DeepXDE v1.3.0

Backend JAX supports ODE forward problems.🎉🎉🎉

Areas of improvement

• Backend TensorFlow supports model.save and model.restore.
• Backend TensorFlow uses @tf.function(jit_compile=True) for faster speed.

DeepXDE v1.2.0

DeepXDE supports physics-informed DeepONet.🎉🎉🎉

Areas of improvement

• Fix the issue when setting random seed via dde.set_random_seed()

New APIs

• Support physics-informed DeepONet
  • Add function spaces: dde.data.PowerSeries, dde.data.Chebyshev, dde.data.GRF, dde.data.GRF_KL, dde.data.GRF2D
  • Add dde.data.PDEOperator
  • Add dde.nn.PIDeepONet

DeepXDE v1.1.4

Areas of improvement

• Fix is_on_line_segment() such as Polygon.boundary_normal() works for float32
• Refactor backend JAX: utilize vmap, and add auxiliary arguments to data.losses

New APIs

• Add NN.num_trainable_parameters()

DeepXDE v1.1.3

API changes

• dde.data.MfDataSet adds a new argument standardize=False, which changes the default behavior of MfDataSet. In the previous version, "standardize" is used by default.

DeepXDE v1.1.2

Areas of improvement

• Backend JAX supports Jacobian
• Backend JAX uses a random random seed, and supports user-defined random seed
• MFNN test loss returns 0 instead of nan
• Bug fix: CSG for float32

API changes

• Move saving and plotting functions to utils.external. Delete deepxde/postprocessing.py

DeepXDE v1.1.1

Areas of improvement

• JAX backend supports different optimizers
• JAX backend: improve FNN to fully support function approximation
• Improve code format, and code cleaning
• Bug fix: float64
• Documentation and examples improvements

DeepXDE v1.1.0

DeepXDE supports JAX backend for function approximation. 🎉🎉🎉

Areas of improvement

• Bug fix: float64 in PyTorch
• Documentation and examples improvements

New APIs

• Add PODDeepONet
• Support MIONet: Quadruple, QuadrupleCartesianProd, MIONet, MIONetCartesianProd

API changes

• Rename dde.icbcs to dde.icbc

DeepXDE First Stable Release v1.0.0

DeepXDE was developed starting from the summer of 2018, self-hosted in Subversion at Brown University, originally under the name SciCoNet (Scientific Computing Neural Networks). On Feb 7, 2019, SciCoNet was moved from Subversion to GitHub, renamed to DeepXDE. The first version v0.1.0 was then released on Jun 12, 2019. After the development of more than three years and a half, DeepXDE with backend TensorFlow 1.x becomes stable, and thus we now release the first stable version v1.0.0! 🎉🎉🎉

Thank you all for taking the time to contribute! A non-exhaustive but growing list needs to mention: @lululxvi @smao-astro @ZongrenZou @Saransh-cpp @anranjiao @pescap @Handi-Zhang @Anilith etc.

DeepXDE already has partial support of backend TensorFlow 2.x (from v0.13.0, Jul 21, 2021) and PyTorch (from v0.13.1, Jul 28, 2021). The multiple backend (also JAX) support will be enhanced in DeepXDE v1.x, and more advanced features (e.g., PINN-DeepONet) will be developed.

---

Here are the new improvements, compared to the previous version.

Areas of improvement

• Stop training if loss becomes nan (#491)
• EarlyStopping callback can monitor testing loss (#501)
• Documentation improvements

New APIs

• Add set_random_seed to set the global random seed (#495)
• Add function dat_to_csv to convert dat files to CSV format (#499)

DeepXDE v0.14.1

Change license from Apache-2.0 to LGPL-2.1

Areas of improvement

• callback MovieDumper supports backend TensorFlow and PyTorch

New APIs

• Add PDE.replace_with_anchors()

DeepXDE v0.14.0

We stop the support of Python 3.6 from this release.

Areas of improvement

• Model.save() and Model.restore() supports backend PyTorch
• Model.predict() supports PDE auxiliary variables for backend TensorFlow 1.x and TensorFlow 2.x
• Bug fix on some double/float issues
• Documentation improvements