Inspiration

Earthquakes are difficult to forecast because seismic activity is inherently uncertain, spatially distributed, and highly imbalanced. Most locations experience no earthquake during a given forecasting window, while the events that do occur can have significant consequences.

For the Earthquake-Q Challenge, our team, Schrodinger's Cats, wanted to explore a question beyond simply training another classifier:

Can we combine traditional earthquake science, modern machine learning, and quantum computing to create better probabilistic seismic forecasts?

Instead of treating earthquake forecasting as a simple yes-or-no classification problem, we focused on estimating a meaningful probability of future seismic activity.

Our task was to estimate

[ P(\text{at least one } M \geq 2.0 \text{ earthquake within 100 km in the next 30 days}) ]

for each geographic grid cell and forecasting period.

This made calibration especially important. A model that predicts an earthquake correctly occasionally but produces unreliable probabilities is not necessarily useful for risk forecasting.

What We Built

Earthquake-Q is a hybrid seismic forecasting framework that combines earthquake-domain features, classical machine learning, probabilistic calibration, and experimental quantum machine learning.

Our pipeline transforms historical earthquake information into spatial and temporal features describing the recent seismic state surrounding each forecast location.

We then compare several modeling approaches, including:

Classical earthquake and statistical baselines
Machine-learning models
Engineered seismic features
Probability calibration methods
Quantum-derived feature representations
Quantum machine-learning experiments
Hybrid classical-quantum models

The final system generates a probability for every scored location and forecasting window rather than simply outputting a binary earthquake/no-earthquake prediction.

Across the challenge evaluation structure, this meant generating predictions