Overcoming Chemical Space Challenges in High-Entropy Argyrodite Li Solid Electrolytes with Fine-Tuned Neural Network Potentials

Argyrodite-type lithium solid electrolytes have attracted significant attention due to their high ionic conductivity, making them promising candidates for next-generation solid-state batteries. The argyrodite structure, typically represented by the general formula Li₆PS₅X (X = Cl, Br, I), exhibits a unique lattice framework conducive to fast Li-ion transport. This property has sparked interest in exploring its potential for enhancing the performance of solid-state batteries. Recent developments have also expanded the focus toward high-entropy argyrodite systems, where multiple atomic species, particularly from groups 14, 15, and 17 of the periodic table, are doped into the structure. High-entropy systems are increasingly recognized for their ability to stabilize complex solid electrolyte materials and further improve ion conductivity, but their broad chemical variability presents new challenges in materials design.
Developing an accurate and reliable neural network potential (NNP) to describe high-entropy argyrodite systems is particularly challenging due to the vast chemical space involved. Conventional bespoke development would require an enormous amount of DFT AIMD data to effectively train the wide range of elements and configurations. In our research, we address this challenge by leveraging fine-tuning techniques to train our NNP efficiently. By fine-tuning SevenNet-0, a pretrained universal potential, we can minimize the data requirements while ensuring the accuracy of potential in describing various doped argyrodite systems. Our approach is designed to develop an NNP that achieves predictive accuracy for various argyrodite-type systems by fine-tuning based on the DFT data of a specific argyrodite-type system. However, a critical issue with fine-tuning is catastrophic forgetting, where the potential loses its ability to generalize across previously learned systems. To address this, we explore strategies such as parameter regularization methods and replay techniques, which enable us to retain accuracy across vast chemical spaces while adapting to the argyrodite systems.
We will present our latest results demonstrating the effectiveness of this fine-tuning strategy. Through DFT validation, we will show that our approach yields an accurate NNP for modeling the diffusivity and structural stability of high-entropy argyrodite materials. Furthermore, we will discuss the potential engineering applications of this research, particularly in the design of optimized argyrodite-type solid electrolytes. By leveraging fine-tuning to explore the vast chemical space efficiently, we provide a tool for identifying ideal dopant combinations that can maximize ionic conductivity and stability in practical devices.