Purpose: The main aim of this paper is development, software implementation and use of the alloys selection method for the design of biocompatible materials in medical production. It is based on the use of Ito decomposition and Logistic Regression. Design/methodology/approach: The technology of machine learning is used to solve the task. The developed classification method is based on the application of multiclass Logistic Regression. In order to reduce the probability of incorrect alloy identification, expansion of the input characteristics based on the Ito decomposition of the second order has been made. On the one hand, it increased the dimension of the input features space, and as a result, it increased the time for training procedure, but on the other, it increased the solution accuracy of the alloys selection task. The accuracy evaluation of the method was carried out using different criteria. In particular, the method accuracy was estimated based on the ratio of correctly classified titanium alloys samples to the test sample dimension. This measure was used to assess the classification accuracy in the training and test modes. For a more detailed analysis of the classification method results, two additional measures were further used: Precision and Recall. Their calculation was based on the constructed confusion matrix. This made it possible to assess the ability of the developed method to find the instances of each individual alloy as a whole, as well as the ability to distinguish instances of one class from representatives on the other. The combination of these indicators allowed to evaluate the classification task accuracy in the conditions of the imbalance dataset for each class of the investigated material separately. Findings: The simulation results confirmed the effectiveness of the use of machine learning tools to solve this task. High indicators of the method’s accuracy based on the experimental results were established. In particular, the overall accuracy of the method is 96.875%, and the average values of Precision and Recall for all four classes are 94% and 98% respectively. Expansion of each vector's features from the training dataset by using Ito decomposition increased the method accuracy by more than 33% compared to the basic Logistic Regression. Research limitations/implications: The Logistic Regression's training procedure, as well as the increase of the space size of the investigated alloy's input features by using Ito decomposition, imposes a number of limitations on the application of the method in tasks that depend on the duration of the work. Practical implications: The proposed machine learning approach foralloys selection allows reducing the time, material and human resources needed to investigate the titanium alloys properties. The developed method increases the accuracy of the selection alloys task compared to the four known methods, an average of 14.5%. It can be used to select materials with appropriate properties for the design of biocompatible medical products. Originality/value: A method and software product for the titanium alloys classification task using a supervised learning technique has been developed. For this aim, the method of Logistic Regression with expanding inputs based on the second-order Ito decomposition is used.

Based on the experimentally established data on the parameters of microstructure, elemental and fractional composition of titanium alloy powders, four classes of their conformity (a material with excellent properties, optimal properties, possible defects in the material and defective material) as source raw materials for the additive technologies are identified. The basic characteristics of the material, which determine its belonging to a certain class, are established. Training and test samples based on 20 features that characterize each of the four classes of titanium alloy powders for the implementation of machine learning procedures were built. The developed method for identification of the class of material, based on the use of the second-order Kolmogorov-Gabor polynomial and the Random Forest algorithm, is described. An experimental comparison of the developed method work results with existing methods: Random Forest, Logistic Regression, and Support Vectors Machines based on the accuracy of their work in the training and application modes was made. The visualization of the results of all the investigated methods was given. The developed supervised learning method allows constructing models for processing a large number of each input vector characteristics. In this case, the Random Forest algorithm provides satisfactory generalization properties while retaining the advantages of an additional increase of the accuracy based on the Kolmogorov-Gabor polynomial. The main advantages of the developed method, in particular, regarding the additional increase of the accuracy of the classification task solution, are experimentally determined. The developed method allows increasing the modeling accuracy by 34.38, 33.34 and 3.13 % compared with the methods: Support Vectors Machine, Logistic Regression, and Random Forest respectively. The obtained results allow one to considerably reduce financial and time expenses during the manufacture of products by additive technologies methods. The use of artificial intelligence tools can reduce the complexity and energy efficiency of experiments to determine the optimum characteristics of powder materials.