Inferring CO₂ saturation from synthetic surface seismic and downhole monitoring data using machine learning for leakage detection at CO₂ sequestration sites

Highlights

• A machine learning workflow is developed for inferring CO₂ saturation from surface seismic and downhole monitoring data.

• The performance of multiple machine learning algorithms is assessed using Kappa statistics.

• Surface seismic monitoring, coupled with downhole measurements, achieves higher accuracy of the CO₂ saturation inversion.

• The impact of seismic noise on the performance of the trained machine learning models is investigated.

Abstract

Inferring CO₂ saturation from seismic data is important when seismic methods are applied at CO₂ sequestration sites for verification and accounting purposes, such as verifying the total injected CO₂ volume, comparing with model predictions for concordance evaluation, tracking the migration of CO₂ plume, and detecting possible leakage from the storage reservoir. In this work, we infer CO₂ saturation levels at three depths from simulated surface seismic, downhole pressure and total dissolved solids (TDS) data using machine learning (ML) methods. The simulated monitoring data are based on 6000 numerical multi-phase flow simulations of hypothetical wellbore CO₂ and brine leakage from a legacy well into shallow aquifers at a model CO₂ storage site. We conduct rock physics modeling to estimate changes in seismic velocity due to the simulated CO₂ and brine leakage at each time step in the flow simulation outputs, resulting in 120,000 forward seismic velocity models. 2D finite-difference acoustic wave modeling is performed for each velocity model to generate synthetic shot gathers, along a sparse 2D seismic line with only 5 shots and 40 receivers. We extract 6 time-lapse seismic attribute anomalies from each trace in the time window relevant to each geologic layer, and use the seismic features, together with downhole pore pressure, TDS features to train the machine learning algorithms. The impact of seismic noise on the performance of the trained machine learning models has also been investigated. Inferred CO₂ saturations from the trained classifiers are in good agreement with observations. Direct pressure and TDS measurements from downhole monitoring can increase the accuracy of the inferred CO₂ saturation class from the forward modeled 2D surface seismic data. Our ML workflow represents a promising way to combine measurements from multiple monitoring techniques, together with seismic monitoring to achieve more accurate seismic quantitative interpretation.

Introduction

Carbon dioxide (CO₂) capture, utilization and sequestration, in which CO₂ is captured from emissions sources, compressed to a dense phase, and injected into deep geologic formations, is a promising approach to help mitigate atmospheric CO₂. Monitoring and verification techniques are needed to ensure safe and long-term storage of CO₂ in the deep subsurface. Operators will build 3D numerical geological models to describe the movement of CO₂ in the underground storage complex, and to demonstrate the conformance of the reservoir storage model prediction with the monitoring data throughout the injection and post-injection periods (Lumley, 2010; USEPA, 2016). Both direct and indirect monitoring techniques have been proposed to track CO₂ plume migration to ensure that CO₂ is contained within the storage complex and does not interfere with other subsurface intervals of concern, such as groundwater aquifer (USEPA, 2013), faults or fracture systems (Zhang et al., 2015). Direct monitoring techniques, such as pressure monitoring and fluid sampling, provide point measurements of primary diagnostic parameters (such as pressure, solution chemistry and CO₂ partial saturation) in the subsurface. Indirect monitoring techniques, usually geophysical approaches, such as seismic, electromagnetic and gravity monitoring techniques, generally sample over a relatively large area and provide volumetrically averaged representations of attributes of the sampled subsurface volume that correlate to primary diagnostic parameters of interest (Daley and Harbert, 2019; USEPA, 2016). Geophysical monitoring methods measure observable changes in the target framework and pore filling fluid parameters as imaged by the particular geophysical monitoring method used, rather than directly measuring the target parameters themselves. Therefore, inversion schemes are needed to infer the target observing parameters from the geophysical monitoring data. Time-lapse surface seismic monitoring is a widely used geophysical monitoring technique, which has been applied at CO₂ sequestration sites for site characterization, imaging the injected CO₂ plume and for conformance assessment (Andrew Chadwick and Noy, 2015; Chadwick et al., 2009; Ivandic et al., 2015; Lüth et al., 2015).

The conventional approach for CO₂ saturation inversion from seismic data involves forward seismic modeling to find the relationship between CO₂ saturation and seismic attributes (such as traveltime, reflector spectral composition or amplitude changes) with the CO₂ saturation values in model domain then inverted to achieve the best fit between the observed and modeled changes in seismic attributes (Meadows, 2008). Chadwick et al. (Chadwick et al., 2004) estimated the total injected volume of CO₂ at the Sleipner Field by inverting the recorded seismic data using the relationships between CO₂ layer thickness with seismic wave reflection amplitude and traveltime differences, separately. Meadows (Meadows, 2008) improved the inversion work at Sleipner by simultaneous inversion of amplitude and traveltime differences. These inversion schemes rely on rock physics modeling and seismic modeling (often 1D approximations) to establish the relationships between the seismic attribute changes with the underlying pressure and fluid saturation changes. Use of simple 1D approximations requires the assumption that the relationships between the seismic attributes and the CO₂ plume thickness established from the 1D modeling are applicable over the entire 3D geological models, not just the location for which the 1D model is designed. Furthermore, such models do not consider seismic noise, and the inversion results are inherently nonunique. Meadows and Cole (Meadows and Cole, 2013) conducted pressure and CO₂ saturation inversion using P- and S-impedance changes calculated from processed 4D seismic data at the Weyburn Field, Saskatchewan. They found that it was more difficult to invert for CO₂ saturation changes than for effective pressure changes and the inversion results for CO₂ saturation changes were noisier, relative to those for pressure changes. Gunning and Glinsky (Gunning and Glinsky, 2004) developed a trace-based inversion routine for conducting Bayesian seismic inversion using Markov Chain Monte Carlo methods. The inverted stochastic model parameters include the fluid type and saturation in each layer. However, the assumed prior distribution of the input model parameters and the layer times initialized in the prior layer stack model can substantially affect the inversion results.

Machine learning techniques, such as Artificial Neural Network (ANN) and Support Vector Machine (SVM), which find patterns in data to make predictions based on statistics, have been used in various fields of environmental science (Bertin et al., 2013; Reid et al., 2015; Saghafi and Arabloo, 2017). Reading et al. (Reading, 2015) used the outputs from machine learning supervised classifications prior to geophysical inversion, which improves detail in the inverted 3D geological models. Liu (Liu, 2017) proposed a machine learning workflow for seismic quantitative interpretation using the Bayesian-based Support Vector Regression, unsupervised Self-Organizing Map and supervised Support Vector Classification algorithms.

Machine learning techniques have also been used to invert for seismic velocity models directly from raw seismic data (Araya-Polo et al., 2017; Lin and Wu, 2018). In the work by Araya-Polo et al., velocity analysis semblance panels (location-specific seismic data processing products designed to increase the resolution of interpretabel signal in the presence of background noise) at multiple commom midpoint locations were used as input features in the Deep Neural Network (DNN) algorithm to produce seismic velocity models. Those previous studies used machine learning methods to invert for the seismic velocity model, which can be considered as a proxy or representation of the resulting seismic image of the CO₂ plume when applied to carbon sequestration sites. To our knowledge, there has been no study that considers inverting to infer CO₂ saturation directly from seismic data using machine learning techniques.

Our objective in this study is to use a large collection of labeled forward models to infer leakage and supercritical and gaseous CO₂ saturation levels at deeper and shallower depths along a leaky well from sparse 2D synthetic surface seismic, downhole pressure, and aqueous total dissolved solids (TDS) data using machine learning algorithms. In this analysis, we compare the performance of multiple machine learning algorithms and identify important features in the machine learning models using statistical methods. The impact of seismic noise on the performance of the machine learning classifier is also investigated.