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

Area of Science:

• Materials Science
• Computational Materials Science
• Semiconductor Device Physics

Background:

• Traditional copper interconnects face challenges in advanced integrated circuits, including increased resistivity and reduced electromigration lifetime.
• Emerging materials with high cohesive energy and low resistivity-density-lifetime product (ρ₀ × λ) are sought as alternatives.

Purpose of the Study:

• To accelerate the discovery of novel binary intermetallic compounds for next-generation interconnect materials.
• To identify materials with improved performance over traditional copper interconnects.

Main Methods:

• Active learning coupled with density functional theory (DFT) computations for accelerated material screening.
• Interpretable machine learning, specifically Shapley additive explanations (SHAP), for physical insights.

Main Results:

• Screening of 100 binary intermetallic compounds after five active learning iterations.
• Achieved a 76% proportion of promising materials, significantly outperforming random screening (4.9%).
• Identified key material descriptors: small cell volumes and similar Mendeleev numbers correlate with low ρ₀ × λ values.

Conclusions:

• Machine learning techniques show immense potential for discovering high-performance interconnect materials.
• Promising intermetallic candidates like VMo, IrRh₃, PtRh₃, NbRu, and CrIr₃ identified as potential replacements for Cu interconnects.
• The study demonstrates the efficacy of active learning and interpretable AI in materials discovery.