Metaheuristic optimized and explainable machine learning for predicting the compressive strength of lightweight foamed concrete
Abstract
Lightweight foamed concrete (LFC) can provide reduced structural dead load and enhanced thermal performance, however the compressive strength of LFC is determined by the complex nonlinear interactions between the mixture constituents, density and testing age, which makes the prediction and proportioning of the product difficult. This study addresses this challenge through leakage-resistant machine learning combined with metaheuristic optimization and explainable artificial intelligence. A literature-derived database of 993 observations was divided by mix-design group into 795 development records representing 276 groups and 198 unseen-test records representing 69 independent groups, with no overlap. Random forest (RF) and Extreme Gradient Boosting (XGB) models, together with their Leopard Seal Optimizer (LSO), Harris Hawks Optimizer (HHO), and Slime Mould Algorithm (SMA) optimized variants, were assessed using 10-fold grouped cross-validation and a completely held-out test set. SHAP and PDP-ICE analyses were subsequently applied to the best-generalizing model. LSO-XGB achieved the strongest grouped out-of-fold performance ((R ² = 0.888); RMSE = 3.93 MPa), whereas HHO-XGB performed best on unseen data ((R ² = 0.875); RMSE = 4.10 MPa), reducing the baseline-XGB RMSE by 4.1%. HHO-XGB also produced a median absolute error of 0.82 MPa, with 95.96% and 97.47% of predictions within 5 and 10 MPa, respectively. Binder was the dominant predictor, contributing 45.4% of the global SHAP importance, followed by measured density at 16.9%. The proposed framework within the represented data domain can support preliminary LFC mixture screening and guide locally validated, multi-objective experimental designs balancing strength, density, cost, and environmental impact.
Related
- Machine learning-based prediction of compressive strength of wastewater-exposed concrete using explainable and optimized models
- Explainable and uncertainty-calibrated stacked ensemble for predicting compressive strength of blended cementitious concrete
- Explainable Bayesian-optimized LSTM for predicting multivariate tool wear in tribological systems
- Interpretable metaheuristic-optimized machine learning for predicting biodiversity indices in the Hyrcanian forest
- β-PruNet: An Explainable, PSO-Optimized Degree-Centrality-Pruned Regularized Extreme Learning Machine for Lightweight Household Appliance Recognition.
BibTeX
@article{elsheikh2026metaheuristic,
title = {{Metaheuristic optimized and explainable machine learning for predicting the compressive strength of lightweight foamed concrete}},
author = {Ahmed Elsheikh and Muhammad Zeeshan Qureshi and Inamullah Inam and Hawreen Ahmed and Ali Ajwad and Abdelmalek H. Aljarah and Adham E Ragab},
journal = {Scientific Reports},
year = {2026},
month = oct,
doi = {10.1038/s41598-026-76016-8},
url = {https://www.nature.com/articles/s41598-026-76016-8},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Metaheuristic optimized and explainable machine learning for predicting the compressive strength of lightweight foamed concrete
AU - Ahmed Elsheikh
AU - Muhammad Zeeshan Qureshi
AU - Inamullah Inam
AU - Hawreen Ahmed
AU - Ali Ajwad
AU - Abdelmalek H. Aljarah
AU - Adham E Ragab
PY - 2026
DA - 2026/10/15
JO - Scientific Reports
DO - 10.1038/s41598-026-76016-8
UR - https://www.nature.com/articles/s41598-026-76016-8
AB - Lightweight foamed concrete (LFC) can provide reduced structural dead load and enhanced thermal performance, however the compressive strength of LFC is determined by the complex nonlinear interactions between the mixture constituents, density and testing age, which makes the prediction and proportioning of the product difficult. This study addresses this challenge through leakage-resistant machine learning combined with metaheuristic optimization and explainable artificial intelligence. A literature-derived database of 993 observations was divided by mix-design group into 795 development records representing 276 groups and 198 unseen-test records representing 69 independent groups, with no overlap. Random forest (RF) and Extreme Gradient Boosting (XGB) models, together with their Leopard Seal Optimizer (LSO), Harris Hawks Optimizer (HHO), and Slime Mould Algorithm (SMA) optimized variants, were assessed using 10-fold grouped cross-validation and a completely held-out test set. SHAP and PDP-ICE analyses were subsequently applied to the best-generalizing model. LSO-XGB achieved the strongest grouped out-of-fold performance ((R ² = 0.888); RMSE = 3.93 MPa), whereas HHO-XGB performed best on unseen data ((R ² = 0.875); RMSE = 4.10 MPa), reducing the baseline-XGB RMSE by 4.1%. HHO-XGB also produced a median absolute error of 0.82 MPa, with 95.96% and 97.47% of predictions within 5 and 10 MPa, respectively. Binder was the dominant predictor, contributing 45.4% of the global SHAP importance, followed by measured density at 16.9%. The proposed framework within the represented data domain can support preliminary LFC mixture screening and guide locally validated, multi-objective experimental designs balancing strength, density, cost, and environmental impact.
SN - 2045-2322
ER -