Monday, 20 January 2025/ Edition 42

Two 7.0-magnitude earthquakes—one in northern California last December and the other in early January in southern Tibet—have caught my attention. While geoscience informs us where earthquakes usually occur, it cannot currently predict the timing of big earthquakes in specific locations. This edition of Core Elements shares some new developments in earthquake prediction, including the application of AI tools.

Rasoul Sorkhabi

Editor, Core Elements

Geomagnetic Signatures

Menur/ Shutterstock.com

Traditional earthquake prediction methods are based on identifying and monitoring seismic precursors. One of these precursors, electromagnetic disturbances, has been actively researched over the past few years.

Here is a look at three recent studies on this topic across the globe:

Turkey: Volvach and colleagues in Physics and Chemistry of the Earth report on the 7.8 and 7.5 earthquakes that shook up Turkey on February 6, 2023.

Key conclusions: Using the wavelet transform data method, the researchers found that pronounced geomagnetic oscillations near Nurdağı occurred between 5 to 6 hours and again 25 to 30 minutes before the earthquake.

China: Polarization is the measure of alignment of magnetic dipoles within a material. Feng and colleagues in Physics of Earth and Planetary Interiors studied geomagnetic signals related to the magnitude 7.4 earthquake that hit Maduo in China on May 22, 2021.

Key conclusions: These researchers suggest a polarization method for extracting geomagnetic anomalies as earthquake precursors. They note that two polarization highs occurred at the epicenter seven months and then 15 days before the earthquake.

Andaman Nicobar Islands: Prajapati and Arora in Nonlinear Processes in Geophysics conducted fractal analysis of the vertical component of the geomagnetic field for data collected from March 2019 to April 2020.

Key conclusions: The researchers identified disturbance signals in the geomagnetic field that occurred 10, 12, and 20 days prior to moderate earthquakes in the region. This study holds promise for short-term earthquake prediction.

A message from AAPG Academy and Eliis

Earthquake Prediction by AI

Deemerwha Studio/ Shutterstock.com

AI integrations around earthquake prediction is a rapidly growing field, shrouded in constantly evolving AI jargon. Let’s look at some case studies.

Case study from Turkey: Biswas and colleagues report an AI model for earthquake magnitude prediction and spatial risk mapping in Turkey published in Decision Making Advances.

Methodology:

• The researchers used earthquake data archived by the U.S. Geological Survey in Turkey from January 2014 to August 2023. The data included earthquake magnitude, location, depth, and distance to monitoring station.

The AI involved:

• Artificial Neural Networks (ANN) algorithms were written to train the AI used in this study: Three ANN architectures, each employing one of three optimizers—RMSprop, Adam, and Stochastic Gradient Descent—were developed.

• Each ANN architecture consisted of five layers of neurons. A single neuron produced the result.

• The researchers produced three ANN-based earthquake risk maps for all of Turkey. These were color coded to indicate high to low forecast/actual percentage.

• High-risk areas such as west-central Anatolia, northwestern Marmara, and the East Mediterranean were successfully captured on the ANN-produced maps.

Key conclusions:

• All three ANNs, each using a different optimizer, performed well in post-training testing. Comparatively, ANN-1 using RMSprop optimizer showed better performance metrics avoiding overfitting or underfitting.

• The researchers suggest that using more sophisticated architectures, such as convolutional neural networks and recurrent neural networks, will enhance this field of research.

More AI Research from California

Joel Hensler/Shutterstock.com

Case study from California: Zhang and colleagues published an AI-based forecast model for earthquake in California in Geophysical Journal International.

Methodology:

• Researchers fed a spatial map of past earthquake magnitudes into a fully convolutional network (FCN) to forecast future earthquakes.

• The input data consisted of earthquakes of shallower than 40 kilometers from January 1980 until December 2020.

• Researchers then constructed an FCN-based alert level map for earthquakes greater than Magnitude 3 and compared it with that of epidemic type aftershock sequence (ETAS) model.

Key conclusions:

• The FCN output map was very close to that of ETAS; however, training and implementing the FCN model was much faster than calibrating the ETAS model to produce its forecasts.

• The FCN forecasts were overly pessimistic: they proposed earthquake occurrence probabilities larger than their empirical frequencies. This possibly indicates imbalanced data classification in the machine learning procedure.    

In another article in JGR Solid Earth, Zhang and colleagues argue that:

• ANN models “claiming superior performances” do not necessarily “surpass simple geophysical models that clearly describe the underlying physical processes.”

• When a test shows that a particular ANN model has better performance, it simply means that it is better than the reference model used. It does not necessarily mean that the tested ANN model is good.

• The best seismologically informed reference models should be used to test the AI-based prediction tests. This requires “integration of data mining experts and seismologists.”

• Spatially varying Poisson distribution informed from statistical seismology should be used as a reference model instead of the commonly used spatially uniform Poisson distribution model.

Go deeper: Read this study by Pwavodi and colleagues in Artificial Intelligence in Geosciences to learn even more about the role of AI in earthquake prediction.

Quiz of the Week

NicoElNino/ Shutterstock.com

Last week’s quiz question was: What is the earliest known mass extinction? (Think of the Precambrian.) 

Here is my answer: The first mass extinction on record occurred 2.46 billion years ago (Archean-Proterozoic boundary) during the Great Oxygenation Event, in which photosynthetic cyanobacteria spread globally and wiped out many anaerobic bacteria which cannot protect their enzymes from oxidants.   

This week’s quiz: Most people think that earthquake magnitudes are still stated on the Richter scale, but this is not true. What is the current scale used to measure earthquake magnitude? How does it differ from Richter’s scale?

Please send your response by January 23 to editorial@aapg.org (subject line: Core Elements Quiz).

👍 If you enjoyed this edition of Core Elements, consider supporting AAPG's brand of newsletters by forwarding to a friend or colleague and signing up for our other newsletters here.

➡️ Was this newsletter forwarded to you? Subscribe to Core Elements here.

✉️ To get in touch with Rasoul, send an email to editorial@aapg.org.