Résumé:
The objective of this thesis is to develop new predictive models for the compressive loadcarrying capacity of concrete columns confined using composite materials, either as internal
reinforcements and/or external fiber-reinforced polymer (FRP) wraps, under various axial
load eccentricity levels. To achieve this, general regression approaches and advanced machine
learning models, such as XGBoost and Random Forest (RF), were employed. We also used
deep learning models, such as BiLSTM and CNN-BiLSTM, to predict the axial compressive
capacity of concrete-filled steel tubes confined with fiber-reinforced polymers. The proposed
confinement models were developed using a bibliographic database composed of 308
reinforced concrete specimens strengthened with FRP bars, and 250 reinforced concrete
specimens confined by external FRP wraps or steel tubes. The results showed that the
proposed models, particularly those based on XGBoost and Random Forest (RF), achieved
high accuracy with a coefficient of determination (R²) of 0.98, along with minimal values for
the root mean square error (RMSE) and mean absolute error (MAE), thus outperforming
conventional empirical models. The CNN-BiLSTM model also demonstrated better
performance than the BiLSTM model. Furthermore, finite element analysis using ABAQUS
showed that the predicted axial loads and deformations closely matched the experimental
results, thereby confirming the accuracy and conservativeness of the finite element model
employed.
