{"id":"arxiv/2609.10983","url":"https://arxiv.org/abs/2609.10983","tags":["math.NA","cs.LG"],"score":0,"title":"A variational physics-informed graph neural network for heterogeneous solid mechanics","domain":"math.NA","source":"arxiv","authors":["Aashay Rajan Yadav","Amiya Prakash Das","Ratna Kumar Annabattula"],"reviews":0,"subject":"math.NA","summary":"Stress localization in heterogeneous solids is governed by the bimaterial interface, where the displacement field remains $C^0$-continuous, while in-plane stresses jump due to the stiffness mismatch. Coordinate-based physics-informed neural networks (PINNs) represent this jump via a prescribed regularization width or a weighted interface penalty, making their accuracy sensitive to how phase-contrast changes are handled. This work presents a variational, label-free physics-informed graph neural network (PI-GNN) in which the heterogeneity is carried by the discretization rather than by the…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-09-10","reactions":{}}