This repository gives the codes for numerical solver independent seismic wave simulation using Task-decomposed Physics-informed Neural Networks. **This repository reproduces the results of the paper "Numerical solver independent seismic wave simulation using Task-decomposed Physics-informed Neural Networks." IEEE Geoscience and Remote Sensing Letters 20, 3002905.
We propose a task-decomposed (TD) training scheme of Physics-informed Neural Network (PINN) to perform the time-domain wave equation modeling. Besides this feature, we use analytical wavefield solutions (Fourier transformed from frequency-domain ones) as the initial condition to avoid the source singularity issue.
Figure (a) shows the wavefield from the TD-PINN, and Figure (b) shows the wavefield from the finite-difference method (FDM).
pre-training.ipynb: Stage1-pretraining
train_full.ipynb: Stage2-Full training
tra_physics.ipynb: Stage3-Physics enhanced training
If you find our codes and publications helpful, please kindly cite the following publications.
@article{zou2023numerical, title={Numerical solver independent seismic wave simulation using Task-decomposed Physics-informed Neural Networks}, author={Zou, Jingbo and Liu, Cai and Song, Chao and Zhao, Pengfei}, journal={IEEE Geoscience and Remote Sensing Letters}, year={2023}, publisher={IEEE} }
If there are any problems, please contact me through my emails: chao.song@kaust.edu.sa;chaosong@jlu.edu.cn