Machine learning prediction of UHPC creep coefficient using optimized extremely randomized trees

Abstract

The increased mechanics, durability, and affordability of ultra-high-performance concrete (UHPC) with regular Coarse Aggregate (CA) make it promising. machine learning (ML) was used to study UHPC Creep. Using feature significance analysis, the Aggregate-To-Cement Ratio (AG/C), compressive strength at loading age (CSLA), Water-To-Binder Ratio (W/B), relative humidity (RH\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$RH$$\end{document}), Steel Fiber Volume Content (SF), Elastic Modulus at Loading Age (EMLA), and Curing Temperature (Tc) were the most essential characteristics. To estimate the Creep Coefficient (CC), this study attempts to demonstrate the potential effectiveness of ML techniques, such as the Extremely Randomized (ExT) model. Reliability is significantly affected by the ExT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ExT$$\end{document} Hyperparameters, which should be selected using metaheuristic optimization (MO) methods. This is achieved through a variety of techniques, including the Genghis Khan Shark (GKS) and manta ray foraging (MRF). The 338 data points on the (CC-UHPC) were gathered from scholarly publications. After evaluating each component, the study concluded that 75% (254) of the dataset would be utilized for learning and 25% (84) for assessing. The ExT(GKS) generated lower U95\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${U}_{95}$$\end{document} measure ratings throughout the Learning Phases and Evaluation Phases, with values of 0.0411 and 0.0364, respectively. With scores of 0.0591 and 0.077, accordingly, the LP and EP of the ExT(MRF) demonstrated a lower level of Dependability than the other results. The credibility and reliability of each model was assessed based on the Residual Value study, common sense and statistical measures that are widely accepted in statistics evaluation. The main contribution of this research is not to present a new algorithm, but to develop an integrated framework for UHPC creep modeling. In this framework, two methods, GKS and MRF, are compared under the same situations for ExT optimization. Also, SHAP examination is used to interpret the effect of input agents and relate the model results to the creep behavior of UHPC.