Multimodal Machine Learning Analysis of GaSe Molecular Beam Epitaxy Growth Conditions

Area of Science:

• Materials Science
• Thin-film deposition
• Machine Learning

Background:

• Autonomous synthesis platforms can revolutionize thin-film growth through real-time optimization.
• Molecular beam epitaxy (MBE) applications using these platforms are underdeveloped.
• Reflection high-energy electron diffraction (RHEED) is a key in situ diagnostic for thin-film growth.

Purpose of the Study:

• To develop a machine learning-guided framework for MBE growth of Gallium Selenide (GaSe) films.
• To leverage RHEED patterns for real-time process optimization and quality assessment.
• To establish a data-driven approach for accelerated thin-film synthesis.

Main Methods:

• Utilized unsupervised learning on RHEED patterns to identify quality boundaries.
• Employed mutual information analysis to correlate RHEED features with ex situ characterization (XRD, AFM).
• Developed supervised learning models to predict film quality metrics (fwhm, RMS roughness), incorporating RHEED embeddings.

Main Results:

• Unsupervised learning on RHEED data revealed distinct high- and low-quality sample features.
• RHEED embeddings strongly correlated with rocking curve full-width at half-maximum (fwhm), weakly with AFM roughness.
• Growth rate primarily impacted fwhm; Se/Ga flux ratio influenced RMS roughness and RHEED embeddings.
• Supervised models improved prediction accuracy with RHEED embeddings; anomaly detection aligned with unsupervised classification.

Conclusions:

• A data-driven framework for machine learning-assisted MBE was established.
• The framework enables real-time process control and accelerated optimization of thin-film synthesis.
• Integrating in situ RHEED diagnostics with machine learning significantly enhances MBE process understanding and control.