Experimental study and machine learning prediction on compressive strength of spontaneous-combustion coal gangue aggregate concrete

Abstract

Spontaneous-combustion coal gangue aggregate (SCGA) has low physical properties and strong water absorption capacity. Previous studies are mostly based on the principle of equal mix proportion, which makes the slump of spontaneous-combustion coal gangue aggregate concrete (SCGAC) lower than that of natural aggregate concrete (NAC). This study obtained the compressive strength of SCGAC at different curing ages under the premise of similar concrete workability, and quantified the effect of SCGA content on the compressive strength of concrete; the meso-structure of SCGAC was obtained by SEM test to reveal the deterioration mechanism of SCGAC compressive performance; based on five independent machine learning (ML) algorithms and three ensemble ML algorithms, SCGAC compressive strength prediction models were proposed, and the prediction performance of each model was evaluated to quantify the importance of each influencing factor on the concrete compressive strength. The results show that by using pre-soaked SCGA and refining the concrete mix proportion, fresh concrete with different SCGA replacement ratios can exhibit similar slump values (70–75 mm); due to the weak interfacial transition zones (ITZs) between cement mortar and SCGA, and the inferior properties of SCGA, the compressive strengths of concrete with 100% replacement ratio at 7 d, 14 d, 28 d and 90 d were reduced by 15.3%, 16.7%, 22.0% and 21.6% respectively compared with NAC; the predictive ability of the ensemble ML models was higher than that of the independent ML models, and the XGB model had the best predictive ability; based on the feature importance analysis of SHAP value, it was found that SCGA density, aggregate-cement ratio and SCGA water absorption were the most important factors affecting the compressive strength of SCGAC.