Predicting and Accelerating Nanomaterial Synthesis Using Machine Learning Featurization.

Area of Science:

• Materials Science
• Machine Learning
• Data Science

Background:

• Traditional materials synthesis relies on time-consuming, manual feedback loops and siloed characterization methods.
• Optimization is often limited by human intuition and the need for extensive experimental data.

Purpose of the Study:

• To automate and generalize feature extraction from reflection high-energy electron diffraction (RHEED) data using machine learning.
• To establish quantitatively predictive relationships from small, expert-labeled datasets for accelerated materials synthesis.
• To demonstrate the application in predicting film properties and estimating dopant concentrations.

Main Methods:

• Developed a machine learning model for automated feature extraction from RHEED data.
• Utilized a materials-agnostic approach, enabling application across different systems without retraining.
• Evaluated the model on W1-xVxSe2 thin film growth on sapphire substrates.

Main Results:

• Successfully predicted grain alignment using pregrowth substrate data.
• Accurately estimated vanadium dopant concentration using in situ RHEED as a proxy for ex situ techniques.
• Achieved potential time savings of up to 80% in a simulated 100-sample synthesis campaign.

Conclusions:

• Automated RHEED analysis with machine learning significantly reduces synthesis time and characterization needs.
• The materials-agnostic approach offers a generalizable solution for optimizing various material synthesis processes.
• This predictive capability guides experimental design, minimizes failed trials, and enhances control over material properties.