---

name: deepxde description: "Use DeepXDE for scientific machine learning, PINNs, DeepONet/operator learning, backend selection, training, and troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: LGPL 2.1

DeepXDE repo skill

Use this skill when a task involves DeepXDE, physics-informed neural networks (PINNs), scientific machine learning for ODE/PDE/IDE/FPDE problems, DeepONet/operator learning, or DeepXDE backend/training troubleshooting.

This skill is self-contained for future use: it distills the package source, docs, examples, and verified PyTorch CPU inspection into bundled routes, references, and smoke scripts. It does not require the original repository checkout.

Verified baseline

• Package/distribution: DeepXDE, import name deepxde.
• Python: package metadata requires Python >=3.9.
• Base package dependencies: matplotlib, numpy, scikit-learn, scikit-optimize>=0.10.2, and scipy.
• Tensor backend requirement: install at least one DeepXDE backend package before use.
• Runtime verified by this construction: PyTorch backend on CPU for import, simple PDE assembly, FNN, Model.compile, one-step Model.train, and Model.predict.
• Optional/alternative paths that require target-environment verification: TensorFlow 1.x compatibility, TensorFlow 2.x + TensorFlow Probability, JAX + Flax + Optax, PaddlePaddle, GPU, Horovod/MPI, and long training examples.

Minimal installation and import check

For a CPU-safe DeepXDE setup, install DeepXDE plus PyTorch in an isolated Python environment, then select the backend before importing DeepXDE:

python -m pip install deepxde torch
DDE_BACKEND=pytorch python - <<'PY'
import deepxde as dde
print(dde.__version__)
print(dde.backend.backend_name)
PY

Use package-manager-appropriate PyTorch wheels for CPU, CUDA, ROCm, or MPS in the target environment. Do not treat a CPU import as proof of GPU or Horovod correctness.

Route by task

If the user asks to...	Read
install DeepXDE, choose DDE_BACKEND, debug missing TensorFlow/JAX/Paddle/PyTorch packages, check GPU visibility, set dtype/autodiff/random seed/XLA/parallel scaling	sub-skills/backend-and-configuration/SKILL.md
build a forward or inverse PINN for an ODE/PDE/IDE/FPDE, define geometry, BC/IC/point-set constraints, residuals, gradients, hard constraints, or adaptive points	sub-skills/pinn-problem-setup/SKILL.md
compile/train a dde.Model, choose optimizers, callbacks, metrics, save/restore, predict residuals/outputs, fit functions/tabular data, or use multifidelity data	sub-skills/training-workflows/SKILL.md
build DeepONet, POD-DeepONet, MIONet, PI-DeepONet, Cartesian-product operator data, function spaces, ZCS, or troubleshoot branch/trunk/data-shape errors	sub-skills/operator-learning/SKILL.md

Cross-cutting references and scripts

• references/backend-and-installation.md summarizes backend packages, selection order, and the verified/unverified backend boundary.
• references/troubleshooting.md triages common install, backend, PDE, training, plotting, data-shape, and optional hardware failures and routes to the owning sub-skill.
• references/repo-provenance.md records the source commit, package version, dirty-state baseline, and evidence paths used to generate this skill.
• scripts/smoke_deepxde.py is a safe PyTorch CPU diagnostic that imports DeepXDE and can optionally run a tiny PDE training smoke.

Run the root smoke before adapting a task when the environment is uncertain:

python scripts/smoke_deepxde.py --backend pytorch --train-steps 1
python scripts/smoke_deepxde.py --backend pytorch --json

Operating rules for future agents

1. Set DDE_BACKEND before importing any DeepXDE module. Backend selection happens during import.
2. Use backend-native tensor operations inside residuals and operators; use dde.grad.jacobian and dde.grad.hessian for derivatives.
3. Keep problem definition separate from training: assemble geometry/icbc/data in the PINN or operator sub-skill, then use training-workflows for Model.compile, train, callbacks, and checkpoints.
4. Do not assume every example supports every backend. Many examples document supported backends; this skill distills the common PyTorch CPU-safe path and marks other paths optional.
5. Do not run long examples, notebooks, benchmarks, GPU/Horovod jobs, or data-download workflows as smoke tests unless the user explicitly requests that cost and the required backend/data are available.
6. If the package commit, public APIs, backend dependencies, or example layout differ from repo provenance, refresh this skill before relying on stale guidance.