[PatrickHenkel1]

Phase 2 of the physXAI 1.0.0 refactoring: one expression language for physics and trainable blocks. Stacked on #71 (base v1/phase-1-data; retarget to refactoring_v1.0.0 after #71 is merged).

Summary

• physXAI.expr replaces constructed features and modular expressions (no global registries any more):

  from physXAI import expr as px
  v = px.Vars()                                   # v.T, v['Func(logistic)'] – no separate declarations
  model = ModularANN(px.Linear([v.mDot, v.load, v.mDot * (v.T_in - v.T), v.T]), solver='ols')
  td = PreprocessingSingleStep(output=px.Change('T')).pipeline(path, model=model)  # inputs from the model

  • Variables, parameters, arithmetic and activations, each with NumPy, Keras and CasADi backends. Names are deterministic; lag() and substitute() are available.
  • Trainable blocks Linear, MLP, MonotoneMLP (CMNN), Polynomial, RBF, Mean, Normalize, Pretrained. Blocks normalize their inputs, arithmetic works on raw physical values.
  • JSON serialization keeps shared nodes shared. The graph is embedded in the model config; the separate modular_expression_config.json is gone.
• Derived inputs inside the model are the default: AgentLib-MPC receives a KerasANN on raw inputs only, so the MPC no longer has to compute e.g. vT.
  • solver='ols' solves a linear regression on expressions in closed form.
  • As a fallback, columns={'vT': expr} adds a data column. This is also used for PINN targets.
• The preprocessing takes its inputs from the model. They are ordered as AgentLib-MPC expects: recursive feature and its lags last, missing intermediate lags added.
• Ported: ConstrainedRegression (fixed: its layer name had no dense, so the CasADi export failed; a second fit duplicated the constraints) and the RNN feature_architecture.
• Executables, AgentLib examples and tests are migrated. The old coupled unittests/modular test is replaced by independent tests.

Contract

• Every block translates to CasADi through AgentLib-MPC (new contract test).
• modular_linreg contract script: vT inside the model. The exported KerasANN has inputs mDot, load, T_in with T recursive; OLS recovers the simulated physics within 0.1 %.

Test plan

• pytest: 91 passed.
• AgentLib-MPC plugin tests pass, and run_example(testing=True) runs.
• All executables/ run (smoke test with 1 epoch).
• CI on Linux and Windows.

Notes

• executables/regression_prototype.py does not use physXAI and needs matplotlib, which is not a dependency, so verify_executables would fail on it once it changes. Should it move out of executables/?

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.