Avaliação de modelos de aprendizado de máquina para estimativas da resistência em concreto autoadensáveis

Abstract

The process of validation of the concrete dosage also depends on laboratorial tests that confirm the characteristics and strengths initially estimated. As it is an experimental method, this process tends to be onerous and laborious, with possible adjustments by trial and error. Thus, recent researches shows applications of technologies that optimize this process, using technologies such as machine learning, with the potential to reduce and optimize the workflow. In this context, this paper evaluated the performance of different learning models for estimating and predicting the compressive strength of self-compacting concrete, investigating the influence of basic materials (cement, aggregates, water) as additives and additions (plasticizers, pozzolans). For this, a database was created with 503 dosages of self-compacting concrete, taken from the literature, with Linear, Lasso, and Ridge Regression models being implemented, and Decision Tree and Random Forest regressors, in addition to more simplified models of neural network, as MLP Regressor and Polynomial Regressor, being these models the most referred in the literature for problems of this nature. Prior to implementation, an exploratory data analysis (AED) was performed, where a correlation map could be generated to guide the evaluation of the generated models. It was possible to verify not only correlations between the components of the self-compacting concrete, but other factors such as precision, standard deviation and computational expense of each implemented model.