GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 3.3 GNGTS 2024 Figure 4: Wiggle plot of the area between 0.6s-1.3s and 120m-1250m of the Viking Graben dataset, focusing on a wedge structure. The input seismic line is shown in A, while the high frequency predicton of the same area is depicted in B. In Figure 4B, the prompt detecton of coherent signals is evident, notably horizontal refectors beneath and within the wedge structure. Despite the limited lefward extent of the secton, there is a remarkable improvement in vertcal resoluton, potentally enabling the identfcaton of pinch- out structures. Conclusions We introduce a novel 1-D approach based on LSTM (Long Short-Term Memory) Neural Networks for addressing the low- and high-frequency gaps in seismic signal processing. The proposed method involves training two distnct neural networks: a low-frequency model, trained to infer lower frequency output from higher frequency signals, and a high-frequency model, trained with reversed input and output. The method's scalability is assured thanks to its ability to operate without direct consideraton of the frequency components, tme length, and sampling informaton. The NN is trained with a custom loss functon that incorporates both amplitude and frequency components. A crucial parameter known as Sample Duraton (SD) governs the frequency content generaton during training, providing fexibility for adjustng input data sampling and, consequently, the generated frequency output. The method's adaptability is demonstrated by rescaling signals to the trained