PSI - Issue 42

Iryna Didych et al. / Procedia Structural Integrity 42 (2022) 1344–1349 Author name / Structural Integrity Procedia 00 (2019) 000 – 000

1345

2

Mohanty et al. (2009), Yasnii et al. (2018, 2020). Machine learning methods predict the integrity and durability of structural materials, particularly, in paper by Seed et al. (1998) the growth of short cracks in the neural network is predicted, in paper by Artymiak et al. (1999) limited endurance and fatigue limits with different load cycle asymmetries, while in paper by Kang et al. (1998) fatigue cracks closing parameters. In paper by Pujol et al. (2011) fatigue life under the conditions of step voltage by neural network is predicted. Subsequently, such approaches predict fatigue life based on the analysis of damage model in the case of broadband Gaussian process of two peaks by Kim et al. (2016) and detect small fatigue cracks by Rovinelli et al. (2018), as well as predict the growth of fatigue cracks by Wang et al. (2017). In addition, the methods of machine learning simulate the diagrams of aluminum alloy AMg6 deformation by Yasniy et al. (2020), by Didych et al. (2022). Experimental research methods of materials are complex and expensive. In particular, with a large amount of experimental data, they need to be generalized, i.e. to find the dependence that most accurately approximates them. Therefore, in this case, it is advisable to use machine learning methods that solve such problems with great accuracy. The constructed model is estimated by means of loss function, which in this paper is selected as the root mean square error (MSE) by Pidaparti et al. (1995): (1) The objective of the investigation is to predict the deformation diagrams of aluminum alloy AL-6061 by machine learning methods, particularly, neural networks and boosted trees at temperatures Т = 343  C and 413  C, without explicit specifying analytical models and compare the results. 2 1 1 ( n  ) prediction true i E y n y = = −

Nomenclature y prediction predicted sample element y true

experimental value of the sample element

n

training sample size

2. Material and methods Effective learning algorithms such as neural networks and boosted trees are used to model Al-6061 aluminum alloy deformation diagram. One of the main approaches used to predict data is teacher training. Its purpose is to minimize the loss function due to the constant change of network parameters. Boosted tree algorithm is a method of learning with the teacher, where the local area is identified in the order of recursive splitting by smaller number of steps by Alpayndin (2010). Boosted tree consists of internal nodes and leaves (Fig. 1).

Fig. 1. Example of a dataset and the corresponding decision tree. Oval nodes are the decision nodes and rectangles are leaf nodes by Alpayndin (2010).