New research highlights the potential to predict powerful earthquakes months in advance by using machine learning to detect early signs of seismic activity. However, the effectiveness and ethical implications of such prediction technology remain a matter of debate.

A study by scientists at the University of Alaska Fairbanks suggests that the public could receive earthquake notifications days or months in advance by identifying previous low-level tectonic activity across large areas. The analysis focused on two major earthquakes in Alaska and California.

The work was supervised by Associate Professor Tarsilo Girona of the UAF Geophysics Institute.

Geophysicist and data scientist Girona investigates the antecedent activity of volcanic eruptions and earthquakes. Geologist Kyriaki Drimoni from the Ludwig-Maximilians-Universität Munich, Germany, is a co-author of the study. The detection method, based on machine learning, was published on August 28. Nature Communication.

“Our paper shows that advanced statistical techniques, including machine learning, have the potential to detect precursors of large-magnitude earthquakes by analyzing datasets obtained from earthquake catalogs,” Girona said.

The authors wrote a computer algorithm to search data to detect anomalous seismic activity. Algorithms are sets of computer instructions that teach a program to interpret data, analyze it, and make informed predictions or decisions.

Case Studies: Anchorage and Ridgecrest Earthquakes

They focused on two large earthquakes: the 2018 7.1 Anchorage earthquake and the 2019 Ridgecrest, California, 6.4-7.1 earthquake sequence. They found that each of the two earthquakes was preceded by approximately three months of anomalous low-magnitude regional seismicity over approximately 15% to 25% of south-central Alaska and southern California.

Their research shows that disturbances preceding major earthquakes are often recorded with seismic activity below magnitude 1.5. The Anchorage earthquake occurred at 8:29 a.m. on November 30, 2018, with its epicenter approximately 10.5 miles north of the city. It caused significant damage to many buildings, as well as some roads and highways.

Results and implications

Using their data-trained program, Girona and Drimoni found that the probability of a major earthquake occurring within 30 days or less, using the Anchorage earthquake as an example, jumped to about 80% about three months before the Nov. 30 earthquake. The probability rose to about 85% just a few days before the event. They reached similar conclusions about the probability of the Ridgecrest earthquake sequence for a period starting about 40 days before the earthquake sequence began.

Girona and Drimoni propose a geological reason for the low activity of the precursors: a significant increase in pore water pressure within the fault. Pore fluid pressure refers to the fluid pressure in the rock. If the pressure is sufficient to overcome the frictional resistance between the rock blocks on either side of the fault, the high pore fluid pressure could potentially cause fault slip.

“The increased pore water pressure in faults that causes strong earthquakes changes the mechanical properties of the faults, which leads to non-uniform fluctuations in the regional stress field,” Drimoni said. “We assume that these irregular oscillations control the anomalous, previous low-magnitude seismicity.” Machine learning has a major positive impact on earthquake research, according to Girona.

“Modern seismic networks produce huge data sets that, if properly analyzed, can provide valuable information about the precursors of seismic events,” he said. “This is where advances in machine learning and high-performance computing can play a transformative role, allowing researchers to identify meaningful patterns that could signal an impending earthquake.”

Problems with earthquake prediction

The authors say their algorithm will be tested in near-real-time situations to identify and resolve potential problems with earthquake prediction. They add that the method should not be applied to new regions without first training the algorithm on the historical seismicity of the region. Creating reliable earthquake predictions is “a very important and often controversial dimension,” Girona said.

“Accurate forecasting has the potential to save lives and reduce economic losses by providing early warnings that allow for timely evacuations and preparations,” he said. “But uncertainty in earthquake prediction also raises important ethical and practical questions.”

“False alarms can cause unnecessary panic, economic disruption and loss of public confidence; miscalculated forecasts can have disastrous consequences,” he said.