Multitask learning-based prediction of fatigue performance in steel-concrete composite deck slabs

This study proposes a multitask learning (MTL)-based framework for the integrated prediction of key structural performance parameters in steel–concrete composite deck slabs, including mid-span deflection and residual fatigue life. Traditional single-task approaches focus on isolated performance indicators and fail to capture the complex nonlinear interactions among design variables and the inherent coupling between structural responses. To address these limitations, a comprehensive hybrid dataset of 152 validated records was compiled by integrating experimental data, finite element analysis (FEA) simulations, and design-code-based augmented data. Input variables were standardized and optimized through feature importance analysis, and a multitask artificial neural network (MT-ANN) was developed to simultaneously predict both target outputs. The model performance was benchmarked against established machine learning algorithms including Random Forest, XGBoost, Support Vector Regression, and Long Short-Term Memory networks. The proposed MT-ANN consistently outperformed single-task baselines in accuracy and generalization. SHAP-based interpretability analysis further revealed that geometric parameters and material properties contribute differently across prediction tasks, confirming the value of the multitask architecture. The findings demonstrate that multitask learning provides a robust and scalable solution for holistic assessment of composite deck structural performance. The developed model supports structural design optimization, service life evaluation, and predictive maintenance planning. Its compatibility with digital twin and BIM-integrated environments further enhances its potential for real-time structural health monitoring in modern infrastructure systems.