Interpretable auto-tune machine learning prediction of strength and flow properties for self-compacting concrete

Abstract

Self-compacting concrete (SCC) is a promising building material, particularly in developed countries, due to its excellent flowability achieved through a relatively high binder content. To reduce cement usage and carbon footprint, supplementary cementitious materials (SCMs) such as fly ash and limestone powder are utilized. However, traditional concrete design relies on empirical trial-and-error methods, which pose challenges when dealing with SCC due to conflicting mechanical and workability requirements. Thus, leveraging the value of existing data, particularly with the aid of machine learning (ML) techniques, becomes essential. The purpose of this study is to establish highly nonlinear relationships between SCC mix proportions and properties that cannot be expressed by explicit mathematical equations. To address these issues, this study developed interpretable ML technologies, to predict 28-day compressive strength (SC) and slump flow (SF) of SCC containing SCMs, aiming to guide SCC design. The proposed models exhibit high accuracy in predicting the key properties of SCC through careful feature engineering, model training, and hyper-parameter optimization. Experimental verification shows that the proposed models can accurately predict the properties of C45, C50, and C55 SCC, with prediction errors of 0.6%, 6.8%, −13.5% for 28-day SC and −2.8%, 3%, 5.4% for SF, respectively. The combination of ML and SHAP interpretable algorithm provides physical rationality, allowing engineers to predict and design SCC by adjusting proportions according to the parameter analysis.