高温熔体与固体陶瓷之间的接触角是表征界面润湿行为的重要参数，对优化冶金过程控制和评估耐火材料性能具有重要意义。为了实现高温熔体接触角的准确预测并提升模型的准确性及可解释性，本文搜集并整合了20种金属与16种氧化物的1427组数据，引入13项关键特征表征材料属性与实验条件。通过扩展的孤立森林（EIF）算法对原始数据进行深度清洗，在剔除异常值的同时保留了93%的高质量样本。随后，对比评估了9种机器学习模型，结果表明集成学习模型在预测精度上整体优于单一模型，其中CatBoost模型表现最优，其决定系数R2达到0.9024。多随机种子测试进一步证实了该模型在不同数据分布下均具有良好的稳定性与泛化能力。最后基于SHAP值的可解释分析显示氧化物生成能与界面能是影响接触角的主要因素，表明模型预测逻辑与物理化学原理的高度一致性。本研究为高温界面润湿性的快速评估提供了新方法，为冶金过程中的界面反应控制及钢液洁净化提供有益参考。

Contact angle between high-temperature metallic melts and solid ceramics is a critical physical parameter for characterizing interfacial wetting behavior. This parameter is intrinsically linked to the erosion resistance and service life of refractory linings, as well as the collision, agglomeration, and removal of non-metallic inclusions in liquid steel. However, traditional measurements, such as the sessile drop method, are limited by high costs, lengthy experimental cycles, and extreme operational demands. Furthermore, existing predictive models often suffer from limited dataset sizes and poor generalization across diverse material systems. To address these limitations, this study established a comprehensive, high-quality database by integrating extensive experimental data from peer-reviewed literature, encompassing 1427 samples across 20 metallic elements and 16 oxide ceramic systems. Thirteen key input features were extracted, representing the intrinsic physicochemical properties of metals and oxides, experimental conditions, and derived composite parameters. During the data preprocessing step, the Extended Isolation Forest (EIF) algorithm was implemented for deep cleaning and anomaly detection. After anomaly removal, 93% of the samples were retained, forming a robust dataset for model training. On this basis, the study systematically evaluated the predictive performance of several advanced ensemble learning algorithms including Categorical Boosting (CatBoost), eXtreme Gradient Boosting (XGBoost), and Random Forest (RF), alongside traditional benchmarks such as Decision Trees (DT), Support Vector Machines (SVM), and Multi-Layer Perceptrons (MLP). Bayesian optimization was employed to automate hyperparameter tuning and accelerate model convergence. Results indicate that CatBoost exhibits a superior precision and stability, achieving a coefficient of determination (R2) of 0.9024. Subsequent validation through multiple random seed tests further confirmed the exceptional robustness of the model and generalization capability on unseen data. To bridge the gap between machine learning and metallurgical knowledge, an interpretability analysis was conducted using the SHAP (SHapley Additive exPlanations) framework. The analysis revealed that oxide formation energy and metallic interfacial energy are the dominant features influencing contact angle predictions. The obtained findings validate the physical consistency of the model’s decision-making logic. In conclusion, this research develops a high-precision, physically interpretable framework for the rapid assessment of high-temperature wettability, providing a transformative tool for interfacial reaction control and the optimization of cleanliness control in advanced steel manufacturing.