Cross-Alloy Transfer Learning With Transferability Assessment for Mechanical Property Prediction of Lead-Free Solders

Abstract

The mechanical properties of lead-free solder alloys, including yield strength (YS), ultimate tensile strength (UTS), and elastic modulus, are critical indicators of reliability. These properties are typically obtained through experimental testing; however, generating comprehensive datasets through such experiments is costly and time-consuming, particularly for newly developed alloy systems. To address this limitation, this study proposes a transferability-guided transfer-learning (TL) framework that leverages data-rich conventional alloys (e.g., SAC305) to predict the UTS of newly developed, data-scarce alloys (e.g., Innolot). The proposed framework is benchmarked against conventional machine learning (ML) models, including artificial neural networks, random forests (RFs), Gaussian process regression (GPR), support vector regression (SVR), and extreme gradient boosting, demonstrating superior accuracy and robustness in low-data regimes. For example, for SACQ, TL reduces the mean absolute error (MAE) by more than 50% relative to the best conventional ML model. Transferability analysis is employed to evaluate the effectiveness of knowledge transfer from data-rich source alloys to data-scarce target alloys, indicating stronger transfer to SACQ than to Innolot. In addition, interpretability analysis confirms that the TL model captures physically consistent trends, including strength degradation with increasing temperature and aging, and strengthening with higher strain rates. Overall, this work demonstrates that transferability-guided TL enables accurate prediction of mechanical properties for newly developed alloys using minimal experimental data, reducing cost and accelerating the design of high-reliability lead-free solder materials.