Reverse Time Migration with Quantified Uncertainty
Abstract
Reverse Time Migration (RTM) is considered a workhorse for seismic imaging in the Oil&Gas industry. In the present work, we assess the presence of uncertainties and how they affect the reliability of the produced images. We particularly emphasize the role of high-performance computing and identify the main bottlenecks of an Uncertainty Quantification (UQ) analysis in such a context. To assess the final impact of uncertainties in the process of building a seismic image, we designed a workflow formed by three axles connected by a high-performance layer responsible for moving data and managing provenance between them. The first axle is dedicated to obtaining a velocity subsurface model conciliating measured signals on the surface and expert knowledge. The scarcity of noisy measurements along with human subjective intervention leads to significant sources of uncertainties. We employ a Bayesian travel time tomography formulation to accommodate such aspects of the velocity model, leading to a set of realizations of a random spatial field, what serve as inputs for the second axle. We employ an Eikonal solver in the tomography to make it computationally feasible within the workflow, what constitutes another uncertainty source due to the limited physics involved. To carry out a consistent probabilistic analysis, the number of realizations tend to be large, and each one of them is supposed to be highly defined to capture the variability of the spatial field. The second axle migrates the input velocity by solving twice (direct and reverse time) an acoustic wave equation, that requires very fine computational grids, and, consequently, very time-consuming. To perform the UQ analysis, we employ a Monte Carlo (MC) algorithm, solving the two-way acoustic migration for each sample of the velocity field. The third axle provides automatic computational tools for supporting the geologists to pick up important features within the images. The biggest computational challenge in that final step of the proposed workflow relies on how to handle a significant number of samples generate in the previous axle by the MC analysis. In the end, we present some examples to illustrate the performance and challenges of the workflow.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.T44A..06R
- Keywords:
-
- 0545 Modeling;
- COMPUTATIONAL GEOPHYSICS;
- 0560 Numerical solutions;
- COMPUTATIONAL GEOPHYSICS;
- 1932 High-performance computing;
- INFORMATICS;
- 3275 Uncertainty quantification;
- MATHEMATICAL GEOPHYSICS