Overcoming sampling limitations using machine-learned interatomic potentials: the case of water-in-salt electrolytes

Abstract

Machine-learned interatomic potentials hold the promise to enable the modeling of highly concentrated liquids over meaningful timescales, far from reach for current ab initio electronic structure methods. Here we evaluate the performances of various MACE potentials in modeling a $21 m$ water-in-salt electrolyte based on lithium bis(trifluoromethanesulfonyl)imide. We test out-of-the-box foundation models, as well as both fine tuning and from scratch training strategies. Our simulations demonstrate that surrogate models allow to overcome sampling limitations of ab initio molecular dynamics, reaching an excellent agreement with experimental observables such as the structure factor. We also demonstrate the benefit of fine tuning a foundation model over training from scratch: in terms of data efficiency, but most importantly as a means to provide information regarding configurations hard to sample, such as short Li$^+$--Li$^+$ distances. Finally, we show that depending on the reference exchange-correlation functional, empirical dispersion correction schemes can be detrimental. All in all, our work shows that machine-learned interatomic potentials are a good fit for the modeling of highly concentrated electrolytes over long timescales.

Figures (5)

• Figure 1: (a) Snapshot of the 20.9 m LiTFSI water-in-salt electrolyte. The bottom panel shows the molecular depictions used throughout: TFSI$^{-}$, Li$^{+}$, and H$_2$O; (b) Workflows of the two different training strategies, training from scratch (TfS) and fine-tuning (FT); (c) Occurrence of unphysical Li-Li dimers over a time span of 1 ns for three different models. A snapshot of such a dimer is displayed in the bottom panel.
• Figure 2: (a,d) Log--log density plots of absolute prediction errors for force components, and per-atom energies as a function of the corresponding r$^2$SCAN reference magnitude for the FT model, evaluated on the same test set. (b,e) Learning curves of RMSE versus training-set size $N$ for forces (b), and energies (e), comparing FT (purple) and TfS (orange). Points are averages over three independent replicas. Dashed lines indicate the RMSE obtained when using the largest training-set size in this sweep. (c,f) Same as (a,d), but for the TfS model.
• Figure 3: Mean mass density $\rho$ (g cm$^{-3}$) of the 20.9 m LiTFSI WiSE. Results are shown for multiple MACE-based potentials, including FT, TfS, and foundation (F) models. Blue squares and red diamonds correspond to bare and D3-corrected models, respectively. Horizontal error bars indicate the estimated uncertainty from the finite sampling time. The vertical black line marks the experimental density.
• Figure 4: Transport properties computed using bare models. From left to right: diffusion coefficients ($D_{\mathrm{Li}}$, $D_{\mathrm{TFSI}}$, $D_{\mathrm{H_2O}}$), and Nernst-Einstein conductivity ($\sigma_{\mathrm{NE}}$). Shaded bands correspond to the experimental ranges.
• Figure 5: Structural analysis. (a) Time- and model-dependence X-ray structure factor $S(q)$, comparing the experimental reference (black symbols, from Ref. zhangWaterinSaltLiTFSIAqueous2021a) with estimates obtained from short (20 ps) and long (2 ns) MD trajectories.Legend labels report the $R$-factor (defined in the text) for each computed curve. The abbreviation FT refers to the fine-tuned FT_MATPES-R2SCAN model throughout. (b) Absolute deviation $|\Delta S(q)| = |S_{\mathrm{sim}}(q)-S_{\mathrm{exp}}(q)|$ for the three simulated curves shown in (a); the inset magnifies the $q=1$--$2~\AA^{-1}$ region. (c,d) Radial distribution functions (RDFs, $g(r)$) for Li--O(total) (solid lines, oxygen atoms from both H$_2$O and TFSI$^{-}$) and Li--Li (dashed lines, scaled by a factor of 18 for visibility), for (c) bare and (d) D3-corrected trajectories. (e,f) Total X-ray $S(q)$ from long-time MLMD trajectories for four selected models compared to experiment: (e) bare, (f) D3-corrected MLMD. Legend labels indicate the model abbreviation and the corresponding $R$-factor. In panels (e) and (f), F denotes the foundation model (F_MATPES-R2SCAN-OMAT-FT), TfS the trained-from-scratch model (TfS_C128_LMAX1), FT-S50 the few-shot fine-tuned model (FT_MATPES-R2SCAN_S50, 50 configurations), and FT the fully fine-tuned model (FT_MATPES-R2SCAN).