Title:
Generalizing Shear Strength Prediction in Reinforced Concrete Beams with Out-of-Distribution Data (Prepublished)
Author(s):
Zecheng Yu and Bing Li
Publication:
Structural Journal
Volume:
Issue:
Appears on pages(s):
Keywords:
Gaussian process; out-of-distribution; reinforced concrete beam; shear strength
DOI:
10.14359/51746759
Date:
4/9/2025
Abstract:
Despite advancements in machine learning (ML) that have boosted structural performance prediction, current ML models can still struggle to generalize to unseen situations, leading to performance degradation. This vulnerability arises from their overreliance on data, neglecting established engineering principles like mechanical priors. Models trained on specific data distributions can suffer significant accuracy degradation when encountering inputs that fall outside those distributions. To overcome the limitations of data-driven models with unseen data, A mechanics-guided Gaussian process (MGGP) for accurate prediction of shear strength in reinforced-concrete (RC) beams is proposed. The complex variation of shear strength in RC beams was captured using a Gaussian process (GP) model with a mean function derived from mechanical principles and a hybrid kernel to account for inherent prediction variability. This combination allows for accurate prediction of shear strength while considering the underlying physical mechanisms. This approach leverages domain knowledge from mechanics by incorporating a relevant design equation into the mean function of a GP model. This integration significantly enhances the model's ability to predict shear strength by capturing the underlying physical principles governing the shear strength. Cross-validation studies have shown that the MGGP offers consistent performance compared to traditional GPs in predicting the shear strength of RC beams.