基于物理引导机器学习的活性多主元合金设计与拉伸屈服强度预测OA
Design of Reactive Multi-principal Element Alloys Based on Physics-guided Machine Learning and its Prediction of Tensile Yield Strength
活性多主元合金兼具优异力学性能与高氧化燃烧热,在含能结构材料领域具有重要的应用潜力.目前针对该类材料的力学性能研究多集中于准静态压缩,决定其结构承载极限及冲击破碎释能特性的拉伸屈服强度数据相对匮乏,且因数据量少、非线性强,传统的试错法难以在广阔成分空间中实现拉伸屈服强度的精准预测与定向设计.为此本研究提出了一种机器学习驱动的设计策略,以解决小样本下活性多主元合金拉伸屈服强度的预测与优化难题.研究基于收集的88组铸态活性多主元合金数据,结合融合领域知识的33个物理描述符,采用5种机器学习算法构建预测模型,并利用遗传算法进行特征降维.结果表明,最优的支持向量回归(SVR)模型在测试集上的决定系数(R2)达到0.928.SHAP可解释性分析揭示,组元熔点差异是影响屈服强度的最关键因素,原子半径和电负性差异亦起重要正向作用.基于模型的成分空间反向设计预测显示,在Ti-Zr-Nb-Ta体系中,通过增加Ta含量并减少Nb含量可显著提升拉伸屈服强度.实验制备的TiZrNbTax系列合金验证了这一规律,证实了该数据驱动范式在高性能活性多主元含能结构材料设计中的有效性与准确性.
Reactive multi-principal element alloys(RMPEAs),combining superior mechanical properties with high heat of oxida-tion,possess significant application potential in the field of energetic structural materials.Currently,research on the mechanical properties of these materials focuses predominantly on quasi-static compression.Data regarding tensile yield strength,which gov-erns the structural load-bearing limit and impact-induced fragmentation and energy release characteristics,remain relatively scarce.Furthermore,due to the limited dataset size and strong non-linearity,traditional trial-and-error methods struggle to achieve precise prediction and targeted design of tensile yield strength within the vast compositional space.This study proposes a machine learning-driven design strategy to address the challenges of predicting and optimizing the tensile yield strength of RM-PEAs under small-sample conditions.Based on a collected dataset of 88 as-cast RMPEAs and incorporating 33 domain-knowledge-integrated physical descriptors,prediction models were constructed using five machine learning algorithms,with a genetic algorithm employed for feature dimensionality reduction.The results demonstrate that the optimal Support Vector Regression(SVR)model achieves a coefficient of determination(R²)of 0.928 on the test set.SHapley Additive explanation(SHAP)interpretability analysis reveals that the difference in melting points of the constituent elements is the most critical factor influencing yield strength,while differences in atomic radius and electronegativity also play significant positive roles.Inverse de-sign of the compositional space based on the model predicts that within the Ti-Zr-Nb-Ta system,increasing Ta content while re-ducing Nb content can significantly enhance tensile yield strength.The experimentally fabricated TiZrNbTax series alloys validat-ed this trend,confirming the effectiveness and accuracy of this data-driven paradigm for the design of high-performance reactive multi-principal element energetic structural materials.
张周然;张龙辉;彭泳潜;李顺;陈荣;白书欣
国防科技大学 空天科学学院,湖南 长沙 410073国防科技大学 空天科学学院,湖南 长沙 410073国防科技大学 空天科学学院,湖南 长沙 410073国防科技大学 空天科学学院,湖南 长沙 410073国防科技大学 理学院,湖南 长沙 410073国防科技大学 空天科学学院,湖南 长沙 410073
军事科技
活性多主元合金设计拉伸屈服强度预测机器学习遗传算法可解释性分析
reactive multi-principal element alloy designtensile yield strength predictionmachine learninggenetic algorithminterpretability analysis
《含能材料》 2026 (4)
338-349,12
国家自然科学基金(U2441214)National Natural Science Foundation of China(No.U2441214)
评论