• An unsupervised contrastive learning framework with temporal similarity constraints enables real-time RHEED analysis during substrate deoxidation, eliminating manual labeling.
• The pre-trained model achieves 88.1% clustering accuracy on GaAs deoxidation samples without additional labels, and 94.3–95.5% accuracy after fine-tuning with only five sample pairs across GaAs, Ge, and InAs.
• The framework is optimized for edge devices, allowing plug-and-play integration with existing MBE systems and rapid adaptation to different materials and equipment.
• This approach enhances automation and reproducibility in semiconductor manufacturing by providing smooth, interpretable feature trajectories for deoxidation state transitions.