GNGTS 2022 - Atti del 40° Convegno Nazionale

GNGTS 2022 Sessione 3.3 509 Fig. 2 - An original deblended common shot gather (a) with its corresponding pseudo-deblended gather (e). Result of deblending and corresponding residual with dictionary learning based deblending (b.f), end-to-end CNN deblending (c,g), deep preconditioners based inversion (d,h). see the blending noise in Fig. 5(b), there is distinct event leakage. Deep preconditioner-based method (Fig. 2h) can effectively remove almost all the blending noise with little useful signal leakage. A good training dataset is key to train an effective AE. Standard solutions consist in building high-quality synthetic dataset or collecting an appropriate field dataset, but neither is easy to implement. We avoid this by relying on the blended dataset itself to construct the training samples: this means that no extra data is required. Based on reciprocity, unblended common shot gathers (CSGs) are similar to unblended common receiver gathers (CRGs). On the other hand, despite the acquired CSGs are blended, we can still find that the data are coherent and thus they share similar characteristic with the unblended data. On the contrary, blended CRGs are corrupted by random blending noise. Training is performed using the Adam optimizer (Kingma and Ba, 2014) with learning rate lr = 10^-3, weight decay regularization eq = 10^-5, and p = 2. Discussion and conclusions . Deep preconditioning is a framework that leverages nonlinear based dimensionality reduction to solve of ill-posed inverse problems. We have demonstrated his effectiveness trough an example of seismic deblending. An important feature of this application is that there is no need for any training data beyond the recorded seismic data itself. The choice of the network architecture, the size of the latent representation and the training strategy seem to be the key factors to find a representative sparse representation of our seismic data. Future work will focus on testing the framework on a wide set of seismic processing problems. References Candes E. 2006, Compressive Sampling, Proceedings of the international congress of mathematicians 3, 1433-1452 Kingma D., J. Ba, 2015 Adam: A Method for Stochastic Optimization. International Conference on Learning Representations Goodfellow I., Y. Bengio, and A. Courville, 2016 Deep learning. MIT press. Ravasi M., 2021, Seismic wavefield processing with deep preconditioners , in First International Meeting for Applied Geoscience & Energy, Expanded Abstracts, pp. 2859-2863.