In this study, we explored deploying deep learning to control the steel spheroidization process and metallography determination. Utilizing raw material images, we refined parameters such as temperature for enhanced spheroidization stability, drawing on expert knowledge. Employing an extensive dataset, we tested the steel spheroidization process at 775 and 745 ℃, comparing spheroidization levels with a 760 ℃ baseline. Our focus spanned preprocess metallography and post-spheroidization assessment, using the EfficientNet transfer learning model. Results revealed high accuracy in preprocess determination. For spheroidization rates, categorizing into two groups yielded a correctness rate of more than 95%, showcasing the model’s proficiency in discerning metallographic characteristics.