Transfer learning on stacked machine-learning model for predicting pull-out behavior of steel fibers from concrete
Torkan Shafighfard, Neda Asgarkhani, Farzin Kazemi, Doo‐Yeol Yoo
Abstract
Fiber-reinforced concrete (FRC) can enhance the concrete's resistance to cracking, impact, and fatigue, making it suitable for high-stress applications and infrastructures. Application of artificial intelligence in engineering problems can boost the design procedure while reducing the experimental cost and time. The aim of this study is to provide a transfer learning-based stacked machine learning (ML) model for predicting average and equivalent bond strengths , pull-out energy, maximum load, and corresponding slip capacity of steel fiber from concrete. A data-rich framework, including 472 data points, was constructed based on a comprehensive literature review. The overall results show that the fine-tuned bagging regression model provided the most accurate predictions, whereas artificial neural networks (ANNs) and gradient boosting ML models were not sufficiently reliable. To overcome the prediction shortcoming, experimental test was done to prepare unseen dataset for validation, and the transfer learning-based stacked ML model proposed to estimate all of the output targets in one model, with the highest accuracy (i.e., R 2 of 99.4 %) and the lowest error values (i.e., average error of 0.25 %) amongst those 18 ML models investigated in this research. Then, a graphical user interface (GUI) was provided to ease similar prospective analyses and obtain predictions, while proposing the update in the database for future studies. The proposed model paves the way for the prediction and evaluation of the fiber pull-out performance.