ArGEnT: Geometry-Conditioned Transformer for Operator Learning
A recent publication on arXiv (2602.11626) presents ArGEnT (Arbitrary Geometry-encoded Transformer), a framework that utilizes geometry-conditioned attention to learn solution operators across various geometries. This innovative method separates geometry encoding from the evaluation of query points, overcoming the challenges faced by current operator-learning techniques that depend on structured discretizations or point-cloud models. ArGEnT includes three types—self-attention, cross-attention, and hybrid-attention—and can function independently or be combined with neural operators to integrate non-geometric physical inputs. The goal is to enhance adaptability in irregular and non-parameterized domains, which is crucial for many-query simulations, physics-informed learning, and changing geometries. This paper is noted as a replace-cross announcement, signaling a revision and contributing to ongoing research in scientific machine learning.
Key facts
- ArGEnT stands for Arbitrary Geometry-encoded Transformer.
- It is a geometry-conditioned attention framework.
- It decouples geometry encoding from query-point evaluation.
- Three variants: self-attention, cross-attention, and hybrid-attention.
- Can be integrated with neural operators for non-geometric physical inputs.
- Targets arbitrary geometries and irregular domains.
- Paper arXiv:2602.11626, version 3, replace-cross.
- Published on arXiv.
Entities
Institutions
- arXiv