Dissertação | Aeronáutica | Artificial Intelligence-Based Prediction of Instability in Stiffened Panel Structures Using Linear Analysis and GFEM

Company Description

Somos um Centro de Engenharia e Desenvolvimento de Produto que concebe, desenvolve e produz novas tecnologias, produtos e serviços para uma sociedade mais sustentável.

Trabalhamos fazendo a ponte entre cidades, indústria e universidades em torno de grandes programas de desenvolvimento e industrialização de produtos e serviços, a partir de Portugal, em sectores de alta tecnologia como a mobilidade, a aeronáutica e o espaço.

Esta oportunidade é na àrea de aeronáutica.

Job Description

Description:
This project aims to investigate and compare the efficacy of Artificial Neural Networks (ANNs) in predicting structural instability using solely results from conventional linear analysis methods and information from the FE mesh (using global FE model modeling assumptions).

Objectives:
To explore and implement a conventional artificial intelligence model in the task of predicting structural instability of stiffened panels, examining only the internal forces results obtained with a conventional linear analysis over the panel (GFEM model).
To develop and train such AI model using the simplest possible Deep Neural Network architecture, to accomplish the task of reasonably predicting a panel's buckling under controlled loading state in order to reproduce and perhaps match the accuracy of analytical methods;
Use such a model to fill some gaps and/or extend the analytical methods, as a proof of the importance of such an approach. For example, extrapolating the limit of applicability of already covered variables or adding new ones, such as new geometric features, boundary conditions, or a more generic or combined loading. Corroborate the predicted results using SOL105.

Project Description:
To comprehensively study the instability phenomena of stiffened panel structures. Define the simplification hypothesis and the adequate assumptions on loading and geometry simplification (following the analytical method test base);
To study the DNNs capabilities and required architecture (input tokenization, number of neurons, layer architectures, activation functions, etc.) in identifying patterns and capturing the complex relationships between input data (geometric tokens and tokenized linear analysis results) and the desired output (buckling percentage prediction);
Generate datasets to train the NN: Generate a collection of FE models and analyze them (under SOL 101 and 105) for a range of load cases. Utilize geometric and load data extraction/simplification for the collection, along with their analytical instability results, to train the NN. Segregate models and load cases for training and testing;
Create a simple tool for entering the parameterized panel and internal loads, feed the trained NN, and predict the panel failure along with the critical load level (Equivalent to the SOL 105’s eigenvalue). Use this tool to generate curves and compare them with the literature.

Qualifications

Licenciatura em Aeronautica ou Aerospacial

NASTRAN SOL 105