Bio-inspired machine learning for strength prediction of cement mortars incorporating reservoir waste silt: A case study from Bologna, Italy

Abstract

This study addresses the challenge of accurately predicting the mechanical performance of cement mortars incorporating waste silt, a byproduct from reservoir sedimentation, which has potential for sustainable construction applications. The aim is to develop predictive models for the flexural strength (FS) and unconfined compressive strength (UCS) of such mortars using advanced machine learning techniques. Two hybrid algorithms were applied: artificial neural networks optimized via the artificial bee colony algorithm (ABC-ANN) and the combinatorial group method of data handling (GMDH-Combi). Input parameters included the proportions of cement, water, sand, silt, and additives, while FS and UCS were treated as separate target outputs. The predictive performance of the proposed ABC-ANN and GMDH-Combi models was evaluated and compared against results from a previously studied design of experiments (DOE) methodology. The ABC-ANN model demonstrated superior predictive accuracy compared to the DOE method, achieving coefficients of determination (R2) values of 0.9948 and 0.9997 and mean absolute percentage error (MAPE) of 1.134 % and 0.319 % for FS and UCS, respectively, while the GMDH model yielded R2 values of 0.9362 and 0.9629 and MAPE of 6.706 % and 6.150 %. Sensitivity and parametric analyses indicated that water content had the greatest influence on strength predictions, whereas cement content had the least, confirming that both models effectively captured the experimental behavior of the mortars.