Speaker
Description
Computational high-throughput screening for superconducting materials relies critically on the accurate assessment of dynamical stability. Systems with strong electron-phonon coupling are among the most promising candidates for high-temperature superconductivity, but are also precisely those most prone to lattice instabilities, including metastable phases such as hydrides at elevated pressures. Anharmonic effects are particularly significant in these and other light-element or structurally complex compounds, where harmonic approximations routinely fail to capture the true lattice behavior. The stochastic self-consistent harmonic approximation (SSCHA) represents an accurate method to capture these effects, but its combination with density functional theory (DFT) carries a substantial computational cost, rendering large-scale screening intractable in most settings.
Recently proposed workflows have addressed this bottleneck by employing system-specific machine-learning interatomic potentials (MLIPs) trained for individual systems, significantly reducing the computational burden while retaining DFT-level accuracy. In this work, we explore whether universal MLIPs (uMLIPs), trained across broad chemical spaces, can achieve comparable accuracy without the overhead of system-specific training. This can further lower the barrier to large-scale deployment and enable exploration of design spaces previously excluded based on harmonic stability analysis. We benchmark the uMLIP-based approach across a set of representative systems and discuss the feasibility of integrating anharmonic stability analysis into high-throughput screening workflows.