GNGTS 2019 - Atti del 38° Convegno Nazionale

558 GNGTS 2019 S essione 3.1 enthusiasm, and immense knowledge. We would also like to thank SEG EVOLVE organization, which gave us the excellent opportunity to participate to this project. We would also like to acknowledge the Department of Physics and Geology of University of Perugia for the support and for providing us the laboratory hardware and software. In particular, we thank IHS Markit for the academic licenses of Kingdom software suite. EVOLVE 2019 was made possible by Halliburton (Founding Sponsor), BRT Energy Advisors, LLC, Michael Forrest (Team Sponsors), Aramco (Meeting Sponsor) and Supporters: Schlumberger, IHS Markit, IHRDC,Rose & Associates, LLP and Ikon Science. References Galloway, W. E.; 2008: Chapter 15 Depositional Evolution of the Gulf of Mexico Sedimentary Basin. Sedimentary Basins of the World, 5, 505–549. Hackley, P.C.; 2012: Geologic assessment of undiscovered conventional oil and gas resources—Middle Eocene Claiborne Group, United States part of the Gulf of Mexico Basin: U.S. Geological Survey Open–File. Report 2012–1144, 87 p., available only at http://pubs.usgs.gov/of/2012/1144/. DEEP-LEARNING-BASED COMPRESSION FOR SEISMIC IMAGING V. Lipari, P. Bestagini, S. Tubaro Dipartimento di Elettronica Informazione e Bioingegneria - Politecnico di Milano, Italy Introduction. Several current imaging technologies, such as Full Waveform Inversion (FWI) and Reverse Time Migration (RTM) rely on the adjoint state method. These imaging algorithms need both forward and adjoint wavefields at each time step for the computation of sensitivity kernels; this is usually obtained by saving the whole forward wavefield and requires huge storage capabilities. Moreover, storing the whole forward wavefield to the disk causes significant I/O overhead and dramatically worsens the performances and introduces an important bottleneck, especially on GPUs. The main strategies proposed to attenuate this problem are checkpointing and compression. Checkpointing techniques reduce the memory requirements but greatly increase the computational cost (Anderson et al. , 2012). The typical overhead caused by this approach is the order of one additional forward simulation. A valid alternative for reducing storage requirements and I/O overhead is given by compression techniques. In the ideal case, if we were able to compress the forward wavefield enough to fit it entirely into the memory, neither disk I/O nor checkpointing would be necessary anymore. Lossless compression allows an exact data reconstruction, whereas lossy algorithms introduce a controlled error, but are more effective in terms of compression ratio (CR). Industrial solutions often use standard lossless compression algorithms. The straightforward solution to recast data into a lower precision format (i.e. from double to single precision), which is often used in industrial solutions, is nothing but a trivial lossy compression scheme. Dalmau et al. (2014) and Boehm et al. (2016) proposed compression algorithms specifically designed for the wavefields simulated in FWI. The main innovation of the last few years, in the signal processing field, is certainly the explosion of techniques based on CNNs and deep learning. Interesting results have been obtained for seismic processing tasks such as interpolation (Mandelli et al. , 2018) and denoising (Li et al. , 2018) suggesting that CNNs have the capability to provide a compact representation of seismic data. These considerations lead us to explore CNNs as a tool for lossy compression of seismic wavefields. In this work we study two CNN based compression strategies for seismic wavefields, namely snapshots compression and snapshots interpolation. These strategies can be also jointly combined to improve compression results. To perform snapshots compression, we leverage a specially designed convolutional autoencoder that projects data into a reduced dimensionality