Research & the Operational Service
The relationship between the peer-reviewed Bikos (2024) methodology and the live prediction system.
The Research Foundation
Seismikon's prediction methodology is grounded in peer-reviewed work: Bikos, A.N. (2024). Seismic Nowcasting: A Systemic Artificial Neural Network Predictive Model. Geoinformatics & Geostatistics: An Overview, 12:4.
The published research describes natural-time analysis, sliding-window feature extraction, an LSTM architecture, the Earthquake Potential Score, and retrospective validation. Its principal Greece dataset is the National Observatory of Athens (NOA) catalogue of events above magnitude 2.0 from 1950–2024; USGS ANSS ComCat was co-utilized as a supplementary global public catalogue for proof-of-concept and to improve catalogue completeness. The paper's four-parameter output covers latitude, longitude, focal depth, and magnitude; occurrence time-frame estimation is discussed separately. Download the peer-reviewed paper (PDF).
From Paper to Production
Translating a research methodology into a live operational service involves engineering decisions not fully specified in the original paper:
- Data source: the current S.A.F. uses California catalogue data extending back to 1932, magnitude-specific catalogue branches, and live regional monitoring; this differs from the published research's principal NOA Greece dataset and supplementary USGS ANSS ComCat use.
- Update cadence: predictions are issued on a defined schedule; the paper describes a batch analysis rather than a real-time pipeline.
- Evaluation protocol: the paper's retrospective analysis uses internal validation; the operational service uses a publicly-published, versioned evaluation protocol that is fixed before predictions are issued.
- Architecture and output: the current S.A.F. adds GNNs alongside LSTMs, uses dual inverse sliding windows, and works with a five-element output including time; these are current developments, not claims about the 2024 paper.
Where They Agree
The research and service share natural-time preprocessing and EPS concepts, but the current S.A.F. has evolved beyond the published configuration. It uses LSTM and GNN models, Dual Sliding Windows moving inversely along the same time axis, and a 5-tuple predictive output (longitude, latitude, focal depth, magnitude, and time). The 2024 paper describes an LSTM-based four-parameter output (latitude, longitude, focal depth, and magnitude), with time-frame estimation discussed separately. Current developments are versioned and disclosed on the Research Methodology page.
Where More Work Is Needed
The paper demonstrates retrospective validation. The operational service is the first large-scale prospective test of the methodology in real-time conditions. Results may differ from retrospective figures due to non-stationarity in the seismic system, distributional shift between training and deployment data, or aspects of the pipeline not fully captured in the paper.
Seismikon's scientific co-founder (Dr. Bikos) is directly involved in the operational service, and any significant divergence from paper predictions prompts methodology review. Seismikon's seismic research is active and ongoing, with further research and scientific dissemination in development.