High performance modeling of the stress-strain state of thin-walled shell structures with the use of deep learning

Computer modeling is one of the most common approaches to the analysis of thin-walled shell structures stress-strain state analysis. It requires considerable time costs and high-performance hardware, especially when it is necessary to conduct a comparative analysis of various shell configurations. In this paper, we propose the use of deep learning methods to improve performance of this process. The purpose of work is to develop methods for high-performance computer simulation of thin-walled shell structures using deep neural networks, allowing to take into account geometric and physical properties of the structure as well as the load applied to it. A training approach and deep neural network architecture were developed to perform computer modeling of the stress-strain state of the shell. To form a training dataset, a computational experiment was carried out to simulate 3904 different configurations of doubly curved shallow shells that differ in linear dimensions, curvature radii, and materials used. Based on this dataset, 30 deep neural networks with different architectures were trained. To determine the optimal architecture in terms of modeling accuracy, mean absolute percentage error with clipping near-zero samples was calculated for each of the neural networks based on the test dataset. A network has been developed that allows calculating the stress-strain state of different shell configurations under an arbitrary uniformly distributed load. This is the first solution in the field of shell neural network modeling that allows us to vary the applied load, geometric and physical parameters of the shell and obtain calculation results at an arbitrary point of its middle surface. Performance measurements were carried out which show that the developed neural network allows simulating the stress-strain state of a shell structure 2117 times faster compared to the duration of solving the same problem by classical simulation. The modeling error using the network is at an acceptable level. An original architecture of a neural network for modeling the stress-strain state of shells was proposed which, through minor modifications, can be adapted for high-performance modeling of other building structure types. In accordance with the described architecture, a deep neural network was trained which reduces the computation time by several orders of magnitude. The results obtained are of high practical importance for researchers in the field of thin-walled shells modeling since they allow us to significantly reduce the time costs associated with conducting computational experiments. One of the possible applications for developed solution is prototyping of various shell configurations. Once prototyping is complete, the most efficient shell configurations can be explored in detail using classical computer simulation techniques.