Improving Decision-Making in 3d Concrete Printing Through Shap-Guided Machine Learning: Predictive Models and Feature Importance for Yield Stress and Viscosity

Abstract

The accurate prediction of yield stress and viscosity plays a vital role in 3D concrete printing, necessitating the development of dependable predictive models to mitigate challenges and ensure correct printing operations. This study focuses on the development and evaluation of three machine learning (ML) models, namely the Gradient Booster Regressor (GBR) optimized with bayesSearchCV, Artificial Neural Network (ANN) optimized using GPyOpt library and Extreme Gradient Boosting (XGBoost), optimized using Hyperopt library. The primary objective was to predict the rheological properties of 3D concrete printing (3DCP) and examine the correlation between mixture design variables and their characteristics. Model performance was assessed using R2, MAE, and RMSE with 5-fold cross-validation. The SHAP technique, which leverages cooperative game theory to estimate rheological properties and feature importance ranks, was employed to provide interpretability to the results. The results show that the XGBoost-based model, coupled with missForest imputation and Hyperopt optimization, outperforms the other model. This forecasting tool combined with SHAP analysis, can predict the rheological properties of 3D printing and support decision-making processes considering influential variables and providing the opportunity to optimize the mixture while minimizing errors.

Keywords: Machine Learning, 3D Concrete printing, BayesSearchCV, Hyperopt, Shapley Additive exPlanations (SHAP), Static Yield Stress, Viscosity of the mixture