Machine Learning-Driven Insights into Oxygen Diffusion Barriers—A Closed-Loop Approach for Optimizing Superconducting Thin Films from First Principles

The formation of surface oxides on superconducting thin films is known to introduce two-level systems (TLSs), which significantly contribute to losses in superconducting qubits. To mitigate this issue, a viable strategy is to introduce an oxygen diffusion barrier using a metal cap with a high affinity for oxygen. We present a machine learning (ML) approach that is iteratively trained on experimental data characterizing the effectiveness of various diffusion barrier metals, facilitating the recommendation of promising candidates for future testing in a closed-loop manner. The model utilizes metal interstitial energies and oxide vacancy energies, either directly calculated from density functional theory (DFT) or inferred from the Materials Project using a novel neural network approach. These data, combined with a set of experimental observations and a quantification of uncertainties inherent in those experiments, are analyzed using logistic regression. This produces a set of predictions with quantified uncertainties on which materials for a given metal cap will prevent oxide formation at the metal/niobium interface. Further experiments were conducted to validate our predictions, demonstrating high reliability in identifying new oxygen diffusion barrier materials. Finally, because the vacancy and interstitial descriptors we use are directly related to fundamental materials processes, our final model provides a natural physical interpretation of which materials make effective diffusion barriers. These findings highlight the effectiveness of our ML technique in closing the theory-experiment loop, accelerating materials characterization and discovery, and providing valuable insights for the rational design of new materials for superconducting and other devices.