SPbGASU graduate student Yury Zgoda developed the OptiShellNeuro program, a web application for neural network modeling of thin-walled shell structures. Certificate of state registration of the computer program No. 2023682977 was received.

– Shells are actively used in architecture and construction, but due to the complex curvilinear geometry and non-trivial mathematical models of deformation, their computer modeling is a long, labor-intensive process. Classical methods of computer modeling involve solving a large number of systems of linear and nonlinear equations, numerical integration of hundreds of thousands of terms and other resource-intensive calculations. Modeling using neural networks (one of the current methods of artificial intelligence) is a relatively new approach to calculating shells and allows you to reduce a complex multi-stage computing process to a series of relatively simple data processing manipulations.

On the topic of high-performance computer modeling of shells, Yury Zgoda is writing a PhD thesis under the scientific supervision of Associate Professor at the Department of Information Systems and Technologies Aleksey Semenov.

The web application was developed by the SPbGASU scientific school "Computer technologies for integrated research of strength, stability and nonlinear vibrations of constructions, buildings and structures", founded by Professor Vladimir Karpov.

– Neural networks can be trained on a certain set of data by “showing” them examples of input data and corresponding solutions. Upon completion of training, the neural network can find a solution with high accuracy based on the “experience” of processing the training data. At the same time, all calculations associated with the use of a trained neural network are a fixed set of arithmetic operations that do not require the use of expensive equipment or significant time costs to perform them. In other words, neural networks inherently allow calculations to be performed much faster compared to classical modeling methods. There are currently several studies that use neural networks to predict the behavior of shells under the influence of external forces, but all of them are limited to some extent. Most often, neural networks are used only to determine the critical loads of shells, so there is no need to talk about full-fledged modeling of the stress-strain state, – the author said.

The most significant stage in preparing a neural network as a whole, without reference to the field of shell modeling, is the formation of a training data set (samples of problems and their solutions on which the neural network is trained), selection of the architecture of the neural network (determining the structure of connections between artificial neurons) and training neural networks with this architecture.

To achieve acceptable modeling accuracy, the training data must include a large number of samples from which the neural network can “extract” the features of the modeled phenomenon. However, calculating shells using classical methods is a lengthy computational process. To speed up calculations, the previously developed OptiShellX software was used, which implements a number of original software optimizations.

Due to the lack of any research related to the choice of neural network architecture in the context of shell modeling, several dozen different configurations were considered during the development of OptiShellNeuro, ranging from the most trivial to complex configurations with subnets, parameter passing and batch normalization. Based on the research results, the most effective configuration in terms of accuracy and duration of modeling was determined.

Yury Zgoda calculated 4410 cylindrical shells, differing from each other in materials and geometry parameters, and also developed the architecture of a neural network for modeling the stress-strain state of a cylindrical shell.

Upon completion of training, the neural network demonstrated high modeling accuracy, while the duration of calculations was reduced by approximately a thousand times, from tens of seconds to hundredths of a second (the exact values vary depending on the parameters of the simulated structures).

Based on a trained neural network, Yury Zgoda created a web application that allows you to perform calculations through a web browser. Due to the high performance of neural network modeling, it is possible to obtain structural modeling results in almost real time. The application has a simple user interface and can visualize the stress-strain state in the form of interactive two-dimensional and three-dimensional graphs.

– The developed program allows both to reduce hardware requirements and to significantly speed up the calculation process, thereby speeding up the work processes of engineers. Also, this development can be effectively used in the educational process.

The author plans to expand the functionality of the OptiShellNeuro web application: to implement support for other types of structures in addition to cylindrical ones and to explore new ways to further reduce the computation time.