Machine-Learning-Accelerated Discovery of Polar Metals

Polar metals exhibit the rare combination of electrical conductivity and broken inversion symmetry, offering exotic properties such as superconductivity and high piezoelectricity. Traditional computational searches for new materials are performed primarily using density functional theory (DFT), resulting in computationally expensive procedures often limited in scope. In this project, we developed a two-stage, machine-learning-integrated workflow to accelerate the process of identifying new polar metals. Classifiers trained on 20,585 materials from the Materials Project database were used to predict stability and metallicity, narrowing a broad pool of 139 compositions to only 11 candidates. Structural identification was then performed using the MAGUS framework, which incorporates machine-learned interatomic potentials and an evolutionary algorithm to generate high-quality candidate structures. Subsequent DFT and phonon calculations revealed a polar, metallic structure with minimal phonon instabilities and energy within 0.1 eV/atom of the convex hull, indicating potential metastability. The workflow achieved results comparable to traditional DFT searches while using less than 5% of the computational resources, demonstrating the efficiency of machine-learning integration.