Deep learning-based modeling and simulation of heat conduction

Abstract

The present study focuses on applying deep neural networks (DNNs) to surrogate modeling of heat conduction problems. Deep learning algorithms, valued for their ability to learn hierarchical data representations through multi-layered networks, excel at identifying complex patterns. In this work, the U-Net architecture – widely recognized for its effectiveness in image segmentation – is adapted to model stationary heat transfer, providing a novel approach to a critical challenge in engineering and physics. Specifically, we propose a deep learning-based surrogate model to predict stationary temperature fields in 2D rectangular domains representing two-phase heterogeneous materials.