Load-Induced Crack Prediction Using Hybrid Mechanistic Machine Learning Models

Abstract

This thesis implements hybrid mechanistic-machine learning models to predict load-induced cracking in concrete beams without transverse (applied along the depth of the beam) reinforcement. Predicting load-induced cracking is crucial for robustly predicting shear capacity. Mechanistic models lack the flexibility to represent load-induced cracking, and machine learning models lack a sufficient data set to learn load-induced cracking relationships. Hybrid models have the best chance of accurately predicting load-induced cracking. To implement hybrid modeling, we developed the Hybrid Learning theory and identified optimal combinations of mechanistic and machine learning models. Additionally, we developed a framework that has low mechanistic bias and sufficient constraint. This framework will allow for mechanistically consistent predictions. Hybrid models have great potential for modeling in Structural Engineering because of their flexibility and interpretability, and robust prediction of shear capacity will lead to increased design efficiency and understanding of concrete beam failure mechanics.