GNGTS 2019 - Atti del 38° Convegno Nazionale

GNGTS 2019 S essione 3.3 751 Fig. 1 - Input profile (A), desired output (B) and the predicted output (C). Because the net does not model data from equations, it has to learn from the training examples how an input (velocity) is related to its response (seismograms). This leads to the crucial importance of the training dataset used: a bad dataset will lead to a bad model because the net will learn all (and only) the features that are in the training data. Fig. 2 shows a comparison of the input, the real output, the prediction of the retrained model and the prediction on the original model. We have a critical level of dispersion and it is interesting that such predictions come from the same trained net: this means that the net is also able to discern when and how dispersion has to be predicted. The source of initial problems in predicting multiple reflections was traced back to the limited number of such events in the training dataset that resulted in inadequate statistics. At the beginning, training was performed on a dataset with 7 ground layers: this led to a very low probability of finding multiple reflections. Then the net seemed not to learn what these signals were due to. To try to solve this problem, we retrained the model on a dataset with a lot of multiple reflections. We imposed a 3-layers velocity model with two thin shallow layers and a thicker third one. This model results in a dataset with lots of statistics on multiple reflections. As we can see in fig. 2 in the retrained model prediction, a multiple reflection at 500 ms is close to the real signal, while the old net did not predict such signal. We made the method more flexible by allowing inserting information like waveform and offset. An example of this is shown in Fig. 3 with old and new prediction using a Ricker wavelet at 10 Hz and 20 Hz respectively. This seemed to work properly because during training the gradient is quite smooth and an absolute minimum is reached in few epochs. This may be due to the fact that we do not need to create new set of weights, because they are already close to the right solution. Prediction time for this net depends only on the number of offsets, the temporal discretization, and the total recording time. The FD generation time does not depend on the number of offsets but is dependent on the model size and its discretization. This leads to the capability for the net, once trained, to produce synthetic seismograms on a wide area with great offset spacing much faster than classical FD methods.