PhononScore: a phonon-aware scoring function for dynamical stability

Abstract

In recent years, crystal generation models have enabled the design of massive numbers of candidate materials. However, the lack of dynamical stability among generated structures has become a major bottleneck preventing their translation into practical materials discovery. To address this challenge, we propose PhononScore, a phonon-aware scoring function for crystal generation. Unlike computationally expensive explicit phonon calculations, PhononScore predicts a unified stability score from crystal structures, enabling ranking of candidate materials dynamical stability with second-level computational cost. We construct a multi-fidelity phonon dataset containing 157,463 crystal structures. On the PhononBench benchmark, PhononScore improves the average dynamical stability rate of candidate pools generated by nine crystal generation models from 30.7% to 83.7%, achieving a 2.72-fold enrichment of stable structures, while the average stability rate of the Top-10 candidates reaches 97.5%. On a high-fidelity DFT-PBE phonon benchmark, the DFT-finetuned PhononScore-DFT increases the Top-100 stability rate to 93.0% and achieves 5-6-fold enrichment of dynamically stable structures under an extremely imbalanced hard-screening scenario. As a materials-screening tool analogous to scoring functions in drug discovery, PhononScore can serve directly as a dynamical-stability feedback signal for crystal generation, active learning, and reinforcement learning, enabling second-level stability-aware reranking without explicit phonon calculations and providing a unified and efficient dynamical stability evaluator for high-throughput materials discovery, active learning, reinforcement learning, and closed-loop inverse design. The online PhononScore platform is available at: http://phononbench.cn/phononscore/