Accurate simulation of seismic wave propagation in complex geological media is essential for assessing the response of critical infrastructures, yet high-fidelity numerical solvers remain computationally expensive. This limitation restricts their use in rapid scenario evaluation and uncertainty analysis. To address this challenge, this study introduces a surrogate modeling framework for seismic displacement signals based on neural networks in the frequency domain. Seismic time-domain responses are first generated by solving the elastodynamic equations using the EFISPEC3D spectral-element code. The resulting signals are then transformed into the frequency domain, and their real and imaginary spectral components are extracted and used to train separate surrogate models. Three machine learning approaches are investigated: a Multilayer Perceptron, a Random Forest model, and a Convolutional Neural Network. Model performance is assessed by reconstructing time-domain signals through inverse Fourier transforms and comparing them with the original numerical simulations. Among the tested approaches, the Multilayer Perceptron demonstrates the most robust performance, achieving a coefficient of determination of \(R^2 = 0.94\) on the validation dataset. The resulting surrogate model provides a high-fidelity approximation of the seismic response while significantly reducing computational cost. This surrogate constitutes a Digital Shadow of the seismic system: a high-fidelity, data-driven representation that mirrors the behavior of the underlying physical model without directly interacting with the physical asset. This Digital Shadow enabling fast and reliable seismic scenario analysis without direct interaction with the physical infrastructure. Although real-time bidirectional data exchange is not considered in this work, the proposed framework lays the foundation for future integration of observational data, representing a key step toward the development of a seismic Digital Twin.