Why Earthquakes Are So Hard to Warn About

After the earthquake today, I saw many posts on Facebook asking why there was no warning. So I’d like to give a short explanation, from the point of view of a physics lecturer who used to work on earthquakes, of why this is hard.

First, (1) the source of an earthquake is usually very deep underground, more than 10 km down, and we can’t really monitor that depth yet. (2) Earthquakes come from the movement of the Earth’s crust, which is hard to predict. Energy (tectonic stress) slowly builds up in the rock, and when there is enough, the crust moves. Many variables are involved, which makes prediction difficult. It’s a bit like snapping the plywood chopsticks you get at a noodle shop: sometimes they break quickly, sometimes slowly. (3) There are still no warning signs we can use for prediction. On some faults, stress builds up for hundreds of years before anything moves. On others it takes only months (the 1992 Big Bear Earthquake).

But to be accurate, warning systems do exist. They just can’t warn very far in advance, perhaps on the order of 10 seconds, depending on how far you are from the epicenter. This is because earthquakes produce two kinds of waves, the P-wave (primary) and the S-wave (secondary). When the crust moves, both waves are created at the epicenter. The P-wave is a longitudinal wave: particles move along the direction of the wave, pushing the ground forward and back. The S-wave is a transverse wave: particles have to move side to side (shear), which takes more energy than pushing them forward and back. So the P-wave arrives before the S-wave.

The S-wave is the one that shakes the ground surface hard. We don’t feel the P-wave, and most earthquake early warning (EEW) systems rely on exactly this: if they detect the P-wave, they warn that the S-wave is coming. Dogs and cats can actually feel the P-wave. You may have seen clips of a cat bolting in a panic five or six seconds before the shaking arrives. It felt the P-wave first.

So what can you do with a warning of only 5 to 10 seconds? You might think it’s not enough to run out of the building, and you’d be right. What you should do is Drop, Cover, Hold On: get under a table and wait until the shaking stops. Other services can also reduce damage with 5 to 10 seconds of notice, such as braking trains, parking the disk heads in a data center, or having a building shut its gas valves automatically.

Predicting exactly how big the shaking will be after the P-wave, and in how many seconds it will arrive, can’t be done that precisely today. In research there are the GMPE (Ground Motion Prediction Equation) approaches, which model these waves from local parameters and use them to forecast, and the machine learning approaches, which learn from large amounts of earthquake wave data.

Another place AI can help is generating earthquake waves for designing and testing building structures. This is work I did during my Ph.D. (generative AI from before the ChatGPT era). Why does it matter? Earthquake data mostly follows a power law: high-magnitude events are rare, while low-magnitude events around magnitude 1 to 2 are extremely common. The problem is that it is the high-magnitude events that bring buildings down, and we don’t have much data on them. So the project used ML to generate synthetic earthquake waves, where we could specify the magnitude and the distance from the epicenter.

In short, the model used is a Generative Adversarial Network. There are two small models inside, fighting each other. One creates earthquake waves that look as real as possible, starting from random noise, and sends them to the other. The other has to tell whether each wave is real or fake. You let them fight for a long time, and in the end you take the first model, the generator, and use it to produce realistic earthquake waves. If you’re interested, the full paper is here: https://pubs.geoscienceworld.org/ssa/bssa/article-abstract/112/4/1979/613199/Data-Driven-Synthesis-of-Broadband-Earthquake

Next, suppose a system like this detects an earthquake. Then what? You need a cell broadcast service that can send a message to every phone in the area. In Thailand, the mobile operators have actually already tested this system, but it hasn’t been switched on for service. So this time, the fastest earthquake early warning ended up being: you had a friend in Myanmar who messaged you, or you went and checked Twitter.