Neural network models of deformation of thin-walled shells

Computer simulation of thin-walled shell structures is a time- and resource-intensive process, especially in those cases when the researcher needs to implement original modeling software. Neural network modeling is a promising approach that can significantly increase modeling performance of such constructions and simplify software development is neural network modeling. Due to the novelty of this approach, its practical application features remain underexplored. In this work, we aim to classify and analyze various approaches to creation of neural network models for describing the deformation of shell structures. To that end, we carry out a literature review of modern studies devoted to the application of this approach in various scientific fields, including modeling shells and plates (which can be considered as a special case of shells). On the basis of the conducted literature review, a system for classifying neural network models of shell deformation is proposed. This system includes three directions with a description of advantages and disadvantages of each of the mentioned approaches. The introduction of such a classification system simplifies the process of selecting neural network architectures and training data sources. In addition, the system represents the existing scientific knowledge in a structured manner. For one of the considered neural network architectures, a computational experiment is carried out with the description of training data preparation, network training, and validation procedures. The analytical properties of the architecture are presented. The most promising subtypes of neural network shell models requiring further research are identified. The high efficiency of neural network research in terms of performance and accuracy is noted.

Authors: Iu. N. Zgoda

Direction: Informatics, Computer Technologies And Control

Keywords: shells, neural network modeling, Julia, Flux, high-performance computing, classification, computational experiment

---

View full article