PSI - Issue 48

Bernadett Spisák et al. / Procedia Structural Integrity 48 (2023) 326–333 Spisák et al / Structural Integrity Procedia 00 (2019) 000 – 000

328

3

2. Identifying damage parameters with the help of ANN The ANN is a network of artificial neurons inspired by the way the human brain works, designed to perform specific tasks. These biological methods of computing are recognized as the next big advance in computing. The application of the ANN technique has also become a widespread method for determining material properties. Bhadeshia (1999) was one of the first researcher who reviewed the usability of artificial neural networks in materials science. For example, Guo et. al. (2005) developed a n ANN model to simulate the non-linear relationship between the beta transus temperature of titanium alloys and the alloy chemistry, Guan and Jin (2012) proposed an ANN model to extract the residual stress and strain-hardening exponent based on spherical indentation. Also there are examples in case of local approach modelling too. Abbassi et. al. (2013) applied the ANN technique for the determination of GTN parameters from tensile test results, and recently Shikalgar et. al. (2020) analyzed pre-cracked small punch test (SPT) specimens, where the GTN parameters were determined with the help of ANN. Due to these articles artificial neural network was used for the determination of the damage parameters of 15H2MFA. The input data in this case was generated with the help of finite element simulation where the MSC.Marc. software was used. The total length of the small sized, flat notched specimen (Fig. 1. (a)) was 27 mm, the thickness was 1 mm, the radius of the notch was 1 mm and the width at the notch was 5 mm. The finite element mesh and the used boundary conditions are shown on Fig. 1. (b). All together 90 GTN parameter sets were generated, where q 1 , q 2 and S n were fixed, therefore the remaining 5 damage parameter was varied. For a more uniform distribution, the Latin Hypercube Sampling (LHS) method was used to create the parameter sets. The essence of this technique is to randomly generate points in a n-dimensional region defined by the range of variables. In contrast to the Monte-Carlo simulation, the distribution of the domain is done according to the number of samples, therefore the probability of the domains is the same, and a more uniform distribution can be obtained.

(a) (b) Fig. 1. (a) Prepared small sized NT specimen; (b) Finite element mesh and boundary conditions.

(a) (b) Fig. 2. (a) generated displacement-force curves with the 90 GTN parameter set; (b) optimized GTN parameter set with ANN.