Aplicação de algoritmos de aprendizagem de máquinas para previsão de abatimento de tronco de cone e resistência à compressão de concretos convencionais

Abstract

Over the last decades, machine learning techniques have been an important subject in concrete research since they achieve an excellent performance in predicting concrete properties. In this regard, this study concentrates on the use of machine learning methods to forecast the Compressive Strength (CS) and slump of conventional concrete. Three distinct ensemble machine learning techniques were used – Artificial Neural Network (ANN), Random Forest (RF) and Support Vector Machine (SVM) – for regression and classification. These methods were evaluated with statistical metrics – Root Mean Squared Error, Mean Absolute Error, Accuracy, Precision, Recall and F1-score – using cross validation technique for hyperparameters tuning and model training. More than 2750 concrete mix proportions were collected from previously published research papers and compiled in two datasets. The inputs are the components contents of concrete, including cement, water, fine aggregate, coarse aggregate, fly ash, slag and plasticizer. In general, the models showed a similar performance in all simulations. ANN performed better in regressions to predict CS (R² = 0.796) and slump (R² = 0.692). Moreover, ANN achieved the best accuracy of 52% to predict the class of CS, and the SVM obtained the best accuracy of 76% in to predict the slump class. The variation in CS and slump were achieved by varying the components contents in a sensitivity analysis. The effect of each component content in the concrete properties was investigated by evaluating the models, and the observed patterns of dosage were in accordance with those described in the literature. Therefore, the proposed models, by a high agreement with scientific studies and good performance, could be used as efficient tools to determine mix proportions that meet requirements of CS and slump.