Laser power modulation for improving laser soldering defects via LSTM and CNN models

Abstract

Achievement of high yields of using solder Sn-3.0Ag-0.5Cu (SAC305) in the large-scale soldering processes was still a formidable challenge for the field of electronic packaging. It was difficult to completely eliminate the defects by straightforward parameters tailoring or metallurgical adjustments. This work novelly proposed a LSTM (Long Short Term Memory) and CNN (Convolutional Neural Network) network to adjust the heat input for the processes optimization by the modulation of waveform. In this work, the seamless transition from long-term time coding to defect classification was realized by using LSTM and CNN models to predict the optimized process. The power data were obtained and fed to the LSTM network to predict the temperature curves. Subsequently, each temperature curve was transferred to a tensor and utilized to identify the defects. Finally, the range of optimized waveforms was obtained. The results demonstrated the LSTM and CNN models had the excellent performance which for LSTM, MAE, MSE, RMSE and R² were 0.03356 °C, 0.001361 °C², 0.036892 and 0.978209, respectively; for CNN, the accuracy exceeded 89 %. Type 1 waveforms were found to consistently yield optimal joint formations by enhancing melting and wetting, albeit with a risk of substrate distortion, whereas Type 3 and Type 4 waveforms were associated with inadequate wetting. High-speed imaging analysis further revealed that waveform modulation could effectively adjust heat input at different stages, promote better wetting and reduce thermally induced defects. This work will provide an innovative method to improve the soldering of SAC305 in the actual production, widen the application of LSTM and CNN in the field of laser soldering and expand the tailoring methodologies to other fields.