Physics-informed neural network for transient behavioral modeling of thermoelectric generators

This study constructs Physics-Informed Neural Network (PINN) for transient behavioral modeling of thermoelectric generators (TEGs) under complex thermal boundary conditions. This meshfree method does not need a labeled dataset, which integrates space–time variables as inputs and considers temperature distribution and electric current density as output parameters. In addition, the time-dependent boundary conditions should also be integrated into the input layer, because the complex thermal boundary is dependent on the input time variable and this physical process should be constructed in the network. This improved PINN structure ensures the physical information accuracy. Based on adapting the weights on the loss terms, PINN is trained and the time-varying thermoelectric coupling field can be obtained. Results show that thermal boundary conditions and TE material parameters can affect thermal stress and output power simultaneously. In order to analyze structural safety and conversion performance of TEGs simultaneously, an optimization purpose function is defined that considers variations in thermal stress and output power. By combining PINN and optimization purpose function, it is possible to optimize TEGs with lower thermal stress and higher output power simultaneously. Some typical numerical examples are presented, which are demonstrated with different time-dependent boundary conditions such as sinusoidal wave, square wave, and gaussian pulse. These research results will offer helpful information for the optimization design and promote the practical application of TEGs.