Welcome to AdePT-ML’s documentation!

AdePT-ML is designed to streamline the training of hybrid physics-informed neural networks by seamlessly combining PyTorch modules with non-differentiable physics solvers written in any framework (NumPy, JAX, SciPy, etc.).

Key features:

  • Three physics backprop modes — full Jacobian, manual VJP, or split-VJP (pullback closure from jax.vjp) — choose based on your solver’s cost.

  • Flexible input routing — serial pipelines or arbitrary fan-in/fan-out between sub-models via HybridConfig.model_inputs.

  • Gradient clipping — built into the training loop for stable physics-informed training.

  • TensorBoard integration — per-epoch train/test losses logged automatically with auto-incrementing run directories.