Our new X account is live! Follow @wizwand_team for updates
WorkDL logo mark

MSPT: Efficient Large-Scale Physical Modeling via Parallelized Multi-Scale Attention

About

A key scalability challenge in neural solvers for industrial-scale physics simulations is efficiently capturing both fine-grained local interactions and long-range global dependencies across millions of spatial elements. We introduce the Multi-Scale Patch Transformer (MSPT), an architecture that combines local point attention within patches with global attention to coarse patch-level representations. To partition the input domain into spatially-coherent patches, we employ ball trees, which handle irregular geometries efficiently. This dual-scale design enables MSPT to scale to millions of points on a single GPU. We validate our method on standard PDE benchmarks (elasticity, plasticity, fluid dynamics, porous flow) and large-scale aerodynamic datasets (ShapeNet-Car, Ahmed-ML), achieving state-of-the-art accuracy with substantially lower memory footprint and computational cost.

Pedro M. P. Curvo, Jan-Willem van de Meent, Maksim Zhdanov• 2025

Related benchmarks

TaskDatasetResultRank
Operator learningAirfoil Structured Mesh (test)
Relative L2 Error0.0051
15
Operator learningPipe Structured Mesh (test)
Relative L2 Error0.0031
15
Operator learningNavier-Stokes Regular Grid (test)
Relative L2 Error0.0632
15
CFD field reconstructionShapeNet Car (test)
Volume Error1.89
15
Operator learningPlasticity Structured Mesh (test)
Relative L2 Error0.001
15
Operator learningDarcy Regular Grid (test)
Relative L2 Error0.0063
15
Operator learningElasticity Point Cloud (test)
Relative L2 Error0.0048
13
CFD field reconstructionAhmedML (test)
Volume Metric2.04
11
Showing 8 of 8 rows

Other info

Follow for update