---

name: training-workflows description: "Guide DeepXDE model lifecycle, optimizers, callbacks, prediction, checkpoints, plotting, function fitting, tabular data, and multifidelity training workflows." disable-model-invocation: true metadata: disco-role: operating license: LGPL 2.1

DeepXDE Training Workflows

Use this sub-skill after the user already has compatible DeepXDE data and net objects and needs to train, checkpoint, resume, predict, or diagnose a dde.Model workflow. This construction verified PyTorch CPU behavior; TensorFlow, JAX, Paddle, GPU, and Horovod behavior should be treated as optional or backend-specific unless separately verified.

Route first

• PDE residuals, geometry, boundary/initial conditions, inverse-variable setup, and adaptive sampling point design: ../pinn-problem-setup/SKILL.md.
• Backend installation, backend selection, dtype, autodiff, seed, XLA, GPU, Horovod, or parallel setup: ../backend-and-configuration/SKILL.md.
• DeepONet/MIONet/PDEOperator data shapes and operator-network construction: ../operator-learning/SKILL.md.

Read these bundled references

• references/model-lifecycle.md: concrete Model(data, net) lifecycle, compile, train, optimizers, callbacks, predict, save/restore, metrics, and plotting.
• references/data-and-function-workflows.md: Function, DataSet, MfFunc, and MfDataSet data contracts and shape checks.
• references/troubleshooting.md: optimizer/backend/plot/checkpoint/convergence failure triage.

Safe smoke check

Run the bundled function-approximation smoke when you need to confirm the training loop and PyTorch CPU backend before adapting a larger problem:

python scripts/smoke_function_approximation.py --iterations 3 --num-train 8 --num-test 16

The script sets DDE_BACKEND=pytorch before importing DeepXDE unless the environment already sets a backend, uses a tiny dde.data.Function + dde.nn.FNN, performs no plotting, writes only to a caller-selected output directory when requested, and prints a JSON summary.

Minimal lifecycle pattern

import deepxde as dde

model = dde.Model(data, net)
model.compile("adam", lr=1e-3, metrics=["l2 relative error"])
losshistory, train_state = model.train(iterations=1000, display_every=100)
y_pred = model.predict(x_eval)

Keep the data and net compatible with the selected backend. For PDE and TimePDE training, do not use batch_size; resample PDE or boundary points with dde.callbacks.PDEPointResampler instead. For ordinary function, tabular, multifidelity, and operator datasets, follow the data-specific batching rules in the references.