A Deep Learning Approach for Reflow Profile Prediction
Yangyang Lai, J. Kataoka, Ke Pan, Jonghwan Ha, Junbo Yang, Karthik Arun Deo, Jiefeng Xu, Pengcheng Yin, Chongyang Cai, Seungbae Park
Abstract
Commercial reflow ovens contain multiple heated zones, which can be individually controlled for temperature. PCB assemblies travel through each zone at a controlled rate to achieve the desired reflow profile. The reflow profiles of all the components on the board are supposed to fall within a safe range. The minimum reflow temperature should be reached for the largest component. Meanwhile, the temperature cannot exceed the threshold temperature that may damage the smallest components. CFD model is widely used to simulate the reflow soldering process. To hit the target reflow profiles, computational expense is required to seek the optimal boundary conditions, which are the preset temperatures of heating zones. The number of zones in the reflow oven determines the complexity of the combination of boundary conditions. To ease the computation cost, the deep learning approach based on the CFD model is employed to predict the reflow profile of a bulky BGA package. The neural network is demonstrated to rapidly predict transient temperatures of the BGA in seconds and provide an average error below 0.5 %.