Engineering Super-Ionic Diffusion in Solid-State Electrolytes via Machine Learning

Leading research teams are actively pursuing solid-state electrolytes (SSEs) for next-generation lithium (Li) batteries. A primary requirement a SSE must have is a high room-temperature ionic conductivity. Since super-ionic SSEs exhibit highly connected pathways that flatten the free-energy landscape, their ionic conductivity must be estimated from finite-temperature molecular dynamics (MD) rather than single-barrier kinetics. However, ab initio molecular dynamics (AIMD) simulations are computationally expensive. This limits accessible simulation times and the number of atoms for conventional first-principles methods. To mitigate these limits, a pretrained state-of-the-art machine learning interatomic potential called MACE-MPA-0 was used. This work tests the pretrained MACE model on vacancy-mediated Li migration barriers in rock salt materials via the nudged elastic band method. The pretrained MACE model underestimated the energy barriers by around 10% when compared to our reference first principles calculations. To improve the model's accuracy, additional first-principles data were incorporated through targeted finetuning. By allowing a Li atom to roll down an energy saddle point, the finetuning process substantially improved the predicted barrier accuracy to within 0.5%. Moreover, the diffusion coefficients in the sulfide SSE Li10Si(PS6)2 were evaluated across a temperatures range of 300 K to 1000 K by conducting MACE MD simulations. Then an exploration of two methodologies were performed to estimate the uncertainty associated with the calculated diffusion coefficients. Furthermore, at most temperatures the pretrained MACE-MPA-0 model overestimates the diffusion coefficients when comparing them to first principles and experimental results. The model’s prediction was especially worse at 300 K. Therefore, two finetuning algorithms were created to further finetune the pretrained MACE-MPA-0 model. One algorithm uses data gathered by AIMD simulations and the other uses data obtained by single point first-principles calculations. These workflows generated data and finetuned the MACE-MPA-0 model to better predict the diffusion coefficients especially at 300 K. Using each of these finetuning workflows, the diffusion coefficient of Li10Si(PS6)2 was calculated at 300 K by running MACE MD simulations and Arrhenius extrapolation methods. The diffusion coefficients obtained from both finetuning workflows are in close agreement and within an order magnitude of the experimental reference value of 5.8 ± 2 x 10-12 m^2/s. However, only the diffusion coefficients obtained directly from MACE MD simulations at 300 K were within experimental uncertainty. Lastly, the author, with the contributions of the research group, introduce an automated fine-tuning strategy based on active learning methods. This approach iteratively queries high-uncertainty configurations for labeling. With this active learning model, both methods of calculating the 300 K diffusion coefficient were within the experimental uncertainty. The closest predicted diffusion coefficient from the active learnt model is 6.41 x 10-12m^2/s at 300 K, obtained by Arrhenius extrapolation. In conclusion, this study provides the basis of concrete workflows to finetune MACE models to accurately predict diffusion coefficients of SSE materials.