConfig, load_config, run_experiment); CLI pypielm/__main__.py (run, sweep, export, list-models subcommands); batch_summary.json for parallel sweeps.

  • Sphinx documentation scaffold (docs/), runnable example scripts (examples/), updated README.md.

Tests

  • 432 tests passing, 17 skipped, 0 failed.

  • MPS (Apple Silicon): 77% benchmark success rate (failures are expected for gradient-based PINN models and unsupported MPS ops).

  • CPU: 98% benchmark success rate.