MS-PINN: Physics-Informed Neural Networks for Modeling Multi-field Coupled Evolution in Metal Solidification
Figure 1: The overall architecture of MS-PINN. The input spatial position
Figure 2: Architecture of the general PDE solving software based on PINN. The architecture demonstrates the integration of a plugin system which enhances flexibility and scalability by allowing user-defined callbacks at various stages of the training process.
... install pytorch
>>> pip install fastapi tqdm rich matplotlib numpy pandas imageio
NeuroPDE/
api.py Provides backend API services based on FastAPI
data/ Contains data files and scripts for generating data files
logs/ Stores training logs and checkpoint file paths
utils/ Utility functions such as CSV handling and metrics calculation
visualize/ Tools for plotting graphs
config/
common.py Configuration settings for NeuroPDE, allowing customization of the training process
train/
pinn.py
callback.py Lifecycle hook class, can be extended to create custom plugins
modules/ Built-in modules, including checkpoints, plotting, logging, and various algorithms
pdes/
gradients Utility functions for calculating gradients
models/ Built-in network architectures, with the option to create custom networks
python ./MS-PINN/train_MS_PINN.py
python ./MS-PINN/train_burgers_dde.py
python ./MS-PINN/train_burgers.py
python ./MS-PINN/train_diffusion_dde.py
python ./MS-PINN/train_diffusion.py
python ./MS-PINN/train_ns_dde.py
python ./MS-PINN/train_ns_inverse.py
This work was sponsored by the National Key Research and Development Program of China (No. 2023YFB4606200), Key Program of Science and Technology of Yunnan Province (No. 202302AB080020), and Key Project of Shanghai Zhangjiang National Independent Innovation Demonstration Zone (No. ZJ2021-ZD-006). The authors gratefully appreciate the anonymous reviewers for their valuable comments.