Precision assessment of supervised machine learning models and parametric optimization for predicting compressive strength of lightweight cellular concrete

Lightweight cellular concrete (LCC) is a cost-effective, sustainable material that can be used in a wide range of construction applications to improve performance and efficiency. In this study, LCC's compressive strength (CS) was assessed and optimized via metaheuristic-optimized machine learning (ML) modeling. Six ML algorithms decision tree (DT), categorial boosting (CatB), extreme gradient boosting (XGB), gradient boosting regressor, random forest, and light gradient boosting were trained and tested based on a dataset including 337 experimental data points from previously published studies. Based on the outcomes of this study, the CatB model outperformed other algorithms with an R 2 score of 0.968 in training and 0.887 for testing. Following CatB, the XGB and DT also showed satisfactory performance with an R 2 value greater than 0.90. In addition, feature importance analysis showed that the most influential features in the strength prediction were the density of concrete, fine aggregate and cement content. The partial dependency analysis revealed that, foam content below 20 kg/m 3 exhibited volatile tends to have an unstable strength response. However, beyond 30 kg/m 3 of foam in mix showed reduction of CS signifying that foam content should be controlled to avoid undesirable air being trapped into the mix. Furthermore, the response surface modeling (RSM) highlighted that the CS is optimized at moderate foam dosage (about 20-40 kg/m 3 ) together with higher cement, density, fine aggregate and SCM levels. The primary contribution of this study is combining well-optimized ML modeling with advanced interpretability and RSM analysis. This is further supported by the development of a graphical user interface (GUI) that enables real-time strength prediction and parameter optimization for practical engineering applications.