The Quest For Earthquake Prediction – an AI problem?

776

A research team in Palo Alto seeks an AI technique to find precursor signals in the weeks ahead of major earthquakes – and to save lives.

One hundred and fifty years ago we couldn’t predict the weather.  Seventy-five years ago we couldn’t predict tornados or hurricanes.  At the time, the established scientific communities doubted prediction of these weather phenomena was possible.  In each case, advancements in instrumentation, communication, modeling, and computation chipped away at the problem until the forecasts became reliable – and ultimately, commonplace.  Earthquakes remain one of the last natural disasters yet to be predictable.  As before, the established earthquake community openly doubts we will ever be able to successfully provide short-term forecasts for large seismic events.

QuakeFinder seeks to change that

In the weeks proceeding past earthquakes, various electromagnetic (EM) behaviors have been observed and captured.  Based on the theory that stressed rock releases a unique form of electrical energy, a growing community of researchers worldwide is attempting to characterize these behaviors as precursors to seismic activity and enable an earthquake forecasting system.  Imagine the lives that will be saved if successful.

QuakeFinder is a privately-funded Humanitarian R&D project based out of Palo Alto, California.  Since 2005, this small team of satellite engineers, geoscientists and technicians has developed and deployed 165 remote sensor stations collecting EM data along fault lines in California, Peru, Chile, Greece, Indonesia and Taiwan.  The sensors, magnetometers, detect extremely slight changes in the earth’s local magnetic field.  These perturbations are brought on by all things EM – lightning, solar storms, machinery, electric fences, cars, trains, power lines, …and pre-earthquake pulses.  Most are noise sources but past earthquakes have shown large increases in unique, nano Tesla-level unipolar pulses believed to be precursory behavior from the stressed subsurface rock.

QuakeFinder has now collected data (GPS time-stamped at a dense 50 samples/sec) containing nearly 1000 earthquakes from the sensors.  The number is highlighted as previous short-term earthquake forecasting attempts by other methods have suffered from a lack of data, and have more than once resulted in public panic due to false alarms.  Given this large data set, algorithm development and prospective testing can provide a statistical basis by which to someday issue forecasts.

With this rich data set in hand, the QuakeFinder team has turned their attention to the detailed data analyses, signal processing and forecasting algorithm development – all in an effort to find the pre-earthquake pulses and patterns.  NASA even recently assisted by co-sponsoring a crowdsourcing campaign named Quest For Quakes.  QuakeFinder provided masked data sets containing earthquakes from which competitors attempted to develop forecasting techniques.  The two-week competition resulted in 2 algorithms that demonstrated statistical significance between the pre-earthquake EM activity and seismic events.  One algorithm was a relatively standard signal processing technique to uncover anomalous behavior above a running average.  The other was an AI method.  The quick competition did not result in a home run ‘start-predicting-earthquakes algorithm’, but definitely provided strong evidence of precursory behavior – and that AI may be able to find it.

The QuakeFinder team has expended considerable effort preparing the data and data sets and understanding the noise sources and other peculiarities.  And a recent partnership with Altiscale, Inc. has provided the cluster computing/Hadoop platform on which to make massive computational runs.  So, together with a handful of carefully selected research partners both domestic and international, advancements are being made on many fronts toward cracking this yet-unsolved problem.  But help is needed.  QuakeFinder is seeking a (no-cost) partner with expertise in time-series, low signal/noise data, Artificial Intelligence and/or Machine Learning.  This partner would provide dedicated resources and work arm-in-arm with the QuakeFinder team to advance the understanding of earthquake electromagnetic precursory behavior, enable an earthquake forecasting system, and save lives.

Interested parties please contact dcoughlin@quakefinder.com