부산대학교 도서관

Based on the deep energy method recently brought forward to handle linearelastic or hyper-elastic finite deformation problems in solid mechanics, in this paper, we propose a deep heat energy method (DHEM) which is specially tailored to deal with structural steady-state heat conduction problems with the help of deep learning techniques. In our work, the deep neural networks are utilized to construct the admissible temperature fields; secondly, the potential energy functional in the heat conduction process which works as the loss function of the deep neural networks is calculated by numerical integration techniques; finally, the parameters of the network including weights and bias, are optimized by the quasi-Newton method to yield the minimal of the potential energy functional which indicates the heat conduction has entered a steady state. Numerical examples with a diversity of materials, including the isotropic and homogeneous material, the orthotropic material, the non-homogeneous materials and temperature dependent materials, are carried out to illustrate the validity and capacity of DHEM in both linear uncoupled and thermal-material coupling heat conduction problems.