EVALUATION OF BOND STRENGTH IN CONCRETE BRIDGES WITH NOVEL MATERIALS THROUGH MACHINE LEARNING MODELS

Conventional reinforced concrete (RC) bridges are usually reinforced with conventional steel reinforcing bars. However, RC bridges are subjected to continuous deterioration due to various environmental agents that can affect their safety when some critical threshold is exceeded. Corrosion, in particular, has become a serious issue in existing bridge structures, limiting their service in many cases. Recently, there has been a significant interest in applying advanced composite materials in bridges with the aim of improving their durability. These materials have shown great advantages for developing RC structures, including the potential to achieve more durable structures with excellent seismic performance. Nevertheless, the detailed seismic behavior and failure mechanism of RC bridges with such materials are still not well defined. A good definition of the bond mechanism and strength, in particular, is critical to ensure the safety of RC bridges with novel materials. The novel construction concept of Titanium Alloy Bars Reinforced with Ultra-High Performance Concrete (TARUHPC) has been recently proposed and different studies are under conduction to validate its seismic performance. However, conducting extensive experimental campaigns with such materials is costly and time-consuming. Then, new approaches can also be explored to guide new studies and optimize the information that can be extracted from laboratory tests. As one alternative, Machine Learning (ML) techniques are gaining significant attention in the structural engineering community since more accurate and robust models can be developed for some complex problems. This paper explores two approaches to develop a data-driven model for the bond strength between titanium alloy bars (TiABs) and ultra-high performance concrete (UHPC). Given that there is limited data from experimental tests for such materials, transfer learning (TL) techniques are also applied to transfer the knowledge from models trained with data from conventional concrete elements. A comparison of the prediction capacity between traditional ML models and TL models is provided. The proposed ML models showed much better performance (R² above 0.8) than conventional design equations when predicting the bond strength of TARUHPC elements. The results indicated that TL can provide accurate predictions and can be used to guide further exploration of TARUHPC structures.