Analysis of Monolayer to Bilayer Silicene Transformation in CaSi₂F_x (x<1) using Universal Neural Network Potential

Silicene is a silicone (Si) counterpart of graphene, a typical two-dimensional (2D) material which consists of a honeycomb structure of carbon. Silicene is predicted to have a Dirac cone-type electronic structure similar to that of graphene, and has been extensively studied as a promising material for ultra-energy-saving, high-mobility devices. In addition, silicene, graphene, and other 2D materials in the form of nanoribbons and nanodots have been investigated for applications in batteries and biological imaging. For practical applications, further development of fabrication techniques for silicene-derived structures is necessary because the monolayer structure is unstable under air due to the lack of vertical sp³ bonds.<br/>Synthesis of 2D materials by chemical exfoliation of three-dimensional (3D) crystals with layered structures is a convenient top-down method. Calcium disilicide (CaSi₂) consists of alternating monolayers of silicene (MLSi) and calcium, where wet-chemical exfoliation can be applied to produce MLSi derivatives. A solid state reaction due to fluorine (F) diffusion into CaSi₂ leads to a phase transformation from MLSi to bi- or tri-layered silicene (BLSi and TLSi) sandwiched between non-conducting CaF₂ layers. This structure prevents exposure to the atmosphere and improves the stability of the materials. The thermodynamic mechanism of the phase transformation was investigated in the previous study using density functional theory (DFT) calculations and cluster expansion for MLSi and BLSi of CaSi₂F_x(0&lt;x&lt;1). However, the analysis of the reaction dynamics, which is important for nano-structure fabrications and battery development, was not feasible due to the high computational cost.<br/>Machine learning potentials have been increasingly applied to various atomistic simulations. The Preferred Potential (PFP) implemented on Matlantis^TM is a recently developed graph neural network potential with its unique feature of universality. PFP is trained on large DFT datasets, including not only stable crystals and molecules, but also surfaces and disordered structures. As a result, it is applicable to reaction analysis and molecular dynamics of various materials without compromising accuracy.<br/>In this study, we have applied PFP to analyze the mechanism of the phase transformation of CaSi₂ due to fluorine diffusion. Firstly, the phase stability of CaSi₂F_x in ML and BL structures is studied as a function of the fluorine ratio x. A few hundred of CaSiF_x structures were generated using the<i> icet</i> package and the convex hull of the formation energy is evaluated using PFP. The results show that PFP reproduces the previous study using DFT: ML structure becomes unstable with increasing fluorine ratio, and BL structure becomes more stable for x &gt; 0.6. This phase transformation is due to the charge imbalance of the ionic CaSi₂. With increasing electronegative fluorine atoms, strong ionic bonds are formed between F^- and Ca²⁺, while the formal charge of Si^- becomes neutral. This result indicates that PFP is applicable to our system with DFT accuracy, although the stability depends on both ionic and covalent sp³ electronic nature. Following this result, we have performed molecular dynamics simulations of stable structures on the convex hull. During the MD simulations at 800 K for more than 1 ns, interlayer diffusion of Si atoms are observed at the defects in the layered structure associated with the formation of CaF₂. These results reveal the dynamics of phase transformation of CaSi₂F_x, and demonstrate the accuracy and universality of PFP for materials with variable compositions, such as battery materials, alloys and solid solutions.