Small Tremors Came Before Nearly Every Injection-Triggered Quake in a Western Canada Study
Earthquake prediction has a long record of not working. Decades of hunting for a reliable precursor, a swarm of small shocks, a change in groundwater, a shift in the ground's electrical properties, have left seismologists with no way to say that a particular fault will break next week rather than in forty years.
Earthquakes that people cause are a different case, because something about them is under human control. When fluid is pumped into rock at an oil or gas site, the added pressure can unclamp a fault that was already close to failing. The pumping can be slowed. It can be stopped. So whether these earthquakes give any warning is not an academic question; it is the assumption underneath the rules that govern the pumping.
A peer-reviewed study published in Science on Aug. 27 puts a number on it. Bei Wang of the Zhejiang University of Technology, Honn Kao of the Geological Survey of Canada and colleagues built an enhanced catalog of seismicity in northeastern British Columbia and western Alberta covering 2014 to 2024, pulled from it 77 injection-induced sequences whose largest shock reached local magnitude 3 or more, and counted how many had been preceded by smaller tremors. Ninety-two percent had.
Those smaller tremors are the raw material of what regulators call a traffic-light protocol: a rule that lets injection continue while the earthquakes recorded nearby stay small, obliges the operator to change something once they cross a first threshold, and halts the work at a second. Foreshock activity, as the paper puts it, is a key consideration in such protocols, and the processes that produce foreshocks at injection sites remain poorly understood.
A figure that high rests on two choices a reader should keep in view: how small an earthquake the network can hear, and how close in space and time a tremor has to be before it counts as a foreshock rather than as background. Move either and the percentage moves with it.
The team's answer to the first was to rebuild the record. The catalog deposited with the paper, open under a CC BY license, combines standard network detections with a machine-learning detector that rereads continuous recordings for signals a conventional trigger would slide past. It brings the British Columbia half down to a completeness of roughly magnitude 0.7 to 1.0, depending on which part of the Montney gas play you are in, far below anything a person standing on top of it would feel. The Alberta half comes from the provincial regulator's open earthquake dataset. Both spreadsheets can be downloaded.
In this part of Canada most of that injection is hydraulic fracturing, though the paper's own language throughout is the broader one: fluid injection, and injection-induced earthquakes.
That leaves the 8%, roughly one sequence in twelve that reached magnitude 3 with nothing smaller in front of it. Where no foreshock arrives, a scheme that watches for foreshocks has nothing to watch. That last step is a reading of the numbers rather than a finding of the paper's, and it carries a qualification: traffic-light protocols also respond to the size of the events they record, the mainshock included, so a missing precursor removes one input and not the whole apparatus.
The larger part of the paper is about the harder question underneath: how one of these ruptures gets started at all. The sequences sort into three nucleation models. In the first, fluid drives slow slip on a patch of fault that has already been weakened. In the second, the same slow slip occurs on a patch that is still intact. In the third there is no slow slip to speak of: one small earthquake loads the next, which loads the next, in a cascade that ends in the mainshock. Fluid is central to all three, enabling both aseismic slip, movement too smooth to radiate a wave anyone can record, and the transfer of stress from one event to another before the main rupture.
Which of the three a given fault follows is not arbitrary, and that is where a percentage turns into something usable. How many foreshocks a sequence produces, and where and when they cluster, track the interplay of three things: the injection itself, the seismogenic index of the formation, a measure of how readily a particular rock converts injected fluid into earthquakes, and how close the fault already sat to slipping. Those differ from place to place, so foreshock behavior does too. The authors' closing implication is that monitoring should be spatially conditioned: tuned to how productive the foreshocks are in a given place, rather than applied as one uniform rule across a basin.
Ground motion is already known to vary sharply over a few kilometers, and a threshold set for a fault that rumbles first will not fit one that stays silent.
None of this amounts to a forecast. Knowing that most of these sequences begin with smaller tremors is not the same as knowing which smaller tremor is the beginning of one, and the paper is explicit that the processes behind induced foreshocks are still poorly understood. What the work adds is a sense of where a warning is likely to arrive at all, and where a rule that assumes one will be waiting for something that does not come.