Deepxde

Areas of improvement

- Improve compatibility with NumPy 2
- Bug fix: `dde.data.QuadrupleCartesianProd`
- Backend TensorFlow 1.x: Improve DeepONet
- Backend PyTorch: FNN supports regularization

Thanks to all the contributors to this release!

anranjiao cwoolfo1 vl-dud enigne lululxvi

Areas of improvement

- Add `clear()` for forward-mode autodiff to prevent memory leak
- Tensorflow 1.x backend: `DeepONet` supports layer-by-layer dropout rate setting
- Bug fix: `SingleOutputStrategy` has unnecessary error checking

New APIs

- `dde.geometry.Hypercube` supports `uniform_boundary_points`

Thanks to all the contributors to this release!

vl-dud Yiii9 MinZhu123 lululxvi HydrogenSulfate

Areas of improvement

- `EarlyStopping` callback supports a new argument `start_from_epoch`
- Backend TensorFlow v1/v2: Fix many codes to match the new TensorFlow APIs and Keras 3
- Backend Tensorflow v1: `DeepONet` and `DeepONetCartesianProd` support dropout
- Backend PyTorch: Fix the L-BFGS code to support PyTorch 2.x
- Backend Paddle: Fix the L-BFGS code
- Backend Paddle: `DeepONetCartesianProd` supports multiple outputs
- Backend JAX: Support callback `VariableValue`
- Documentation improvements

New APIs

- `dde.data.PDEOperator` supports `resample_train_points`

Thanks to all the contributors to this release!

bonneted vnikoofard vl-dud tjboise HydrogenSulfate agniv-the-marker lululxvi lijialin03 DecoderLiu anranjiao

Areas of improvement

- Add 2D interface boundary condition `dde.icbc.Interface2DBC`
- Backend JAX: Support loss weights
- Backend JAX: Support `dde.nn.PFNN`
- Backend JAX: Support `dde.callbacks.OperatorPredictor`
- Backend JAX: Fix input and output transform
- Add new examples in docs

Thanks to all the contributors to this release!

lululxvi kuangdai HydrogenSulfate bonneted jdellag vl-dud SebastianCobaise

- DeepXDE stops the support of Python 3.8 from this release.
- Many exciting new functions of automatic differentiation (AD) are added.

Areas of improvement

- `dde.grad` supports forward-mode AD for backends TensorFlow 1.x and 2.x, PyTorch, JAX. Use `dde.config.set_default_autodiff` to select.
- `dde.grad.jacobian` allows both `i` and `j` are None
- Backend PyTorch: DeepONet supports multiple outputs

New APIs

- Support new AD method in `dde.zcs`: Zero Coordinate Shift (ZCS), see https://arxiv.org/abs/2311.00860

Areas of improvement

- Refactor `dde.grad` module
- Backend TensorFlow 1.x and 2.x: `DeepONet` & `DeepONetCartesianProd` support multiple outputs
- Backend TensorFlow: Add regularization to `DeepONet`
- Backend PyTorch: Bug fix of `MIONet` `input_transform`
- Backend JAX: Support more PINN examples
- Backend JAX: Bug fix of `dde.grad`