Desenvolvimento de algoritmos genéticos acoplados à técnicas de machine learning para otimização de geometrias de clusters atômicos e/ou moleculares com ênfase em metodologias ab initio

• Ano de defesa: 2024
• Autor(a) principal: Umar Lucio Esper Mucelli Rezende
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; ICX - DEPARTAMENTO DE QUÍMICA; Programa de Pós-Graduação em Química; UFMG
• Palavras-chave em Português: Clusters; Ab initio; Algoritmo genético; Operadores genéticos; Otimização; Machine learning; Físico-química; Algoritmos genéticos; Funcionais de densidade; Aprendizado do computador; Redes neurais (Computação); Química quântica
• Link de acesso: http://hdl.handle.net/1843/72178; https://orcid.org/0009-0002-1795-2326

Clusters are small aggregates of particles that can exhibit vastly different properties depending on their size, composition, and other characteristics ranging from resembling individual atoms or molecules to properties observed in the bulk limit of the material. Due to this differentiation in properties, clusters can be considered as a new class of materials, attracting significant interest within the scientific field. To study clusters, it is crucial to find or determine the geometry, or conformation, that they would assume in nature under the conditions they are being studied. To determine the geometry of these clusters, genetic algorithms have been implemented and developed over the past decades, proving to be essential tools, although still far from perfect in solving this type of problem. In this work, a new genetic algorithm, the NQGA (New Quantum Genetic Algorithm), along with a set of new genetic operators, is proposed to be a more efficient tool in locating the global minimum of potential energy surface for atomic and molecular clusters. The doppelgänger predator (DGP) and Machine Learning Prediction (MLP) operators stand out as the main contributors to drastically reduce the number of samples required from the quantum energy surface, enabling more efficient structure optimizations. The NQGA is capable of performing structure optimization using classical energy calculations (parametrized interatomic potentials) or directly through ab initio methods, by being coupled with the well developed quantum packages GAMESS-US and ORCA. The NQGAmethodology was validated for the classical method using clusters of copper, gold, and copper-gold and gold-silver nanoalloys for different cases, and in all cases, it yielded coherent results with those found in the literature. Small clusters of lithium were studied using the CCSD(T) methodology to test the NQGA’s ability to operate in high-level ab initio methods. The NQGA demonstrated great capability to find the geometry of minimum energy of molecular clusters, as in the case of the (H2O)11 cluster, and achieved improved results for the global minimum of the Mg6H4 cluster.