Towards Structure-Preserving Deep Learning for Welding: A GENERIC-based Graph Network Approach
Please login to view abstract download link
Simulating industrial welding processes is essential for manufacturing quality, yet classical Finite Element Methods (FEM) are often too computationally expensive for real-time applications like Digital Twins. While Deep Learning offers acceleration, ensuring physical consistency remains a significant challenge. To address this, we present a framework based on MeshGraphNets enhanced by the GENERIC formalism. By imposing a structural bias, this formalism restricts the learning process to identify operators that satisfy the GENERIC bracket structure, guaranteeing thermodynamic consistency—strictly conserving energy and ensuring non-negative entropy. Applied to a nonlinear thermal problem using a Goldak double-ellipsoid heat source, the model achieves inference-speed simulation capabilities. It generalizes effectively to unseen geometries, material parameters, and source trajectories, paving the way for integrating fast, physics-aware surrogates into complex industrial workflows.