**ForgeNet: A Deep Learning-Based Surrogate Model for Cold Forging Simulation with Adaptive Remeshing**
Please login to view abstract download link
Cold forging simulation relies on nonlinear FEM to capture large plastic deformations and high contact pressures, but adaptive remeshing breaks the node correspondence that standard sequence models require. We address this by projecting FEM output at every timestep onto a fixed 512 × 512 Eulerian grid via Delaunay triangulation with adaptive alpha-shape filtering and exact barycentric interpolation, converting each remeshed simulation into a temporally consistent spatial sequence. On this representation, we train two independent Transolver-based models on cold forging trajectories generated in ANSYS: a node-type model that predicts the binary metal/air activity mask and a stress model that predicts the equivalent von Mises stress field. We present a framework that combines fixed-grid projection of an adaptively remeshed simulation with a two-model decomposition treating boundary evolution and stress evolution as separate prediction problems. For cold forging surrogates, this combination has not appeared in prior work to our knowledge. Prior to full model coupling, isolated autoregressive rollouts demonstrate that both models are highly stable: the node-type model sustains IoU above 0.99 across all 2000 simulation timesteps, and the stress model exhibits bounded MAE drift under 2% of peak observed stress with an 𝑅² value of 0.924. When coupled in full cascade, this stability does not transfer, and the joint rollout degrades within the earlier time steps. We identify this coupled autoregressive instability as the central open problem from this work and outline directions for follow-up investigation.
