Comparison among four techniques to predict the compressive strength of concrete: Extreme Gradient Boosting, Support Vector Regression, Artificial Neural Networks, and Gaussian Process Regression

Resumo

The compressive strength (Rc) of concrete is an important feature that influences the safety, durability,
and cost of a structure. To achieve the desired Rc, professionals generally use mix design methods based on empirical
tables. Then, the Rc must be confirmed in laboratory with tests that cost time and resources. To mitigate this issue,
this study proposes and compares the use of four Machine Learning (ML) techniques to predict the Rc of concretes
from their components. The techniques are: Extreme Gradient Boosting, Support Vector Regression, Artificial Neural
Networks, and Gaussian Process Regression. Initially, a dataset vastly used in the literature for this purpose was used
as input. Secondly, a dataset built by the authors was used to validate the models’ generalization ability. All models
were cross-validated (10-fold) and their accuracies were measured by R2, MAE, and RMSE. XGBoost and GPR
presented the best performance, while SVR presented the worst. Despite the positive performances measured in all
models with the first dataset, the metrics dropped sharply in the validation step involving the second dataset. Thus,
the ML techniques are promising tools for the mix design of concretes, but attention must be taken to guarantee that
models are not overfitted because of the homogeneity of the input data.