High-throughput search of topological materials for interconnects using first-principles transport calculations and machine learning

The performance of semiconductor-based computing technologies is increasingly hindered by the resistivities of back-end-of-line (BEOL) interconnect materials as device dimensions continue to shrink. Relative surface imperfections intensify in narrow wires, and conventional metals such as copper suffer strong nanoscale resistivity increases from surface scattering. Topological materials may address this bottleneck via topologically protected surface states. However, previously studied candidates have not reached interconnect-competitive conductance. To address this challenge, we combine first-principles electron-transport calculations with an active-learning workflow that couples Bayesian optimization (BO) and semi-supervised learning. A semi-supervised crystal graph neural network leverages a scarce amount of labeled data and an abundant number of unlabeled structures to predict bulk conductance and surface conductance ($G{_surf}$). BO accelerates materials discovery by intelligently deciding on the next best material to investigate. These predictions are integrated into a high-throughput, ML-based active-learning framework to accelerate the discovery of topological materials for BEOL interconnects. Our study aims to identify candidates with the potential to outperform copper at sub-10-nm dimensions.