Electronic Paddlewheels Impact the Dynamics of Superionic Conduction in AgI

Harender S. Dhattarwal, Richard C. Remsing · 2025

Read Electronic Paddlewheels Impact the Dynamics of Superionic Conduction in AgI by Harender S. Dhattarwal, Richard C. Remsing (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-16704.txt.

Introduction

Harender Dhattarwal and Richard Remsing model a superionic conductor in a chemistry paper. Silver iodide, in its hot superionic phase, lets silver ions flow almost like a liquid. That trait is wanted for solid-state batteries, but the electronic side of the motion is easy to miss in models.

The authors train a neural network potential to sample ion motion at near quantum exactness, then compute charge traits. Silver drift couples to swirling electron density on the nearby iodides, which they call electronic paddlewheels. Those wheels plainly change the friction a moving ion feels. Plain force fields miss the effect fully.

Catalog record

CREATORHarender S. Dhattarwal, Richard C. Remsing
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE45 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-16704

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Electronic Paddlewheels Impact the Dynamics of Superionic Conduction in AgI Harender S. Dhattarwal, Richard C. Remsing Abstract Solid-state ion conductors hold promise as next generation battery materials. To realize their full potential, an understanding of atomic-scale ion conduction mechanisms is needed, including ionic and electronic degrees of freedom. Molecular simulations can create such an understanding, however, including a description of electronic structure necessitates computationally expensive methods that limit their application to small scales. We examine an alternative approach, in which neural network models are used to efficiently sample ionic configurations and dynamics at ab initio accuracy. Then, these configurations are used to determine electronic properties in a post-processing step. We demonstrate this approach by modeling the superionic phase of AgI, in which cation diffusion is coupled to rotational motion of local electron density on the surrounding iodide ions, termed electronic paddlewheels. The neural network potential can capture the many-body effects of electronic paddlewheels on ionic dynamics, but classical force field models cannot. Through an analysis rooted the generalized Langevin equation framework, we find that electronic paddlewheels have a significant impact on the time-dependent friction experienced by a mobile cation. Our approach will enable investigations of electronic fluctuations in materials on large length and time scales, and ultimately the control of ion dynamics through electronic paddlewheels. I Introduction Increasing world energy demands requires increasing energy availability in a renewable and sustainable manner to avoid detrimental environmental effects 1 . Renewable energy sources are often intermittent, creating a need for efficient energy storage systems like battery technologies 2 , 3 . Most currently available battery technologies rely on liquid electrolytes, typically with Li + as the relevant electrolyte cation. These conventional liquid electrolytes exhibit several problems, including flammability, leakage, and limited electrochemical stability windows 4 , 5 . Alternatives such as water-in-salt electrolytes 6 , 7 , 8 , ionic liquids 9 , and their hybrids 10 , 11 are being developed to overcome some of these limitations. Solid-state ion conductors can mitigate many of these issues within solid-state batteries 12 , 13 . However, because of the strong interactions within purely ionic solids, solid-state ion conductors have their own limitations, such as slow ion conduction at room temperature. It is therefore advantageous to design new solid-state ion conductors with higher conductivity at intended operating conditions, and the knowledge-based design of these materials requires a detailed, atomic-scale understanding of ion conduction mechanisms. We recently demonstrated the existence of electronic paddlewheels in solid-state ion conduction within AgI 14 . In the electronic paddlewheel mechanism, the translational motion of one ion type (here Ag + ) is coupled to rotations of the electron density on the other ions (here I - ). This is akin to the paddlewheel mechanism proposed for molecular solid-state ion conductors like LiSO 4 , except electronic degrees of freedom are orientationally disordered instead of a molecule 15 , 16 . The electronic-structure based perspective on ion conduction provided by electronic paddlewheels opens new avenues for developing design principles, but several fundamental questions need to be addressed before the electronic paddlewheel perspective can be used for materials design. Here, we address two of these questions: (i) The ab initio simulations required to investigate ion conduction are prohibitively expensive and limited to small system sizes and short timescales, so how can these simulations be scaled up? (ii) How do electronic paddlewheels impact ionic dynamics? To address the first question of scalability, we consider machine learning potentials (MLPs) as efficient alternatives to ab initio simulations, but these approaches typically lack an explicit description of electronic structure 17 , 18 , 19 , 20 , 21 , 22 . However, within the Born-Oppenheimer approximation, the electronic structure and its fluctuations are determined by the instantaneous configuration of the nuclei (though the forces on the nuclei and the resulting trajectory depends on both nuclear and electronic degrees of freedom). Therefore, one may attempt to use MLPs to efficiently access long time scales and even large system sizes, and then use these configurations to examine electronic structure fluctuations. Here, we assess the ability of trajectories generated by a machine learning potential to describe the static and dynamic electronic fluctuations responsible for electronic paddlewheels in \alpha -AgI. We find near perfect agreement between the ab initio and MLP results, because of the ability of the MLP to capture the complex many-body effects arising from electronic paddlewheels. We contrast the MLP results by also studying a classical, empirical force field (FF) model, which, despite exhibiting structural averages similar to the ab initio simulations, cannot capture the effects from electronic paddlewheels. To demonstrate the inability of the FF model to capture electronic paddlewheel effects on ion dynamics, we address the second question regarding how electronic paddlewheels impact ion dynamics. Through an analysis of ionic dynamics within the framework of the generalized Langevin equation, we demonstrate that the classical FF model cannot capture the impact of electronic paddlewheels on ion dynamics. In particular, we show that electronic paddlewheels within the ab initio and MLP simulations lead to higher frequency force fluctuations than found in the classical FF model, and these additional fluctuations significantly alter the time-dependent friction acting on a silver cation. This electronic paddlewheel-induced change in friction enhances cation diffusion. II Methods We compared the structure and dynamics of AgI from density functional theory based molecular dynamics (DFT-MD) simulations with classical molecular dynamics simulations performed using the Neural Equivariant Interatomic Potential 19 (NequIP) MLP and the Vashishta-Rahman empirical pair potential 23 . DFT-MD trajectories were taken from our previous work 14 . We then selected configurations from trajectories at different temperatures to map atom positions and chemical species to the total potential energy and interatomic forces. These configurations correspond to a cubic simulation cell of \alpha -AgI with a length of 15.2565 Å and 108 atoms 24 . We used a total of 3600 configurations (500 at 100 K, 1050 at 300 K, 1050 at 750 K, 500 at 1000 K, and another 500 at 1600 K), out of which 2000 configurations were used for training, 1000 for validation, and 600 for testing the model, all sampled uniformly from the full data set. The locality for the features was imposed with a radial cutoff of 6 Å. We used an initial learning rate of 0.005 and a batch size of 5 for the training. The learning rate gets rescaled by a factor of 0.5 if the validation loss of atomic forces has not improved over 100 consecutive epochs. The training was terminated after around 2100 epochs, once the errors in the forces and energies were sufficiently small; we used energy and force convergence criteria of 8\times 10^{-4} eV/atom 9\times 10^{-3} eV/Å, respectively. Correlation plots between the MLP with DFT energies and forces are shown in Fig. S1 of the SI and validate the accuracy of the MLP model. Moreover, the good agreement between the structural and dynamic properties computed from the MLP and DFT simulations indicates that the MLP is able to capture the relevant physics of the system. We used the Vashishta-Rahman potential for the classical force field-based MD simulations of \alpha -AgI, and we refer the reader to the original work for a listing of the

Read the complete text