Investigation of Multi-Output Regression Modeling in Predicting Concrete Mix Design

Abstract

Current research is a piece of an innovative approach to concrete mix-design prediction by implementing advanced regression techniques, that addressed the limitations of traditional methods in IS 10262 and ACI 318 standards which rely heavily on empirical relationships and require extensive trial batching. The study investigates eight important mix-design parameters namely water-cementratio, cement content, flyash content, fine aggregate content, 10 mm and 20 mm aggregate content, water content, and superplasticizer content. The methodology utilizes comprehensive Multioutput Regression with gradient boosting and decision tree regressors, chosen, for their ability to capture complex non-linear relationships between material properties and mix-design proportions that traditional methods often oversimplify. Through k-cross validation using a 70-30 train-test split on a dataset of 180 actual laboratory samples, the Multioutput Regression achieved a coefficient of determination (R-squared) of 0.99, significantly outperforming both the Decision Tree Regressor (R-squared: 0.89), and traditional design methods. While traditional methods exhibit 15-20% prediction errors, the current model reduced this error margin to 3-5%, leading to potential material cost savings of 8-12% and reducing trial batching by up to 30%. Furthermore, Mean Squared Error was calculated across all predicted parameters which helped in verifying the model’s robustness and good performance in aspects like water-cement ratio (MSE: 0.051) and cement content (MSE: 1:52). Unlike conventionally used methods which require multiple trial and errors of mix-design, the current method simultaneously optimizes all mix-designingredients while taking into account 27 distinct material properties offering a more efficient and accurate design process.