Scientific Research
The peer-reviewed work that Seismikon's earthquake nowcasting is built on
A methodology grounded in the scientific literature
Seismikon does not rely on proprietary black-box methods. The core methodology - natural-time analysis combined with LSTM neural networks - is fully described in a peer-reviewed publication that is freely available for download. Every claim about how the system works can be verified against the published paper.
The primary research was conducted by Dr. Anastasios N. Bikos at the University of Patras, Greece, and validated on decades of seismic catalogue data. Seismikon is the operational implementation of that methodology, extended to a live public demonstration using California seismic data.
This page provides the full citation, an overview of the methodology, and a summary of the validation results. The paper itself is the definitive reference.
Primary Publication
Seismic Nowcasting: A Systemic Artificial Neural Network Predictive Model
This paper introduces an earthquake nowcasting framework that combines natural-time statistical mechanics with Long Short-Term Memory (LSTM) neural networks to produce prospective, falsifiable seismic predictions. The Earthquake Potential Score (EPS) is derived from natural-time analysis of regional seismic catalogues and fed alongside sliding-window feature extractions into an LSTM architecture trained to forecast the 4-tuple: event latitude, longitude, focal depth, and magnitude. The methodology was originally validated on 74 years of Greek seismic data from the National Observatory of Athens and subsequently applied to the California seismic catalogue, achieving 98% precision in retrospective back-tests on earthquake sequences from 2000-2023.
Note on the 98% figure: This is a retrospective back-test result from the Bikos (2024) paper, not a live prospective performance claim. For forward-issued prediction outcomes, see the Track Record page.
About the Researcher
Research Areas
- Natural-time seismology
- LSTM neural networks for geophysical prediction
- Earthquake Potential Score (EPS) methodology
- Prospective prediction evaluation frameworks
Dr. Anastasios N. Bikos is a seismologist and data scientist at the University of Patras, Greece. His research focuses on earthquake nowcasting - the application of natural-time statistical mechanics and machine learning to prospective seismic event prediction.
Unlike retrospective pattern-matching or early warning systems, the nowcasting methodology Dr. Bikos developed produces predictions before the fact: a specific location, magnitude class, focal depth range, and time window, issued prospectively with defined tolerances. This allows independent evaluation against real seismic catalogue data, treating each prediction as a testable scientific hypothesis.
The methodology originated in work on Greek seismic sequences, using 74 years of high-quality catalogue data from the National Observatory of Athens (NOA). The same framework was then applied to the California seismic catalogue (2000-2023) to demonstrate geographic portability and to establish the California demonstration that Seismikon runs publicly.
Dr. Bikos serves as Co-Founder and Chief Scientist of Seismikon and is directly involved in the operational prediction pipeline - the same pipeline that produces the predictions published on this website.
Publications
Geoinformatics & Geostatistics: An Overview, 12:4 - Primary research paper describing the methodology implemented in Seismikon. Peer-reviewed.
Additional publications will appear here as they are released.
Methodology at a Glance
Four interconnected components form the prediction pipeline. Each is described in full in the Bikos (2024) paper. The Research Methodology page and the How It Works guide provide plain-language explanations for non-specialist readers.
Natural Time
Seismic event count replaces clock time as the unit of analysis. This event-based representation reveals patterns in the seismic cycle that are invisible in ordinary calendar time - specifically, the accumulation of stress and strain between successive large earthquakes in a defined geographic region.
Originally introduced by Varotsos et al. in the context of complex system physics. Applied to earthquake nowcasting in Bikos (2024).
Earthquake Potential Score (EPS)
A 0-100 index that quantifies how far through the seismic cycle a given region currently is. Derived from natural-time entropy statistics (specifically the variance of the natural-time probability distribution weighted by seismic energy release). A high EPS signals that the region is approaching conditions historically associated with a significant event.
Formulated in Bikos (2024). Methodology detailed in full in the paper.
Sliding Window Feature Extraction
A dynamic sliding window technique divides seismic catalogues into overlapping natural-time segments and extracts statistical features from each: earthquake count, mean and maximum magnitude, seismic energy release, spatial clustering metrics, and temporal derivatives of the EPS. These features form the input to the LSTM network.
Described as the stochastic filter component in Bikos (2024).
LSTM Neural Network
A Long Short-Term Memory recurrent network trained on historical seismic sequences learns the temporal and spatial dependencies that precede significant events. Training uses Connectionist Temporal Classification (CTC) loss to handle variable-length seismic sequences. Output is a 4-tuple forecast: latitude, longitude, focal depth, and magnitude.
Network architecture and training protocol described in full in Bikos (2024).
Validation Results
From Research Paper to Live System
The Bikos (2024) methodology has been implemented as a live, continuously running operational system. Dr. Bikos is directly involved in the pipeline that generates each prediction - this is not a third-party implementation of the paper. The key differences between the paper and the live system are:
Paper (Bikos 2024)
- Retrospective back-test on historical data
- Greece: National Observatory of Athens catalogue (1950-2024)
- California: USGS/FDSN catalogue (2000-2023)
- Results evaluated after the fact against known earthquake sequences
- Fixed training and test splits
- 98% precision (location and magnitude class)
Live System (Seismikon)
- Prospective predictions - issued before the evaluation window opens
- California public demonstration using live FDSN seismic feed
- Every prediction is timestamped before the event window
- Results evaluated by independent FDSN catalogue in your browser
- Model continuously updated as new seismic data arrives
- Live performance visible on the Track Record page
Scientific Enquiries
We welcome critical engagement from the seismological community. If you are a researcher interested in the methodology, would like to discuss the evaluation framework, or want to propose independent validation studies, please get in touch.