POSEIDON: Building the Largest Physics-Informed Earthquake Prediction System

When my brother Fedor and I set out to tackle earthquake prediction, we knew we wanted to do something different. Too many machine learning models in seismology treat physics as an afterthought—something to validate against, not something to learn from. That never sat right with us.

Today, I'm excited to share POSEIDON (Physics-Optimized Seismic Energy Inference and Detection Operating Network), the culmination of our work combining 2.8 million earthquake events with physics-informed deep learning. You can read the full technical details in our paper on arXiv, and there's also coverage in The Ritz Herald.

Why Physics-Informed AI?

Here's the thing about earthquake prediction: we already know a lot about how earthquakes behave. The Gutenberg-Richter law tells us how magnitude and frequency relate. The Omori-Utsu law describes aftershock decay. These aren't just empirical observations—they're fundamental patterns that have held true across decades of seismological research.

So why would we build AI systems that ignore all of that knowledge?

That was our starting question. Instead of treating these physical laws as external constraints, we embedded them directly into the neural network architecture as learnable parameters. The model doesn't just have to respect physics—it actively learns to optimize within physical principles.

Three Problems, One Solution

POSEIDON tackles three interconnected challenges that researchers usually study separately:

Aftershock prediction - Will this earthquake trigger a sequence of follow-up events?

Tsunami assessment - Does this earthquake have the potential to generate a devastating tsunami?

Foreshock detection - Is this earthquake a precursor to something larger?

These aren't isolated problems. They're deeply connected aspects of seismic behavior, and treating them as a unified multi-task learning problem actually improves performance across the board.

The Dataset That Made It Possible

Building POSEIDON required something that didn't exist: a comprehensive, ML-ready earthquake dataset with proper feature engineering. So we built the Poseidon dataset—2.8 million seismic events spanning 30 years of global observations.

Every event includes:

• Core seismic properties (magnitude, depth, location, timing)

• Pre-computed energy features based on the Gutenberg-Richter relation

• Spatial grid indices for efficient processing

• Quality metrics from reporting networks

• Multi-scale temporal context (7, 30, and 90-day windows)

We've made it publicly available on HuggingFace because progress in this field requires collaboration. If you're working on earthquake prediction, seismic hazard assessment, or physics-informed ML, this dataset is for you.

The Results

I'll be honest—I was nervous about whether respecting physics would compromise accuracy. The conventional wisdom in ML is that constraints limit what the model can learn.

We proved that wrong.

POSEIDON achieved state-of-the-art performance across all three tasks. For tsunami detection, we hit an AUC of 0.971 despite tsunamis representing only 1.14% of events—that's extreme class imbalance. We outperformed gradient boosting, random forests, and standard CNNs.

But here's what really excites me: the learned physics parameters converged to scientifically meaningful values. Our Gutenberg-Richter b-value settled at 0.752. The Omori-Utsu parameters reached p = 0.835 and c = 0.1948 days. The parameter values aren't arbitrary numbers—they fall squarely within established seismological ranges.

The model didn't just learn to make accurate predictions. It learned the physics.

What's Next

This is just the beginning. Fedor and I are already working on integrating real-time waveform data and extending the system to continuous probabilistic forecasting. We're exploring crustal stress transfer physics and how that might improve prediction windows.

I've put together a video walkthrough if you want to dive deeper into the architecture and see the results in detail.

The big vision? Earthquake early warning systems that combine the pattern recognition power of deep learning with the reliability and interpretability that comes from respecting physical law. Systems that scientists can trust because they understand not just what the model predicts, but why.

Try It Yourself

All the code and data are open source. Whether you're a seismologist curious about ML or an ML researcher interested in physics-informed approaches, I'd love to see what you build with this.

And if you're working on similar problems—physics-informed models for climate, materials science, fluid dynamics, whatever—reach out. The principles we developed for POSEIDON apply far beyond earthquakes.

The future of scientific ML isn't choosing between physics and data. It's making them work together.