为突破传统斗齿用铸钢成分设计与热处理工艺优化过程中依赖经验试错、开发周期长且多性能难以协同调控的局限性，本工作采用机器学习与非支配排序遗传算法(NSGA-Ⅱ)，构建了斗齿用铸钢成分与热处理工艺的优化设计方法，实现了具备高强度高硬度和良好塑性的新型斗齿用铸钢的化学成分及热处理工艺参数的快速设计。基于以往有关铸钢的研究，形成了包含成分、热处理工艺以及力学性能(洛氏硬度、抗拉强度和延伸率)的800条铸钢样本数据集，通过研究构建五种机器学习模型(决策树(DT)、随机森林(RF)、支持向量机(SVM)、梯度提升决策树(XGBoost)和多层感知机(MLP))，预测铸钢的洛氏硬度、抗拉强度以及延伸率。结果表明，XGBoost模型在洛氏硬度和延伸率上整体预测效果最佳，SVM模型在抗拉强度上整体预测效果最佳。随后，利用训练好的最佳机器学习模型结合NSGA-Ⅱ对铸钢的成分空间进行多目标寻优，在所得的可行解中选出一种铸钢成分进行实验验证。实验结果表明，该新型斗齿用铸钢的洛氏硬度为50.54 HRC，抗拉强度为1736.9 MPa，延伸率为14%，与模型优化目标高度吻合。

As a key and vulnerable component of excavator engineering machinery, the performance of bucket teeth directly determines the machine's working efficiency, economic cost, and operational safety. In traditional material development models, reliance has primarily been placed on trial-and-error methods and the experience of experimental personnel. This approach not only results in lengthy development cycles and high costs but also often fails to achieve a balance among the multiple performance indicators of cast steel, significantly limiting the efficiency of developing new high-performance materials. In recent years, with the rapid advancement of artificial intelligence technology, particularly machine learning methods, revolutionary tools have been provided to address the mapping relationship between material composition, processing, and performance. This study combines machine learning and optimization algorithms to construct a composition space for cast steel, thereby screening cast steel materials with high hardness and strength. A dataset of 800 cast steel samples, including composition, heat treatment processes, and mechanical properties (Rockwell hardness, tensile strength, and elongation), was compiled from the literature. Five machine learning models (SVM, XGBoost, RF, DT, MLP) were developed to predict the Rockwell hardness, tensile strength, and elongation of cast steel. The results show that the trained gradient boosting model (XGBoost) achieved the best overall prediction performance for Rockwell hardness and elongation, with mean absolute errors of 2.146 and 1.6442, and coefficients of determination (R2) of 0.9082 and 0.9166, respectively. The support vector machine model (SVM) performed best in predicting tensile strength, with a mean absolute error of 70.1750 and a coefficient of determination (R2) of 0.9234. Subsequently, the trained machine learning models were combined with an optimization algorithm (NSGA-Ⅱ) to search the composition space of cast steel. One cast steel composition was selected from the feasible solutions obtained through the search and experimentally validated. The experimental results showed that the cast steel had a Rockwell hardness of 50.54 HRC, a tensile strength of 1736.9 MPa, and an elongation of 14%, which aligned with the model's search objectives. This study provides a more efficient new strategy for the development of high-strength, high-hardness cast steel for bucket teeth.