GNGTS 2023 - Atti del 41° Convegno Nazionale

Session 3.3 ______ ___ GNGTS 2023 between this kind of neurons and the encoder-decoder geometry grants us that each value of the hidden layer can be directly linked to the correct position in time or, in some way, to what happened at previous times. This assures the causality typical of seismic waves propagation. Furthermore, the introduction of the pooling layer, allows us to reduce the amount of information that the NN can bring to the encoder, i.e., by decimating time axis: this is a crucial point for the method. The proposed workflow is: I. Initialize an ED-LSTM Neural Network. II. Train the model on the data of interest. III. Use the HL predictions as set of additional information (i.e., DA) for improved seismic analysis, interpretation purposes or quantitative estimation. As we have defined the NN as Encoder Decoder geometry, we have to let the NN to Encode the output in simple input and decode this input to restore the data. We can describe this process, thinking about a cake, Figure 1: the Encoding part would be both the understanding of the ingredients of the cake and their amount. One the methodology has the encoded ingredients, it has to restore than by understanding how to merge and how to bake the cake, in order to restore it properly. Figure 1: Encoding and decoding of a cake. While the Encoder part leads us to the ingredients, the Decoding part is the actual cooking of the cake.