Banca de QUALIFICAÇÃO: João Paulo Silva Lima

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
DISCENTE : João Paulo Silva Lima
DATA : 31/03/2023
HORA: 14:00
LOCAL: Plataforma Teams
TÍTULO:

A NOVEL HYPERPARAMETER-OPTIMIZED BI-FIDELITY MULTI-TASK DEEP NEURAL NETWORK USING BAYESIAN OPTIMIZATION FOR STRUCTURAL RELIABILITY ASSESSMENT

PALAVRAS-CHAVES:

structural reliability, uncertainty quantification, neural networks, multi-fidelity 

PÁGINAS: 107
RESUMO:

This thesis presents a novel Bi-Fidelity Multi-task Learning Model based on a Deep Neural Network (BFMT-DNN) to address the computational challenge of structural reliability analysis applied to complex structures. The main contribution is developing a novel hyperparameteroptimized BFMT-DNN, considering the advantages of Bayesian Optimization, focusing on prediction accuracy, stability, and computational efficiency. For constructing this surrogate model, the study proceeded through three stages before. Firstly, a hyperparameter-optimized Multi-task Deep Neural Network (MT-DNN) is presented with an assessment realized in a stiffened panel under axial load and lateral pressure case to provide the ultimate strength prediction and the collapse mode classification. Following this approach, the problem can be transformed into a multi-objective optimization problem associating the trained MT-DNN to an adaptive Bayesian approach multi-objective evolutionary optimization algorithm to obtain a multi-objective optimization considering the weight vs. ultimate strength of the stiffened panel. The second step presented a method based on a Bi-fidelity Kriging model for structural reliability analysis. The bi-fidelity Kriging surrogate model's efficiency as a surrogate model will be assessed for a stiffened panel reliability problem that demands high computational costs, such as nonlinear finite element analysis structural models. The next step proposed a two-stage Bi-Fidelity Deep Neural Network surrogate model to quantify the uncertainty of structural analysis and assess the probability of failure prediction using the Multi-Fidelity Model surrogate for the performance function. In the three steps, the surrogate models can reproduce the nonlinear behavior in the variable's uncertainty analysis, reducing the high computational demand of these problems. Furthermore, the BF-DNN and MT-DNN surrogate models used Bayesian optimization to fine-tunning the hyperparameters. The multi-fidelity models used low-fidelity data samples added to the model to predict high-fidelity responses, and, when presenting a good correlation between the fidelities, the assessment of the proposed method showed that the proposed Multi-fidelity Method is a good strategy because it can provide an accurate probability of failure estimation with a lower computational cost. The results show that the proposed multi-fidelity methods can give a precise failure probability estimation with less computational cost.

MEMBROS DA BANCA:
Presidente - 2047809 - FRANCISCO EVANGELISTA JUNIOR
Interno - 404840 - WILLIAM TAYLOR MATIAS SILVA
Externa ao Programa - 1310273 - MARCELA RODRIGUES MACHADO
Externo à Instituição - MARCOS PAULINO RORIZ JUNIOR
Externo à Instituição - MAURO DE VASCONCELLOS REAL - FURG