This is the implementation of DONIS: Importance Sampling for Training Physics-Informed DeepONet.

In this work, we introduce a two-step importance sampling framework that sequentially applies importance sampling to the function and collocation point inputs of DeepONet, which prioritizes mini-batch samples with greater influence (measured by the loss magnitude) on the learning objective, enabling faster convergence and better accuracy.

This repository is licensed under the MIT License.
Note: This project uses DeepXDE, which is licensed under the GNU Lesser General Public License v2.1.