Data-Driven Analysis of Nuclear Resonance Vibrational Spectra with Machine-Learning Potentials

5 Jan 2026, 16:44
3m
Siemens Auditorium (ETH Zurich)

Siemens Auditorium

ETH Zurich

Campus Hönggerberg
Poster and Flash Presentation Flash Presentations

Speaker

Alexey Rulev (Empa)

Description

Nuclear Resonance Vibrational Spectroscopy (NRVS) is a synchrotron-based inelastic X-ray scattering technique that probes the vibrational density of states projected onto a Mössbauer isotope. NRVS spectra can be transformed into an element-projected phonon density of states (PDOS) integrated over the Brillouin zone. While the forward problem - calculating the PDOS from a known structure and interatomic forces - is straightforward, the inverse problem of extracting structural and kinetic information from experimental PDOS is considerably more challenging and requires advanced modelling and data-driven analysis.
In this work, we report the first operando NRVS measurements of a LiFePO4 (LFP) electrode in a Li-ion cell. During cycling, lithium is extracted from LFP, forming FePO4 via an intermediate metastable phase that is only rarely reported in the literature. Vibrational spectroscopy probes the local curvature of the Born–Oppenheimer surface, which is crucial for understanding phase-transformation kinetics and ionic transport. Using machine-learning-based techniques such as principal component analysis, we detect the signature of a third, metastable phase formed during charge–discharge. Non-negative matrix factorization enables us to decompose the raw spectra into contributions from individual phases, including the metastable intermediate.
To interpret these components structurally and kinetically, we perform ab initio phonon calculations for candidate structures and match the computed PDOS to experiment. For the metastable phase, we move beyond conventional DFT to a neural-network-based universal interatomic potential, that we fine-tuned on our DFT dataset, which allows us to simulate substantially larger supercells with diverse defect arrangements. We first tune the DFT-NRVS agreement for the stable phases and then identify configurations that best reproduce the experimental spectra of the intermediate. This workflow yields insight into the structure, thermodynamic stability, and transformation kinetics of the metastable phase. Finally, we apply the same data-driven NRVS–simulation framework to vibrational spectra of ceramic proton conductors, establishing quantitative links between phonons and transport of light ions such as Li+ and H+.

Author

Alexey Rulev (Empa)

Co-author

Artur BRAUN (Empa)

Presentation materials

There are no materials yet.