External Total Materia AG Kirchenfeld 71 Zurich, 8052 Switzerland Predictor 2.2 Whitepaper 1. Summary This paper presents a machine learning system aimed at predicting material properties of a wide range of diversified materials, such as steels, non-ferous alloys and polymers. By using copious training sets provided from a very large database and proprietary methodology for taxonomy, data curation and normalization, the developed system is able to predict physical and mechanical properties for hundreds of thousands of materials, at various temperatures and various heat treatments and delivery conditions. The accuracy achieved in terms of relative error is in most cases above 90%, and frequently above 95%, thus being higher than readily available, statistically derived data sources such as MMPDS B-Basis values, which are routinely used in aerospace industry. 2. Introduction The importance of accurate material properties information for engineering calculations and simulations, such as CAE (Computer Aided Engineering) and FEA (Finite Element Analysis), can never be overstated. Conventional mechanical properties such as yield strength, tensile strength, hardness, and ductility may vary more than tenfold for structural steels at room temperature, depending on the variations of alloying elements, heat treatment and fabrication. With even a moderate change in working temperature, the property’s variations and changes can become even more profound and their approximation using the typical property values for some groups of alloys may lead to very serious errors. Contemporary large material databases and material selection software can substantially help with these challenges in engineering and simulation, however it is unfortunately technically impossible to have all properties for all materials readily available from experiments and standards. Recent developments in artificial intelligence and machine learning however provide an opportunity to overcome this gap. Machine learning (ML) has been increasingly applied in material science, including property modelling and prediction. Some examples include modelling the correlation between processing parameters of maraging steels [1], predicting mechanical properties of AISI 10xx steel bars [2] and predicting crack propagation in stainless steels [3]. A comprehensive view of ML applications in material science, notably for metallic materials, is provided in [4]. While possibly of interest for some particular applications and material groups, in most cases these models are not helpful to provide material properties needed for practical CAE and FEA usage. Their main flaw is the lack of needed capability to simulate the behavior of thousands of diversified structural materials from carbon and stainless steel, to special alloys, nonferrous 1 External metals and polymers. Development of such a universal system for material properties prediction would require, besides application of modern ML algorithms, an access to a large pool of datasets for the learning and testing of ML models, as well as innovative and consistent methodology to handle and curate such a quantity of data. 3. Development Methodology The required breadth of material property data needed for the development of universal ML system for predicting material properties can be in principle obtained from a large database containing structural material properties, such as Total Materia Horizon [5], which comprises more than 500,000 materials. However, there is no single ML model capable of effectively using its 20 million property records for dozens of material properties to deliver predictions of any acceptable quality. Therefore, classifying and grouping materials, normalizing their properties and conditions (a combination of thermic-, mechanical- and other processing patterns of a material), and then applying dataset splitting and division, is essential to enable effective ML on such comprehensive and diversified datasets. 3.1 Data Curation Methodology To get to the scope of data feasible for the application of ML, it is necessary to divide the problem space into smaller subspaces. Criteria for division and classification may vary from selfevident, such as type and family of materials, properties to product forms and thermomechanical treatments of material. Diversified taxonomies already applied in the industry, such as VDA 231-200 [6], can be helpful guideline in this process. The division of the problem space leads the creation of dozens and even hundreds of specific ML models for certain families of materials and properties, as it can be seen in Figure 1, for the example of aluminum alloys. For each of them, it is then necessary to prepare datasets appropriate for ML purposes. This data curation process is actually very complex and includes the elimination of redundant and repetitive inputs, exclusion of extreme or inconsistent values and, handling missing data or intervals (e.g. in chemical composition). Fig 1. An example of the classification of ML models for aluminum alloys. 2 External One of the particularly challenging parts is the transformation of categorical values, such as heat treatment, into consistent numerical values that can be effectively used by ML models. To achieve that, a nine-digit classification system was used, where the first digit determines larger group of treatments or temper type, e.g. 1 for F state (as fabricated), 2 for H state (Strain Hardened), 3 for O state (Annealed) and 4 for T state (Thermally Treated), and other digits are determined by subdivisions of tempers. For example, T62 temper gets code 462000000, H111 temper products gets 211100000, whilst T8E30 gets 480000130. The meaningfulness and consistency of applied classifications is of crucial importance for the quality of models and their results. The representativeness of datasets for training and testing is also important for the performance of the ML models. Even though the datasets for each ML model have a considerably more manageable size than the whole initial database, still obtaining uniform coverage of the application domain of several thousand points in n-dimensional space is not a trivial problem. A combination of random splitting and Kennard-Stone algorithm has been applied to achieve that, Figure 2. Fig 2. Preparation of datasets for training and testing of ML models. 3.2 Machine Learning Architectures From a large number of possible ML concepts and architectures, XGBoost [7] has been selected from tree boosting ML methods, owing to its proven performance in solving similar problems, see [8-9] and references therein. XGBoost, which a scalable, distributed gradientboosted decision tree algorithm, was first introduced in 2016 by Tianqi Chen and Carlos Guestrin. The structure of XGBoost includes multiple root nodes, internal nodes, leaf nodes, and branches (Figure 3a). Then, the internal nodes make subsequent decisions, the branch points point directly to the decision to be made, and the leaf nodes represent the prediction results of a single three. Finally, the results of all leaf-pointing nodes are combined to obtain the prediction results of the XGBoost model [8]. In the search for the best leaf node segmentation, XGBoost uses the basic exact greedy algorithm and the corresponding approximate algorithm to enumerate all the features to ensure the accuracy [9]. 3