IR Spectra for the EMIM-TFSI Ion Pair Using Deep Potentials.

Abstract

Despite advances in the characterization of ionic liquids (ILs), elucidating their infrared (IR) spectra remains challenging due to the computational demands of ab initio methods. In this study, we employ a framework that integrates deep potential (DP) and deep Wannier (DW) models to investigate the configuration, dipole moment, and IR spectra of a 1-ethyl-3-methylimidazolium bis(trifluoromethylsulfonyl)imide ([EMIM]⁺-[TFSI]^-) pair. The accuracy and reliability of these models are evaluated by benchmarking against ab initio molecular dynamics (AIMD) across structural, dipolar, and spectral features. Our results demonstrate overall agreement while emphasizing the importance of achieving well-converged dipole distributions─typically requiring tens to hundreds of picoseconds of simulation─to enhance spectral resolution. Such convergence is essential for minimizing noise or bias arising from specific ionic configurations (referred to as "on-top" or "in-front" in the current study) and is enabled by the computational efficiency of DW- and DP-based molecular dynamics (DW/DPMD), which supports long simulation time scales. The DW/DPMD approach reproduces both the dipole moment range (7-16 D) and the average (∼10 D) observed in AIMD while yielding smoother and better-converged distributions. Furthermore, the IR spectrum obtained from DW/DPMD closely aligns with that of AIMD, faithfully capturing key vibrational features such as v_{S - N - S, as} < v_CF3 < v_SO2, as, consistent with experimental observations. In contrast, classical IR spectra tend to underestimate or overestimate the intensities of specific bands and fail to reproduce the correct relative wavenumbers compared to AIMD and experimental data. This study highlights the capability of deep learning potentials and dipole models─particularly DP and DW─to address systems involving charged species and complex ionic interactions while illustrating the limitations of classical approaches. Our findings pave the way for the development of more advanced surrogate models and their application to increasingly complex systems, including bulk materials and interfaces.