PaddlePaddle/PaddleScience

What it solves

PaddleScience is a scientific computing suite designed to solve complex problems in physics, chemistry, and meteorology. It addresses the challenge of solving differential equations and simulating physical systems by combining deep learning with traditional scientific principles, allowing for more efficient and flexible modeling than traditional numerical methods alone.

How it works

Built on the PaddlePaddle framework, the suite leverages deep neural networks and automatic (high-order) differentiation to compute solutions. It supports three primary solving methodologies:

1. Physics-Driven: Uses physical laws and mechanisms to guide the learning process (e.g., PINNs).
2. Data-Driven: Relies on supervised learning from existing datasets to predict outcomes.
3. Hybrid (Physics-Data Fusion): Combines both physical mechanisms and data to improve accuracy and robustness.

It also integrates the sympy library for symbolic equation representation and supports various boundary conditions (Dirichlet, Neumann, Robin) and complex geometry sampling via STL files.

Who it’s for

It is intended for researchers and engineers in scientific computing, computational physics, fluid dynamics, materials science, and meteorology who want to use AI to accelerate simulations or solve inverse problems.

Highlights

• Broad Application Range: Includes 20+ cases across fluids, structures, and weather forecasting.
• Flexible Solving Modes: Supports purely physics-driven, purely data-driven, and hybrid approaches.
• Advanced Geometry Support: Handles simple and complex STL geometries with boolean operations.
• Research Efficiency Tools: Features experiment source code tracking and one-click parallel experiment launching.
• Symbolic Integration: Uses sympy for equation representation and solving systems of equations.