Integrating Machine Learning and Symbolic Regression for Transparent Prediction of Ultra-High Performance Concrete Strength

Abstract

Ultra-High-Performance Concrete (UHPC) is characterized in modern high performance building materials as being extremely heavy-duty safe and optimized mix design. However, the interdependence of its ingredients is very nonlinear, which makes the accurate prediction of compressive strength a complicated task. The current research presents a combination of Machine Learning (ML) algorithms with a Symbolic Regression as a means to predict the compressive strength of UHPC based on 810 samples of a reliable open dataset. To increase generalizability, 5-fold stratified cross-validation was used to train 7 ML algorithms including K-Nearest Neighbors, Decision Tree, Random Forest, Gradient Boosting, AdaBoost, Neural Network and Linear Regression. To tune each model hyper-parameters and assess both performance based on MAE, RMSE and R2 measure, Orange Data Mining software was used. To improve the explanation, the four models that had the highest accuracy, which is Gradient Boosting, AdaBoost, Neural Network, and Random Forest, were subsequently understood by means of Multi Expression Programming X (MEPX), to come up with explanatory equations. These equations as well as the equation of the original dataset have been tested through mathematical simplicity and practical engineering application. The results revealed that interpretability and predictive soundness could be presented in symbolic models, presenting a viable choice of engineering use. The study thus combines information-based intelligence and clear decision-making over the optimization of UHPC mix and makes it easier, cheaper, and more understandable to design materials in structural engineering disciplines.