A framework based on physics-informed neural networks and extreme learning for the analysis of composite structures. (June 2022)

• Record Type: Journal Article
• Authors: Yan, C.A.; Vescovini, R.; Dozio, L.
• Is Part Of: Computers & structures. Volume 265(2022)
• Publication Date: 2022-06
• Subjects: Physics-informed neural networks -- Extreme learning machine -- Structural analysis -- Shell structures; Structural engineering -- Data processing -- Periodicals; Electronic data processing -- Structures, Theory of -- Periodicals; 624.171
• DOI: 10.1016/j.compstruc.2022.106761
• Languages: English
• ISSNs: 0045-7949

Abstract

Highlights: Physics-Informed Neural Networks are presented for analyzing composite structures. Extreme Learning is proposed for fast training of the network. Subdomain decomposition allows assemblies of shell elements to be modeled. Parametric studies illustrate the effect of network architectures and hyperparameters. Improved effiency is achieved with respect to Gradient-Based Learning. Abstract: This paper presents a novel approach for solving direct problems in linear elasticity involving plate and shell structures. The method relies upon a combination of Physics-Informed Neural Networks and Extreme Learning Machine. A subdomain decomposition method is proposed as a viable mean for studying structures composed by multiple plate/shell elements, as well as improving the solution in domains composed by one single element. Sensitivity studies are presented to gather insight into the effects of different network configurations and sets of hyperparameters. Within the framework presented here, direct problems can be solved with or without available sampled data. In addition, the approach can be extended to the solution of inverse problems. The results are compared with exact elasticity solutions and finite element calculations, illustrating the potential of the approach as an effective mean for addressing a wide class of problems in structural mechanics.