Active Learning for the Discovery of Binary Intermetallic Compounds as Advanced Interconnects

Abstract

The scaling of advanced integrated circuits has posed significant challenges for traditional Cu interconnects, including increased resistivity and reduced electromigration lifetime. Materials with high cohesive energy and low ρ₀ × λ values are emerging as promising alternatives. In this work, active learning coupling density functional theory (DFT) computation is employed to accelerate the discovery of binary intermetallic compounds for interconnect materials. Following five active learning iterations, 100 compounds are screened out. Among them, the proportion of promising materials reaches an impressive 76%, in sharp contrast to a paltry 4.9% under traditional random screening. Moreover, this research adopts an interpretable machine learning method to provide further physical insights. The Shapley additive explanations (SHAP) analysis revealed that binary intermetallic compounds featuring small cell volumes and similar Mendeleev numbers tend to possess low ρ₀ × λ values. Several promising intermetallic candidates were also identified, including VMo, IrRh₃, PtRh₃, NbRu, and CrIr₃, as potential alternatives to traditional Cu interconnects in future technology nodes. The findings in the study highlight the immense potential of machine learning techniques to accelerate the discovery of novel high-performance interconnect materials.