Universal and Task‐Specific Machine‐Learning Interatomic Potentials With Eleven‐Element Database for Thermal‐Transport Modeling of Semiconductors and Interfaces

ABSTRACT The continuous miniaturization and integration of electronic devices have established thermal transport of semiconductors and crossing interfaces as a critical bottleneck for heat dissipation. The reliable machine‐learning interatomic potentials (MLIPs), together with tractable and dedicated training datasets, offer an accurate and efficient computational approach to simulate these challenging systems. Here, we construct a Semiconductor and Interfaces Database (SemiCID), covering most of the III‐IV‐V semiconductors and their heterointerfaces. Using the ultrasmall and superlinear SUS 2 ‐MLIP framework ( Proc. Natl. Acad. Sci. U.S.A . 122 , e2503439122 (2025) ), we strategically developed a family of ready‐to‐use MLIPs comprising one universal model (SUS 2 ‐SemiCID‐U) for general‐purpose applications and 60 task‐specific models (SUS 2 ‐SemiCID‐TS) for domain‐targeted studies. With these potentials, we accurately reproduce the thermal conductivity of 19 bulk semiconductors, showing very good agreement with experiments. Simulations of the BAs‐based heterostructures—including BAs/BP, BAs/BSb, BAs/AlN, BAs/GaN, and BAs/InN—together with the GaN/AlN reveal a positive correlation between interfacial thermal conductance (ITC) and acoustic‐phonon spectral matching, consistent with physical understanding and measured data for some of them. Taking AlAs/GaAs superlattices as examples, we quantitatively reproduce the observed suppression of thermal conductivity due to interfacial roughness. This work establishes a scalable platform for advancing thermal transport studies of semiconductors and interfaces in realistic electronic devices.