Design of a floating offshore structure by a deep neural network.

• Ano de defesa: 2022
• Autor(a) principal: Esteves, Fillipe Rocha Leonel
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado computacional; Deep neural networks; Estruturas offshore flutuantes; Inteligência artificial; Offshore design; Redes neurais; Response surface; Superfícies de resposta
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/3/3135/tde-22052023-110138/

The Deep Neural Network (DNN) is a machine learning algorithm that principle is to concatenate nonlinear operations involving matrices. These artificial networks can achieve reasonable transformations of input to output data by updating a matrix of randomly initialized weights. It is necessary to provide a training dataset to minimize a loss function during the network training. Validation and test procedures guarantee the quality of the trained network. The offshore design requires complex modeling that reflects the nature of the ocean environment. To produce a mapping of the hydrodynamic response of the offshore system, an extensive volume of simulations is often necessary, which elevates the computational cost of the design process. At this point, the opportunity to converge the deep learning potentialities and the challenges of offshore design emerges. This work proposes a framework to assess deep neural networks used as response surfaces of the semi-submersible platform dynamic models in waves: a mass-spring-damper model and an analytical hydrodynamic model validated with reference data. The low computation cost of these models allowed the generation of large datasets. The N-dimensional response hypersurface in each case is a combination of input parameters. An appropriate study elucidated the correct parameters definition of the DNN: the number of layers and the number of neurons per layer, targeting the configuration that provides the minimum mean squared error. The response surface represented by the DNN can easily be coupled to an optimization algorithm that evaluates hundreds of viable solutions and finds the optimal design. Using neural networks as a response surface has excellent cost-benefit in preliminary design dynamic modeling, in cases where the available time before the optimization tasks is long enough to prepare a training dataset, and in cases subjected to requisites updates throughout the conceptual design phase.