A Deep-Learning Potential for Crystalline and Amorphous Li-Si Alloys

This work investigates the ability of the deep-learning potential (DP) to describe structural, dynamic and energetic properties of crystalline and amorphous Li-Si alloys. Li-Si systems play an important role in the development of high energy lithium ion batteries. One challenge in simulating Li-Si systems is to balance the proper description of complex Li-Si interactions and the system size. Machine-learning potential paves an avenue to achieve this balance by describing complex interactions using machine-learning models. The machine-learning models enable us to investigate complex systems beyond the capability of the classical force fields using molecular dynamics simulations. We develop a DP for Li-Si systems with Li/Si ranging from 0 to 4.2 based on a vast dataset generated using the quantum mechanical calculations in an active learning procedure. Then we investigate the structural and dynamic properties of several crystalline and amorphous Li-Si systems using this developed DP. The DP can predict bulk densities, the radial distribution functions and diffusivity of Li in amorphous Li-Si systems at a level similar to the quantum mechanical calculations, while the molecular dynamics simulations are 20 times faster than the quantum mechanical calculations. We also discuss several issues when developing DP.