Machine-Learning-Assisted and Real-Time-Feedback-Controlled Growth of InAs/GaAs Quantum Dots

Chao Shen, Wenkang Zhan, Kaiyao Xin, Manyang Li, Zhenyu Sun, Jian Tang, Zhaofeng Wu, Bo Xu, Zhongming Wei, Chao Zhao, Zhanguo WangarXiv 2023 · DOI · arXiv

Abstract

This study introduces a significant advancement in semiconductor manufacturing by integrating deep learning directly into the growth of nanostructures. Researchers from the Chinese Academy of Sciences have developed a real-time, closed-loop feedback system that utilizes 3D computer vision to automate the growth of InAs/GaAs quantum dots (QDs). By analyzing Reflection High-Energy Electron Diffraction (RHEED) videos in situ, the system achieves precise control over dot density—a task that has traditionally relied on expert intuition and exhaustive trial-and-error. The system demonstrates the ability to tune QD densities across three orders of magnitude with high accuracy, paving the way for autonomous semiconductor synthesis.

Introduction

Growing high-quality quantum dots via Molecular Beam Epitaxy (MBE) is a multidimensional optimization challenge. The density and quality of these dots are critical for specific optoelectronic applications: high-efficiency lasers require high dot density, while single-photon sources for quantum communication require extremely low density.

Traditionally, researchers find the optimal growth parameters—such as substrate temperature, flux ratios, and growth rates—through time-consuming trial-and-error. While RHEED is commonly used to monitor the surface in real-time, the resulting diffraction data is noisy and complex, often requiring human experts to interpret subtle transitions. Previous machine learning attempts in this field often relied on static images or "posteriori" analysis, which lacks the capability for mid-growth corrections. This research addresses the need for an intelligent, in-situ connection between real-time characterization and active growth control.

Methodology

The core innovation of this work is the application of a 3D ResNet-50 architecture to RHEED data. Unlike standard 2D Convolutional Neural Networks (CNNs) that analyze single images, 3D convolutions capture spatiotemporal features by processing sequences of frames.

  • Input Processing: The model takes sequences of 8 consecutive RHEED frames (captured at 8 fps) as a single input. This allows the model to "see" the temporal evolution of the surface, specifically the transition from a streaky pattern (indicating a flat surface) to a spotty pattern (indicating dot formation).
  • Dual-Model System: The researchers implemented two distinct models: a "QDs formation model" to detect the exact onset of dot formation (binary classification) and a "density classification model" to categorize the surface into four density regimes: zero, low, middle, and high.
  • Closed-Loop Control: These models were integrated into a LabVIEW environment. If the real-time machine learning output deviates from the user-defined target density, the system automatically adjusts the substrate temperature. Once the target density is reached, the system automatically closes the material shutters to terminate growth.

Experimental Setup and Dataset

The researchers constructed a custom dataset from their MBE operations to train and validate the system. They grew 30 unique QD samples with densities ranging from 0 to $1.69 \times 10^{11} \text{ cm}^{-2}$. Ground truth for training was established using post-growth Atomic Force Microscopy (AFM) images to count the dots and assign labels.

Figure 1: (a) Schematic of the InAs/GaAs QD sample structure. (b) Table of the 30 samples used for training, categorized by density. (c-f) AFM images and (g-j) corresponding RHEED patterns for the four density labels: zero, low, middle, and high.

As shown in Figure 1, the dataset covers a wide range of morphological outcomes. The 3D ResNet model was trained on approximately 360,000 augmented data arrays derived from 120 RHEED videos. The use of 3D convolutions proved essential for accounting for substrate rotation and the transient nature of surface phase changes during growth.

Key Results and Discussion

The integrated system demonstrated remarkable precision and reliability:

  • Density Tuning: The system successfully tuned QD densities across three orders of magnitude, from a low of $3.8 \times 10^8 \text{ cm}^{-2}$ to a high of $1.4 \times 10^{11} \text{ cm}^{-2}$.
  • High Accuracy: The formation detection model achieved a validation accuracy of 94.4%, while the density classification model reached 95.5%.
  • Predictive Intervention: A key insight was the model's ability to predict final post-growth density based on RHEED patterns before the dots had fully formed. This predictive capability allows the system to adjust growth parameters early enough to steer the process toward the desired outcome.
  • Hardware Robustness: The model proved effective across different MBE systems with varying mechanical characteristics (such as substrate "wobble") without requiring retraining.

Conclusion

This work represents a significant step toward the "intelligent manufacturing" of semiconductor nanostructures. By replacing human intuition with a real-time, deep-learning-based feedback loop, the researchers have significantly improved the reproducibility of QD growth and reduced the optimization cycle. Future work will focus on reducing communication latency between the ML server and MBE controllers and moving toward finer-grained numerical density predictions rather than broad categorical labels. This methodology holds potential for expansion into other complex thin-film growth processes, including real-time defect detection and repair.

Summary updated

BibTeX
@misc{shen2023machine,
  author = {Chao Shen and Wenkang Zhan and Kaiyao Xin and Manyang Li and Zhenyu Sun and Jian Tang and Zhaofeng Wu and Bo Xu and Zhongming Wei and Chao Zhao and Zhanguo Wang},
  title = {{Machine-Learning-Assisted and Real-Time-Feedback-Controlled Growth of InAs/GaAs Quantum Dots}},
  year = {2023},
  eprint = {2306.12898},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2306.12898},
}
Figure