How AI Is Used
LSTM neural networks and GNNs learn spatio-temporal dependencies in earthquake sequences - here is what that means in practice.
The Role of AI in Seismikon
Seismikon's current S.A.F. uses Long Short-Term Memory (LSTM) neural networks and Graph Neural Networks (GNNs) to learn temporal and spatial relationships in historical earthquake sequences and generate forward predictive outputs.
The AI is one component of a pipeline that also includes natural time preprocessing, a Dual Sliding Window technique (two windows moving inversely along the same time axis to increase dynamic convolutional sensitivity), and the EPS calculation. These models do not run directly on raw seismic waveforms; they run on structured features derived from the earthquake catalogue.
Why LSTM Networks and GNNs for Seismicity
Earthquake sequences have long-range temporal and spatial dependencies: the likelihood of a large event may be influenced by activity patterns months or years earlier. Standard feedforward networks struggle with this because they have no memory of past inputs. LSTM networks are specifically designed to retain and selectively use information from long sequences.
- Long-term dependencies: LSTM memory cells retain information from many time steps back - critical for seismic cycle patterns.
- Spatio-temporal learning: LSTMs and GNNs learn correlations between earthquakes at different locations and times.
- Multi-scale patterns: the Dual Sliding Window input exposes the model to both short-term precursory signals and longer-term background trends.
- Sequence output: Connectionist Temporal Classification (CTC) training allows the model to produce sequences of predictions rather than single-event forecasts.
What the Model Predicts
The current S.A.F. works with a 5-tuple predictive output: predicted longitude, latitude, focal depth (km), magnitude, and time. Each prediction also carries a model confidence score and the EPS value at prediction time. The current public service uses a 48-hour target window.
The current S.A.F. uses California seismic catalogue data with historical coverage extending back to 1932, supplemented by magnitude-specific catalogue branches and continuous live regional monitoring.
Limitations of the AI Approach
Machine learning models learn from historical patterns. If the seismic system enters a regime genuinely unlike the training data - an entirely new fault segment activating, a rare large-magnitude event with no historical analogue - the model's predictions may be less reliable. This is a known limitation of all data-driven seismic forecasting approaches.
The AI is a tool for surfacing patterns in data; it is not a physical model of fault mechanics. See Limitations & Open Questions for the full account.