Exploring uncertainty in glass phase transitions through machine learning

Abstract Machine learning methods have shown significant potential and are widely used in modern physics research. However, the uncertainty linked to machine learning, arising from the opacity of its workflow, demands attention and consideration. This study investigates the application of machine learning models in analyzing the glass transition of Cu50Zr50 metallic glass. By employing supervised learning techniques with ResNet50, MobileNetV3, and GoogleNet image extraction models, the study reveals that while machine learning can capture variations in the disordered atomic structure during the transition process, different models may yield divergent results in determining the glass transition temperature. Moreover, variations in atomic sizes within the images can lead to fluctuations in the predicted transition temperatures. These findings highlight the inherent uncertainties associated with using machine learning to analyze continuous phase transitions and stress the importance of ensuring that the extracted structural features align with the physical characteristics of the transition process.