Machine-learned potential for amorphous Indium-Tin-Oxide alloys

TL;DR

This paper develops a machine-learned potential for amorphous Indium-Tin-Oxide, enabling large-scale molecular dynamics simulations that are much faster than traditional ab initio methods, aiding device design.

Contribution

The authors created a Gaussian approximation potential for ITO, allowing accurate and efficient atomistic simulations of amorphous and crystalline phases, surpassing AIMD limitations.

Findings

MLMD simulations agree with AIMD results

ML potential captures minority atomic interactions

Simulations are 3-4 orders of magnitude faster

Abstract

Machine-learned potential-driven molecular dynamics (MLMD) simulations are of great value in guiding the design and optimization of memory devices. Amorphous indium-tin-oxide (ITO) is widely used as transparent conducting oxide for flat-panel display and solar cell applications, and also as a capping layer in phase-change-materials-based reconfigurable color display devices. However, atomistic simulations of ITO using ab initio molecular dynamics (AIMD) are limited to systems of a few hundred atoms due to expensive computational costs, which prevents the device-scale modelling of real-world applications. In this work, we develop a machine-learned potential for ITO and its parent phase In2O3 based on the Gaussian approximation potential (GAP) framework. We generate a comprehensive training dataset using an iterative training protocol. Our MLMD simulations of crystalline, liquid and melt-quenched amorphous ITO models are in great agreement with the AIMD reference. In particular, the ML potential well captures the minority atomic interaction, such as Sn-Sn bonds, which have poor statistics in small-scale AIMD simulations. We demonstrate that the MLMD simulations are 3-4 orders of magnitude faster than AIMD. The training dataset and the fitted GAP potentials for ITO and In2O3 are openly accessible.