[Research Briefing] Multimodal ML Analysis Framework for GaSe MBE Growth

Introduction

MBE (Molecular Beam Epitaxy) processes for synthesizing 2D chalcogenides, essential materials for semiconductors and next-generation optoelectronic devices, allow for atomic-level precision. However, the cost of experimental trial-and-error to identify optimal conditions is prohibitively high. In particular, subtle environmental fluctuations during growth directly impact the crystallinity and surface roughness of samples, creating a limitation where traditional ex-situ analysis cannot enable real-time response.

The multimodal ML analysis framework presented in this study demonstrates an approach that transforms real-time diagnostic data, such as RHEED, into quantitative features and integrates them into process prediction models, moving beyond simple monitoring tools. This represents a critical engineering step toward evolving into a data-driven Autonomous Synthesis Platform in the field of high-precision material synthesis.

What the Source Says

Researchers sought to optimize MBE process conditions using data from 37 GaSe thin-film samples grown on GaAs(111)B substrates. Input variables were set as Ga flux (growth rate), Se:Ga flux ratio, and substrate temperature, while output variables were measured as XRD FWHM (crystallinity) and AFM RMS roughness. Methodologically, UMAP was used for unsupervised learning of RHEED patterns to generate low-dimensional embeddings (Z0, Z1), Mutual Information (MI) analysis was performed to analyze relationships between variables, and characteristics were finally predicted using a Random Forest regression model.

As a result, high-quality and low-quality samples were clearly distinguished based on the RHEED embedding Z0=1. For the FWHM prediction model, the coefficient of determination (R²) improved to 0.55 when low-quality samples were removed via RHEED embeddings compared to using the entire sample set. Furthermore, MI analysis quantified that the growth rate had the greatest impact on FWHM, while the Se:Ga ratio had the greatest impact on RMS roughness. The RMS roughness prediction model showed a predictive power of R²=0.56 based on LOO cross-validation.

Root Failure Mechanisms

First, there is an overfitting risk due to the limited scale of the dataset. Because this study used a small set of 37 samples, Random Forest and LOO CV were selected over deep learning; however, there remains a risk that the influence of specific variables may be overestimated given the low data volume.

Second, there is a potential mismatch in correlation between in-situ data and ex-situ metrics. As seen where RHEED embeddings showed a strong correlation with FWHM (crystallinity) but a weak relationship with RMS roughness, incorrect process correction decisions may be made if a specific real-time metric fails to represent all quality characteristics.

Third, model distortion caused by stochastic anomalies. When low-quality samples generated by random factors, such as Se flake interference or temperature fluctuations, are included in the training data, the model exhibits vulnerability where the R² value drops sharply and predictive reliability collapses.

Source: Photo by Steve A Johnson on Unsplash

MBE Process ML Modeling and Data Validation Checklist

To preemptively prevent such system collapses in practical engineering pipelines, the following core defense mechanisms must be enforced during the operational stage:

• Real-time Embedding Construction: Has image data, such as RHEED patterns, been converted into low-dimensional vectors with physical meaning via UMAP or similar methods?
• Mutual Information (MI) Analysis: Have key features been selected by quantitatively analyzing the non-linear correlations between input process variables and output quality metrics?
• Outlier Filtering Strategy: Has a criterion (e.g., Z0 > 1) been established to exclude low-quality/anomalous samples before model training using in-situ diagnostic data?
• LOO Cross-Validation Application: Given the nature of small datasets, has the generalization performance of the model been strictly evaluated using Leave-One-Out CV?

My Perspective & Field Notes

In actual field operations, there is often a reliance on qualitative analysis where experts visually inspect RHEED patterns to determine if ‘streaks have appeared well.’ This paper takes a highly practical approach by converting this into quantitative coordinates through UMAP embeddings, effectively replacing expert intuition with mathematical boundaries.

However, a Random Forest performance of R²=0.55 is still insufficient for full automation. Ultimately, rather than ML solving everything, its value is greater as an ‘accelerator’ that drastically reduces the number of researcher experiments by rapidly filtering out outliers.

Executive Governance Guide

For internal AI projects to transition from PoC-stage early demos to actual business value, the following governance principles must be established from the planning phase:

1. Data Integrity First Principle: Define the data refinement stage via physical diagnostic tools (in-situ) as a mandatory process prior to supervised learning.
2. Explainable Feature Selection: Verify that the basis for model predictions aligns with physical growth mechanisms through MI analysis and Feature Importance analysis.
3. Conservative Performance Evaluation: For models based on small experimental datasets, determine deployment based on strict validation metrics such as LOO CV rather than simple training accuracy.

References & Source