Surrogate Modeling for Radiative Heat Transfer Analysis using Graph Neural Network

Abstract

This study aimed to develop an effective approach for fast simulation of material processing by surrogate modeling with graph neural network. Specifically, radiation problems which imitate heating process of thermoplastic materials were addressed. The training data for surrogate model was generated through calculations using finite element method as steady-state heat transfer analysis. The simulation model consisted of single material and six heaters. And the material was exposed to radiative heat from the heaters under different heating conditions with different material shapes. A structure of neural network model suitable for these problems was proposed, incorporating surficial graphs and edges to express pairs for radiative heat transport. The model predicted view factor on each edge using the information of relative coordinates between relevant nodes. It demonstrated good results in predicting radiative flux for various material shapes and conditions. Although there is still room for improving accuracy, fast prediction of radiation was achieved successfully.