Prediction of Compressive Strength of UHPC Mix Designs Based on Mix Components Using Machine Learning Techniques

Abstract

Listen

Translate article

The high cost and limited availability of commercial ultra-high performance concrete (UHPC) products have motivated researchers and several state departments of transportation to develop non-proprietary UHPC mixes using locally available materials. Several non-proprietary mixes following different mixing and curing regimens have been developed in different geographic locations across the United States. Although these existing mixes can be utilized as starting points for developing suitable UHPC mix design for a project, fine-tuning localized mix designs is still a cost, labor, and time intensive task. A limited number of studies in the existing literature have utilized machine learning techniques to predict strength of inputted UHPC mixes. However, practitioners typically determine the concrete mix design after selecting the target performance criteria such as compressive strength, and workability for the project. To address this need, this study is a first step towards utilizing machine learning techniques for reverse-engineering UHPC mix proportions based on selected performance criteria and incorporating readily available materials. A database of 215 UHPC mixes extracted from 24 published manuscripts is compiled. This database is used to train ensemble learning algorithms such as Random Forest, Support Vector Machine (SVM), and Gradient Boosting. A comparative assessment on the most promising ML models trained with selected mix design parameters is presented. The tool developed in this study will be beneficial for practitioners to establish a locally sourced non-proprietary mix design of UHPC based on the required attributes.