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

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 can handle meshes with over 160 million cells, achieving impressive performance across three challenging simulation benchmarks, including aircraft and automotive design tasks. Code is available at https://github.com/thuml/Transolver-3.

When engineers design a car or an aircraft, they rely on simulations that predict how air flows around the vehicle, which determines its fuel efficiency, stability, and safety. These simulations are accurate but slow, sometimes taking days for a single design. Artificial intelligence promises to deliver the same predictions in seconds, but it faces a major obstacle: a realistic vehicle must be broken into a grid of more than a hundred million tiny pieces, and current AI models run out of computer memory long before they reach that size, forcing engineers to settle for coarser, less reliable predictions. We developed Transolver-3, an AI model that overcomes this memory barrier. We reorganized its internal calculations to use far less memory, taught it the underlying physics by showing it small random samples of the full grid rather than the whole grid at once, and let it make predictions by processing the grid in manageable pieces while still drawing on every detail. With these advances, Transolver-3 handles grids roughly twenty times larger than earlier models in its family, while also making more accurate predictions. This brings fast, dependable AI simulation meaningfully closer to everyday use in real engineering design.