Heat source field inversion and detection based on physics-informed deep learning

Abstract

Heat source field inversion and detection (HSFID) has drawn increasing attention as the exponentially growing application for integrated circuits, which offers promising way for determining the system's unnormal operation condition. In the HSFID, embodying the physical constraints in the neural networks could significantly reduce the data demand for training and offer higher reconstruction accuracy. In present study, the physics-informed neural network (PINN) is employed to achieve the goal of HSFID. The problem of reconstructing the heat source field is transformed into the challenge of temperature field reconstruction. The PINN is employed to conduct the HSFID with various locations, shapes, sizes and power densities under multi-heat source configurations. For the two-source configuration, the heat source shape and position similarity (HSSPS) for detecting triangular heat sources is 98.9 %, meanwhile for four heat source configurations, the HSSPS is 93.5 %. In complex heat source systems where the location, shape, size and power density change randomly and simultaneously, the maximum temperature mean absolute error (TMAE) value is around 0.003 K, the maximum value of the temperature absolute error (M-TAE) value fluctuates in the range of 0.02 K, and the HSSPS is not less than 92 %.