AI Outperforms Traditional Seismic Prediction Model
Researchers implement AI to pinpoint the location of earthquake aftershocks

One of the most important aspects of scientific research is the ability to predict. For centuries, scientists have built empirical or physical models to achieve this task. Once a model is able to explain the pattern and variability in the observed dataset, it’s used to predict outcomes under different circumstances. However, for complex data, finding the appropriate model is often the thorn in any scientists’ foot. Nonetheless, since the development of Artificial Intelligence (AI), that problem is somewhat curbed. Using thousands of data points to train itself, AI is capable of finding intricate patterns that normally misses the human eye. Now, using AI methods like deep learning, Harvard and Google researchers are able to predict the location to earthquake aftershocks – far better than existing models.

In the field of seismology, an initial earthquake is followed by several aftershocks. While scientists have devised models to forecast when they will occur, predicting the location of these aftershocks remained beyond their scope. Pinpointing those locations are vital – for a structure weakened by the initial earthquake has a greater probability of collapsing if hit by an aftershock. Aftershocks usually persist in the affected area for months, causing mild quakes that brings already-hit structures to the ground.

Now, thanks to AI, that problem might be solved. Harvard Seismologists, collaborating with Google’s AI experts, have developed a deep learning network that understands how numerous variables, including the ground type and seismic plates, interact to define the ways energy propagates in waves through the Earth’s crust. They have trained a neural network on a database of more than 131,000 “mainshock-aftershock” events. The research, published in the journal Nature on 29th August, 2018, iterates that the reason for the accuracy of the algorithm lies in its usage of two complex metrics – maximum shear stress change and the von-Mises yield criterion. They were previously applied in material sciences – to test the ‘bendability’ of copper and aluminum. By incorporating them in aftershock prediction, the AI-based algorithm was found to be more accurate than pre-existing models.

One of the most useful existing model, the “Coulomb failure stress change” was previously used to predict the location of aftershocks. On a scale of accuracy running from 0 to 1, where 1 depicts a perfectly predicting model and 0 indicates a model that fails to predict every single time – the Coulomb model usually hit a score of 0.583. This rendered the model slightly better than predicting the side of a coin, which is 0.5. Testing the prediction of the AI based model on a database of 30,000 pairs of observed data, it gave a score of 0.849.

Predicting the location of an earthquake and its aftershocks is vital to mitigate casualties during the incident. The west coast of the United States, situated on top of the San Andreas Fault, is particularly prone to seismic activities. The 1906 San Francisco earthquake, measuring 7.9 on the Richter scale, collapsed 80% of the city and resulted in 3,000 deaths. Its predicted that California has a 30% probability of being hit by an earthquake of greater magnitude by 2040. Thus, such AI-based models are critical in minimizing the loss of life and property during the incident.

However, Phoebe DeVries, co-author of the paper, has said, “We’re quite far away from having this be used in any operational sense at all. We view this as a very motivating first step.”