<p>Arsenic, as a harmful impurity in copper matte during the matte-smelting process, directly affects the copper grade and final quality of electrolytic copper. However, traditional production generally suffers from delayed analysis and limited control measures, making it difficult to achieve real-time prediction and optimal control of the copper matte grade and arsenic content in copper matte. To address this challenge, a Bayesian optimization-eXtreme Gradient Boosting prediction model integrated with a Bayesian optimization algorithm and interpretability mechanisms is proposed in this paper, aiming at accurate modeling and efficient prediction of key quality indicators of copper matte. The results demonstrate that the model achieves <i>R</i>² values of 0.910 and 0.943 for the prediction of copper matte grade and arsenic content, respectively. Simultaneously, a mean absolute error of 0.0098 and 0.294, root-mean-square error of 0.0114 and 0.3859, and a mean absolute percentage error of 0.435% and 2.86%, respectively, are achieved. All performance metrics are significantly better than those of other comparative models. Furthermore, the Shapley additive explanations interpretability framework is introduced to quantify the importance of each feature variable, and in combination with analysis of metallurgical mechanisms, the effects of key process parameters such as temperature, Fe/SiO₂, and CaO on the copper matte grade and arsenic content are thoroughly illustrated, thereby offering a guideline for an industrially applicable model for optimizing the mixing of raw materials and process parameter control in practical matte-smelting production.</p> Graphical Abstract <p></p>