Machine learning-based bond-strength prediction model for ultra-high-performance concrete

This study experimentally evaluated the bond performance of precast bridge-deck joints made with ultrahigh-performance concrete (UHPC) and aimed to improve the structural efficiency and constructability by addressing the limitations of existing design codes. Traditional connection methods, such as the loop splice method, reduce the constructability and structural efficiency. The applicability of UHPC as an alternative filling material was examined to shorten the splice length while securing structural performance. Fourteen splice specimens were fabricated with varying compressive strengths (80 MPa, 100 MPa, and 120 MPa) and splice lengths (8db, 10db, and 15db, where db is the bar diameter), and splice tests were conducted using a four-point loading method. The test results confirmed that the high compressive strength and fiber reinforcement effect of UHPC secured sufficient structural performance and crack control even at short splice lengths, whereas the existing design codes such as Eurocode 2, Korean Highway Bridge Design Code, ACI318, and Structural Design Guideline for Fiber-Reinforced SUPER Concrete revealed limitations in sufficiently reflecting the UHPC characteristics. To solve these problems, machine learning techniques were applied, and new bond strength and development-length prediction models employing random forest and SHAP for factor analysis and symbolic regression (PySR) were proposed. The proposed models achieved high prediction accuracy, demonstrating their potential for optimizing the bond performance of UHPC structures to ensure structural safety and improving the design efficiency of precast construction methods.