Integrated kinematic time-lapse inversion workflow leveraging full-waveform inversion and machine learning

Abstract

We demonstrate that a workflow combining emergent timelapse full-waveform inversion (FWI) and machine learning technologies can address the demand for faster time-lapse processing and analysis. During the first stage of our proposed workflow, we invert long-wavelength velocity changes using a tomographically enhanced version of multiparameter simultaneous reflection FWI with model-difference regularization. Short-wavelength changes are inverted during the second stage of the workflow by a specialized high-resolution image-difference tomography algorithm using a neural network. We discuss application areas for each component of the workflow and show the results of a West Africa case study. Introduction One challenge of conventional time-lapse processing and analysis techniques is that these methods rely on significant QA/QC efforts, survey cross-equalization, and manual parametrization. Maximizing the net present value in the current environment requires shorter 4D data turnaround times. Our objective is to fill this opportunity space with novel methodologies that can expedite and automate the delivery of interpretational 4D products without compromising their quality. A variety of existing modeland image-space time-lapse analysis techniques lend themselves to automation but are constrained by limited resolution, repeatability requirements, and the accuracy of underlying elastodynamic modeling (Maharramov and Albertin, 2007; Chu, 2013; Shragge and Lumley, 2013; Qu and Verschuur, 2016). Model-space full-waveform inversion (FWI) methods that have no explicit requirements for data conformity across multiple acquisitions can be less sensitive to survey repeatability and can be applied in an automated inversion of time-lapse attributes requiring minimal data preprocessing (Asnaashari et al., 2012; Routh et al., 2012; Raknes et al., 2013; Willemsen and Malcolm, 2015; Alemie and Sacchi, 2016; Maharramov et al., 2016). During the first stage of our kinematic time-lapse inversion workflow, we conduct a tomographically enhanced version of our 4D FWI method for frequencies up to 30 Hz (Maharramov et al., 2017b). This step resolves large-scale 4D velocity effects that cause time displacements in excess of the data sampling rate. During the second stage, we supply the resulting full or partial stacks to our new image-difference tomography method with a model-difference Musa Maharramov1, Bram Willemsen2, Partha S. Routh1, Emily F. Peacock1, Mark Froneberger2, Alana P. Robinson2, Glenn W. Bear3, and Spyros K. Lazaratos2 regularization. Our approach departs from earlier image-difference tomography techniques (Maharramov and Albertin, 2007; Yang et al., 2014), since it employs a kinematic misfit functional of two images based on a neural network. The misfit functional is designed to be robust with respect to nonrepeatability and dynamic effects in the two images by training the underlying neural network on noisy synthetic data with poor repeatability. We discuss separate and joint deployment of 4D FWI and image-difference tomography in 4D projects and describe the application of our workflow to resolve gas-ex-solution (i.e., gas coming out of solution when pressure drops below the bubble point), water sweep, and pressure effects in a West Africa case study. Theory Kinematic FWI objective functions based on phase misfits in time and frequency domains are easily computable and yield explicit expressions for the adjoint source (Fichtner, 2011). Such phase misfits as well as the normalized L2 objective function (Routh et al., 2011) use the phase as a proxy for traveltime information. The phase-only objective function with implicit spectral shaping (Maharramov et al., 2017a; Fu et al., 2018) fully separates kinematic and dynamic information for a single transmission or reflection event in dispersive viscoelastic media. An FWI method using a phase-misfit objective function provides true kinematic FWI when applied to refractiononly (diving-wave) data that are cleanly separated into individual events (Fichtner, 2011). While application of joint 4D FWI to diving-wave data is of potential interest for shallow anomaly identification or long-offset repeat acquisitions (Maharramov et al., 2016), our objective is 4D FWI of reflection data because of the typically limited offset range of conventional monitor acquisitions. In our approach, we use the normalized L2 and phase-only objective functions in a tomographically enhanced reflection 4D FWI. To reduce the leakage of reflection amplitudes into the extracted phase, we use time-windowed objective functions (Maharramov et al., 2017a). The 4D FWI is preceded by the inversion of a baseline pseudodensity (with no long-wavelength density components) that populates the forward-modeled baseline data with reflections that appear at the right depth for the baseline survey. The inverted baseline pseudodensity is then supplied as a fixed parameter for baseline and monitor forward modeling and used in a joint 4D FWI with a model-difference regularization. The following sections describe this method and the subsequent step of residual image-difference tomography. This approach is similar to the one taken by 1ExxonMobil Upstream Research Company, Spring, Texas, USA. E-mail: [email protected]; [email protected]; [email protected]. 2ExxonMobil Upstream Integrated Solutions Company, Spring, Texas, USA. E-mail: [email protected]; [email protected]; [email protected]; [email protected]. 3ExxonMobil Services and Technology, Bangalore, India. E-mail: [email protected]. https://doi.org/10.1190/tle38120943.1 D ow nl oa de d 12 /0 2/ 19 to 1 58 .2 6. 2. 16 9. R ed is tr ib ut io n su bj ec t t o SE G li ce ns e or c op yr ig ht ; s ee T er m s of U se a t h ttp :// lib ra ry .s eg .o rg /