AI Framework for Ground-Motion Prediction Across the Italian Seismic Network

Estimating ground‐motion intensity at individual seismic stations is a fundamental task in seismology, with direct implications for seismic hazard assessment.
Ground‐motion intensity measures (IMs) such as Peak Ground Acceleration (PGA), Peak Ground Velocity (PGV), and Spectral Acceleration (SA) at selected periods are commonly used to quantify shaking severity and relate it to building structural response and seismic hazard.

Here, we present an AI-driven framework for predicting multiple IMs at the station level, building on previous graph-based approaches for the Italian seismic network.
Starting from the INSTANCE dataset, we applied filters on event magnitude (≥ 3) and waveform quality, resulting in 3076 events recorded across 565 stations. For waveform analysis, we used a 10-second time window starting 1 second before the first P-wave arrival, balancing prediction speed and accuracy.

Seismic waveforms are first encoded using a pre-trained PhaseNet model to extract compact temporal representations. Spatial dependencies are modeled with a masked Graph Convolutional Network (GCN) based on Delaunay triangulation, which links each station to its nearest neighbors while avoiding long or crossing edges. This structure allows identification of border stations (at mesh edges) and coastal stations (within 15 km of the coastline). A binary mask distinguishes these nodes, helping the model account for areas with high azimuthal gap, which can make IMs estimation more challenging.

Before the final prediction layer, the model concatenates the maximum waveform amplitude across stations with event metadata predicted by a fine-tuned LLM for magnitude and location estimation.
This enables joint exploitation of temporal waveform features, network geometry, and global event information.

The framework predicts PGA, PGV, and SA at periods of 0.3, 1, and 3 s at all stations. We obtained preliminary results for multiple configurations. The baseline model served as reference, which included waveform representations and the GCN but not LLM metadata, border features, or weighted loss. Including LLM-derived metadata consistently improved performance across all regions and parameters, reducing relative errors by 14% in Southern Italy and over 27% in Northern Italy. The addition of border and coastal features provided only minor improvements, confirming that metadata was the main factor driving gains.

Applying a weighted loss emphasizing stations closer to the epicenter further improved predictions, particularly in Northern and Southern Italy. In Central Italy, where network coverage is denser, improvements were smaller, suggesting that local contributions were already well captured. For the best configuration (LLM metadata and weighted loss), global mean absolute errors across stations were 0.511 (PGA), 0.423 (PGV), 0.514 (SA0.3), 0.447 (SA1.0), and 0.397 (SA3.0), demonstrating the model’s predictive accuracy.

Overall, these preliminary results show that combining waveform features, network topology, and LLM-informed event metadata can substantially enhance station-level IMs estimation, achieving better results in challenging conditions (north or south Italy) with respect to previous similar approaches. This method has potential for rapid earthquake characterization and early warning, where timely and accurate ground-motion predictions are essential.