Microstructure-sensitive machine learning for smart metallurgical manufacture

神经网络和机器学习(ML)算法在材料科学和工程中的应用已有多年，在开发新材料和新的制造方法方面也取得了成功。然而，这项研究的大部分是基于学习数据集，这些数据集试图将数值材料属性数据与制造过程变量联系起来。这种方法的潜力有限，因为没有明确考虑材料的微观结构，而微观结构实际上决定了材料的性能及其在加工过程中的演变。因此，训练好的最大似然模型能够很好地内插属于训练数据范围的可能性，但往往无法在训练数据范围之外做出可行的预测。该项目将通过开发一种ML方法来解决这一能力差距，该方法不仅包括合金性能和加工路线信息，而且还包括来自材料表征实验和计算机模拟的相应微结构数据。训练集还将纳入微观结构演变的既定物理定律，并将寻找训练数据中可能需要进一步科学审查的偏差和细微差别，以及允许在训练集域之外进行预测性外推。也就是说，除了能够“连接点”之外，该方法还将允许“在未知领域绘制新的点”。研究计划：该项目将以数据库的形式开发最大似然算法的训练集，该数据库是根据现有的关于高温合金开发和工业加工的已公布研究数据建立起来的。该数据库将整理合金成分、完整的生产路线、机械性能、抗环境降解性、成本和能源消耗，并将这些与显微结构的图像和其他实验数据联系起来。将对图像进行处理和分析，以测量微观结构的关键几何参数，例如控制基础机械性能的颗粒尺寸分布和沉淀物尺寸分布，以及优化处理可能导致的不良特征，例如对产品寿命和可靠性产生不利影响的不良有害相和缺陷的分布。来自计算机模拟的研究数据也将与实验数据一起在数据库中使用。然而，最终用户将能够选择仅根据实验数据、理论模拟或两者都进行的ML预测。项目目标：(1)建立一个数据库，其中包含各种高温合金独特的材料“指纹”，将数值成分、性能和材料生产数据与来自显微镜和X射线散射的二维和三维微结构成像数据结合起来。(2)将数据库与能够解决相关回归和/或分类问题的适当的ML算法接口。生成灵活而强大的元数据结构将对此至关重要。(3)使用数据库作为ML训练集，以建立数据中的趋势和模式。(4)针对微结构演化和相场计算机模拟的已知物理规律，以及适当的实验，验证训练集的预测能力并建立数学趋势。

Neural networks and machine learning (ML) algorithms have been used in materials science and engineering for some years now and have even yielded successes in developing new materials and novel manufacturing methods. However, the majority of this research is based upon learning data sets that try to link numerical materials property data to the manufacturing process variables. Such approaches have limited potential, because the microscopic structure of the materials that actually determines the properties and its evolution during processing is not taken into account explicitly. As a result, the trained ML models are able to interpolate well the possibilities that fall with the domain of the training data, but often fail to make viable predictions outside of it. Thus, potentially superior novel processing methods and materials with improved properties can remain undiscovered.The project will address this capability gap by developing a ML methodology that will incorporate not only the alloy properties and processing route information, but also the corresponding microstructural data from materials characterisation experiments and from computer simulations. The training set will also incorporate established physical laws for microstructural evolution and will seek out deviations and nuances in the training data that may warrant further scientific scrutiny, as well as allow predictive extrapolation outside the training set domain. I.e. in addition to being able to "join the dots" the approach will allow "plotting new dots in uncharted territory". Research plan: The project will develop a training set for ML algorithms in the form of a database built up from available published research data on the development and industrial processing of superalloys. The database will collate alloy compositions, full production routes, mechanical properties, environmental degradation resistance, cost and energy consumption and will relate these to the images of the microstructure and other experimental data. The images will be processed and analysed to measure key geometrical parameters of the microstructure such as the grain size distribution and precipitate size distribution which govern base mechanical performance, as well as undesirable features that may result from optimal processing such as the distribution of undesirable detrimental phases and defects including cracks, voids, tears, etc. which adversely affect the product lifetime and reliability. Research data from computer simulations will also be used in the database alongside the experimental data. However, the end user will be able to chose of the ML predictions are made from solely experimental data, theoretical simulations or both. Project objectives:(1) Creation of a database containing the unique material "fingerprints" of a broad range of superalloys combining numerical composition, property and material production data with 2d and 3d microstructure imaging data from microscopy and x-ray scattering. (2) Interfacing the database with suitable ML algorithms that will be able to solve the relevant regression and/or classification problems. Generation of a flexible and powerful metadata structure will be essential to this. (3) Use of the database as the ML training set to establish trends and patterns in the data.(4) Validation of the predictive capability of the training set and established mathematical trends against known physical laws for microstructure evolution and phase field computer simulations, as well as experiments where appropriate.