Phonon-Driven Limitations of AI-Predicted Stability in Li-Ion Anode Materials: A Computational Meta-Analysis of Vibrational Instabilities

Abstract

Machine learning interatomic potentials (MLIPs) and high-throughput density functional theory (DFT) screening have dramatically accelerated the discovery of Li-ion battery anode materials; however, a critical gap persists between computational stability predictions and real-world synthesis feasibility. This study presents a systematic computational meta-analysis of phonon-driven instability among 79 Li-ion anode candidate structures spanning silicon-based (n = 34), carbon-based (n = 21), and lithium-metal phases (n = 24), all predicted to be stable by the Materials Project AI-assisted DFT pipeline (energy above hull, Ehull < 0.05 eV atom⁻¹). Using the finite-displacement method implemented in Phonopy and an empirical Born–von Karman force-constant model sampled at four high-symmetry Brillouin-zone points, 49 of the 79 structures (62.0%) were classified as dynamically unstable within the adopted phonon-screening framework. Carbon-based anodes exhibited the highest failure rate (95.2%), followed by lithium-metal phases (70.8%) and silicon-based anodes (35.3%). Instabilities were overwhelmingly concentrated at the Gamma point (98.0% of unstable structures), identifying the zone center as the predominant location of phonon instabilities across all three anode families. Chi-squared analysis confirmed that failure rates differed significantly among material families (χ² = 20.94, df = 2, p < 0.0001). These findings highlight important limitations in current AI-based screening frameworks for anode materials and support the incorporation of phonon instability screening as an additional post-processing step in high-throughput materials-discovery workflows.