PaddleScience

Developed with PaddlePaddle

👀Description

PaddleScience 是一个基于深度学习框架 PaddlePaddle 开发的科学计算套件，利用深度神经网络的学习能力和 PaddlePaddle 框架的自动(高阶)微分机制，解决物理、化学、气象等领域的问题。支持物理机理驱动、数据驱动、数理融合三种求解方式，并提供了基础 API 和详尽文档供用户使用与二次开发。

推荐使用 DeepWiki 辅助编程 👉

✨Feature

• 支持 实验源码跟踪，一键启动并行实验，提高科研效率。
• 支持简单几何和复杂 STL 几何的采样与布尔运算。
• 支持包括 Dirichlet、Neumann、Robin 以及自定义边界条件。
• 支持物理机理驱动、数据驱动、数理融合三种问题求解方式。涵盖流体、结构、气象等领域 20+ 案例。
• 支持结果可视化输出与日志结构化保存。
• 完善的 type hints，用户使用和代码贡献全流程文档，经典案例 AI studio 快速体验，降低使用门槛，提高开发效率。
• 支持基于 sympy 符号计算库的方程表示与联立方程组计算。
• 更多特性正在开发中...

📝Case List

MathematicsEngineering ScienceMaterials ScienceEarth SciencesChemical Sciences

Problem Type	Case Name	Optimization Method	Model Type	Training Method	Dataset	References
Helmholtz Equation	SPINN(Helmholtz3D)	Physics-driven	SPINN	Unsupervised Learning	-	Paper
Phase Field Equation	Allen-Cahn	Physics-driven	MLP	Unsupervised Learning	Data	Paper
Differential Equation	Laplace Equation	Physics-driven	MLP	Unsupervised Learning	-	-
Differential Equation	Burgers Equation	Physics-driven	MLP	Unsupervised Learning	Data	Paper
Differential Equation	Nonlinear PDE	Physics-driven	PIRBN	Unsupervised Learning	-	Paper
Differential Equation	Lorenz Equation	Data-driven	Transformer-Physx	Supervised Learning	Data	Paper
Differential Equation	Rössler Equation	Data-driven	Transformer-Physx	Supervised Learning	Data	Paper
Operator Learning	DeepONet	Data-driven	MLP	Supervised Learning	Data	Paper
Differential Equation	Gradient-Enhanced Physics-Informed PDE Solving	Physics-driven	gPINN	Unsupervised Learning	-	Paper
Integral Equation	Volterra Integral Equation	Physics-driven	MLP	Unsupervised Learning	-	Project
Differential Equation	Fractional Differential Equation	Physics-driven	MLP	Unsupervised Learning	-	-
Optical Soliton	Optical soliton	Physics-driven	MLP	Unsupervised Learning	-	Paper
Optical Rogue Wave	Optical rogue wave	Physics-driven	MLP	Unsupervised Learning	-	Paper
Domain Decomposition	XPINN	Physics-driven	MLP	Unsupervised Learning	-	Paper
Brusselator Diffusion System	3D-Brusselator	Data-driven	LNO	Supervised Learning	-	Paper
Symbolic Regression	Transformer4SR	Data-driven	Transformer	Supervised Learning	-	Paper
Operator Learning	Latent Neural Operator LNO	Data-driven	Transformer	Supervised Learning	-	Paper

Problem Type	Case Name	Optimization Method	Model Type	Training Method	Dataset	References
Vehicle Surface Drag Prediction	Transolver	Data-driven	Transolver	Supervised Learning	Data	Paper
Vehicle Surface Drag Prediction	DrivAerNet	Data-driven	RegDGCNN	Supervised Learning	Data	Paper
1D Linear Convection Problem	1D Linear Convection	Data-driven	ViT	Supervised Learning	Data	Paper
Unsteady Incompressible Flow	2D Buoyancy-Driven Cavity Flow	Data-driven	ViT	Supervised Learning	Data	Paper
Steady Incompressible Flow	Re3200 2D Steady Cavity Flow	Physics-driven	MLP	Unsupervised Learning	-
Steady Incompressible Flow	2D Darcy Flow	Physics-driven	MLP	Unsupervised Learning	-
Steady Incompressible Flow	2D Pipe Flow	Physics-driven	MLP	Unsupervised Learning	-	Paper
Steady Incompressible Flow	3D Intracranial Aneurysm	Physics-driven	MLP	Unsupervised Learning	Data	Project
Steady Incompressible Flow	Arbitrary 2D Geometry Flow	Data-driven	DeepCFD	Supervised Learning	-	Paper
Unsteady Incompressible Flow	2D Unsteady Cavity Flow	Physics-driven	MLP	Unsupervised Learning	-	-
Unsteady Incompressible Flow	Re100 2D Cylinder Flow	Physics-driven	MLP	Semi-supervised Learning	Data	Paper
Unsteady Incompressible Flow	Re100~750 2D Cylinder Flow	Data-driven	Transformer-Physx	Supervised Learning	Data	Paper
Compressible Flow	2D Air Shock Wave	Physics-driven	PINN-WE	Unsupervised Learning	Data	-
Aircraft Design	MeshGraphNets	Data-driven	GNN	Supervised Learning	Data	Paper
Aircraft Design	Rocket Engine Vacuum Plume	Data-driven	CNN	Supervised Learning	Data	-
Aircraft Design	Deep-Flow-Prediction	Data-driven	TurbNetG	Supervised Learning	Data	Paper
General Flow Simulation	Aerodynamic Shape Design	Data-driven	AMGNet	Supervised Learning	Data	Paper
Fluid-Structure Interaction	Vortex-Induced Vibration	Physics-driven	MLP	Semi-supervised Learning	Data	Paper
Multiphase Flow	Gas-Liquid Two-Phase Flow	Physics-driven	BubbleNet	Semi-supervised Learning	Data	Paper
Multiphase Flow	twophasePINN	Physics-driven	MLP	Unsupervised Learning	-	Paper
High-Resolution Flow Field Reconstruction	2D Turbulent Flow Field Reconstruction	Data-driven	tempoGAN	Supervised Learning	Train Data Eval Data	Paper
High-Resolution Flow Field Reconstruction	2D Turbulent Flow Field Reconstruction	Data-driven	cycleGAN	Supervised Learning	Train Data Eval Data	Paper
High-Resolution Flow Field Reconstruction	Global Field Reconstruction from Sparse Sensors via Voronoi Embedding-Assisted Deep Learning	Data-driven	CNN	Supervised Learning	Data1 Data2 Data3	Paper
Flow Field Prediction	Catheter	Data-driven	FNO	Supervised Learning	Data	Paper
Solver Coupling	CFD-GCN	Data-driven	GCN	Supervised Learning	Data Mesh	Paper
Force Analysis	1D Euler Beam Deformation	Physics-driven	MLP	Unsupervised Learning	-	-
Force Analysis	2D Plate Deformation	Physics-driven	MLP	Unsupervised Learning	-	Paper
Force Analysis	3D Bracket Deformation	Physics-driven	MLP	Unsupervised Learning	Data	Tutorial
Force Analysis	Structural Vibration Simulation	Physics-driven	PhyLSTM	Supervised Learning	Data	Paper
Force Analysis	2D Elastoplastic Structure	Physics-driven	EPNN	Unsupervised Learning	Train Data Eval Data	Paper
Force Analysis and Inverse Problem	3D Vehicle Control Arm Deformation	Physics-driven	MLP	Unsupervised Learning	-	-
Force Analysis and Inverse Problem	3D Heart Simulation	Physics-Data Fusion	PINN	Supervised Learning	-	-
Topology Optimization	2D Topology Optimization	Data-driven	TopOptNN	Supervised Learning	Data	Paper
Topology Optimization	2/3D Topology Optimization	Physics-driven	DenseSIRENModel	Unsupervised Learning	-	Paper
Thermal Simulation	1D Heat Exchanger Thermal Simulation	Physics-driven	PI-DeepONet	Unsupervised Learning	-	-
Thermal Simulation	2D Thermal Simulation	Physics-driven	PINN	Unsupervised Learning	-	Paper
Thermal Simulation	2D Chip Thermal Simulation	Physics-driven	PI-DeepONet	Unsupervised Learning	-	Paper

Problem Type	Case Name	Optimization Method	Model Type	Training Method	Dataset	References
Material Design	Diffuser Design (Inverse Problem)	Physics-driven	Transformer	Unsupervised Learning	Train Data Eval Data	Paper
Crystal Material Property Prediction	CGCNN	Data-driven	GNN	Supervised Learning	MP / Perovskite / C2DB / test	Paper
2D Material Generation and Database	ML2DDB	Data-driven	GNN/Diffusion	Supervised Learning	Coming Soon	Paper

Problem Type	Case Name	Optimization Method	Model Type	Training Method	Dataset	References
Meteorological Downscaling	KMCast	Data-driven	Diffusion	Supervised Learning	GFS	-
Weather Forecasting	Extformer-MoE Weather Forecasting	Data-driven	Transformer	Supervised Learning	enso	-
Weather Forecasting	FourCastNet Weather Forecasting	Data-driven	AFNO	Supervised Learning	ERA5	Paper
Weather Forecasting	NowCastNet Weather Forecasting	Data-driven	GAN	Supervised Learning	MRMS	Paper
Weather Forecasting	GraphCast Weather Forecasting	Data-driven	GNN	Supervised Learning	-	Paper
Weather Forecasting	GenCast Weather Forecasting	Data-driven	Diffusion+Graph transformer	Supervised Learning	Gencast	Paper
Weather Forecasting	Fuxi Weather Forecasting	Data-driven	Transformer	Supervised Learning	-	Paper
Weather Forecasting	FengWu Weather Forecasting	Data-driven	Transformer	Supervised Learning	-	Paper
Weather Forecasting	Pangu-Weather Weather Forecasting	Data-driven	Transformer	Supervised Learning	-	Paper
Atmospheric Pollutants	UNet Pollutant Diffusion	Data-driven	UNet	Supervised Learning	Data	-
Atmospheric Pollutants	STAFNet Pollutant Concentration Prediction	Data-driven	STAFNet	Supervised Learning	Data	Paper
Weather Forecasting	DGMR Weather Forecasting	Data-driven	GAN	Supervised Learning	UK dataset	Paper
Seismic Waveform Inversion	VelocityGAN Seismic Waveform Inversion	Data-driven	VelocityGAN	Supervised Learning	OpenFWI	Paper
Remote Sensing Image Segmentation	UNetFormer Image Segmentation	Data-driven	UNetformer	Supervised Learning	Vaihingen	Paper
Traffic Prediction	TGCN Traffic Flow Prediction	Data-driven	GCN & CNN	Supervised Learning	PEMSD4 & PEMSD8	-
Weather Forecasting	Meteoformer Multi-Meteorological Element Prediction	Data-driven	Transformer	Supervised Learning	ERA5	-
Weather Forecasting	Preformer Short-Term Precipitation Prediction	Data-driven	Transformer	Supervised Learning	ERA5	Paper
Weather Forecasting	Climateformer Climate Prediction	Data-driven	Transformer	Supervised Learning	ERA5	-
Generative Model	Gradient Penalty Application in Image Generation	Data-driven	WGAN GP	Supervised Learning	Data1 Data2	Paper
Remote Sensing Image Segmentation	UTAE Remote Sensing Time Series Semantic/Panoptic Segmentation	Data-driven	UTAE	Supervised Learning	PASTIS	Paper

Problem Type	Case Name	Optimization Method	Model Type	Training Method	Dataset	References
Chemical Molecule Generation	Moflow	Data-driven	moflow	Supervised Learning	qm9/ zink250k	MoFlow: An Invertible Flow Model for Generating Molecular Graphs
Chemical Reaction Prediction	IFM	Data-driven	IFM-MLP	Supervised Learning	tox21/sider/hiv/bace/bbbp	Understanding the Limitations of Deep Models for Molecular property prediction: Insights and Solutions

🚀Quick Installation

Method 1: Source Code Installation [Recommended]Method 2: pip Installation

git clone -b develop https://github.com/PaddlePaddle/PaddleScience.git
# 若 github clone 速度比较慢，可以使用 gitee clone
# git clone -b develop https://gitee.com/paddlepaddle/PaddleScience.git

cd PaddleScience

# install paddlesci with editable mode
python -m pip install -e . -i https://pypi.tuna.tsinghua.edu.cn/simple

python -m pip install -U paddlesci

Complete Installation Guide: Installation and Setup

🕘Recent Updates

• 沐曦MetaX 和 太初元碁Tecorigin 完成与 PaddleScience 的第一阶段适配工作，详见：多硬件支持。
• 基于 PaddleScience 的 ADR 方程求解方法 Physics-informed neural networks for advection–diffusion–Langmuir adsorption processes 被 Physics of Fluids 2024 接受。
• 添加 IJCAI 2024: 任意三维几何外形车辆的风阻快速预测竞赛，track A, B, C 的 paddle/pytorch 代码链接。
• 添加 SPINN(基于 Helmholtz3D 方程求解) helmholtz3d。
• 添加 CVit(基于 Advection 方程和 N-S 方程求解) CVit(Navier-Stokes)、CVit(Advection)。
• 添加 PirateNet(基于 Allen-cahn 方程和 N-S 方程求解) Allen-Cahn、LDC2D(Re3200)。
• 基于 PaddleScience 的快速热仿真方法 A fast general thermal simulation model based on MultiBranch Physics-Informed deep operator neural network 被 Physics of Fluids 2024 接受。
• 添加多目标优化算法 Relobralo 。
• 添加气泡流求解案例(Bubble)、机翼优化案例(DeepCFD)、热传导仿真案例(HeatPINN)、非线性短临预报模型(Nowcasting(仅推理))、拓扑优化案例(TopOpt)、矩形平板线弹性方程求解案例(Biharmonic2D)。
• 添加二维血管案例(LabelFree-DNN-Surrogate)、空气激波案例(ShockWave)、去噪网络模型(DUCNN)、风电预测模型(Deep Spatial Temporal)、域分解模型(XPINNs)、积分方程求解案例(Volterra Equation)、分数阶方程求解案例(Fractional Poisson 2D)。
• 针对串联方程和复杂方程场景，Equation 模块支持基于 sympy 的符号计算，并支持和 python 函数混合使用(#507、#505)。
• Geometry 模块和 InteriorConstraint、InitialConstraint 支持计算 SDF 微分功能(#539)。
• 添加 MultiTaskLearning(ppsci.loss.mtl) 多任务学习模块，针对多任务优化(如 PINN 方法)进一步提升性能，使用方式：多任务学习指南(#493、#492)。

🎈Ecosystem Tools

除 PaddleScience 外，Paddle 框架同时支持了科学计算领域相关的研发套件和基础工具：

工具	简介	支持情况
DeepXDE	方程求解套件	全量支持
DeepMD-kit	分子动力学套件	部分支持
Modulus-sym	AI仿真套件	全量支持
NVIDIA/warp	基于 Python 的 GPU 高性能仿真和图形库	全量支持
tensorly	张量运算库	全量支持
Open3D	三维图形库	全量支持
neuraloperator	神经算子库	全量支持
paddle_scatter	张量稀疏聚合库	全量支持
paddle_cluster	几何采样聚合库	全量支持
paddle_sparse	张量稀疏计算库	全量支持
paddle_harmonics	球面谐波变换库	全量支持
deepali	图像、点云配准库	全量支持
DLPACK(v0.8)	跨框架张量内存共享协议	全量支持

💬Support

如在使用过程中遇到问题或想提出开发建议，欢迎在 Discussion 中提出，或者在 Issue 页面新建 issue，会有专业的研发人员进行解答。

👫Contribution

PaddleScience 项目欢迎并依赖开发人员和开源社区中的用户，会不定期推出开源活动。

在开源活动中如需使用 PaddleScience 进行开发，可参考 PaddleScience 开发与贡献指南 以提升开发效率和质量。

• 🎁快乐开源

  旨在鼓励更多的开发者参与到飞桨科学计算社区的开源建设中，帮助社区修复 bug 或贡献 feature，加入开源、共建飞桨。了解编程基本知识的入门用户即可参与，活动进行中： PaddleScience 快乐开源活动表单

🎯Collaboration

PaddleScience 作为一个开源项目，欢迎来各行各业的伙伴携手共建基于飞桨的 AI for Science 领域顶尖开源项目, 打造活跃的前瞻性的 AI for Science 开源社区，建立产学研闭环，推动科研创新与产业赋能。点击了解 飞桨AI for Science共创计划。

❤️Thanks

• PaddleScience 的部分模块和案例设计受 NVIDIA-Modulus、DeepXDE、PaddleNLP、PaddleClas 等优秀开源套件的启发。

• Part of PaddleScience's code is contributed by the following outstanding developers (sorted by Contributors):

🤝Partner Organizations

📜License

Apache License 2.0