Estudo numérico do problema de arrancamento direto utilizando redes neurais

Abstract

The application of computational simulation using Finite Element Methods (FEM) has intensified in contemporary engineering problem-solving. By understanding the input data and boundary conditions of the physical problem, it is possible to use a combination of computational techniques to optimize the solution. Optimization is a numerical resource that can be applied to solving an inverse problem. Deep learning is a subset of artificial intelligence aimed at learning specific tasks and improving performance with each iteration. It is also known for being a neural network that simulates the human neuron, due to the use of artificial neurons connected by artificial synapses.The goal of this work is to apply deep learning techniques to optimize engineering parameters in a problem involving the adhesion between steel and concrete. The methodology involves numerically solving the physical problem using FEM with the ANSYS solver, and performing optimization in MATLAB. The research investigates the application of deep neural networks to adjust parameters and minimize discrepancies between simulations and experimental data, focusing on the pull-out system and steel-concrete adhesion. The ReLU function in MATLAB was used to introduce non-linearities and enhance the precision of the results. The methodology was validated based on experimental results from pull-out tests on steel-concrete bars.