Significance for secure CO₂ storage of earthquakes induced by fluid injection

Given the similarities involved between the injection of large volumes of produced water into deep saline aquifers and proposed CCS operations, the nascent CCS industry can learn much from these incidents. For example:

• What (if anything) controls the maximum earthquake magnitude that could be induced by injection?
• Do events typically occur solely within, or below, or above the injection interval? This has obvious implications for storage security; an event in the caprock would be of greater concern with respect to caprock integrity than an event several kilometres below the target interval.
• How are events distributed laterally? This again has implications for storage security because an event close to the injection point is likely to have a greater impact on wellbore casing integrity, while an event at significant distance may never actually reached by injected CO₂, even though it is triggered by the resulting pressure increases. An important question when developing CCS sites is determining a 'radius-of-influence', defining the lateral extent of site characterization studies and baseline monitoring efforts carried out prior to injection. The maximum extent of induced seismic events might provide a minimum bound for this radius.

While there have been a number of reviews of injection induced seismicity in recent years, these specific questions have not necessarily been addressed with respect to their implications for CCS. In this paper we consider the case examples listed in table 1, considering the above questions, with specific focus on their implications for secure CO₂ storage.

Fluid injection into an aquifer will cause an increase in pore pressure to propagate from the injection point into the formation. During CO₂ sequestration, this pore pressure front may propagate well in advance of the actual CO₂ plume (e.g., Birkholzer et al 2009). The effect of an increase in pore pressure will reduce the effective normal stresses ($\sigma _{n}^{^{\prime} }$) acting on a fault plane, leaving the effective shear stresses (τ') unchanged. The Mohr–Coulomb criterion dictates that a fault will fail when

$$\begin{eqnarray*}&&\tau ^{\prime} \geqslant m\sigma _{n}^{^{\prime} }+c,\end{eqnarray*}$$

where m is the coefficient of friction, and c is the cohesion of the fault plane. Therefore an increase in pore pressure created by fluid injection into a reservoir, by reducing effective normal stresses, has the potential to trigger failure on a fault, thereby generating earthquakes.

It is also possible to induce earthquakes on faults that do not have a direct hydraulic connection to the injection zone (for example faults that are kilometres below the injection interval). Changes in the stress field created by inflation of the target reservoir can propagate through the layers of rock above and below, and can potentially change the effective stresses acting on such faults by a sufficient amount to trigger a seismic event.

For obvious reasons, the faults most likely to be triggered by injection are those optimally oriented in the present day stress field such that τ' is maximized and $\sigma _{n}^{^{\prime} }$ is minimized. The natural occurrence of intra-plate seismicity indicates that throughout the Earth's crust the in situ state of effective stress is such that a subset of optimally oriented faults are close to this critical failure criterion (e.g., Zoback and Healy 1992). That is to say that the crust is critically stressed (e.g., Zoback et al 2002), which in turn means that small changes in stress are capable of triggering earthquakes (e.g., Main 1996).

Determining how commonly large volume fluid injection wells induce seismic activity is difficult to establish, as it depends on the detection thresholds of the various seismic networks available, as well as the minimum injection volume considered. For example, there are over 100 000 Class II injection wells in the USA, of which only a small handful have been associated with seismic events, implying that injection-induced seismicity is a rare occurrence.

However, the recent increase in injection-induced events in the USA (Ellsworth 2013) has lead to a reassessment of this assumption. For example, Frohlich (2012) re-analysed Earthscope USArray data in the Barnett shale region, identifying 8 groups of seismic events that were potentially associated with injection wells. These smaller (2.0 ⩽ M ⩽ 3.0) events had previously gone undetected on national seismic arrays. While there are thousands of fluid disposal wells in the region, Frohlich identified that all eight seismicity clusters were associated with wells injecting at rates of 24 000 m³ month⁻¹ or greater, of which there were only 161 in the region of interest. Although events of magnitude 2.0 ⩽ M ⩽ 3.0 are too small to cause damage at the surface, if they were to occur in the caprock of a CCS operation they would pose questions regarding sealing integrity. In addition they have the potential to be felt by the public, which is likely to lead to local opposition that can pose a risk to ongoing operations.

Given the possibility of inducing seismic events during CO₂ sequestration, a key question is to ascertain what is the maximum earthquake magnitude (M_MAX) that could be induced by CO₂ injection into the subsurface.

The mechanism typically invoked for induced seismicity is that a small stress or pore pressure change occurs on a pre-existing fault that, being optimally oriented within the current stress field, is close to a critical failure threshold. Therefore small tectonic stress changes in the recent past may have triggered seismic events on the same (or at least similar) faults. Given this, it is possible that the historical seismic magnitudes of a region could provide some indication of the maximum magnitude induced by injection.

In some cases, M_MAX during injection appears to correlate with nearby (within 25 km) historical earthquakes. For example, the M4.5–4.7 Lake County earthquake occurred in 1943 approximately 15 km from the epicentre of the M5.0 Perry earthquake (Wesson and Nicholson 1986); an M4.2 earthquake was recorded in 1973 within 25 km of Trinidad prior to the 2001 M4.6 event; the Enola earthquake swarm, with a maximum of M4.4, occurred in 2001 less than 20 km from the Guy–Greenbrier fault (Rabak et al 2010).

In some cases, past earthquakes can be identified, albeit at greater distance. For example, the aforementioned 1943 Lake County earthquake is within 50 km of the Ashtabula M4.3 earthquake, and in 1998 an M5.1 earthquake was recorded in western Pennsylvania, within 50 km of the Youngstown M4.0 earthquake. For the remaining case studies, there is either no historical record of seismicity in the area, or recorded earthquakes within 50 km are at least 1 magnitude unit smaller. This indicates that while historical seismic magnitudes can be useful as an indicator for M_MAX, this is not a metric that can be relied upon completely. This might be in part because historical records of seismicity are not sufficiently long or detailed enough to fully characterize the natural seismicity of an area.

Alternatively, relationships have been developed that relate the maximum magnitude induced by injection to either the volume of fluid injected (McGarr 2014) or the length-scale of an ellipsoid of rock stimulated by injection (Shapiro et al 2011). McGarr (2014) modifies an equation first developed in McGarr (1976) to relate the cumulative seismic moment released due to a subsurface volume change, adapting it to describe M_MAX as a function of the injection volume, ΔV:

$$\begin{eqnarray*}&&{{M}_{{\mbox{MAX}}}}=\left( 2/3 \right)({{{\rm log} }_{10}}(\mu \Delta V)-9.1),\end{eqnarray*}$$

where μ is the rock shear modulus.

Alternatively, Shapiro et al (2011) argue that M_MAX will be controlled by the size of the zone stimulated by injection. The maximum size of the fault that can be ruptured scales to this length, and M_MAX can then be calculated from this length using generic fault scaling relationships (e.g. Abercrombie 1995). By treating the stimulated volume as an ellipsoid, Shapiro et al (2011) find that M_MAX can be scaled to the length of the smallest principal axis, L_MIN:

$$\begin{eqnarray*}&&{{M}_{{\mbox{MAX}}}}\approx 2{{{\rm log} }_{10}}\left( {{L}_{{\mbox{MIN}}}} \right)-{\mbox{1}}{\mbox{.}}\end{eqnarray*}$$

Shapiro et al (2011) use spatial distributions of microseismic events to determine the dimensions of the ellipsoid stimulated by injection. We perform similar calculations for our case examples. We begin by defining discrete clusters of events using a density-based spatial clustering algorithm, and compute the dimensions of an ellipsoid containing 95% of the events of the cluster that has the largest number of events, selecting L_MIN as the length of the smallest principal axis of this ellipsoid. We have only performed this calculation for case examples where data from local seismic networks is available, providing more accurate event locations (see section 2.3 and supplementary material), and where the largest cluster contains more than 15 events.

Figure 2 shows observed seismic moment released by the largest earthquake(s) of each case study, and the volumes injected at the time each event occurred (black circles). Also shown is the McGarr (2014) relationship (black dashed line), and the values computed via the Shapiro et al (2011) method for each case study using observed event locations (red squares).

Zoom In Zoom Out Reset image size

Given the relative parsimony of the above models, they both appear to provide a reasonable envelope for observed values of M_MAX. It should be kept in mind that these relationships describe an envelope of maximum potential magnitude that could be induced for a given injection volume, rather than the actual magnitude that will be induced. Hallo et al (2014) modify the maximum moment released by a seismic efficiency factor, S_eff.

We envisage that installing monitoring networks to record microseismicity during early stages of injection may allow S_eff to be estimated, providing better constraint on the expected M_MAX for a given injection volume. Similarly, such a microseismic monitoring network would allow a 'stimulated ellipsoid' to be determined, which is required to determine M_MAX following the Shapiro et al (2011) method.

Whether or not an event induced during CO₂ injection might pose a risk of leakage will be determined by the event hypocentre. For example, an induced event triggered well below the target storage formation will not pose as much of a risk to leakage as an event in the sealing caprocks. Similarly, the pulse of increased pressure created by CO₂ injection will extend far ahead of the actual CO₂ plume. As such, a fault some distance from the injection well could be re-activated by the pressure pulse, but never pose a leakage risk because it will not be reached by the zone of CO₂ saturation. Equally, active faults near to the wellbore may pose a risk to wellbore integrity through shearing, whereas a more distant fault may not.

At this juncture, therefore, accurate event locations, and especially hypocenter depths, are required. Many of the events discussed so far have been detected only by national and/or regional seismic networks, where the nearest seismic station may be many tens of kilometres away. Such networks are not able to provide locations of sufficient accuracy for our purposes. In the following sections we therefore consider only events recorded on local seismic networks. The specific data sources that we use, and estimates of event location errors, are listed in the supplementary material.

In many of the case examples considered, local networks were deployed after the occurrence of the largest event. As such it is possible, even likely, that seismicity was occurring prior to these deployments, and that these events have gone undetected. Therefore our data are probably biased in that for many cases they may not capture the earliest stages of induced seismicity.

In figure 3(a) we plot histograms of event depths, scaled such that 0 on the 'y'-axis corresponds to the deepest injection depth at each site. Vertical red lines at the centre of each dataset illustrate the extent of injection intervals, while dashed green lines show the approximate basement depths, where this information is available.

Zoom In Zoom Out Reset image size

It has been suggested that injection either directly into basement rocks, or into basal layers with no underlying confinement, is a key factor in injection-induced seismicity (Person et al 2013). Dashed green lines in figure 3(a) show the approximate basement depths, where such data is available. In a number of cases (RMA, PER, ASH, and YOU) injection does indeed occur in close proximity to the basement, with the majority of events located in the basement rocks. Furthermore, for GUY and PRA, the majority of events appear to occur in basement rocks, even though deepest injection depths are more than 500 m above the basement.

In figure 3(b) we plot the cumulative distribution of event depths relative to the injection interval. Only 2% of events occur more than 500 m above the shallowest injection interval, while 80% of events occur at least 500 m below the deepest injection depth, and 50% of events are located more than 2.5 km below the injection interval. It is not clear whether the triggering of seismicity several kilometres below the injection interval is a result of hydraulic communication along faults or through dynamic (e.g., Kilb et al 2000) or static (e.g., King et al 1994) changes in the stress field transmitted through the rock frame.

These observations are, at first glance, encouraging for the future viability of CCS in one respect, but they also pose a challenge. The fact that the majority of events occur below the injection interval should be considered favourable in terms of storage integrity. However, our ability to mitigate induced seismicity depends in part on identifying potential faults during the site selection stage, so that they can be included in geomechanical models. Prior to injection, we anticipate that seismic reflection surveys with sufficient resolution to detect faults within the reservoir interval will be conducted. However, the resolution of such surveys is a function of depth, meaning that the detection of a fault more than 2 km below the injection zone may be a more challenging prospect.

A caveat to these observations is the potential difference between wastewater and CO₂ as injected fluids. While we consider these differences more generally in the discussion section below, with respect to earthquake depths a key potential difference is the density of the different fluids, with injected water having a similar density to the in situ fluid, while CO₂ is less dense, and therefore more likely to move upwards through overlying formations if a permeable pathway is available.

However, Mori and Abercrombie (1997) show that the increases in stress, rock strength and rock homogeneity leads to an increase in the number of large earthquakes with depth. These increases will be particularly significant moving from shallow sediments to crystalline basement rocks. We postulate that the downward migration of induced events may be a result of these factors. Similarly, Sumy et al (2014) argue that the largest event at Prague was triggered by static Coulomb stress transfer promoted by earlier events that were directly triggered by pore pressure increases. If this is the case, then the density of the injected fluid will not necessarily play a role in determining the locations of induced earthquakes.

Moreover, we note that Rinaldi et al (2014) have modelled the effects of CO₂ injection into faulted reservoirs. Even though the CO₂ was buoyant, and permeable pathways through the caprock were included in the models, they observed that the largest amounts of fault slip, and thereby it must be inferred the majority of the induced seismicity, occurred in layers bounding the base of the reservoir, rather than in the overburden.

The pulse of pressure increase produced by CO₂ injection will extend far in advance of the plume of CO₂ itself. The 'radius-of-influence', which determines the limit of site characterization and baseline monitoring studies conducted prior to injection, should be determined by the zone of pressure increase, which may extend many tens of kilometres from the injection point, rather than the footprint of the CO₂ plume, which may only extend a few kilometres.

With this in mind we consider the radius of influence for our case examples, as demarcated by the lateral distance of seismicity from the injection point. The injection rates and volumes anticipated for CCS operations exceed that of all the cases considered in this paper, so our observations should be considered to represent a lower bound estimate for the potential radius of influence of a CO₂ sequestration site.

In figure 4(a) we plot histograms of epicentral distances from the injection wells, while in 4(b) we plot the cumulative distribution. In all cases bar Perry, some seismicity occurs within 2 km of the well. The mean epicentral distance from the nearest injection well is 5 km, while 99% of events occur within 20 km. These observations imply that a radius of influence of 20 km, as suggested in previous studies (e.g. Sminchak et al 2002), is a reasonable lower bound for the radius of influence.

Zoom In Zoom Out Reset image size