How lululxvi/deepxde works
Project type and purpose
• DeepXDE is an open-source Python library (not a web or mobile app) for scientific machine learning—especially physics-informed neural networks (PINNs), DeepONet, and multifidelity neural networks—to solve differential equations.
• It provides a concise user API, plugs into multiple tensor-computation backends (TensorFlow v1 & v2, PyTorch, JAX, PaddlePaddle), and exposes modular components for data, geometry, gradients, networks, optimizers, and utilities.
File structure
• Top-level “deepxde/” directory holds the core library.
– deepxde/backend: adapter layer and Tensor/ops uniform API + subdirs for each backend
– deepxde/data, geometry, gradients, icbc: problem definition (domains, PDEs/BCs, sampling)
– deepxde/nn: network abstractions (FNN, DeepONet, etc.) with per-backend implementations
– deepxde/optimizers: optimizer wrappers with per-backend code
– deepxde/utils: cross-backend compatibility utilities
– deepxde/model.py, callbacks.py, losses.py, metrics.py: training pipeline and high-level abstractions
– deepxde/zcs: zero-coordinate-shift AD strategy
• Infrastructure folders: docs/ (documentation), examples/ (demo scripts), docker/, CI workflows.
README for additional insights
• Confirms plug-and-play backend support, lists algorithms (PINN, DeepONet, MFNN), AD methods, sampling, geometries, BCs, and training features (callbacks, data-parallel).
• Contains architecture snapshots (images/pinn.png, backend.png) that hint at layered design.
Essential elements
– Notes indicating “same structure for other backends”