Avaliação de parâmetros de ensaios de compactação a partir de propriedades índices do solo usando técnicas de machine learning

2026-10-202026-10-202026-06-09LEÃO, Max Cantão de. Avaliação de parâmetros de ensaios de compactação a partir de propriedades índices do solo usando técnicas de machine learning. Orientador: Gérson Jacques Miranda dos Anjos. 2026. 49 f. Trabalho de Curso (Bacharelado em Engenharia Civil) – Faculdade de Engenharia Civil, Instituto de Tecnologia, Universidade Federal do Pará, Belém, 2026. Disponível em: https://bdm.ufpa.br/handle/prefix/10398. Acesso em:.https://bdm.ufpa.br/handle/prefix/10398The determination of soil compaction parameters, specifically optimum moisture content (wót) and maximum dry unit weight (γd,máx), through the Proctor test is crucial for technological control in earthworks and paving. However, this laboratory method is costly and time-consuming, posing challenges in large-scale projects. Therefore, this study proposes an innovative predictive model, based on the Extreme Gradient Boosting (XGBoost) machine learning technique, to estimate these parameters from easily obtainable soil index properties, aiming for greater agility and efficiency. The choice of XGBoost is justified by its ability to model complex non-linear relationships and multidimensional interactions between variables, overcoming the limitations of traditional statistical models. For its development, a robust consolidated database of 2,730 soil samples was used (74% Brazilian and 26% international), covering Atterberg limits (LL and PL) and grain size composition (coarse and fine fractions). The methodology followed a rigorous data science workflow in Python, including data preparation with DMatrix, hyperparameter tuning via grid search, and 5-fold cross-validation, ensuring the model's robustness and generalization, while mitigating the risk of overfitting. The results demonstrated high predictive performance of the XGBoost model. For (γd,máx), the coefficient of determination (R2) reached 0.870, while for (wót), the (R2) was 0.891. These values indicate that the model explains most of the data variability with a reduced error margin, providing high reliability. Feature importance analysis revealed that the Plastic Limit (PL) and Liquid Limit (LL) are the most influential predictors, corroborating the fundamental principles of soil mechanics. It is concluded that XGBoost is a promising and effective tool for pre-design and quality control in geotechnical engineering, optimizing processes, reducing operational costs, and aligning geotechnics with modern artificial intelligence approaches.Acesso AbertoCompactação do SoloParâmetros de compactaçãoMachine LearningXGBoostGeotecniaEngenharia CivilModelagem PreditivaSoil CompactionCompaction ParametersMachine LearningGeotechnicsCivil EngineeringPredictive ModelingCNPQ::ENGENHARIAS::ENGENHARIA CIVILAvaliação de parâmetros de ensaios de compactação a partir de propriedades índices do solo usando técnicas de machine learningEvaluation of compaction test parameters from soil index properties using machine learning techniquesTrabalho de Curso - Graduação - MonografiaAttribution-NonCommercial-NoDerivs 3.0 Brazil