Transfer learning for intelligent design of lightweight Strain-Hardening Ultra-High-Performance Concrete (SH-UHPC)

Abstract

The design of lightweight Ultra-High-Performance Concrete (UHPC) requires pursuing superior material efficiency, which involves striking a delicate balance between ultra-high compressive strength and reduced material density. This paper compiled a comprehensive dataset of 176 ordinary UHPC and 72 lightweight UHPC and proposed a framework that integrates both transfer learning and Bayesian Optimization-enhanced Extreme Gradient Boosting (BO-XGBoost) for the design of lightweight UHPC. The BO-XGBoost model was pre-trained through hyperparameter tuning, laying a solid foundation for predicting material efficiency. Transfer learning was incorporated to address data limitations in lightweight UHPC while capturing its unique properties. The framework achieved 98.2 % accuracy in forward prediction and 94.7 % in reverse design. Notably, a lightweight UHPC using local materials was successfully developed and exhibited strain-hardening behavior based on the proposed approach, pushing the performance envelope of existing UHPC materials. This approach provided a solution for the design of lightweight strain-hardening UHPC towards superior material efficiency.