A new model for physics problems that works regardless of coordinates

The ISLA model uses slice attention to handle boundary-driven partial differential equations.

Researchers introduced ISLA, which stands for Invariant Slice Attention. It is a slice-attention transformer for steady, boundary-driven partial differential equations (PDEs).

Current models fail when a problem is rotated or re-sampled. Rotating a PDE can raise surface-pressure error from 4% to more than 100%. Re-sampling the boundary raises error to about 30%.

In ISLA, no learned operation receives a coordinate. Positions enter only as distances and relative angles to learned slice anchors. Lengths are measured in a unit computed from the boundary.

Interior points query the boundary through a tree-accelerated learned kernel. Each query is independent of all other queries. The model is equivariant to rigid motions and reflections.

On AeroZoo, ISLA velocity and pressure errors are at least 3.1 and 3.3 times lower than any baseline. On AirfRANS, errors are at least 1.6 times lower than Transolver and GeoTransolver.

Key facts

  • ISLA is equivariant to rigid motions and reflections.
  • The model is invariant to the units of its declared data.
  • ISLA is consistent under mesh refinement.

All sections

Edition