Banca de QUALIFICAÇÃO: JOAO PAULO DE FREITAS GAMA

Uma banca de QUALIFICAÇÃO de MESTRADO foi cadastrada pelo programa.
STUDENT : JOAO PAULO DE FREITAS GAMA
DATE: 08/04/2026
TIME: 09:00
LOCAL: https://meet.google.com/okk-vibt-vyq
TITLE:

Analysis of the fatigue behavior of composite materials using modular neural networks: evaluation of training algorithms

KEY WORDS:

Composite Materials, Artificial Neural Networks, Fatigue, Levenberg-Marquardt.

 

PAGES: 75
BIG AREA: Engenharias
AREA: Engenharia Mecânica
SUBÁREA: Mecânica dos Sólidos
SPECIALTY: Análise de Tensões
SUMMARY:

The modeling of fatigue life in polymer composite materials represents a significant challenge due to the highly nonlinear mechanical behavior under different load ratios. This study evaluates the application of a Modular Neural Network architecture for predicting the fatigue strength of glass fiber- and carbon fiber-reinforced laminates. Different training configurations were investigated, considering the Backpropagation (BP) and Levenberg–Marquardt (LM) algorithms, as well as two weight initialization strategies: conventional random initialization and the Glorot method. Model performance was assessed using cross-validation, employing Mean Squared Error (MSE) and BMEF as evaluation metrics, in addition to repeatability analysis. The results showed that the Glorot initialization significantly reduced the dispersion of minimum error values, increasing solution stability. The BP algorithm exhibited greater sensitivity to noise when using periteration weight updates, a behavior that was mitigated through batch training. Overall, the BP combined with Glorot initialization provided the best global performance, yielding the lowest MSE and BMEF values for most of the analyzed materials. The Goodman diagrams generated from the best-performing architectures demonstrated strong agreement between predicted and experimental curves, with larger deviations observed only under severe loading conditions. The results indicate that the Modular Neural Network presents strong generalization capability and constitutes a promising tool for fatigue life prediction of polymer composite materials.

COMMITTEE MEMBERS:
Presidente - 1338331 - RAIMUNDO CARLOS SILVERIO FREIRE JUNIOR
Externo ao Programa - 1086657 - JOELTON FONSECA BARBOSA - UFRNExterno à Instituição - BRUNO DA CUNHA DINIZ - UFBA