Reflow Recipe Establishment Based on CFD-Informed Machine Learning Model
Yangyang Lai, Jong Hwan Ha, Karthik Arun Deo, Junbo Yang, Pengcheng Yin, Seungbae Park
Abstract
This article proposes a computational fluid dynamics (CFD)-based machine learning (ML) model to set the temperature of a convection reflow oven and control conveyor speed. Due to the thermal mass variation of printed circuit boards (PCBs) and surface mount components, the reflow recipe should be specified for each PCB assembly individually. Instead of relying on process engineers’ skills and experience, the ML model supported by heat transfer mechanisms provides a clear and logical path in the reflow recipe setup. An ideal recipe should let the largest component reach the minimum reflow temperature; meanwhile, the temperature cannot damage the small or heat-sensitive components. To simulate the reflow soldering process, the CFD model was used after being validated with experiment results. A system was programmed to feed the output of the CFD model as the input to the ML model. Benefiting from the interpolation and fast response brought by ML models, we can rapidly obtain the results subjected to millions of recipes, which are the combination of preset temperatures of all the heating zones and the conveyor speed in the oven. Using a weighting filter function, the optimal recipe can be obtained. This methodology is especially beneficial to reflowing smaller batches of complex PCB assemblies with a big heat capacity.