Reliability Analysis and Mechanical Characterization of Lead-Free Solders Using Interpretable Machine Learning

Abstract

Predicting solder joint reliability and efficiently characterizing the mechanical properties of lead-free solder materials remain critical challenges in the electronics industry. Accurate reliability prediction is complicated by the large number of design and process parameters, as well as the complex nonlinear interactions among thermal, mechanical, and material-related factors that govern solder behavior under harsh operating conditions. Traditional reliability assessment methods, including analytical and numerical approaches, often rely on simplifying assumptions that limit their ability to capture these nonlinear relationships, thereby reducing prediction accuracy. At the same time, conventional mechanical characterization techniques, such as tensile, creep, and shear testing, are experimentally intensive, time-consuming, and resource-intensive. These challenges highlight the need for efficient, interpretable, and data-driven methodologies to improve both reliability prediction and mechanical characterization of lead-free solder materials. Recently, machine learning (ML) techniques have emerged as promising tools for addressing complex reliability and material characterization problems in electronics manufacturing. Their ability to model nonlinear relationships and incorporate numerous interacting variables has enabled more accurate prediction of solder joint behavior compared to conventional approaches. However, despite their predictive capabilities, many ML models suffer from limited interpretability, which reduces confidence in their predictions and limits their adoption in reliability-critical applications. Therefore, developing interpretable and robust data-driven frameworks is essential for enabling trustworthy decision-making in solder reliability assessment and material development To address these challenges, this dissertation develops a sequential data-driven framework through three interconnected studies. The first study establishes an interpretable reliability-prediction framework for solder joints subjected to thermal cycling by integrating linear mixed-effects (LME) models and artificial neural networks (ANNs) with sensitivity analysis. By leveraging both experimental and simulated reliability data, this study demonstrates how interpretable ML models can accurately predict solder joint life while simultaneously identifying the critical factors governing reliability behavior. The extracted insights align with empirical knowledge and experimental observations, providing a transparent foundation for data-driven reliability assessment. Building on this foundation, the second study addresses one of the major limitations in accelerated life testing (ALT): the handling of censored reliability data. In this study, advanced survival ML methods, including random survival forests (RSF), survival support vector machines (SSVM), and extreme gradient boosting (XGB), are employed and compared against traditional survival analysis models such as Weibull regression and the Cox proportional hazards model (Cox-PHM). Furthermore, Shapley additive explanations (SHAP)-based interpretability analysis is incorporated to explain the influence of input variables on reliability predictions. The results demonstrate that survival ML methods provide superior predictive performance and improved modeling of complex nonlinear reliability behavior under censored conditions. Finally, the third study extends these data-driven methodologies to efficiently characterize lead-free solder alloys by introducing a transferability-guided transfer learning (TL) framework for predicting tensile strength. In this framework, knowledge learned from data-rich conventional solder alloys is transferred to newly developed, data-scarce alloys, enabling accurate mechanical property prediction with limited experimental data. The proposed TL framework is benchmarked against multiple conventional ML approaches and demonstrates improved robustness and predictive accuracy in low-data scenarios. In addition, transferability and interpretability analyses provide insights into the effectiveness of knowledge transfer and the physical consistency of the learned relationships. Collectively, these three studies establish a comprehensive, interpretable, data-driven framework that progresses from solder joint reliability prediction to censored reliability modeling, and finally to mechanical property characterization and alloy development. The proposed methodologies accelerate the evaluation, optimization, and adoption of high-reliability lead-free solder materials in advanced electronic packaging applications.