Generative Adversarial Network‐Based Microstructure Prediction of Crystallization in Sodium Borosilicate Glass

ABSTRACT

Glass‐ceramics produced through controlled crystallization exhibit properties that are strongly governed by their microstructures. However, quantitative characterization of microstructural evolution is often labor‐intensive and subject to operator‐dependent variability. In this study, a generative adversarial network (GAN)‐based framework was developed to predict the microstructures of sodium borosilicate glass‐ceramics prepared from commercial Pyrex 7740 under different heat‐treatment conditions. The proposed model, designated as a kinetically guided regression‐based conditional Wasserstein GAN with gradient penalty (KG‐RC‐WGAN‐GP), integrates a kinetics‐guided surrogate model with auxiliary regression branches to ensure that the generated microstructures are consistent with both experimental crystallization kinetics and prescribed processing parameters. The model was validated using experimental scanning electron microscopy images and successfully reproduced the crystallization trends and the temperature‐dependent variation in Avrami exponents. Morphological analysis further confirmed that the generated images agreed well with the experimental results in terms of crystal size, aspect ratio, circularity, and spatial distribution. This framework provides an efficient approach for establishing process−microstructure relationships and accelerating the data‐driven design of glass‐ceramic materials.