Transfer learning enabled geometry, process, and material agnostic RGNN for temperature prediction in directed energy deposition

Abstract

Thermal simulation in directed energy deposition is an area of active interest, as the evolution of the temperature field during deposition is a key variable that affects resulting build properties. Surrogate machine learning models enable an accelerated alternative to finite element analysis, but often face challenges in generalizing new inference conditions. In this work, we present a geometry- and process-agnostic recurrent graph neural network (RGNN) that preserves the mesh connectivity through graph nodes and edges. Compared to finite element analysis, the RGNN model provides thermal predictions up to 405 times faster in computational speed. Furthermore, we demonstrate the effectiveness of transfer learning (TL) via fine-tuning to adapt the pretrained model for a new material system with distinct heat transfer characteristics. We evaluate the TL model in a challenging scenario, where it successfully predicts thermal behavior over an extended 1000 time steps using a new geometry. The TL model exhibits much lower error accumulation over time compared to the pretrained model, while requiring only a fraction of training data and training time. The TL predictions show a good match with an experimental temperature field obtained from IR imaging, demonstrating its robustness and adaptability.