The coefficient of thermal expansion (CTE) significantly influences the performance of concrete. However, CTE measurements are both time consuming and expensive; therefore, CTE is often predicted from empirical equations based on historical data and concrete composition. In this work we demonstrate the application of linear regression and random forest machine learning methods to predict CTE and other properties from a database of Wisconsin concrete mixes. The random forest model accuracy, as assessed by cross-validation, is found to be significantly better than the American Association of State Highway and Transportation Officials (AASHTO) recommended prediction methods for CTE, denoted as level-2 and level-3.

Introduction

The objective of this study was to demonstrate the usefulness of using machine learning in predicting a range of concrete properties with an emphasis on concrete coefficient of thermal expansion (CTE). The ability to predict concrete properties using other indicator properties or descriptors can save time and costs related to making and testing materials. While prediction of concrete strengths has been extensively studied using both non-machinelearning and machine-learning techniques [1–۵], there have been fewer studies on prediction of concrete coefficient of thermal expansion (CTE), and none exploring machine learning methods for this property. Concrete CTE is an important input in pavement design, as detailed in the American Association of State Highway and Transportation Officials’ (AASHTO) Mechanistic-Empirical Pavement Design Guide (MEPDG) [6], and has significant effects on slab cracking, joint faulting, and surface roughness [7]. The AASHTO’s MEPDG describes three levels of design input for concrete CTE. Level-1 input is site- or project-specific, and requires testing for concrete CTE using the same materials that would be used for a specific paving project.