6-10. Machine learning of alloy compositions and properties embedded by a cluster structural unit model

Qing Wang, Zhen Li, Chuang Dong

School of Materials Science and Engineering, Dalian University of Technology, Dalian, China, 116024

Abstract: A much higher requirement on the specific alloy composition and microstructure has been proposed for the high performance under rigorous environments. From the composition viewpoint, it could be realized by the co-addition of multiple alloying elements and these elements must be matched well together. In our previous works, we proposed a “cluster-plus-glue-atom” model based on the chemical short range orders of solute distributions in the several nearest-neighbor shells. This model considered the interactions among elements, in which each element could be well matched and thus the amount of each element could be obtained. Thereof, a cluster formula approach would be achieved to design compositions of multi-component alloys with prominent properties. The present work will focus on the relationship of compositions and properties of maraging stainless steels and -Ti alloys with low Young’s modulus (E), where the former requires body-centered-cubic (BCC) martensite matrix and high strength/micorhardness, and the latter the BCC structure and lower E. Both need multi-component alloying to achieve their specific structure and properties. We tried to construct the relationship between alloy compositions and properties in light of machine learning (ML), in which the cluster structural unit and some characteristic parameters including the equivalent parameter for structural stability and electron concentration, etc., are embedded for better fitting. Thus, at an given composition, we can predict alloy properties well via this relationship achieved by ML (high strength/microhardness for martensite stainless steels and low E for -Ti alloys). On the other side, for any required property, we could search for the optimized alloy compositions easily within the constraint of the cluster formula in the ML. Therefore, the ML approach embedded by the cluster structural formula would open a new way to develop complex high-performance alloys, in which the research efficiency could be improved on a large extent.

Keywords: Engineering alloys; Machine learning; Cluster-formula design approach; Alloy properties

团簇结构单元成分式嵌入的机器学习：工程合金成分与性能

王清，李震，董闯

大连理工大学材料科学与工程学院，大连116024

摘要：特殊工况条件下的高性能需求对工程合金的成分和组织提出了更高的要求，从成分角度出发，需要多个溶质元素共同合金化、且合金化元素之间达到最佳含量匹配。在我们前期工作中，基于溶质元素分布的化学短程序特征，提出了“团簇加连接原子”结构单元模型，该模型考虑了组元间的交互作用，可合理配分多个合金化元素，以此给出添加的合金化元素的含量，从而形成了团簇成分式设计方法，并获得系列性能优异的多组元成分复杂合金。本工作以马氏体时效不锈钢和低弾性模量-Ti合金为例，需要多元合金化才能为实现其独特的结构和性能，前者要求BCC马氏体基体和高强度/硬度，后者要求BCC -Ti结构和低弹性模量。我们借助机器学习方法、并嵌入团簇结构单元成分式、结构稳定性当量参数、电子浓度参量等因素，可很好地建立合金成分与性能(强度、模量)之间的关联。由此，在给定成分下，可准确预测合金的性能(马氏体时效不锈钢要求高强度/硬度，-Ti合金要求低弹性模量)；反之，给定性能时，在团簇结构单元成分式的约束下，可以更好地搜索出满足性能要求的多元合金的成分。因此，这种团簇结构单元成分式嵌入的机器学习方法有望大大提升高性能工程合金研发的效率。

关键词：工程合金；机器学习；团簇式成分设计方法；材料性能