Predictive Machine Learning Molecular Dynamics of SEI Formation in Concentrated LiTFSI and LiPF6 Electrolytes for Lithium Metal Batteries

High-energy-density lithium metal batteries require electrolytes that enable fast ion transport and form a stable solid-electrolyte interphase (SEI) to sustain high-rate cycling, a process that remains challenging to capture experimentally. Here, we develop a Deep Potential-based machine learning molecular dynamics (MLMD) framework, trained on extensive ab initio datasets and validated against experimental transport properties, to resolve early-stage SEI nucleation at lithium metal interfaces with quantum accuracy. We find that at the Li-metal interface, 3.5 M LiTFSI/DMC induces spontaneous, thermally activated reduction reactions, yielding rapidly growing thick anion-derived SEIs enriched in O/F-containing species. In contrast, 1.5-2.5 M LiTFSI/DMC and 1 M LiPF6/EMC/DMC/EC form thinner, LiF-dominated interphases with slower growth kinetics. Our modeling results are consistent with experimental observations, where 3.5 M LiTFSI enhances cycling stability and rate capability, while lower concentrations result in weaker passivation. Our MLMD framework efficiently captures the electrolyte transport and early-stage SEI formation mechanisms in LMBs.

💡 Research Summary

This paper presents a Deep Potential (DP)–based machine‑learning molecular dynamics (MLMD) framework that achieves near‑ab initio accuracy while retaining the scalability needed to study solid‑electrolyte interphase (SEI) formation on lithium metal anodes. The authors generated extensive ab initio molecular dynamics (AIMD) datasets for four electrolyte compositions: 1 M LiPF₆ in a mixed EC/EMC/DMC solvent, and 1.5 M, 2.5 M, and 3.5 M LiTFSI in dimethyl carbonate (DMC). Using DeepMD‑kit, they trained neural‑network potentials that reproduce AIMD energies and forces with low root‑mean‑square errors and R² values of 0.96–0.99, confirming quantum‑level fidelity.

Validation against experimental transport data shows that the DP models predict ionic conductivities and viscosities across the concentration range with markedly better agreement than classical force fields (CMD). Radial distribution functions (RDFs) and coordination numbers reveal that CMD systematically overestimates Li–O distances in LiTFSI and compresses Li–F distances in LiPF₆, reflecting the inability of non‑polarizable force fields to capture many‑body and polarization effects. In contrast, the DP potentials capture tighter Li–O coordination in LiTFSI and more realistic Li–F separations in LiPF₆, leading to accurate structural descriptions of concentrated electrolytes.

The core of the study focuses on early‑stage SEI nucleation at a lithium metal slab. Large‑scale DP‑MD simulations (up to ~100 ps) show that 3.5 M LiTFSI/DMC undergoes rapid, thermally activated reduction at the metal surface. Li–O and Li–F bond counts increase logarithmically with time, reaching a nucleation rate of ~10 Å⁻¹ per 100 ps. This fast growth yields a thick, inorganic‑rich SEI dominated by oxygen‑ and fluorine‑containing species (e.g., Li₂O, LiF, mixed Li‑O‑F phases). By contrast, 1 M LiPF₆/EC‑EMC‑DMC forms far fewer Li–O bonds and a slower increase in Li–F bonds, resulting in a thinner, LiF‑rich interphase. Intermediate concentrations of LiTFSI (1.5 M, 2.5 M) display behavior between these extremes, with modest bond formation and thinner SEI layers.

Spatial density profiles along the surface normal (z‑direction) illustrate a two‑stage process: an initial sharp peak of O and F atoms at the interface within 2–5 ps, followed by broadening and penetration of these species into the lithium substrate by 100 ps. High‑concentration LiTFSI maintains multiple high‑intensity fluorine layers, whereas LiPF₆ shows a more diffuse fluorine distribution. The authors correlate these atomistic findings with experimental electrochemical performance: the 3.5 M LiTFSI electrolyte enables stable cycling at 2 C and superior capacity retention at 1 C, attributed to its robust inorganic SEI despite higher initial impedance. Lower‑concentration electrolytes and LiPF₆ deliver lower impedance but suffer from poorer long‑term protection and faster capacity fade.

The study acknowledges limitations: the simulations span only sub‑nanosecond timescales, capturing initial nucleation but not the multi‑layer growth and aging of SEI observed experimentally over many cycles. Moreover, while the training set is extensive, highly decomposed, defect‑rich configurations that may appear at later stages could lie outside the model’s interpolation domain. Future work should extend the DP‑MD approach to longer times, incorporate applied potentials to mimic electrochemical bias, and integrate multi‑scale modeling of charge transfer across the evolving SEI.

In summary, the authors demonstrate that a Deep Potential‑based MLMD framework can simultaneously predict bulk electrolyte transport properties and resolve atomistic SEI formation pathways with quantum accuracy. The approach provides a powerful predictive tool for rational electrolyte design, highlighting the trade‑off between high‑concentration LiTFSI (thick, inorganic‑rich SEI, excellent high‑rate stability) and low‑concentration or LiPF₆ systems (thin, LiF‑dominated SEI, lower impedance but reduced durability). This work bridges the gap between computational efficiency and chemical fidelity, advancing the understanding of interfacial chemistry in lithium metal batteries.