SRP-PINN: A Physics-Informed Neural Network Model for Simulating Thermal Profile of Soldering Reflow Process

Because of the tendency toward downsizing and the rising complexity of printed circuit board (PCB) design, monitoring and optimizing the soldering reflow process (SRP) has become an important but challenging task for surface mount technology (SMT). To ensure a PCB’s quality, the thermal behavior of the solder joints, which connect the electronic components to the PCB, should be precisely controlled to match the thermal profile specified by the solder joint manufacturer. Previous studies have been mapped the relationship between the process parameters and thermal profile by using physics-based techniques or traditional machine learning (ML) algorithms. However, those approaches require substantial computational costs or large data samples to capture the thermal profile accurately. This article proposes a physics-informed neural network (PINN), which is the first effective physics-based deep-learning framework for modeling the continuous nonlinear thermal behavior of a PCB during the SRP. Specifically, the governing equations of the system—including the general heat transfer and the convection partial differential equations (PDEs)—are solved by optimizing the parameters of a deep neural network (DNN) using a physics-based loss function. The performance of the proposed SRP-PINN is compared with benchmark ML algorithms to demonstrate its effectiveness and efficiency in predicting the thermal profile with limited experimental data for training, resulting in a surrogate model that can simulate the real-time thermal behavior of the board.