Machine Learning Aided Computational Discovery of Boride Substrates for Borophene Synthesis
Borophene, a monolayer of boron, is distinctive among two-dimensional materials owing to its unique properties, including polymorphism and potential applications in quantum technologies and energy conversion and storage. Nevertheless, its applications are hindered by the strong binding of borophene to metal substrates, which impedes the realization of freestanding borophene. Alternative substrates with tunable binding to borophene are therefore highly desirable. Since boride layers arise during borophene synthesis on metals, borides provide potential alternative substrates for borophene synthesis. In this work, we investigate the binding energies and interactions between borophene and 1356 boride surfaces for two borophene phases (v1/6 and v1/9), using density functional theory (DFT) calculations and machine learning (ML). Importantly, incorporating ML-predicted charge transfer as an input feature to a subsequent ML model enables physics-informed predictions of binding energies. Additional DFT calculations on four representative cases were performed to validate the ML predictions. SHAP analysis reveals that the majority of key descriptors are associated with delocalized electrons, highlighting the role of electronic properties in governing binding strength. Combined DFT and ML screening identifies 65 and 73 promising boride surfaces for the v1/6 and v1/9 phases, respectively. Al- and Ag-based borides, including AlB2 and AgB2, emerge as predominant candidates for both phases. The computationally predicted candidate substrates and the deduced design guidelines provide a charge-transfer-informed ML strategy for substrate selection and represent promising steps toward the experimental realization of freestanding borophene.