AI-guided design framework for bond behavior of steel-concrete in steel reinforced concrete composites: From dataset cleaning to feature engineering
Küçük Resim Yok
Tarih
2025
Yazarlar
Dergi Başlığı
Dergi ISSN
Cilt Başlığı
Yayıncı
Elsevier
Erişim Hakkı
info:eu-repo/semantics/closedAccess
Özet
This research focuses on establishing an artificial intelligence (AI)-guided design approach for predicting bond behavior of profiled steel-concrete in steel-reinforced concrete (SRC) composites. For that, an extensive literature survey was undertaken, and datasets for three main characteristic bond stresses-bond stress at initial slip (zs), ultimate bond stress (zu), and residual bond stress (zr) -were gathered. In total, it was gathered data points 150 for zs, 251 for zu, and 215 for zr. In addition, the isolation forest algorithm was used to detect and clean the anomalous data in the dataset, resulting in exhaustive and trustworthy data for training the models. As AI models, four popular machine learning algorithms like RF, XGBoost, LightGBM, and CatBoost are adopted. To improve the prediction performance of the models, three cases are established by Shapley additive explanations (SHAP)-based feature engineering. Additionally, SHAP and feature importance analyses were used to examine the impact of each feature on the bond behavior in SRC composites to ensure the explainability of the model. Meanwhile, to enhance the applicability of the study in real-world applications, a graphical user interface (GUI) was designed. According to the results, the CatBoost model proved its superior predictive ability specifically for zs and zr output values; in the test phase, the RMSE values were 0.07 and 0.05, R2 values were 0.904 and 0.934, MAPE were 10.02% and 8.33%, and MAE values were 0.04 and 0.03, respectively. On the other hand, the XGBoost model had the best predictive efficiency in the test phase for the zu output value with RMSE = 0.06, R2 = 0.833, MAPE = 8.32 % and MAE = 0.04. Lastly, based on SHAP and feature importance assessments, the most impactful features on bond behavior were identified as follows: the ratio of side cover to steel section height (cv/ hs), the compressive strength of concrete (fcu), and the ratio of bonded length to steel section height (lb/hs), stirrup ratio (psv), and the yield strength of profiled steel (fy). This knowledge can guide engineers in paying focus to specific features in their design and evaluation processes, resulting in more reliable and optimized outcomes.
Açıklama
Anahtar Kelimeler
Steel reinforced concrete, Bond behavior, Machine learning, Unsupervised anomaly detection, Feature engineering
Kaynak
Materials Today Communications
WoS Q Değeri
Q2
Scopus Q Değeri
Q2
Cilt
42
