• Fuse ground motion and structural parameters using cross-modal attention. • Enable real-time probabilistic analysis by replacing expensive FEA with a surrogate model. • Predict full probabilistic time-histories at multiple structural locations. • Transform risk assessment from binary classification to continuous quantification. • Achieve high fidelity ( R 2 > 0.98) with a mean prediction error below 1.4%. Ensuring the safety and reliability of critical equipment in nuclear power plants (NPPs) requires accounting for uncertainties in structural parameters. However, probabilistic methods based on finite element analysis (FEA) can be computationally expensive and impractical for real-time safety assessment. In this study, we propose a surrogate framework based on deep learning that directly generates full time-history acceleration responses at multiple structural locations to reduce the computational cost of probabilistic analysis by replacing FEA. The architecture is organized into two encoders, including a structural parameter encoder and a seismic encoder enabling feature extraction from two distinct input modalities. Their latent representations are fused and mapped using a response decoder that reconstructs multi-output acceleration responses. The results of experimental validation show that the model achieved an average maximum mean absolute percentage error (mMAPE) of 1.37%. Beyond deterministic surrogate predictions, the proposed framework generates probabilistic responses to capture the variability of structural response over time and instantly estimate exceedance probabilities with respect to equipment-level thresholds. This probabilistic inspection capability extends conventional deterministic safety checks to a reliability-informed paradigm to offer a computationally efficient solution for the assessment of important equipment in NPPs.
Lee et al. (Sun,) studied this question.