Uso de machine learning para predição da resistência a compressão do concreto

Abstract

Due to its properties such as compression strength, durability, and versatility, concrete has become widely used in the construction industry. Recently, machine learning techniques have stood out in predicting concrete compression strength, offering advantages such as considering multiple variables and identifying complex patterns in data. This work aims to analyze machine learning techniques as a model for analyzing predictive variables and their influence on concrete compression strength. Using a dataset with 1234 compression strength values, 8 and 6 predictive variables were analyzed, selected based on their relevance in SelectKBest. Linear correlation studies were conducted using simple linear regression and non-linear correlation studies using Support Vector Regression, Gradient Boosting, and Artificial Neural Networks. The results obtained were as follows: Support Vector Regression with a coefficient of determination of 0.85 and mean squared error of 30.9051 MPa; Gradient Boosting with a coefficient of determination of 0.90 and mean squared error of 25.5979 MPa; Artificial Neural Networks with a coefficient of determination of 0.87 and mean squared error of 5.781 MPa. The comparison between machine learning methods such as Support Vector Regression, Gradient Boosting, and ANN revealed distinctions between the models. Gradient Boosting achieved a higher coefficient of determination, demonstrating its ability to explain variability in the data. On the other hand, Artificial Neural Networks presented the lowest mean squared error, indicating accuracy in predictions. The choice between these approaches implies considerations regarding the balance between explainability and accuracy, in which Artificial Neural Networks achieved the most satisfactory performance among the models analyzed.