Year: 2026
Type: article
Abstract: Graph-based neural networks have emerged as efficient and versatile surrogates for mesh-based numerical simulations. However, achieving high predictive accuracy remains challenging due to over-smoothing, under-reaching during message passing, and disregard for mesh connectivity. To address this, we propose Global Weight Deep Equilibrium Attention (GDEA). GDEA constructs a sparse global weight matrix from mesh connectivity via the direct stiffness method, introducing an inductive bias toward diagonal dominance. We apply the resulting matrix iteratively until convergence, allowing information to propagate globally without adding extra message-passing layers. We formulate the iterative process as a Deep Equilibrium Model (DEQ) to reduce memory consumption, and we use Anderson acceleration to speed convergence. Extensive experiments across multiple numerical simulation datasets, including the newly introduced Deformed ABC dataset with diverse geometries, materials, and boundary conditions, demonstrate that GDEA outperforms other methods on Deformed ABC while remaining competitive on established benchmarks. An accurate surrogate model that can handle diverse design configurations can facilitate efficient early-stage design exploration.
Source: Communications Engineering
Authors: Junghun Lee, Conrad Tucker
Institution: Carnegie Mellon University
Language: English
Subfield: Statistical and Nonlinear Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,190
OpenAlex ID: W7154256022
