MBSE Net: Addressing Data Scarcity and Model Interpretability in Predicting UHPC Performance for Bridge Infrastructure

Bridges serve as critical infrastructure, enabling the movement of people and goods across regions, thus underscoring the importance of their durability and reliability. Different key components of bridges demand varied material performance characteristics. Traditional materials like standard concrete often exhibit deficiencies in strength, crack resistance, and durability, leading to cracks and damage in essential bridge parts. Ultra-High Performance Concrete (UHPC) addresses these shortcomings, but its complex component interactions pose new challenges for predicting mechanical properties. Machine learning models designed for this purpose often lack generalization capability, while deep learning models face issues related to data scarcity and model interpretability. To tackle the challenge of data scarcity, we introduced a self-supervised pre-training approach that divides the prediction task into two stages: feature learning followed by prediction. Designing an encoder capable of extracting multi-scale effective feature representations from limited data is particularly challenging. After extensive tuning, we developed the Multi-Branch Self-Encoding Network (MBSE Net). This network achieved <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$R^{2}$</tex> values exceeding 0.99 in predicting both compressive strength (CS) and flexural strength (FS), significantly outperforming other methods. To enhance model interpretability, we implemented two visualization techniques to improve transparency and credibility from different perspectives. These techniques elucidate the key factors influencing UHPC performance and provide insights into the optimal material proportion design to meet the performance requirements of critical bridge components.