A simplified machine learning empirical model for biomimetic crack healing of bio-inspired concrete

Abstract

Bio-inspired self-healing is an effective technique to extend the durability and longevity of concrete infrastructures. However, factors influencing biomimetic crack healing need to be understood for effective strategies. This study aims to develop an interpretable empirical prediction model for self-healing concrete using Multivariate linear regression (MVLR). Other machine learning (ML) techniques like deep neural networks (DNN), extreme gradient boosting (XGBoost), and convolutional neural network (CNN) were also employed for comparison. Although the XGBoost model performed best with R2 of 0.95 and the least statistical errors, however, MVLR offered a more interpretable approach. The developed empirical formulation revealed that immobilizer and bacterial strain had the strongest impact on biomimetic crack healing, followed by initial crack width and healing duration. The formulated mathematical expression was moreover experimentally validated, yielding minimal Mean Squared Error and Root Mean Squared Error values of 0.0057 and 0.0588, respectively. Thus, these findings can be utilized for the effective implementation of bio-inspired concrete in the field of construction.