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

2026-02-04

AI Summary

This work addresses the severe memory bottleneck faced by neural PDE solvers when applied to ultra-large-scale industrial geometries—exceeding 100 million mesh cells—where high-resolution grids overwhelm GPU memory. To overcome this challenge, the authors propose a highly scalable Transformer-based framework that integrates amortized training over randomly sampled high-resolution subgrids, a physics-informed state caching mechanism, memory-efficient slicing and unslicing operations leveraging the associativity of matrix multiplication, and a geometry-aware tiling strategy. This architecture dramatically reduces memory consumption while maintaining solution fidelity. The method successfully handles industrial simulations with up to 160 million mesh cells and achieves high-fidelity field predictions across three challenging benchmarks in aerospace and automotive design, demonstrating exceptional scalability and accuracy.

Technology Category

• Machine Learning: Hardware-aware ML
• Data Mining & Knowledge Management: Scalability, Parallel & Distributed Systems
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

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.

Problem

• neural PDE solvers
• industrial-scale geometries
• high-resolution meshes
• memory complexity
• scalability

Innovation

• scalable Transformer
• neural PDE solver
• geometry slice tiling
• amortized training
• high-resolution mesh simulation