PaddleScience 1.0.0 Release Notes
The PaddleScience 1.0.0 version brings a completely new design of APIs and modules to support different training paradigms in one toolkit: mechanism-driven, data-driven and mechanism-data mixed case. The key features are as follows:
- Newly designed modules
- Built-in high-frequency usage API, such as `PDE`, `BoundaryConstraint`, `InteriorConstraint`.
- Decoupled training/evaluation/visualization.
- Automatic result visualization.
- Support resume training, transfer learning, **automatic mixed precision**, **gradient accumulation** and **distributed parallell training**.
- Enhanced features
- Enhanced geometry module, including interior/boundary sampling and boolean operations, as well as STL-based SDF(signed distance field) weighing.
- Support dirichlet, nuemann, robin and **custom** boundary condition.
- Multiple model architecture such as MLP, Transformer.
- Support high-order differentiation related API, such as jacobian/hessian and L-BFGS optimizer.
- Varied examples
- 10+ classical examples covering fluid, structure, meteorology and more areas.
- Detailed user document and end-to-end tutorial.
- Quickly example experience in AIStudio.
- User friendly
- Structured logger.
- Comprehensive docstrings and type hints.