Abstract: The endpoint timing of copper-converter blowing directly affects the quality of blister copper, furnace stability, and blowing efficiency. Therefore, enhancing the digitalization and intelligence levels of this process has significant practical importance. This study employed a deep learning algorithm that integrated a convolutional neural network (CNN) and graph attention network (GAT). It utilized CNNs to extract image features from the cooling samples of high-temperature melts. Subsequently, by fusing these image features with various production condition data and constraints through the GAT, a model was constructed to determine the best endpoint and predict the product composition. This model could predict the main elemental content of furnace products and estimate the required blowing time. A dataset comprising 5172 production parameters and images of high-temperature cooling samples from a furnace was established. The model was trained and validated using this dataset, and the results indicated that the model achieved endpoint judgment accuracies of 96.73% and 97.85% for the slag-making and copper-making periods, respectively, on the test set. The average prediction error for the composition across four cycles of copper-converter blowing was as low as 0.705wt%, and the average error in estimating the required blowing time was only 1.94 min. The results of this study provide new methods and insights for the development of intelligent endpoint judgment technologies for copper-converter blowing.