Machine Learning Potential Powered Insights into the Mechanical Stability of Amorphous Li-Si Alloys

Abstract: Understanding the mechanical properties of solid-state materials at the atomic scale is crucial for developing novel materials. For example, amorphous LiSi alloys are attractive anode materials for solid-state Li-ion batteries but face mechanical instabilities due to significant volume variations with changing Li content. A fundamental grasp of the mechanical behavior in such systems is essential to address their poor mechanical integrity. Experimental methods offer insufficient information to elaborate on dynamic mechanical degradation mechanisms at the atomic scale, and computationally demanding first-principles methods, like DFT, struggle to access the system sizes needed for modeling mechanical phenomena. Machine learning potentials (MLPs) can overcome the computational constraints of traditional DFT-based simulations, enabling large-scale, accurate, and efficient simulations. Here, we provide a concise tutorial on developing and applying MLPs to investigate mechanical properties in materials systems, ranging from bulk to nanoparticles, using Li-Si alloys as an example. Trained on a comprehensive dataset (~45,000 DFT structures) with the aenet package accelerated by PyTorch, a robust MLP is constructed to reproduce results consistent with previous experimental observations. We demonstrate applying the MLP to realistic structures to visualize the deformation mechanism and determine the origin of mechanical instabilities caused by fracturing. This work aims to establish MLP-based simulations as a tool to understand the atomic-scale mechanical behavior in different materials systems.