Metamodelo com arquitetura em duas camadas para a previsão de propriedades mecênicas de misturas de concreto

Abstract

The mechanical properties of concrete, such as modulus of resilience and compressive strength, are of great importance in the design stage. Determining these properties through information about the mix components and physical characteristics of the concrete is not trivial. This task becomes complex because the relationships between the quantities and types of concrete mix components and their mechanical properties have a nonlinearities. The objective of the present work is to employ machine learning methods to create regression models that are able to predict the mechanical properties of a concrete mix. In order to investigate the possibility of improving the results already obtained with the use of machine learning techniques, a joint learning technique, Stacking, will be used. This technique consists of a layer model, where the first is composed of individual models and the second by a metamodel that has the function of using the predictions made by the first layer models as input data, aiming to make a final forecast. Four different experimental databases obtained from the literature were used, each database having different samples, number and types of attributes, allowing to verify the generalizability of the created architecture. With the use of this technique it is intended that the metamodel presents a better predictive capacity than the one presented by the individual methods. The entire computational framework was implemented using the Python language and a set of libraries. Statistical tests were used to assess whether there was a significant difference between Stacking performance and individual models. At the end of the analysis of the results of the evaluation metrics and the statistical tests it was possible to realize that the results of Stacking are better or at least similar to those presented by the best performing individual models.