Earthquake prediction has never been very precise. In fact, areas shown as reasonably safe on seismic maps have been where the world’s strongest earthquakes have hit. In 2008, China’s magnitude-7.9 earthquake killed 90,000 and injured 375,000 more. In 2010, Haiti’s magnitude-7.0 earthquake killed an unknown amount, ranging from 85,000 to much higher. In Japan, the earthquake of 2011 left 19,300 dead or missing. Not one of these locations, according to seismic hazard maps, was considered an area of risk. They’ve shown further problems with accuracy in addition to this. The fault where an earthquake struck Los Angeles, according to these maps, didn’t even exist
Once an earthquake hits, wherever that is, many things must be done. Scientists must make predictions about aftershocks before they strike, that can cause similar damage, if not more. Fortunately, scientists can usually determine the strength of these aftershocks as well as what time they’ll occur. However, like seismic maps, their area of weakness has been predicting just where they’ll be.
The good news is researchers may finally solve this problem with the use of artificial intelligence technology. Scientists believe that AI will give them insights into how earthquakes work and the ways they act, so they can warn sooner of what’s to come. In the case of an earthquake, or a tremor, each second of notice matters.
AI has already demonstrated improvements in aftershock prediction, not that it was hard to do. The old method used now that AI will hopefully replace is known as the Coulomb failure stress change model. This model works by looking at the sequence of changes in rock after an earthquake has struck. Based on this data, it determines where tremors would likely strike next. As history knows, this model has been wrong almost as often as it’s been right. In fact, the accuracy of this failure-stress approach is just slightly above chance.
This is why scientists turned to AI. To first train the machine, it was given data on 131,000 earthquakes and aftershocks that really took place. The machine processed this information quickly, over 500 times more quickly than a group of scientists can. It looked for patterns or sequences in what it was given, such as repeated motions in the movements of the ground.
It then worked to make conclusions based on what it knew. This is called deep learning, possible because its neural network is capable of a type thought much like the human brain. This allows it to compute many outcomes while considering several variables, weighing effects like a brain would do.
It works well for the complex process of earthquake prediction because it can take into account so many things. These include the power of an earthquake, the ways that plates align, and the composition of the surrounding ground. The machine can even apply additional information that is new to the field of earthquake prediction and has never been considered before. An example of this is the breaking point of different materials, like metal, based on exposure to stress.
Following training, scientists put the technology to the test. They gave the AI engine new information to use to make predictions. AI was then compared to the failure-stress model and the results were very clear. If 1 is 100% accuracy and 0.5 is the probability of chance, the failure-stress approach got 0.583. The AI system, the obvious winner, scored a 0.849.
However, the system is not prepared for real-world use just yet. It is still learning and lacks the speed to predict the first tremors that follow an earthquake. However, as it continues to learn it will continue to improve with time. Hopefully we will soon have earthquake predictions that really predict what will come.
Victoria Liset is strategic business & technology consultant to SMEs. She helps businesses improve their performance by using data more efficiently, and helping them to understand the implications of new technologies such as AI, Machine Learning, Big data, blockchain and IoT.