Multi-Output Neural Network for Elastic and Collapse Multipliers Prediction in Structural Vulnerability Assessment

Seismic vulnerability assessment remains a critical challenge in structural engineering, especially in predicting elastic and collapse multipliers under combined permanent and seismic loading in multi-story buildings. In the current investigation, a Multilayer Perceptron neural network is developed as a multi-output regression model to approximate six key seismic response parameters corresponding to permanent load, seismic X-direction, and seismic Z-direction multipliers in both elastic and collapse states. The model is trained on a physics-informed dataset generated using a finite element-based lower bound limit analysis framework and enhanced through systematic data preprocessing. Hyperparameter sensitivity analyses show that the proposed network configuration achieves high predictive accuracy, with test R 2 values exceeding 0.94 across all outputs, indicating strong generalization capability. As key innovations, permutation importance is employed to identify influential structural features, Monte Carlo dropout is utilized to quantify prediction uncertainty through confidence intervals, and correlation analysis is applied to capture interdependencies between seismic responses. The outcomes confirm that the proposed framework accurately reflects the underlying structural behavior while providing reliable uncertainty estimates. In general, the study improves the interpretability and reliability of the neural network-based seismic assessment and offers a practical, data-driven tool for the design of resilient structural.