PaddlePaddle/PaddleCFD v0.3.0

v0.3.0

Repository: PaddlePaddle/PaddleCFD

Tag: v0.3.0

Published: 2025-09-15T11:42:39Z

Prerelease: no

Release notes:

PaddleCFD 0.3.0 Release Notes

中文版：参见 release_notes_cn.md.

PaddleCFD 0.3.0 Overview

PaddleCFD is a deep learning toolkit based on the PaddlePaddle deep learning framework, focused on Computational Fluid Dynamics (CFD) tasks. It is used for fluid mechanics equation discovery, numerical simulation acceleration, flow shape optimization, and flow control strategy discovery. The PaddleCFD 0.3.0 release mainly focuses on accelerating CFD numerical simulations, constructing fluid computation surrogate models based on the PaddlePaddle framework. It has achieved breakthroughs in model accuracy and computational efficiency, providing high baseline models for researchers in the field of fluid mechanics and turnkey tool libraries for enterprise users. The main features of the PaddleCFD 0.3.0 release are as follows:

• Coverage of mainstream and cutting-edge surrogate models: The PaddleCFD 0.3.0 release includes specific models for scientific computing, such as Fourier Neural Operators (FNO), DeepONet, and cutting-edge deep learning models like Transformer, diffusion models, and KAN. These models possess powerful operator learning, prediction, and generation capabilities and represent the forefront of AI + fluid mechanics research (Data source: Clarivate Analytics Essential Science Indicators).
• Leading precision/computational efficiency: PaddleCFD 0.3.0 has improved and adapted models from publicly available papers, achieving a dual increase in model accuracy and computational efficiency. The PPFNO model, with integral correction, has an average relative error of less than 3% for drag coefficient in the test dataset (baseline model ~8%); via operator fusion, training speed is increased by 188%, and inference speed by 159%, achieving second-level inference for tens of millions of grids. The PPTransfomer model, through PaddlePaddle's dynamic-to-static conversion and neural network compiler, boosted training speed by 29.4%, with dual-card parallel efficiency reaching 90.2%, supporting parallel inference for models with tens of millions of grids. The PPKAN model, compared to traditional MLP neural networks, has a 30% improvement in accuracy with an equivalent number of parameters. The PPDifusion model, through data parallelism, achieves scalable acceleration, with single-machine multi-card parallel efficiency of over 99.4%. The PPDeepONet model, based on the MultiONet network structure and through the second-order optimizer SOAP, achieves an approximate 10% improvement in model accuracy.
• Full-scenario support for industry implementation: PaddleCFD 0.3.0 places greater emphasis on industrial application implementation, enhancing model accuracy and computational efficiency through real industrial business scenarios and improving model functional modules. For example, the PPFNO model developed a complete set of functional modules for the drag coefficient prediction task, including training/inference data (volume mesh & surface mesh) preprocessing, distributed training, offline inference, online inference, etc., enabling the model to be containerized for training, inference, and deployment, with applications in leading high-speed train industry enterprises. Additionally, PaddleCFD 0.3.0 provides data parsing modules for various CFD data formats, enabling seamless integration with multiple traditional CFD simulation software.
• Single-folder strategy to enhance usability: PaddleCFD 0.3.0, referencing the successful single-file strategy of the HuggingFace AI suite, places all modules related to the model in the same folder, and avoids over-wrapping of deep learning framework APIs, reducing the learning cost and maintenance cost for users.

PaddleCFD 0.3.0 Updates

New Features

• The PPDeepONet model has added the PirateNets network structure, improving the stability and efficiency of model training.
• The PPFNO model has open-sourced the custom operator fused-segment-csr, which provides higher model training and inference throughput, decouples memory usage from sample size, and supports training and inference for models with tens of millions of grids.

Bug Fixes

• The PPTransformer model has fixed a model import error issue.
• By installing custom operators from source code, the issue of the toolkit environment lacking a specific dynamic library is fixed.
• Updated the PaddleCFD installation documentation, fixing potential errors encountered during installation.
• Fixed an issue where the PPFNO model inference-data preprocessing module imported an unnecessary package causing an error.

Contributors

guhaohao0991, HydrogenSulfate, KaiCHEN-HT, liaoxin2, lijialin03, wangguan1995, XiaoguangHu01