Multi-objective materials bayesian optimization with active learning of design constraints: Design of ductile refractory multi-principal-element alloys
Multi-objective materials bayesian optimization with active learning of design constraints: Design of ductile refractory multi-principal-element alloys
复制标题
主动学习设计约束的多目标材料贝叶斯优化：延性难熔多主元合金的设计
DOI：
10.1016/j.actamat.2022.118133
复制
发表时间：
2022
期刊：
影响因子：
9.4
通讯作者：
Arróyave, Raymundo
中科院分区：
文献类型：
--
作者：
Khatamsaz, Danial;Vela, Brent;Singh, Prashant;Johnson, Duane D.;Allaire, Douglas;Arróyave, Raymundo
贝叶斯优化（BO）作为一种有效探索和利用材料设计空间的强大框架已经出现。迄今为止，大多数材料设计的BO方法都集中在材料发现问题上，就好像它是一个单一的查询昂贵的“黑匣子”，其中目标是优化单个目标（即材料属性或性能指标）。此外，这种方法往往是约束不可知论的。在这里，我们提出了一种新的多信息BO框架，能够主动学习作为多目标和约束问题的材料设计。我们通过优化探索难熔多主元素合金（MPEA）空间来证明这一框架，具体来说，这里是Mo-Nb-Ti-V-W体系。MPEAs旨在优化密度泛函理论（DFT）导出的两个延性指标（皮格比和柯西压力），同时学习与高温燃气轮机部件制造相关的设计约束。使用DFT分析BO Pareto-front合金，以深入了解其优越性能的基本原子和电子基础，并在此框架内进行评估。
Bayesian Optimization (BO) has emerged as a powerful framework to efficiently explore and exploit materials design spaces. To date, most BO approaches to materials design have focused on the materials discovery problem as if it were a single expensive-to-query ‘black box’ in which the target is to optimize a single objective (i.e., material property or performance metric). Also, such approaches tend to be constraint agnostic. Here, we present a novel multi-information BO framework capable of actively learning materials design as a multiple objectives and constraints problem. We demonstrate this framework by optimally exploring a Refractory Multi-Principal-Element Alloy (MPEA) space, here specifically, the system Mo-Nb-Ti-V-W. The MPEAs are explored to optimize two density-functional theory (DFT) derived ductility indicators (Pugh’s Ratio and Cauchy pressure) while learning design constraints relevant to the manufacturing of high-temperature gas-turbine components. Alloys in the BO Pareto-front are analyzed using DFT to gain an insight into fundamental atomic and electronic underpinning for their superior performance, as evaluated within this framework.