Using data to predict earthquakes
Natural disasters represent some of the most startling and devastating instances throughout human history. Often taking place with little to no warning and entirely outside of mankind’s control. Numerous examples of such deadly events stretch from antiquity to the modern-day. While the ability to prepare and minimize the damage has greatly improved, the speed with which action is taken to recover represents a challenge that has not been overcome.
In China, the AETA team at the Peking Shenzhen University Graduate School is determined to use AI and big data as a method to mitigate the loss of human life from earthquake disasters. By deploying a number of three-part sensory systems in a mesh grid network across geographic regions with high seismic activity, the team has created a system that gathers key seismological data before, during, and after an earthquake occurs. This information is used to generate an accurate prediction, which can allow local governments to evacuate cities, shut down high-value infrastructure, and organize relief teams days before an earthquake strikes, thereby minimizing the risk to lives and livelihoods as much as possible.
However, for leaders to act based on this information, the predictions must achieve a consistently high level of accuracy. So even after the AETA team has proven the efficacy of the system, the project still needs further improvement to be made ready for public deployment and official use.
But how will it be done?
“Building on our work means gathering experts from all over the world,” explains Dr. Yong Shanshan, Senior Engineer at the Peking University Shenzhen Graduate School. “At the same time, we wanted to motivate people. So, we brought both of these goals together by setting up a competition where teams worked to find the best way to analyze the data and submit weekly predictions, creating a healthy rivalry among the teams to be the best.”
Competition yields a new prediction model
Together, the AETA team and Capgemini’s AIE Shenzhen set up a seven-month-long competition that brought together 600 teams from 28 different countries. During the competition, every Sunday, the contestants predicted three elements of earthquakes greater or equal to 3.5 magnitudes in a target area for the following week. This was then compared against the numbers generated by the actual phenomena. All of this data was managed through a competition system that took in forecast results and real data before determining prediction deviation.
“For seven months, we had experts pushing the boundaries of what could be done with seismic data,” describes Dr. Shanshan. “The competition had everyone working even harder than usual to try and make gains in their predictive power, which ultimately gave us a variety of exciting options to choose from.”
In the end, WALKTREE, a public security solutions provider from China, was named the overall winner of the 2021 competition. This team developed a deep neural network model that divides the problem into two parts. First, the system makes a distinction between regions and generates predictions based on the approximate longitude and latitude. Secondly, the model classifies the magnitude of a predicted earthquake before transforming this forecast into a classification problem.
Once these two steps are finished, the WALKTREE method constructs a lightweight neural network model that takes the maximum, minimum, and average of the absolute values of electromagnetic and geo-acoustic data as inputs. The final model is then obtained based on training with a sample set before the team translates it into a prediction result including magnitude and its general location.
“After all of the great work that was done by the many participating teams, we ultimately awarded WAKLTREE as the winner, as their model managed the greatest growth in accuracy,” explains Dr. Shanshan. “This represents a major step forward in being able to reliably predict upcoming earthquakes and offer governments a new tool to support their citizens.”
Building towards a more protected future
By the end of the competition, the WALKTREE model had achieved an 88% daily accuracy and a 75% weekly accuracy, both a substantial improvement over the starting point of less than 30%. In addition to optimizing the prediction methodology with a customized lightweight network, the AETA team also improved the durability of the sensors to reduce the maintenance efforts required to keep the data flowing. These gains are essential to changing the approach to preparing for earthquakes in China and have the potential for a broader impact.
“At present, there is no other earthquake prediction solution based on hardware sensors and a data AI algorithm. Instead, we have to rely on real-time warning when earthquakes occur,” says Dr. Shanshan. “Our success in predicting earthquakes can significantly reduce casualties and social and economic losses such as insurance pay-outs, mining disasters, advanced maintenance, and response plans for the country’s vital nuclear, hydropower, and natural gas infrastructure.
“And, eventually, this technology has obvious potential for use in other countries that suffer from regular earthquake activity.”
Turning an eye to the future, the AETA team intends to continue driving gains in the accuracy of earthquake prediction. Following the success of the 2021 competition, the team will continue working with Capgemini and its AIE Shenzhen to host and manage regular events to attract talent to the cause and support the drive for innovation.
Someday, earthquake prediction may become a global practice helping to protect lives and property from the devastation of these natural disasters. At that point, the AETA team and its competitions will likely be regarded as a key stepping stone leading to such revolutionary change.