Transolver-3: Scaling Up Transformer Solvers to Industrial-Scale Geometries

TL;DR

Transolver-3 is a scalable neural PDE solver framework that efficiently handles industrial-scale geometries with over 160 million cells, enabling high-fidelity physics simulations for complex engineering tasks.

Contribution

The paper introduces architectural optimizations and strategies that allow neural PDE solvers to scale to over 160 million cells, addressing memory and computational challenges.

Findings

Handles meshes with over 160 million cells

Achieves high accuracy on complex engineering benchmarks

Demonstrates significant performance improvements

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.