RHEED pattern classification by a convolutional neural network for the growth of chalcogenide thin films and nanostructures

TL;DR

This paper develops a CNN model for real-time classification of RHEED patterns in chalcogenide thin film growth, achieving high accuracy and enabling automation of growth process control.

Contribution

It introduces a material-agnostic CNN approach for classifying RHEED patterns and growth modes, facilitating real-time monitoring and control in thin film deposition.

Findings

Achieved 94.9% accuracy in RHEED pattern classification.

Successfully distinguished three growth modes: Volmer Weber, Stransky-Krastanov, Frank-van der Merwe.

CNN is material-agnostic and can be integrated into growth automation.

Abstract

The use of reflection high energy electron diffraction (RHEED) plays a critical role for in-situ characterization in molecular beam epitaxy, pulsed laser deposition and sputtering. While sensitive to crystal symmetries and morphology, it is used ubiquitously to determine the growth modes of thin films. However, analysis of RHEED patterns depends on skilled experts and is therefore difficult to incorporate into the growth strategy in real-time. The development of machine learning (ML) processes, specifically convolutional neural networks (CNNs), presents a unique opportunity towards real-time RHEED pattern recognition. In this study, we develop a CNN model that can accurately classify four common and distinct RHEED patterns present in chalcogenide thin film growth. Its reached accuracy reached 94.9% for single run and 91.2% when averaged over 20 seeds. Our network is able to distinguish the nucleation of three common growth modes encountered in epitaxy, namely Volmer Weber, Stransky-Krastanov and Frank-van der Merwe, potentially enabling future automation of substrate temperature and shutter control informed by RHEED data. The network is material-agnostic and distinguishes the VW process with greater than 98% accuracy but is somewhat more limited in its ability to properly classify roughening and the initiation of Stransky-Krastanov growth. Our findings show that ML techniques can be successfully implemented even in cases where there is no detailed knowledge of growth chemistry providing an avenue towards real-time incorporation of ML to control nanostructure nucleation and thin film morphology.