Predicting Crystal Structures and Ionic Conductivities in Li$_{3}$YCl$_{6-x}$Br$_{x}$ Halide Solid Electrolytes Using a Fine-Tuned Machine Learning Interatomic Potential

TL;DR

This study develops a data-efficient workflow to model Li$^+$ transport in halide solid electrolytes $Li_{3}YCl_{6-x}Br_{x}$ by fine-tuning the Crystal Hamiltonian Graph Network (CHGNet) on halide-specific data. It combines an enumeration-ranking approach to identify physically meaningful ground-state structures with an iterative, temperature-aware fine-tuning loop to achieve near-DFT accuracy at orders-of-magnitude lower cost, enabling nanosecond-scale molecular dynamics. The authors quantify how composition and halide substitution modulate phase stability and diffusion pathways, revealing anisotropic diffusion in LYCB and isotropic diffusion in LYB, and show that pressure and composition tune ionic conductivity. The framework provides a transferable, scalable route for predicting structure-property-transport relationships in complex, partially disordered halide electrolytes, accelerating design of next-generation solid-state batteries.

Abstract

Understanding ionic transport in halide solid electrolytes is essential for advancing next-generation solid-state batteries. This work demonstrates the effectiveness of fine-tuning the Crystal Hamiltonian Graph Network (CHGNet) universal machine learning interatomic potential to accurately predict total energies, relaxed geometries, and lithium-ion dynamics in the ternary halide family Li$_{3}$YCl$_{6-x}$Br$_{x}$ (LYCB). Starting from experimentally refined disordered structures of Li$_{3}$YCl$_{6}$ and Li$_{3}$YBr$_{6}$, we present a strategy for generating ordered structural models through systematic enumeration and energy ranking, providing realistic structural models. These serve as initial configurations for an iterative fine-tuning workflow that integrates molecular dynamics simulations and static density functional theory calculations to achieve near-ab initio accuracy at four orders of magnitude lower computational cost. We further reveal the influence of composition (varied x) on the predicted phase stability and ionic conductivity in LYCB, demonstrating the robustness of our approach for modeling transport properties in complex solid electrolytes.

Figures (10)

• Figure 1: Global workflow combining the enumeration-ranking strategy (black boxes) with the fine-tuning framework (blue boxes). Starting from experimentally resolved structures LYB_LYC_exp$-$ which exhibit partial occupancies $-$ symmetry-distinct ordered configurations are enumerated and ranked using the . Additional DFT calculations are performed on the lowest-energy configurations to identify the ground-state ordered configuration $-$ subsequently used as a starting configuration in the fine-tuning procedure. The fine-tuning procedure involves a series of -driven simulations (with LAMMPS) at finite temperatures ($T_{\mathrm{i}}$). For each run (1,000,000 steps), DFT calculations (with VASP) are performed on a subset of structures to obtain additional training data and fine-tune . Computational details are provided in the Computational Methods.
• Figure 2: (a) Total energies predicted with (colored circles), compared to DFT reference data (black crosses), for a series of finite-temperature structures generated during the fine-tuning procedure in Figure \ref{['fig:1 (workflow)']}. (b) Benchmark of the pretrained and fine-tuned potentials for total energy prediction of , , and , compared to DFT data. (c) MAEs on energy (in 0.1 eV/atom), forces (in eV/Å), and stress (in GPa). (d) and (e) Average volume/atom resulting from the $NpT$ equilibration runs, using the pretrained (circles) and fine-tuned _600K potentials (empty orange triangles: equilibrated with p=1 bar; solid orange triangles: p=10 kbar), compared to (purple diamonds) park_sevennet_mf_ompa_2025, and experimental data (black and gray markers) from Refs. LYB_LYC_expliu_tuning_2024asano_solid_2018. Note the blue star in (e) indicates an unstable simulation with unphysical cell volume, which is scaled by a factor of 10 in the plot.
• Figure 3: For (a) and (b) , we report the activation energy ($E_A$) and room-temperature ionic conductivity ($\sigma_{300\text{K}}$), obtained from simulations in the $NVT$ ensemble, using , _600 K, and . The simulations were conducted in $NpT$-equilibrated cells under p=1 bar (empty orange triangles) and p=10 kbar (filled orange triangles). Error bars originate from the variances in the diffusion coefficients that are estimated according to Ref. he_statistical_2018. We compare our predictions to AIMD wang_lithium_2019 and experimental values: exp, exp*, exp** (*different synthesis) LYB_LYC_exp; exp_2 liu_tuning_2024; exp_3, exp_3* (*different synthesis) asano_solid_2018; exp_4 van_der_maas_investigation_2023.
• Figure 4: Li$^+$ transport properties of (a)-(e) and (f)-(j) , obtained from simulations in the $NVT$ ensemble, starting from $NpT$-equilibrated structures (p=10 kbar), and using the _600K potential. (a) and (f) MSD as a function of $\Delta t$ at T=500 K; (b) and (g) temperature-dependence of the diffusion coefficient, in logarithmic scale; and (c)-(e) and (h)-(j) Li$^+$ probability density maps from simulations at T=500 K projected on the $ab$-, $ac$-, and $bc$-plane, respectively (Li: green, Y: gray, Cl: blue, Br: brown). Error bars are derived following the procedure in Ref. he_statistical_2018.
• Figure 5: Li$^+$ transport properties for obtained from _600K simulations (starting from $NpT$-equilibrated structures under p=10 kbar), including (a) activation energy $E_A$ and (b) room-temperature ionic conductivity $\sigma_{300\text{K}}$. We consider $C2$-symmetry structures of with $\text{x}=3$ to 6. Experimental values (exp_4 van_der_maas_investigation_2023) are shown with black crosses.