A comparative study of artificial intelligent methods for explosive spalling diagnosis of hybrid fiber-reinforced ultra-high-performance concrete

Abstract

Explosive spalling is a major concern to ultra-high-performance concrete at elevated temperatures. Previous physics-based numerical models for predicting explosive spalling of concrete are not validated sufficiently and still impractical for industrial applications. In this work, six machine learning models are developed to assess the thermal spalling risk of UHPC with hybrid polypropylene fibers and steel fibers, as well as to determine the minimal dosage of fibers required to eliminate spalling. Among the six models, five models are based on artificial neural network, support vector machine, decision tree, random forest, and extreme gradient boosting, respectively. Furthermore, a voting ensemble model is proposed based on the five individual machine learning models. To test the effectiveness of these six machine learning models, 36 groups of heating tests are conducted on hybrid fiber-reinforced UHPC specimens. The results show that among the six models, the XGBoost model shows the best performance with an accuracy of 97.2% and F1 score of 0.933. Parametric analyses are performed using the XGBoost model to study the influences of various parameters on the minimal dosage of fibers to prevent spalling. According to the analysis, PP fibers play a primary role in preventing explosive spalling of UHPC, and limiting the silica fume content reduces the minimal PP fiber dosage for spalling prevention. As the silica fume/binder ratio decreases from 0.25 to 0.05, the minimal PP fiber dosage decreases from 3.5 to 0.5 kg/m3.