Application of computational intelligence methods for the heterogeneous material stress state evaluation

The use of surrogate models providesgreat advantages in working with computer-aided design and 3D modeling systems, which opens up new opportunities for designing complex systems. They also allow us to significantly rationalize the use of computing power in automated systems, for which response time and low energy consumption are critical. This work is devoted to the creation of a surrogate model for approximating the finite element solution of the problem of dispersion–strengthened composite plane sample deformation. An algorithm for constructing a parametric two–dimensional model of a composite is proposed. The calculation model is created using the ANSYS Mechanical computer-aided design and analysis program using the APDL scripting model builder. The parameters of the stress-strain state of the material microstructure are processed using a convolutional neural network. A neural network based on the U–Net architecture of the encoder-decoder type has been created to predict the distribution of equivalent stresses in the material according to the sample geometry and load values. A direct sequence of layers is takenfrom the specified architecture. To increase the speed and stability of training, the type of part of the convolutional layers has been changed. The architecture of the network consists of serially connected blocks, each of which combines layers such as convolution, normalization, activation, subsampling, and a latent space that connects the encoder and decoder and adds load data. To combine the load vector, such a neural network architecture as a concatenator is created, which additionally includes the Dense, Reshape and Concatenate layers. The model loss function is defined as the root mean square error over all points of the source matrix, which calculates the difference between the actual value of the target variable and the value generated by the surrogate model. Optimization ofthe loss function is performed using the first–order gradient local optimization method ADAM. The study of the model learning process is illustrated by plots of loss functions and additional metrics. There is a tendency for the indicators to coincide between the training and validation sets, which indicates the generalizing capability of the model. Analyzing the output of the model andthe value of the metrics, a conclusion is made about the sufficient quality of the model. However, the values of the network weights after training are still not optimal in terms of minimizing the loss function. And also, to accurately reproduce the solution of the finite element method (FEM), the proposed model is quite simple and requires clarification. The speed comparisonof obtaining results by the FEM and using the architecture of the neural network is proposed. The surrogate model is significantly ahead of the FEM and is used to speed up calculations and determine the overall quality of the approximation of problems of mechanics of this type