1 4 A D V A N C E D M A T E R I A L S & P R O C E S S E S | A P R I L 2 0 2 3 Fig. 1 — Machine learning technology stack. The application of machine learning (ML) in the field of materials science and engineering has rapidly matured over the past decade. However, the full potential of this methodology has yet to be unleashed. This article starts with a succinct synopsis of ML and explores its many diverse characteristics. The results of two recently completed research projects investigating the potential use of ML to establish additive manufacturing (AM) materials property allowables are described. Although continued research and development (R&D) work is required, the results are very promising. The authors’ thoughts on the maturation of ML for this application are then delineated. BACKGROUND The technologies involved in ML are illustrated in Fig. 1 and may be conveniently divided into seven broad areas: (i) data, (ii) ML categories, (iii) environment/infrastructure, (iv) data science and ML libraries, (v) algorithms, (vi) quality, and (vii) models. For the qualification of AMmaterials, the goal of ML is the development of accurate predictive models. Models may be thought of as human artifacts, i.e., representations of reality within prescribed limits. The generation of ML models is based on the analysis of data. AM is a digitally intense process generating an abundance of data. Data quality, type, and quantity is important, and data must be scrubbed to ensure its pedigree and provenance. This is not trivial and can consume 85% of a data scientist’s time[1,2]. The type of data (continuous or categorical/discrete) must be established. Further, prior to deciding upon an ML approach, the quantity of data needed to derive a meaningful model and the required data environment and infrastructure should be considered carefully. Mathematical algorithms are then used to transform data into models. There are many algorithms used in ML, and each has their appropriate applications, strengths, and weaknesses. Table 1 provides a representative partial list of ML algorithms and their notional characteristics[3,4]. Common types of algorithms include regression, neural nets (NN), deep learning (DL), decision trees (DT), k nearest neighbors (KNN), and support vector machines (SVM)[5]. This article discusses regression, which has broad application, and polynomial regression algorithms, which were used in the research reported by the authors in this work. Regress i on analysis encompasses a large variety of statistical methods to estimate the relationship between input variables and their associated features. Typical regression methods include a) linear regression, b) multilinear regression, c) polynomial regression, and d) logistic regression. Regression analysis is not a new technique but its application to big data sets with large numbers of independent variables to establish (with statistical confidence) a set of dependent materials properties allowables is novel. ML can be applied to AM in a variety of ways in varying stages of maturation. Currently the lowest hanging fruit TABLE 1 — GENERAL CHARACTERISTICS OF MACHINE LEARNING ALGORITHM[3,4] Algorithm Algorithm type* Learning type** Data required Computational time to learn Linear regression R Low Low Polynomial regression R Low Low Logistic regression C Low Low Naïve Bayes C S Low Low Neural network C S, U, R High High Deep learning C S, U, R High High Decision tree C & R S Low Low k-Nearest neighbor I I Low Low Support vector machine C S Low High k-Means C U Low *Type of algorithm: Regression (R), Classification (C), Instance (I) **Learning types: Supervised (S), Unsupervised (U), Reinforced (R), Instance (I)
RkJQdWJsaXNoZXIy MTYyMzk3NQ==