PSI - Issue 59

Oleh Yasniy et al. / Procedia Structural Integrity 59 (2024) 271–278 Author name / Structural Integrity Procedia 00 (2023) 000 – 000

275

of a particular shape coefficient depending on the applied loading to the model, as well as a fragment of the design model during the destruction of inclusions with a ɑ12 =10.5 shape coefficient (the stress applied to the model is 285 MPa).

a)

b)

Fig. 4. Histogram of dispersoid destruction under uniaxial tension of calculation model (a) and fragment of calculation model during inclusions destruction (b). FEM modeling made it possible to identify certain patterns of influence of the cyclic component of the loading during tension test at a constant rate of growth of the average stress in the cycle. It is found that the high-frequency component of the loading causes the occurrence of higher stresses on the inclusions and increases the number of destroyed inclusions. Destruction of inclusions intensifies deformation under cyclic loading in comparison with static one. The accumulated deformation during the cyclic component superposition is significantly higher than the same level of maximum loading in comparison with experiments on linearly increasing load. The results of the FEM calculations of the jump diagram of the AMg6 alloy deformation and the experiment are satisfactorily consistent. According to the proposed model, the average values of relative error of plastic deformation prediction are 12%. 3.1. Method of neural networks Developing the latest approaches to analyzing and predicting fracture mechanics processes involves the possible application of machine learning, a subset of artificial intelligence. ML models are the algorithms that can learn from available data. In particular, its task is to find the hidden dependencies in the available dataset. It is known that the process of initiation of discontinuous flow is random. Therefore, given the large amount of experimental data, it is essential to learn how to solve such problems using ML methods, particularly neural networks. Neural networks are powerful methods of machine learning that demonstrate a high level of accuracy in solving mechanics-related tasks by Pidaparti and Palakal (1995), and Mohanty et al. (2009). In particular, neural networks learn from examples. A prevalent approach in data prediction is supervised learning, where neural networks are trained by providing inputs and corresponding target outputs to establish a relationship between them. In this process, often called learning with a teacher, the teacher provides target outputs corresponding to specific input signals. The objective is to minimize signal error, defined as the absolute value of the difference between the output signal and the target signal. This is achieved through the continuous adjustment of neuron weights, as discussed by Goodfellow et al. (2016), Brunton and Kutz (2019), Harrington (2012), Russell and Norvig (2020), Alpaydin