May/June 1999


In recent Opinions in Seismological Research Letters Paul Silver has suggested that earthquake prediction is not possible, whereas Lowell S. Whiteside suggests that it is. Earthquake prediction means many things. There are at least three types of (conventional) earthquake prediction. Deterministic prediction is where the behavior before the earthquake (the stress interactions with the surrounding rocks, say) can be calculated (by whatever techniques are available) so that the time, place, and magnitude of the future large earthquakes can be estimated within well defined windows. Statistical prediction is where seismicity in the past can yield estimates of seismicity in the future. The third and most common type is where some key precursory phenomenon or a group of phenomena indicate that a large earthquake is imminent. I suggest that all three types cannot predict time, place, and magnitude of a future large earthquake. It is complexity and heterogeneity that prevents it each time.

(1) The difficulty as far as deterministic prediction is concerned is that the Earth is immensely complex and heterogeneous at all scales. Heterogeneous features include stress fields, geology, and water saturation, as well as the dimensions, frictional properties, orientation, and focal mechanisms of fault planes, among many others. These range in size from submillimeter grains to faults and photo-lineaments of thousands of kilometers (more than nine orders of magnitude). If the stress-induced interaction of every grain-boundary crack in a stressed fluid-saturated rockmass before a large earthquake could be evaluated, then I dare say it might be possible in principle to predict deterministically the time, place, and magnitude of the future large earthquake. Since the strained rockmass before a large earthquake is at least a million cubic kilometers (and could even be hundreds of millions of cubic kilometers for the largest earthquakes), the required interaction involves possibly up to some 1024 millimeter-diameter grains each with grain-boundary cracks. Even if I am exaggerating by a factor of 1012, say, the remaining 1012 plus interactions (a fault plane of single grains, say) are still far too great for any conceivable computation and far too great for conceivable evaluation of initial starting conditions around each grain. So although in principle deterministic prediction may be possible, in practice, computing and foreknowledge clearly fail by some very substantial orders of magnitude.

This type of behavior is common in even quite simple nonlinear systems where negligibly small differences in initial values can lead to substantially different results. It is just not possible to get sufficient resolution in initial conditions to predict the final result. This is known as deterministic chaos, and it is one of the extraordinary properties of such systems that deterministic equations can lead to effectively chaotic results. Even more extraordinary, these chaotic systems may impose some form of order on the behavior which may be common across totally different systems and can in some instances be calculated. It is suggested that stress-induced manipulations of fluid-saturated microcracks is one such calculable system (see below). Such systems are commonplace and embrace a huge range of phenomena including weather systems, turbulent flow, convection currents, geomagnetic reversals, cardiac fibrillations, dripping taps, chemical reactions, percolation systems, clustering on freeways, and water drops on window panes, among many others. It is not surprising that fluid-rock interactions in the crust are also such a nonlinear, potentially chaotic system.

(2) Statistical analysis of seismicity in the past in order to attempt to predict future behavior again fails because of complexity and heterogeneity. It is clear from seismicity (and from episodes of orogeny, say) that earthquake occurrences vary enormously and cluster in time and in space (more deterministic chaos), so that the seismicity of any area varies with time on all scales from seconds to millions of years (again ~10 orders of magnitude). Assumptions of statistical stability can be expensive. The Parkfield Earthquake Prediction experiment, where a sixth earthquake was expected after a sequence of five M=6 earthquakes repeated regularly every twenty-two years, is already at least five years beyond its sell-by date. Previous seismicity is clearly important, and if one wanted a really aseismic place to live then history would suggest that Saharan Africa, much of Arctic Canada, and Arctic Asia would appear to be seismically safe, although very uncomfortable (is there a message?). Elsewhere, even in the centers of what are presumed to be stable continental shields, there are occasional large earthquakes which would not be expected from historical seismicity.

My favorite example of unpredictability was near my home village. The U.K. has low to moderate seismicity. Probably the most damaging British earthquake in the past 400 years was the 1884 ML 4.6 Colchester earthquake, which had a shallow focus and produced high intensities. Many houses lost chimneys and roof tiles and some churches were damaged, but direct casualties were limited to a few people injured possibly by falling tiles. Apart from a small foreshock a few days previously and doubtful accounts of aftershocks, there is no other known earthquake within 50 km of Colchester in at least 500 years. Not easy to predict!

(3) The difficulty as far as the third type of prediction, seeking and evaluating precursors, is again the immense complexity and heterogeneity of the Earth at all scales. Although large numbers of different precursors have been observed before earthquakes, their occurrences are extremely variable. No two earthquakes are exactly the same, and it is extremely unlikely that there will be some common recognizable unique phenomenon or phenomena that signals the time, place, and magnitude of an imminent earthquake. The fact that we have been searching for such earthquake precursors since John Milne in the 1880's, and that no large earthquake has been successfully and unquestionably predicted (even with hindsight), suggests that the hopes for consistent precursors can be discounted. Only an incurable optimist could believe that there is still a "magic" precursor packed with information about future earthquakes awaiting discovery.

So what can be done? Paul Silver suggests up to 2,000 "plate boundary deformation sites" at 10 km intervals over California each with GPS receivers, borehole strainmeters, borehole three-component seismometers, and borehole accelerometers. This would certainly be a powerful tool for investigating plate boundary deformation, but there is no direct evidence that it would be any use at all in mitigating earthquake hazards. We would not know what to look for in such an enormous flood of data. Strain evaluated from GPS observations is deformation of the largely destressed unconsolidated surface layers of the Earth, which may have behavior substantially different from the strain at the depths of earthquake generation. In addition and more importantly, earthquakes are the result of the release of stress, not strain. Strain is the primary cause of stress, but strain can be released aseismically by creep (and other processes such as APE; see below). To recognize that earthquakes are imminent requires that stress is monitored at close to seismic generation depths in the crust. Two recent developments in understanding and instrumentation make such monitoring feasible.

A new understanding of rock deformation shows that the splitting of seismic shear waves observed in almost all crustal rocks is controlled by the same stress parameters that control deformation. The (prefracturing) deformation mechanism of rock is the result of anisotropic poro-elasticity (APE) with fluid migration along pressure gradients between neighboring grain-boundary cracks and low-aspect ratio pores at different orientations to the stress field. Consequently, the immediate effect of any deformation, however small, is to modify the geometry of the intergranular fluid-saturated microcracks by reorganizing the distribution of crack aspect ratios. This modifies the anisotropic elastic symmetry and means that changes in stress-induced deformation can be directly monitored by analyzing changes in shear-wave splitting. In the past, such changes have been recognized before only a small number of earthquakes.

The reason for the small number of observations is that monitoring changes in shear-wave splitting before large earthquakes requires very stringent source-to-recorder geometries almost immediately above small-magnitude seismicity in the neighborhood of the epicenter of a larger earthquake. Until recently, such changes had only been observed (with hindsight) on four occasions around the world when by chance these geometrical conditions were met.

The rapidly expanding SIL seismic network in Iceland combined with strong but localized seismicity has meant that changes in shear-wave splitting are now seen routinely (but again with hindsight) before larger earthquakes in a small area of southwest Iceland. Such changes were also seen and recognized before the 1996 volcanic eruption beneath the Vatnajökull ice cap. Had we had more confidence in our results, we could have predicted that something was going to happen, although at that time we would not have expected an eruption. It was the first five months of data from the first seismic station in Iceland which had been examined for shear-wave splitting, and we did not believe our luck!

Although we have demonstrated that shear-wave splitting can identify increases in stress before earthquakes, small-scale seismicity is far too intermittent to be relied on for routine monitoring. What is needed is some form of controlled source seismology. Standard exploration techniques, such as reflection surveys and vertical seismic profiles, although producing reliable and consistent observations of shear-wave splitting, lack resolution because of the scattering and attenuation of the near-surface rocks. They also do not readily sample the crucial range of angles of incidence, between 15° and 45° to the vertical plane of the approximately parallel cracks, where stress-induced changes of shear-wave splitting are principally observed. What is needed are crosswell observations between three wells below about 1 km in a relatively homogenous rockmass within a few tens of kilometers of the vulnerable site. This would need one vertical receiver well and two deviated source wells.

Even this would predict only the time and magnitude, but it would reliably indicate that stress was building up and that a larger earthquake (or in some areas an eruption) would happen sometime in the future. The longer the build-up continued the larger would be the potential earthquake. We have called this procedure stress forecasting and would argue that such a crescendo of increasing alarm could best mitigate hazard to life and property by giving time to prepare evacuation procedures, service shut-downs, and property safeguards. With a large earthquake expected, it would also stimulate the search for and more realistic interpretation of short-term precursory phenomena.

Until recently such monitoring would have been technically impossible, but suitable borehole recording and shear-wave sources systems for such stress-forecasting sites are now being developed. Drilling deep deviated wells in preferably crystalline rock would be extremely expensive (cheaper, more simple, source-to-receiver geometries have now been identified), but once the monitoring site had been established, the monitoring itself would be comparatively cheap, depending on the frequency of sampling required. It would be an appropriate insurance policy for a sufficiently vulnerable site.

Note that stress forecasting escapes the impossibility-of-earthquake-prediction conundrum because shear-wave splitting is examining the effects of stress on the rockmass and is independent of the source of the impending earthquake. Stress forecasting merely assesses when the increase of stress in the rock will reach fracture criticality.


There has now been a successful "stress forecast" earthquake based on the changes in shear-wave splitting observed at three stations in a 70 km line in Iceland. Routine observations of changes in shear-wave splitting in one area of southwest Iceland allow the level of fracture criticality to be inferred. In October 1998, it was recognized that the behavior of shear-wave splitting, at two stations 40 km apart, indicated an increase of stress that was approaching fracture criticality in real time before the earthquake had occurred. Consequently, preliminary stress forecasts were issued to the Icelandic National Civil Defense Committee on 27 and 29 October. On 10 November 1998, it was recognized that a third station also showed changes in shear-wave splitting, and a final stress forecast was issued that there would be a M

5 earthquake soon or a M
6 earthquake before the end of February 1999. Three days later, on 13 November, a M = 5 earthquake was reported 2 km from the central station.

The uncertainty in the forecast is because both inferred rates of increase and inferred levels of fracture criticality are subject to errors due to scatter of the data, so at best an early-smaller-magnitude to later-larger-magnitude window can be defined. Industrial seismology suggests that controlled-source measurements would substantially improve the accuracy of such estimates.

Note that variations in shear-wave splitting can be used to stress-forecast the time and magnitude of future large earthquakes but cannot identify the location of the earthquake focus within the large rock volume where stress builds up and variations in shear-wave splitting are observed. However, once the earthquake had been stress-forecast, Ragnar Stefánsson of the Icelandic Meteorological Office correctly predicted the fault on which the earthquake was to occur from local geophysical and geological analyses.

Stuart Crampin
Department of Geology and Geophysics
University of Edinburgh

To send a letter to the editor regarding this opinion or to write your own opinion, contact Editor John Ebel by email or telephone him at (617) 552-8300.

Posted: 20 May 1999