Predicting Concrete Strength: using machine learning models

Concrete remains the most widely used construction material in the world, due to its ease of use, availability, and low cost. Its compressive strength is its key mechanical property and the primary performance criterion for structural design. An experimental database was used to compare the performance of several machine learning techniques: Lasso regression, quadratic polynomial regression, decision trees, support vector machines, artificial neural networks, random forests, bagging, gradient boosting, and Gaussian processes. Four covariance functions were considered for the GPR model. Systematic hyperparameter optimization was applied to each algorithm. A comparative analysis based on statistical indicators and graphical tools was conducted. A sensitivity analysis highlights the relative influence of the formulation parameters, emphasizing in particular the decisive role of cement content. Finally, an uncertainty propagation analysis reveals a significant dispersion in the predictions.