Universal Deoxidation of Semiconductor Substrates Assisted by Machine-Learning and Real-Time-Feedback-Control

Chao Shen, Wenkang Zhan, Jian Tang, Zhaofeng Wu, Bo Xu, Chao Zhao, Zhanguo WangarXiv 2023 · DOI · arXiv

Introduction

In semiconductor fabrication, achieving atomic-level cleanliness is a prerequisite for high-quality device performance. Before the growth of epitaxial layers via Molecular Beam Epitaxy (MBE), the native oxide layer that forms on substrates like Gallium Arsenide (GaAs) must be removed. Traditionally, this "deoxidation" process has been a manual, subjective task. Experienced operators monitor Reflection High-Energy Electron Diffraction (RHEED) patterns—visual cues on a fluorescent screen—to determine when the oxide has evaporated. This reliance on human judgment is slow, difficult to standardize, and prone to errors that can damage the substrate or introduce defects.

Methodology

To address these challenges, researchers developed an automated, "universal" system that integrates real-time computer vision with machine learning to control the deoxidation process.

Schematic of the AI-driven automated deoxidation system using RHEED and real-time feedback control.

Hybrid CNN-ViT Architecture

The core of the system is a hybrid model combining Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs). The CNN component extracts local features, such as the fine details of diffraction spots, while the ViT captures global relationships across the entire RHEED image. To ensure the system is "hardware agnostic," a novel upsampling layer was implemented, allowing the model to process RHEED videos of varying resolutions from different camera setups without retraining.

Real-Time Feedback Control

The system operates in a closed loop, processing 12-frame sequences from the RHEED video at a rate of 9 results per second. If the model identifies the substrate as oxidized, it triggers a temperature increase. To ensure robustness against noise or uneven oxide thickness, the system requires a 95% confidence score of "deoxidized" over 24 consecutive judgments before concluding the process and maintaining the target temperature.

Experimental Setup and Datasets

The model was trained on a massive dataset of 320,000 NPY files derived from RHEED videos of GaAs, Germanium (Ge), and Indium Arsenide (InAs) substrates. To test the "universality" of the approach, the researchers evaluated the model on Gallium Antimonide (GaSb)—a material entirely absent from the training set—using a different MBE machine (Riber C21 vs. Riber 32P) with varying substrate rotation speeds.

Key Results and Discussion

The AI-driven system demonstrated exceptional performance and generalization capabilities:

  • High Precision: The model achieved a validation accuracy of 99.95%.
  • Cross-Platform Generalization: When tested on unseen GaSb data from a different machine, the model maintained 96.3% accuracy for deoxidized states and 88.9% for oxidized states.
  • Successful Automation: In-situ experiments successfully automated the deoxidation of multiple materials, identifying precise deoxidation temperatures (e.g., 405 °C for GaAs) without human intervention.
  • Explainable AI: Attention heatmaps confirmed that the AI focuses on the "specular spot" of the RHEED pattern, aligning with the visual cues used by human experts and validating the model's physical relevance.

Conclusion

This work represents a significant step toward the fully automated, "plug-and-play" manufacturing of semiconductors. By replacing subjective human observation with a robust, hardware-agnostic AI controller, the system standardizes a critical fabrication step. Future research will focus on extending this universal modeling approach to other stages of thin-film growth, such as real-time monitoring of growth rates and surface reconstructions.

Summary updated

BibTeX
@misc{shen2023universal,
  author = {Chao Shen and Wenkang Zhan and Jian Tang and Zhaofeng Wu and Bo Xu and Chao Zhao and Zhanguo Wang},
  title = {{Universal Deoxidation of Semiconductor Substrates Assisted by Machine-Learning and Real-Time-Feedback-Control}},
  year = {2023},
  eprint = {2312.01662},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2312.01662},
}
Figure