Exploring effective charge in electromigration using machine learning

TL;DR

This paper employs machine learning to model the effective charge in electromigration, achieving high accuracy and demonstrating potential for predicting alloy behavior and reliability in electronic interconnections.

Contribution

The study introduces a machine learning model that predicts effective charge as a linear function of elemental properties, with validated accuracy and extrapolation capabilities.

Findings

RMSE/$\sigma$ of 0.37 ± 0.01 indicating high model accuracy

R^2 value of 0.86 demonstrating strong predictive power

Limited but useful extrapolation to new alloys

Abstract

The effective charge of an element is a parameter characterizing the electromgration effect, which can determine the reliability of interconnection in electronic technologies. In this work, machine learning approaches were employed to model the effective charge (z*) as a linear function of physically meaningful elemental properties. Average 5-fold (leave-out-alloy-group) cross-validation yielded root-mean-square-error divided by whole data set standard deviation (RMSE/$\sigma$) values of 0.37 $\pm$ 0.01 (0.22 $\pm$ 0.18), respectively, and $R^2$ values of 0.86. Extrapolation to z* of totally new alloys showed limited but potentially useful predictive ability. The model was used in predicting z* for technologically relevant host-impurity pairs.