Two Stages Machine Learning-Based Design Optimization for Electromigration Reliability in Solder Joints

Abstract

For high-performance electronic components to function well and last a long time, trustworthy solder joint design is essential, particularly in situations with high current densities that cause electromigration (EM). Variations in solder joint shape, such as width, height, and radius, lead to uneven current densities and faster degradation, which exacerbate EM-induced defects, such as voids and cracks. In order to reduce EM effects and improve solder joint reliability, this article presents a two stages machine learning (ML) surrogate modeling framework for the inverse design of solder joints. The framework combines ML-based surrogate models, particularly polynomial regression and k-nearest neighbors, to approximate computationally demanding physics-based simulations. Specifically, a first ML model finds a subset of candidate designs in the input space that closely match the desired outcomes in the initial stage. In order to remove erroneous predictions and improve the outcomes, a second ML model integrated with conformal inference assesses these candidates under the guidance of a user-specified coverage level. This framework guarantees reliable and accurate designs by combining two stages ML models with uncertainty quantification using conformal inference. The proposed method delivers high predictive accuracy with reduced computational cost, offering an effective and practical solution for advancing electronic packaging design.