All-solid-state batteries (ASSB) are widely recognized as a safer and potentially more energy-dense alternative to conventional lithium-ion technologies. Their performance critically depends on fast ionic conduction within solid electrolytes. Traditional methods to identify such materials involve labour-intensive synthesis and characterization processes, often hampered by the limitations of existing computational models in accurately capturing disordered, high-temperature ionic behaviours.
The detection and prediction of liquid-like ion motion in crystalline materials has remained a major challenge, particularly because conventional computational methods for calculating material properties in dynamically disordered systems are prohibitively expensive in computational terms.
This study presents a machine learning (ML) accelerated pipeline that integrates ML force fields with tensorial ML models to predict Raman spectra, demonstrating that significant low-frequency Raman intensity is a reliable spectroscopic indicator of liquid-like ionic conduction. When ions move in a liquid-like manner through a crystal lattice, they dynamically disrupt its symmetry, effectively relaxing Raman selection rules. This symmetry breaking produces characteristic low-frequency Raman scattering that can be directly linked to high ionic mobility. The new apporach enables near-ab initio accuracy in simulating vibrational spectra of complex, disordered materials at finite temperatures while reducing computational costs significantly. By applying this workflow to sodium-ion conductors like Na3SbS4, the researchers demonstrated that pronounced low-frequency Raman features, signatures of symmetry-breaking induced by rapid ion transport, serve as reliable indicators of fast ionic conduction. The approach not only rationalizes previous experimental observations but also opens avenues for high-throughput screening and discovery of superionic materials.
The approach was validated using sodium-ion conducting materials, where the method successfully identified clear Raman signatures associated with liquid-like ion conduction. Systems exhibiting strong low-frequency Raman features were found to correlate directly with high ionic diffusivity and relaxational host-lattice dynamics, while materials dominated by hopping-based conduction did not display such signatures.
By generalizing the breakdown of Raman selection rules beyond canonical superionic systems, the study establishes a unifying framework for interpreting diffusive Raman scattering across diverse material classes. The ML-accelerated Raman pipeline bridges atomistic simulations and experimental observables, enabling more efficient discovery and characterization of fast-ion conductors.
This work opens a new avenue for data-driven materials discovery in energy storage, offering a powerful tool for accelerating the development of high-performance solid-state battery technologies.
The research has been recently published in the online edition of AI for Science, a prominent international journal in the field of interdisciplinary AI research.
Reference: Manuel Grumet, Takeru Miyagawa, Olivier Pittet, Paolo Pegolo, Karin S Thalmann, Waldemar Kaiser, David A Egger. Revealing fast ionic conduction in solid electrolytes through machine learning accelerated Raman calculations[J]. AI for Science, 2026, 2(1): 011001. DOI: 10.1088/3050-287X/ae411a
