Comparison of machine learning techniques to predict the compressive strength of concrete and considerations on model generalization

Abstract The compressive strength of concrete is an essential property to ensure the safety a structure. However, estimating this value usually laborious and uncertain process since mix design based on empirical methods its confirmation in laboratory demands time resources. In context, work aims evaluate Machine Learning (ML) models predict from constituents. For purpose, dataset literature was used as input four ML models: Extreme Gradient Boosting (XGBoost), Support Vector Regression (SVR), Artificial Neural Networks (ANN) Gaussian Process (GPR). accuracy evaluated through 10-fold cross-validation, quantified by R2, Mean Absolute Error (MAE), Root-Mean-Square (RMSE) metrics. Subsequently, new put together with mixtures validate previous models. model creation step, all algorithms obtained similar positive results, MAE between 1.96-2.26 MPa R2 varying 0.79 0.83. validation dropped sharply, growing 3.04-4.04 decreasing 0.37-0.59. ANN GPR showed best while SVR had worst predictions. This that tools are promising techniques concrete. care must be taken data guarantee not overfitted given region, set materials, or type