Machine Learning Assisted Interpretation: Integrated Fault Prediction Extraction Case Study from the Groningen Gas Field, Netherlands

Seismic fault detection using machine learning techniques, in particular the convolution neural network (CNN), is becoming a widely accepted practice in the field of seismic interpretation. Machine learning algorithms are trained to mimic the capabilities of an experienced interpreter by recognizing patterns within seismic data and classifying them. Regardless of the method of seismic fault detection, interpretation or extraction of 3D fault representations from edge evidence or fault probability volumes is routine. Extracted fault representations are important to the understanding of the subsurface geology and are a critical input to upstream workflows including structural framework definition, static reservoir and petroleum system modeling, and well planning and de-risking activities. Efforts to automate the detection and extraction of geological features from seismic data have evolved in line with advances in computer algorithms, hardware, and machine learning techniques. We have developed an assisted fault interpretation workflow for seismic fault detection and extraction, demonstrated through a case study from the Groningen gas field of the Upper Permian, Dutch Rotliegend; a heavily faulted, subsalt gas field located onshore, NE Netherlands. Supervised using interpreter-led labeling, we apply a 2D multi-CNN to detect faults within a 3D pre-stack depth migrated seismic dataset. After prediction, we apply a geometric evaluation of predicted faults, using a principal component analysis (PCA) to produce geometric attribute representations (strike azimuth and planarity) of the fault prediction. Strike azimuth and planarity attributes are used to validate and automatically extract consistent 3D fault geometries, providing geological context to the interpreter and input to dependent workflows more efficiently.

Incorporating dwelling mounds into induced seismic risk analysis for the Groningen gas field in the Netherlands

In order to inform decision-making regarding measures to mitigate the impact of induced seismicity in the Groningen gas field in the Netherlands, a comprehensive seismic risk model has been developed. Starting with gas production scenarios and the consequent reservoir compaction, the model generates synthetic earthquake catalogues which are deployed in Monte Carlo analyses, predicting ground motions at a buried reference rock horizon that are combined with nonlinear amplification factors to estimate response spectral accelerations at the surface. These motions are combined with fragility functions defined for the exposed buildings throughout the region to estimate damage levels, which in turn are transformed to risk in terms of injury through consequence functions. Several older and potentially vulnerable buildings are located on dwelling mounds that were constructed from soils and organic material as a flood defence. These anthropogenic structures are not included in the soil profile models used to develop the amplification factors and hence their influence has not been included in the risk analyses to date. To address this gap in the model, concerted studies have been identified to characterize the dwelling mounds. These include new shear-wave velocity measurements that have enabled dynamic site response analyses to determine the modification of ground shaking due to the presence of the mound. A scheme has then been developed to incorporate the dwelling mounds into the risk calculations, which included an assessment of whether the soil-structure interaction effects for buildings founded on the mounds required modification of the seismic fragility functions.

Defining the usable bandwidth of weak-motion records: application to induced seismicity in the Groningen Gas Field, the Netherlands

Seismic hazard and risk analyses are increasingly tapping into the previously underused resource of local weak-motion records. This is facilitating the development of local- or even application-specific models for the characterisation of earthquake ground motion. In turn, this offers the opportunity to derive non- or partially non-ergodic models and significantly reduce bias and uncertainty. However, weak-motion data, while carrying important information about local earthquake source, path and site effects, are susceptible to noise. We show that high-frequency noise has a record-, or region-specific, impact on pseudo-spectral acceleration (PSA). This impact depends on the shape of the records’ Fourier amplitude spectrum (FAS): PSA from moderately to highly damped ‘soil’ records (e.g. Groningen, the Netherlands) is much less susceptible to high-frequency noise than PSA from weakly damped ‘rock’ records (e.g. Eastern North America). We make use of simulated ground motion records to develop a parametric model for the lower usable period of PSA (Tmin). The model accounts for the impact of high-frequency noise on PSA, conditional on easily measured parameters characterising the shape of a record’s FAS. We then present a workflow, describing processing undertaken for records of induced seismicity from the Groningen gas field. The workflow includes the definition of maximum and minimum usable frequencies and periods of FAS and PSA, respectively. As part of the workflow, we present an approach that considers multiple estimates of Tmin. These include the parametric model and, additionally, record-specific hybrid simulations that artificially extend or modify time series’ FAS beyond the noise floor to assess subsequent impacts on PSA.

Physics-Based Estimates of the Maximum Magnitude of Induced Earthquakes in the Groningen Gas Field

<p>The induced seismicity in the Groningen gas field, The Netherlands, has led to intense public concerns and comprehensive investigations. One of the main challenges for assessing future seismic hazard in the Groningen gas field is to estimate the maximum possible earthquake magnitude (Mmax) that could be induced by gas extraction. Previous methods are strongly rooted in empirical and statistical approaches that are inherently limited by the scarcity of data. Here, we combine a physics-based dynamic rupture model based on the 3D theory of fracture mechanics with field-based and lab-based constraints to estimate Mmax in the Groningen gas field. If earthquakes in the reservoir have a rupture depth extension constrained by the reservoir thickness, the largest earthquakes should develop a large aspect ratio (longer horizontally than vertically). The model is thus an extension of the 3D theoretical rupture model on long faults with uniform stress and strength developed by Weng & Ampuero (2019), in which we have incorporated spatial heterogeneities, such as along-strike variable fault width, depth-dependent initial stresses and friction properties. The essential parameters that control rupture propagation and earthquake magnitude are the stored elastic energy and the fracture energy. Our method requires estimates of the stored elastic energy on reservoir faults as a result of the stresses induced by differential reservoir compaction during depletion. The fracture energy is constrained by laboratory experiments and theoretical frictional models. Coupling physics-based rupture models with field and lab observations provides an estimate of Mmax in the Groningen gas field and serves as a practical step toward physics-based seismic hazard assessment for other gas fields in the world.</p><p> </p><p>Citation:</p><p>Weng, H. and J. P. Ampuero (2019). "The Dynamics of Elongated Earthquake Ruptures." Journal of Geophysical Research: Solid Earth.</p><p><br><br></p>

Do spatial and/or temporal b-value variations exist within the Groningen induced earthquake catalogue?

<p>We present a data-driven analysis to derive whether statistically significant spatial and/or temporal Gutenberg-Richter b-value variations exist within the induced earthquake catalogue of the Groningen gas field. We utilize the method developed by Kamer and Hiemer (2015; J. Geophys. Res. Solid Earth, 120, doi:10.1002/2014JB011510 ) which is based on optimal partitioning using Voronoi tessellation, penalized likelihood, and wisdom of the crowd philosophy. Our implementation derives both the magnitude of completeness and the b-values simultaneously. The magnitude of completeness is computed with the maximum curvature method with a correction applied to avoid bias due to catalogue incompleteness. Finally, following Marzocchi et al. (2020; Geophys. J. Int. 220, doi: 10.1093/gji/ggz541) the b-values computed are corrected for bin size and small sample sizes.</p><p>In a first step we have limited the analysis to spatial variations in the b-values. A significant advantage of the approach taken is that it is feasible to also derive b-values in regions of very low data density. We will show that a statistically significant variation in b-values is obtained. Very low b-values (b<0.8) are observed in the central-northern part of the gas field. However, in the west near the production cluster Eemskanaal (EKL) and in the east near the city of Delfzijl significantly higher b-values (b>1.1) are observed. A Kolmogorov-Smirnov test of frequency-magnitude distributions for the two areas obtains a p-value of 1.5 10-13 and 2.3 10-12 for the EKL region and Delfzijl regions, respectively, rendering the difference more than statistically significant at the 99% confidence level.</p><p>In a second step we extended the spatial analysis to a spatial-temporal analysis. The results of the analysis show that the Groningen earthquake database is too small to derive meaningful spatial results for the full Groningen gas field based on multiple random temporal nodes.  We divided the dataset in two almost equal datasets: both containing roughly 50% of the data and of comparable spatial resolution. Spatial analysis of these two subsets of the catalogue shows a significant decrease of the b-values in the central and southern regions. Particularly in the western EKL region the b-value decreases from 1.2 to 0.92. The decrease is close to significant at the 90% confidence level. The northern region exhibits comparable low b-values in both periods. As the data in the first decade is primarily concentrated in the northern region, we have attempted to assess the spatial b-value here in the period prior to 2005. We find the high b-value area is significantly smaller and the minimum value is higher (b = 0.96 pre-2005 versus b = 0.88 post-2012). The difference is significant only at the interquartile level, but the model resolution is low.</p><p>Based on our results, we could conclude a spatial and temporal variation in b-value is observed. However, despite our efforts to limit bias in the derivation, variations could still result from the presence of a truncation. Hence, we will extend the current analysis by a comparable analysis assuming a constant b-value and estimating the corner magnitude of a taper truncation.</p>

Relating spatial-temporal b-value patterns of induced earthquakes of the Groningen gas field to differential stress changes due to reservoir extraction

<p>The Gutenberg-Richter law describes the frequency-magnitude distribution of seismic events where its slope, the 'b-value', is commonly used to describe the relative occurrence of large and small events. Statistically significant b-value variations have been measured in laboratory experiments, mines, and various tectonic regimes (Wiemer & Wyss, 2002). An inversely proportional dependency of the b-value on the differential stress has been observed across different scales (Amitrano, 2003; Schorlemmer et al., 2005). Layland-Bachmann et al. (2012) have shown that this could explain the observed pattern of induced seismicity spatial-temporal b-value variations in Enhanced Geothermal Systems. In our study, we look for a similar relation applied to the Groningen gas field in the Netherlands.</p><p>It is well known that the poroelastic changes in differential stress during gas extraction are influenced by the offset of the reservoir layer across the fault. Recently, Jansen et al. (2019) and Lehner (2019) proposed an analytical solution for stress changes on offset faults due to reservoir depletion. In a parallel study, we extended this solution to include the development of aseismic slip under slip weakening and the derivation of the onset of seismic slip.<br>We utilize this formulation to derive the onset of seismic slip on theoretical faults of variable fault offset, dip, and reservoir thickness. Subsequently, we map our theoretical faults onto the pre-existing faults in the Groningen gas field, deriving fault segment-specific depletion levels at which the segment would become seismically active. We then simulate reservoir depletion conditions over time and assign an event magnitude to fault segments that move past their seismic activation depletion. To assign a magnitude, we use the observation that b-values are inversely proportional to differential stress, which is governed by the pore pressure depletion. Hence, we assume a simple inverse linear relation with pore pressure depletion. Each event magnitude is then randomly drawn from the probability density function of the Gutenberg-Richter distribution with the b-value assigned.<br>We aim to compare the obtained catalogue and its b-value distribution both in time and space to the observed event-size distribution of the Groningen gas field as derived by Muntendam-Bos and Güdük (EGU abstract 2021).</p>