Domain features-informed two-step machine learning: accelerating the search for superlubric heterostructures

Searching for superlubric heterostructures composed of transitional metal dichalcogenides monolayers is challenging due to the variety of constituent elements. In this study, a two-step machine learning approach based on domain features is employed to efficiently tackle this challenging task. Machine learning models are trained to predict complex domain features from structural features. Bayesian optimization is then used to search for superlubricants. Machine learning models are iteratively rechained based on a small number of high-accuracy calculations, saving computational time and ensuring accuracy. MoS2/WS2, MoS2/VS2, and NiS2/NbSSe heterostructures have been identified as superlubric heterostructures and confirmed through theoretical calculations. Under 1 ~ 5 N, the experimental friction coefficients at the interface of MoS2/WS2 are 12% ~ 36% lower compared to MoS2/MoSe2, which has previously been proven to exhibit superlubricity. These results validate the effectiveness of the two-step machine learning approach in searching for superlubric heterostructures in a significantly reduced time.