ThermoDynaSMT: A Physics-Informed Machine Learning Model for Component Displacement in Surface Mount Technology

Abstract In surface mount technology (SMT), the precision of micro-scale passive components (MSPC) placement on printed circuit boards (PCBs) is critical, influenced by the self-alignment behavior after the soldering reflow process (SRP). It is desired to accurately predict the displacement of MSPC after the self-alignment to better control the process parameters of SMT that includes multiple manufacturing stages. However, existing approaches cannot accurately explain systems governed by multiple physical phenomena, specifically heat transfer and surface energy, which are coupled through a sequential interaction during SRP. This study introduces a multiphysics-informed machine learning (ML) model named ThermoDynaSMT, which combines the dynamics of heat transfer of SRP with the mechanics driven by surface tension. Specifically, the proposed ThermoDynaSMT sequentially integrates a two-dimensional physics-informed neural network for thermal profiling on PCBs with a physics-informed ML model to predict MSPC displacement. This approach adopts adaptive loss weights to optimize convergence and reduce training epochs effectively using only one SRP oven recipe and one postreflow PCB inspection dataset, thereby satisfying all boundary conditions efficiently. The model significantly predicts thermal profiles at three PCB locations under two testing recipes and achieves an average mean absolute error (MAE) of 10 μm in width and 7 μm in length across 11 PCBs for displacement predictions.