I really want to write up a long response to this... I used to be closely adjacent to this field, and I do ML work in other domains. ML is mostly irrelevant for this, though.
Here's the short version:
We're good at predicting where. We're still not good at when.
Understanding what segments of a fault zone have built up the most elastic strain is something we're really quite good at these days. Given historical earthquake data, GPS measurements of deformation, and knowledge of the geometry and makeup of the subsurface, we can reasonably accurately model what segments of a fault are most likely to fail next. No ML needed or wanted.
The issue with the "when" is what the precursor signals are.
The precursor signals we're most likely to be able to reliably measure involve drilling down and putting sensors on both sides of the fault zone and measuring the sonic velocity across the fault zone repeatedly. As rocks or even a fault gouge begin to fail, they have to dilate slightly so that grains can rotate into position to slip (bad explanation, sorry). This dilation causes a drop in sonic velocity. Stresses on the rock mass as a whole also have velocity signals. Lab experiments have observed this fairly reliably. We've actually set things up to measure this in the field as well -- look up SAFOD. There were some likely precursor signals (a bit contentious, but it's fair to say there were interesting signals in the data). However, this type of instrumentation just isn't practical to deploy at scale.
Other possible precursor signals involve hydrologic changes (e.g. well waters rising / changes in pore pressure) or Radon spikes, both due to the same dilation. Again, though, these are local effects and difficult to deploy a system to measure. (e.g. you need a deep borehole for the hydrologic changes, and the radon spikes are most reliably observed at depth as well)
There is one more controversial category that is easily observable: ionospheric perturbations. GPS data is sensitive to the ionosphere, and one of the things you need to do in geodetic GPS station operations is correct for changes in the ionosphere. We tend to have tons of geodetic GPS stations deployed along fault zones. Folks have noticed some very interesting ionospheric perturbations before very large earthquakes. There have been a lot of papers published on this, but it's still not too widely accepted, last I checked. The proposed mechanism (piezoelectric effect and/or dislocation glide/creep building up charge) and the observed magnitude of change don't match. It also has only been documented as a precursor around very major earthquakes (which are rare) and only in retrospect.
At any rate, tons of data don't help if it's all noise. The data that we think should have the most signal is _really_ hard to get. Things are still in the "promising, but a long ways off" state when it comes to predicting "when", and seismologists are _very_ careful to avoid doing work that could be taken as predictions due to past PR SNAFUs.
Lovely response, much more than I'd hoped for. Thank you sir.
I suppose another issue that then arises, even if we did have decent
prediction, is how to present and use predictive info. A "traffic
light" system leading up to a countdown? How not to scare people,
destroy economies with false positives and off-timing?
A strange conclusion from my study (for sound) is that where
eathquakes are concerned the next best thing to none at all is
regular earthquakes.
Here's the short version:
We're good at predicting where. We're still not good at when.
Understanding what segments of a fault zone have built up the most elastic strain is something we're really quite good at these days. Given historical earthquake data, GPS measurements of deformation, and knowledge of the geometry and makeup of the subsurface, we can reasonably accurately model what segments of a fault are most likely to fail next. No ML needed or wanted.
The issue with the "when" is what the precursor signals are.
The precursor signals we're most likely to be able to reliably measure involve drilling down and putting sensors on both sides of the fault zone and measuring the sonic velocity across the fault zone repeatedly. As rocks or even a fault gouge begin to fail, they have to dilate slightly so that grains can rotate into position to slip (bad explanation, sorry). This dilation causes a drop in sonic velocity. Stresses on the rock mass as a whole also have velocity signals. Lab experiments have observed this fairly reliably. We've actually set things up to measure this in the field as well -- look up SAFOD. There were some likely precursor signals (a bit contentious, but it's fair to say there were interesting signals in the data). However, this type of instrumentation just isn't practical to deploy at scale.
Other possible precursor signals involve hydrologic changes (e.g. well waters rising / changes in pore pressure) or Radon spikes, both due to the same dilation. Again, though, these are local effects and difficult to deploy a system to measure. (e.g. you need a deep borehole for the hydrologic changes, and the radon spikes are most reliably observed at depth as well)
There is one more controversial category that is easily observable: ionospheric perturbations. GPS data is sensitive to the ionosphere, and one of the things you need to do in geodetic GPS station operations is correct for changes in the ionosphere. We tend to have tons of geodetic GPS stations deployed along fault zones. Folks have noticed some very interesting ionospheric perturbations before very large earthquakes. There have been a lot of papers published on this, but it's still not too widely accepted, last I checked. The proposed mechanism (piezoelectric effect and/or dislocation glide/creep building up charge) and the observed magnitude of change don't match. It also has only been documented as a precursor around very major earthquakes (which are rare) and only in retrospect.
At any rate, tons of data don't help if it's all noise. The data that we think should have the most signal is _really_ hard to get. Things are still in the "promising, but a long ways off" state when it comes to predicting "when", and seismologists are _very_ careful to avoid doing work that could be taken as predictions due to past PR SNAFUs.