Transolver-3: Scaling Up Transformer Solvers to Industrial-Scale Geometries
Authors:
- Hang ZhouHZHang Zhou
- Haixu WuHWHaixu Wu
- Haonan ShangguanHSHaonan Shangguan
- Yuezhou MaYMYuezhou Ma
- Huikun WengHWHuikun Weng
- Jianmin WangJWJianmin Wang
- Mingsheng LongMLMingsheng Long
Preprint
Abstract
Deep learning has emerged as a transformative tool for the neural surrogate modeling of partial differential equations (PDEs), known as neural PDE solvers. However, scaling these solvers to industrial-scale geometries with over $10^8$ cells remains a fundamental challenge due to the prohibitive memory complexity of processing high-resolution meshes. We present Transolver-3, a new member of the Transolver family as a highly scalable framework designed for high-fidelity physics simulations. To bridge the gap between limited GPU capacity and the resolution requirements of complex engineering tasks, we introduce two key architectural optimizations: faster slice and deslice by exploiting matrix multiplication associative property and geometry slice tiling to partition the computation of physical states. Combined with an amortized training strategy by learning on random subsets of original high-resolution meshes and a physical state caching technique during inference, Transolver-3 enables high-fidelity field prediction on industrial-scale meshes. Extensive experiments demonstrate that Transolver-3 is capable of handling meshes with over 160 million cells, achieving impressive performance across three challenging simulation benchmarks, including aircraft and automotive design tasks.
Keywords
All Related Versions
This publication has 9 references indexed in Scilit:
- Towards Foundation Models for Scientific Machine Learning: Characterizing Scaling and Transfer Behavior2023
- Geometry-Informed Neural Operator for Large-Scale 3D PDEs2023
- Symbolic Discovery of Optimization Algorithms2023
- FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness2022
- When and why PINNs fail to train: A neural tangent kernel perspectiveJournal of Computational Physics, 2022
- Graph U-NetsIEEE Transactions on Pattern Analysis and Machine Intelligence, 2021
- A review of hybrid RANS-LES methods for turbulent flows: Concepts and applicationsProgress in Aerospace Sciences, 2020
- Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equationsJournal of Computational Physics, 2019
- A New Version of Detached-eddy Simulation, Resistant to Ambiguous Grid DensitiesTheoretical and Computational Fluid Dynamics, 2006