Uso de ferramentas de inteligência artificial para previsão de propriedades de resistência à compressão de concretos de alto desempenho e de ultra alto desempenho

Abstract

Advances in concrete production are accelerating, making it possible to make more resistant concretes. High performance concrete (HPC) is already part of the daily life of the main concrete companies, as expected for ultra-high performance concrete (UHPC) in the future. However, its dosage methods did not follow the progress in its development, demanding time and resources in the formulation processes. Since most dosage methods use graphs and tables with limited predictive capacity for new concretes, these require a succession of mix adjustments. In this way, it is necessary to create new tools capable to analyse the influence of the materials in the mixture and predict the behaviour of concrete from a given composition. For this task, machine learning models are potential options, as they have been applied in the most diverse areas for prediction from complex sets of input variables. Health, economics and civil engineering are examples of those that already frequently use these models. In civil engineering, a relevant problem is precisely that of predicting the behaviour of the most diverse types of concrete based on their composition. The most studied characteristics are strength, workability and durability. HPC and UHPC in general are very fluid and the literature does not always detail the behavior in the fresh state, so the focus of the present work was the compressive strength. Therefore, this work developed compressive strength prediction tools for both databases updated by this research from studies published in the literature. In the preliminary exploratory analysis, it was found that these two types of concrete have great differences and a joint analysis is not recommended. Separately, the correlations and influences of the constituent materials on the compressive strength were analysed. For the creation of the prediction tool, Linear Regression, Regression Tree, Random Forest and Artificial Neural Network (ANN) models were tested with the cross-validation process. The ANN model presented better prediction performance with a R² values equal to 0.85 for HPC and 0.86 for UHPC. The best models were then analysed by its sensitivity to investigate the interpretation of concrete behaviour as a function of its composition and compare with the expected behaviour according to the literature. The models showed good prediction performance and adequate physical interpretation.