SRNL Earthquake Probability Team 3 0xDE4DB33F

Inspiration

For this project, we wanted to examine how we could improve existing models for earthquake probability prediction to inform decision-makers of the human risks in the aftermath of major earthquakes.

What it does

This project calculates the probability of an earthquake with at least a 2.0 magnitude happening at a point in a given time window with previous seismic activity and geological data as input, using a classical-quantum hybrid approach.

How we built it

The first step in improving a given baseline model was adding a wide variety of data streams, in particular using a physics-driven calculated metric for the expected earthquake rate of a given point given surrounding seismic activity, Epidemic-Type Aftershock Sequence (ETAS), as well as geographic features like distance to coast, elevation, and distance from fault line to the data set. These geographic parameters were largely found using public-domain data from USGS. Next, we had to determine which parameters added to our predictive capabilities, and which were simply noise. We did this by analyzing the information gain in bits of our models using different parameter configurations compared to the given baseline model. Once we had an idea of what data streams were valuable, perhaps the most important part of our workflow, we worked on improving our quantum machine learning algorithm in the slower but more accurate quantum simulator.

Challenges we ran into

The biggest challenges we ran into was understanding what data sources to implement within our predictive model. While we had access to tidal, solar, and geomagnetic data, we found these to simply add noise to our data set and decrease its predictive capabilities. Instead, publicly available geographic features like distance to coast, distance to fault lines, and elevation, along with physics models like the ETAS model were found to be much higher correlated with earthquake frequency. The biggest challenge we faced actually implementing the quantum processing side of things was figuring out how to manipulate input data to best utilize the limited amount of qubits we had access too.

Accomplishments that we're proud of

We are very proud of our model having an Information Gain (IG) of 0.3366 bits compared the baseline model's judge IG of just 0.1451 bits in our quantum simulator when taking ETAS, distance to fault, distance to coast, and depth of earthquakes as input parameters. That model also produced in ECE value of 0.049.

What we learned

We learned that more parameters does not always result in more accurate models, as longitude, latitude, tidal information, solar information, and geomagnetic information all added noise to our ML models and reduced the information gain of our model. Instead, our path to success focused on a few highly relevant parameters and pre-processing to nudge our model towards expected physical relationships (like the ETAS model).

What's next for SRNL Earthquake Probability Estimation

To improve the estimation model, one could look into how hydrological patterns influence earthquakes. Studies have suggested that fluidic pore pressure, and to a lesser degree rainfall, can affect earthquake frequency. If a reliable dataset were found to model these hydrological characteristics, they could be a useful data stream for improving the model's predictive capabilities. It would also be worth expanding on the risk assessment for policy makers, potentially by having a simple risk assessment product based on population density and odds of a significant earthquake in the area occurring.