A multi-output machine learning framework for predicting rheological and mechanical properties of fresh concrete

The fresh concrete rheology is a key factor in workability, pumpability, and construction performance. An accurate prediction of the rheological properties and engineering performance indicators can help to optimize the mixture design with a minimum number of experimental tests. The purpose of this study is to present an integrated evolutionary machine learning (ML) approach for joint prediction of rheological properties of fresh concrete, yield stress, and plastic viscosity, and engineering-derived properties, namely, elastic modulus, compressive strength, and tensile strength from the parameters of concrete mixture design. The Gradient Boosting Machine (GBM), Light GBM (LGBM) and K-Nearest Neighbors (KNN) have been optimized using the Pufferfish Optimization Algorithm (POA). Model performance was analyzed using several statistical metrics and feature importance, and SHAP analyses were used to increase the interpretability of the model. Among the investigated models, GBM achieved the best overall predictive performance on the testing dataset, yielding Coefficient of Determination ( R 2 ) values of 0.9377, 0.9082, 0.9050, 0.8991, and 0.9088 for yield stress, plastic viscosity, elastic modulus, compressive strength, and tensile strength, respectively. The corresponding Root Mean Square Error (RMSE) values were 5.0873, 3.3374, 443.57, 1.2565, and 0.0087, with Prediction Coverage Degree (PCD) values ranging from 0.8571 to 0.9286, demonstrating robust prediction accuracy and generalization capability across all output variables. The proposed framework offers an efficient and comprehensible data-driven solution for initial performance evaluation and specific mixture design, aiding in the reliable decision-making process within civil engineering applications.