Ab Initio, Classical and Machine-Learning MD Simulations for Modeling Lithium Borosilicate Glasses

Lithium borosilicate (LBS) glass is a typical oxide glass applicable for all solid electrolyte to reduce interfacial resistance between the electrolyte and electrode. Because understanding in the microstructure is essential to improve the ion conductivity, accurate modeling in nanoscale is demanded. However, it is often difficult to reproduce boron coordination change from three-fold (³B) to four-fold (⁴B) coordinate using a conventional functional type of force-field, which hinders designing high ion-conducting glasses using molecular simulations.<br/><br/>This study, therefore, tried to apply machine-learning potential (MLP) for modeling a series of LBS glasses using DeepMD. The MLP was trained to reproduce a number of data on force and energy calculated by density functional theory (DFT) for crystals and amorphous glass models. The MLP successfully reproduced ³B/⁴B ratio in the LBS glasses with a few exceptions. Specifically, three-membered rings including boroxol ring, which is composed of three ³B, were generated in high boron-containing LBS glasses in contrast to classical MD simulations with the functional force-field. Ion conductivity and glass transition temperature were also qualitatively reproduced by the MLP, which proved the applicability of MLP to investigate oxide glasses composed of network formers, which flexibly change their coordination.<br/><br/>In this talk, <i>ab initio</i> and classical MD simulations will be compared with the MLP-MD. For the classical MD simulations, a Buckingham type of force-field was developed to model LBS glasses by introducing a composition-dependent correction. In comparison with the CMD, the MLP was found to construct energetically more stable glass structures including the three-membered rings more in the microstructures. These results highlighted the advantages of the MLP. In contrast, five-hold coordinated silicon atoms were unreasonably generated, and the glass structures were disrupted at high temperature. This talk will also share these issues when applying MLP for modeling amorphous materials.