Integração algoritmo genético com machine learning para design de ligas de alta entropia com propriedades mecânicas otimizadas

• Ano de defesa: 2025
• Autor(a) principal: Binde Stoco, Caroline
• Orientador(a): Coury, Francisco Gil
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência e Engenharia de Materiais - PPGCEM
• Palavras-chave em Português: Ligas de alta entropia; Algoritmo genético; Método CALPHAD; Propriedades mecânicas
• Palavras-chave em Inglês: High entropy alloys; Genetic algorithm; Machine learning; CALPHAD Method; Mechanical properties
• Área do conhecimento CNPq: ENGENHARIAS::ENGENHARIA DE MATERIAIS E METALURGICA::METALURGIA FISICA
• Link de acesso: https://hdl.handle.net/20.500.14289/22002

The rapid pace of technological advancement is driving the need for the development of novel materials that can meet the evolving demands across various service sectors. In this context, High Entropy Alloys (HEAs) have emerged as a promising solution. However, a significant challenge has been the selection of optimal compositions within a vast and complex multicompositional space. To address this challenge, this study developed a genetic algorithm capable of designing HEAs by optimizing multiple, potentially antagonistic, objectives. Through processes of genetic selection, crossover and mutation, the algorithm generated new alloy generations, progressively aligning their properties with the desired parameters. The optimization process aimed to achieve a single-phase face-centered cubic (FCC) structure, assessed through the integration of the CALPHAD method with machine learning techniques for classification using Support Vector Machines (SVM) and active learning. Additionally, the algorithm sought to maximize the Hall-Petch constant (K) and the critical resolved shear stress (τ_Y), enhancing mechanical strength through grain refinement and solid solution strengthening. These parameters were evaluated using empirical equations. The effects of twinning-induced plasticity (TWIP) and transformation-induced plasticity (TRIP) were also incorporated into the algorithm, with stacking fault energy (SFE) predictions made using Support Vector Regression (SVR). The final output of this genetic algorithm is a set of optimized HEA compositions, including two selected for future experimental analysis, demonstrating the genetic algorithm effectiveness in navigating the multicompositional space and addressing conflicting design objectives.