Near surface full waveform inversion via deep learning for subsurface imaging

Parasyris, A. and Stankovic, L. and Pytharouli, S. and Stankovic, V.; Anagnostou, Georgios and Benardos, Andreas and Marinos, Vassilis P., eds. (2023) Near surface full waveform inversion via deep learning for subsurface imaging. In: Expanding Underground - Knowledge and Passion to Make a Positive Impact on the World : Proceedings of the ITA-AITES World Tunnel Congress 2023 (WTC 2023), 12-18 May 2023, Athens, Greece. CRC Press, GRC, pp. 2829-2836. ISBN 9781003348030 (https://doi.org/10.1201/9781003348030-341)

[thumbnail of Parasyris-etal-WTC-2023-Near-surface-full-waveform-inversion-via-deep-learning-for-subsurface-imaging]
Preview
 Text. Filename: Parasyris_etal_WTC_2023_Near_surface_full_waveform_inversion_via_deep_learning_for_subsurface_imaging.pdf
Final Published Version
License: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 logo
Download (1MB)| Preview

Abstract

In order to meet increasing safety standards and technological requirements for underground construction, the estimation of Earth models is needed to characterize the subsurface. This can be achieved via near-surface or standard Full-Waveform Inversion (FWI) velocity model building, which reconstructs the Earth model parameters (compressional and shear wave velocities, density) via recordings obtained on the field. The wave function characterizing the Earth model parameters is inherently non-linear, rendering this optimization problem complex. With advances in computational power, including graphics processing units (GPUs) computing, data driven approaches to solve FWI via Deep Neural Networks (DNN) are increasing in popularity due to its ability to solve the FWI problem accurately. In this paper, we leverage on DNN-based FWI applied to field data, to demonstrate that instead of depending on observed data collected from multiple boreholes across a large distance, it is possible to obtain accurate Earth model parameters for areas with varied geotechnical characteristics by using geotechnical data as prior knowledge and constraining the training models according to a single borehole to map the large geological earth cross section. Also we propose a methodology to simulate acoustic recordings indirectly from laboratory tests on soil samples obtained from boreholes, which were analysed for compressive strength of intact rock and Geological Strength Index. Layers’ geometry and properties for a section of total 3.0 km are used for simulating 15 2D elastic spaces of 200 m width and 50m depth assuming receivers and Ricker-wavelet sources. We adopt a Fully Convolutional Neural Network for Velocity Model Building, previously shown to work well with synthetic data, to generate the 2D predicted Earth model. The results of this study show that the velocity model can be accurately predicted via DNN through the appropriate training with minimum demands for borehole data. The performance is evaluated through both metrics focused on image quality and on velocity values giving a multifaceted understanding of the model’s true ability to predict the subsurface.

ORCID iDs

Parasyris, A., Stankovic, L. ORCID logoORCID: https://orcid.org/0000-0002-8112-1976, Pytharouli, S. ORCID logoORCID: https://orcid.org/0000-0002-2899-1518 and Stankovic, V. ORCID logoORCID: https://orcid.org/0000-0002-1075-2420; Anagnostou, Georgios, Benardos, Andreas and Marinos, Vassilis P.
CORE (COnnecting REpositories)