Review: How DFTMD, machine-learning force fields and generative AI are used to model aqueous batteries

The review organizes molecular modelling methods for aqueous batteries into a spectrum from quantum to data-driven: DFT and DFTMD provide structural and reactive descriptions, empirical force-field MD extends accessible time and length scales, machine-learning force fields (MLFF) bridge the gap with near-quantum accuracy at lower cost, and generative AI is used for inverse design of molecules and materials. Compared with prior method-specific introductions, this work places electronic-structure, classical force-field, MLFF and generative models in one framework, noting that DFTMD is limited to tens of picoseconds and about 1000 atoms, while MLFF can push beyond those scales. A methodological review argument citing representative works such as Car-Parrinello, Behler-Parrinello, SOAP, SchNet and NequIP; it is a consensus-style synthesis rather than a single experiment.

In the water-in-salt electrolyte (WiSE) case, the review collects simulation-plus-experiment structural pictures: at the nanoscale, water-rich channels coexist with an ion-cluster scaffold, and Li+ diffusion in water channels is markedly higher than in ion clusters (e.g. 0.385 vs 0.05 Ų/s), which is used to explain the elevated Li+ transference number and the widened stability window linked to scarce interfacial water. It juxtaposes simulation results from different groups (Borodin, Lim, Yu, Li, Zhang and others) on 21 m LiTFSI structure and transport, and notes that the picture of whether solvation splits into two populations differs across studies. Mainly classical MD cross-validated with 2D-IR, SAXS and IR experiments; some conclusions, such as the diffusion-coefficient difference, are supported by quantitative values.

In the cathode case, the review shows how DFT and DFTMD are used to assess proton/cation intercalation energies, oxygen-vacancy formation, OER/LOER overpotentials and the role of structural water—for example, the lowest LOER overpotential of 0.56 V on LiCoO2, structural water in V2O5 nearly halving the Zn2+ diffusion barrier, and a theoretical capacity of 251 mAh g⁻¹ versus an experimental 381 mAh g⁻¹. It brings together simulation findings across Co, Mn, Ru, W and V oxides as well as transition-metal dichalcogenides and MXenes, highlighting how structural water and interlayer confinement affect intercalation thermodynamics and kinetics. Primarily DFT+U, SCAN, PBE, NEB and GCMC calculations, with some results compared against XRD, SAXS and electrochemical measurements.

In the Zn-ion battery and organic flow battery cases, the review illustrates applications of machine-learning force fields, reactive force fields and generative models: neural-network potentials for Zn2+ hydration, DFTMD observation of organic SEI formation, reinforcement learning to generate organic radicals with target redox potentials, and MLFF simulation of proton transport in Nafion membranes. It extends generative AI and MLFF from the methods chapter into concrete battery systems, emphasizing their potential in molecular and membrane materials discovery. A case-study synthesis involving combined AIMD, MD, DFT and experimental characterization, though most results come from individual systems.

This is a review aimed at researchers who want an overview of molecular modelling methods and applications for aqueous batteries, especially students entering the field and theory-experiment crossover researchers. Its conclusions apply to the discussed systems—water-in-salt electrolytes, transition-metal oxide cathodes, Zn-ion batteries and organic redox flow batteries—and the methodological framework can inform other electrochemical interface studies.

Different studies in the review give different pictures of the same system (e.g. solvation structure in concentrated LiTFSI), indicating that conclusions still depend on force fields and sampling; the time/length-scale limits of DFTMD, extrapolation risks of MLFF and the effectiveness of generative models in chemical space are open questions readers should keep in mind when adopting these methods. In addition, as a review it provides no new experimental data, so specific numbers should be verified against the original literature.