Artificial Intelligence in geoscience applications. AI models for seismic denoising, earthquake source characterization, and ground-motion prediction using the Italian Seismic Network

Earthquakes remain challenging to analyze in real time because waveforms are noisy, source pa- rameters are uncertain, and shaking patterns evolve rapidly. This thesis advances the seismic monitoring pipeline with deep learning (DL) methods that improve data quality, reduce latency, and raise predictive accuracy on operational datasets and networks. First, we introduce CDiffSD, a cold diffusion model for seismic denoising that learns non-Gaussian degradation processes directly from station noise. On the STEAD-based test, CDiffSD increases median signal-to-noise ratio and cross-correlation over a strong deep-denoiser baseline, while markedly increasing P-pick recall within ±50 samples (from ∼ 0.60 to ∼ 0.90 at T =300). Second, we propose a masked Graph Convolutional Network for rapid shaking prediction in Italy. Using only the first 10 s of three- component waveforms from national-scale networks (INSTANCE), the model yields up to 6% and 5.5% mean-squared-error reductions for PGA and PGV, respectively, near-zero median residuals across intensity measures, and mitigates biases relative to GMM (an Italy-specific ground-motion model). Third, we develop LLM4Geo, a domain-adapted large language model that estimates epi- center, hypocenter, origin time, and magnitude within five seconds of first detection. On 3,094 test events, average errors are ∼4 km (epicenter), ∼6.8 km (hypocenter), ∼0.6 s (origin time), and ∼0.2 magnitude units, outperforming early-stage automated INGV solutions while using fewer picks. Fi- nally, we present SeismicAE, a waveform autoencoder that transfers audio-domain representation learning to seismology. The learned embeddings improve few-shot classification and reduce error in ground-motion regression relative to models trained on raw traces. In the end, we build a pro- totype AI agent that combines different AI models for specific tasks (expert) in a single workflow (orchestration), giving compatible results for the evaluation of operational seismic risk.