PSI - Issue 48

2

Oleh Yasniy et al. / Procedia Structural Integrity 48 (2023) 183–189 Yasniy et al / Structural Integrity Procedia 00 (2019) 000 – 000

184

modes loading showed. In the paper by Yasnii et al. (2004), a procedure for predicting the jump-like deformation in the alloy based on the histograms of the distribution of dispersed particles in the initial material is put forward. In general, it is important to evaluate the behavior of the material under different types of stress states with a limited number of experiments. Therefore, in this case, it is advisable to use machine learning methods. Machine learning is an area of artificial intelligence by Smola et al. (2010). It is used for studying input data and constructing a model by continuously evaluating, optimizing, and settings parameters. In particular, due to the ability to interpret non-linear relationships between input and output data, machine learning methods can solve problems of fracture mechanics with high accuracy by Pidaparti et al. (1995), Mohanty et al. (2009). In particular, in the paper by Yasnii et al. (2018) using methods of machine learning (neural networks, boosted trees, random forests, support-vector machines, and k -nearest neighbors), diagrams of fatigue fracture of D16T aluminum alloy under regular loading with a stress ratio R = 0, 0.2, 0.4, and 0.6 were constructed, and it was found that the method of neural networks gives the slightest prediction error equal to 3.2 and 2.5% in tested samples. In papers by Seed et al. (1998), Pujol et al. (2011), the growth of short fatigue cracks and fatigue lifetime under step-stress conditions by method of neural networks are predicted, respectively. It is known that machine learning is often used to predict the dependencies of shape memory alloys (SMAs). In particular, in the paper by Hmede et al. (2022), functions of the SMAs are effectively modeled using machine and deep learning methods, whereas, in the paper by Trehern et al. (2022), an AI-enabled materials discovery framework was successfully used to identify both SMA chemistries and the associated thermo-mechanical processing steps that result in narrow transformation hysteresis and transformation range under applied stress. In addition, stress strain diagrams of aluminum alloy AMg6 by Yasniy et al. (2020), Didych et al. (2022), and aluminum alloy AL-6061 by Didych et al. (2022) were predicted by machine learning methods. Therefore, it is advisable to use them for numerical modeling of stress-strain diagrams of 6061-T651 aluminum alloy at different temperatures. The aim of this study is to predict the stress-strain diagram of 6061-T651 aluminum alloy at six different temperatures (20, 100, 150, 200, 250, 300 ºС) using machi ne learning methods, in particular, the method of k -nearest neighbors and random forests, and to compare the obtained results. 2. Material and methods It is known by Haykin (2006) that there are several main approaches that are widely used in the field of data prediction, that is, supervised learning, unsupervised and mixed. In the first case, it interprets the teacher's participation as the knowledge presented in the form of input-output couples. In particular, the network parameters are corrected by the difference between the desirable and output signals of the network. In contrast, the network with unsupervised learning cannot know the correct answers to each sample of the training sample. In the mixed case, part of the weights is determined by supervised learning, while the other is obtained when the network is self-learning. It is known that at the stage of machine learning model development, preparation of data, algorithm construction, training on the learning data, and verification using the test data are essential. The k – nearest neighbors method algorithm is based on comparing known elements with new ones. Its basic idea is that the new object to be predicted belongs to the class that is most common among k – nearest neighbors of the training sample (Fig.1). The distance between k -nearest neighbors is usually Euclidean. This machine learning method is the algorithm for supervised learning, so it requires a marked data set. In particular, regression problems are concerned with the result prediction of the dependent variable given a set of independent variables. The prediction results are supposed to be the average of the results of its k – nearest neighbors.

Fig.1. Examples of data with plus and minus signs and the query point marked by a blue triangle by Smola et al. (2010)