GNGTS 2019 - Atti del 38° Convegno Nazionale

GNGTS 2019 S essione 3.3 753 Neural Networks provide a good alternative to classical FD method. They are able to rapidly and accurately solve the task. Waveform and interferences are well reproduced and the net manages to predict multiple reflections, which are not directly linked to the interface position. We propose also a method based on retraining to reduce this problem. This method gives good results and it is able to make adjustments of the net parameters in a short time (less than one hour on ENI’s HPC4) and with a small dataset. This technique may open new perspectives of development: future research could test the net in the elastic 1-D approximation prediction and in the solution of 2-D direct and inverse problems. Acknowledgements. The first author wishes to acknowledge ENI SpA for their assistance and for allowing him the use of their computer cluster (HPC4). The Devito project [2] is also acknowledged for the highly optimized finite difference kernels used in this work. References [1] Hochreiter S. and Schmidhuber J.; 1997: Long Short-term Memory. Neural computation 9, pp. 1735–80. [2] Louboutin M., Lange M., Luporini F., Kukreja N., Witte P. A., Herrmann F. J., Velesko P. and Gorman G. J.; 2019: Devito (v3.1.0) an embedded domain-specific language for finite differences and geophysical exploration. Geoscientific Model Development, Volume 12, p 1165-1187, 2019 [3] Murphy K. P.; 2012: Machine learning : a probabilistic perspective ISBN 978-0-262-01802-9, MIT Press. [4] Sherstinsky A.; 2018: Fundamentals of Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) Network. ArXiv abs/1808.03314. A DATA-DRIVEN TRANSDIMENSIONAL INVERSION APPROACH TO INCLUDE LATERAL CONSTRAINTS INTO TARGET-ORIENTED AVA INVERSION A. Salusti 1,2 , M. Aleardi 2 1 University of Florence, Earth Sciences Department, Florence, Italy 2 University of Pisa, Earth Sciences Department, Pisa, Italy Introduction. The inclusion of lateral constraints into the inversion framework is the most popular strategy devoted at attenuating the ill-conditioning of the seismic inversion. The Tikhonov approach is by far the most popular regularization approach even if it has several disadvantages, for example this method often leads to unfocused layer transitions. Other more advanced regularization strategies exist, such as the inclusion of geostatistical constraints in the form of isotropic model correlation functions (Buland et al. 2003), or stratigraphic constraints (Tetyukhina et al. 2010). The main limit of all these approaches is that they rely on an a-priori structural knowledge of the investigated area and force the recovered model to honor such a-priori constraint. These are essentially model-driven regularization strategies that could provide biased model parameter estimations in case of erroneous a-priori assumptions. To overcome these issues more advanced, adaptive, regularization strategies have been proposed (e.g. Aleardi et al. 2018). The goal of these approaches is to locally adapt the structural constraint to the local structural characteristics of the subsurface model that can be iteratively inferred form the local characteristics (i.e. variability) of the observed data. On the line of these data-driven approaches, we present a transdimensional reversible jump Markov Chain Monte Carlo (rjMCMC) algorithm for target-oriented amplitude versus angle