ELFQuake is a research project examining whether Extremely/Very Low Frequency radio observations can augment seismic and astronomical data enough to support useful predictive modeling of earthquakes.
The core hypothesis is that natural ELF/VLF radio anomalies may contain signal that is not present in seismic event history alone. At least one study suggests this may be a viable approach.
We wish to exploit more modern machine learning/AI techniques to create a predictive model based on the transformer architecture. A key consideration is the availability of real-world data for training. The plan is to build a system (based around an avalanche model) with similar characteristics to the geological system and use this to generate synthetic data for baseline training. After validation of this stage, real-world data will be used for fine-tuning.
To be useful any claims must be demonstrated against reproducible seismic-only baselines, held-out time periods, and multimodal ablations.
Status: the first reproducible synthetic-to-real transfer suite is now running on Italy-scoped data. On a chronological 80/20 holdout for seven-day M2.5+ events in fixed geographic cells, synthetic pretraining reached 0.680722 balanced accuracy and 0.307054 precision, below the historical spatial-rate baseline at 0.686013 and 0.343915; no earthquake prediction capability is claimed.
The current focus is on rolling chronological validation, expanding the synthetic corpus, and collecting enough Cumiana VLF data to obtain both positive and negative target classes. The latest transfer suite contains 79,976 synthetic records and 190 weekly spatial samples; VLF and astronomy remain explicit missing inputs in the real holdout because their validated historical overlap is not yet long enough. See report.md, model-comparison.md, and docs/2026-07-15_elfquake-progress.md.
The project also maintains a separate Japan research track using ISEE Moshiri CDF VLF spectra and Japanese seismic catalogs. Japan data is used for scientific comparison and signal-shape analysis only; it is not mixed into the Italy operational scope. See vlf-japan-isee.md and japan-earthquake-vlf-window.md.
This is a diagnostic map from a randomly selected held-out week: blue markers are actual INGV events and red markers are independently predicted cell centres. It is not an operational forecast.
This work was initially prompted by the tragedy of the 2009 L'Aquila earthquake. Around the same period I was aware of developments in Deep Learning and coincidentally had stumbled on material related to natural radio signals occurring as precursors to seismic events (see vlf.it). I made a start on research and began a blog about it : ELFQuake (which I wound up using for general blogging). At the time it seemed possible but very difficult. But since then predictive models have improved in leaps and bounds, meaning that prediction has a much better chance of working. Not only that, but the ability to delegate much of the coding work to an intelligent assistant means that the difficulty in building the system has been drastically reduced.
- INGV seismic event acquisition, normalization, and Italy/Central Italy training windows, currently backfilled from 2024-01-01 through 2026-07-07.
- Cumiana VLF live spectrogram capture through systemd, with pixel-derived image features.
- Astronomical and space-weather archive connectors and normalization.
- Prospective VLF-anchored feature rows with pending target labels.
- A CPU PyTorch self-supervised autoencoder over real Cumiana VLF image sequences.
- A modality-aware masked-patch Transformer evaluation comparing synthetic, real VLF, sequential-transfer, joint, and random initialization over repeated seeds.
- Dependency-light logistic checks and a CPU PyTorch tabular MLP for synthetic aligned rows.
- A first CPU PyTorch GRU sequence model over synthetic avalanche and piezo/VLF sensor tensors.
- A CPU PyTorch patch Transformer scaffold for synthetic sequence engineering checks and event-list target pretraining.
- Summary comparison, sequence sweep, and missing-modality scripts for model diagnostics.
- A compact real-vs-synthetic comparison wrapper for central-Italy seismic baselines and synthetic sequence reports.
- Larger-model scale checks for row count, class balance, sequence size, and CPU-only model limits.
- Real Cumiana VLF image features materialized into the same sequence shape as synthetic piezo/VLF inputs.
- Real prospective aligned model inputs scaffolded for all-Italy and central-Italy rows; real deep-model fine-tuning wrappers refuse to train until both target classes exist.
- VLF model feature roles that allow synthetic piezo/VLF analogue data to exercise the same PyTorch VLF path before real labels mature.
- Synthetic-inlier transfer diagnostics that train on real-like synthetic piezo/VLF descriptors and evaluate reconstruction against held-out real VLF descriptors.
- Mixed real/synthetic VLF descriptor alignment with CORAL loss and centroid/random/full synthetic controls.
- Label-free real VLF anomaly smoke forecasts while supervised real labels remain blocked.
- A deterministic trial weekly event-list forecast that combines historical INGV rates/locations, real VLF context, astronomy context, and synthetic avalanche spatial priors into a downstream-ready CSV/JSON contract.
- A first swappable learned scorer trained on synthetic aligned rows, with the same weekly event-list CSV contract and historical INGV rate calibration metadata.
- Synthetic event-list target and model heads for occurrence, count, magnitude, and centroid, currently useful as an engineering adapter but not yet temporally robust.
- A Transformer target adapter and sweep wrappers for training against the richer h6 synthetic event-list target table.
- Drift-aware synthetic validation and shorter-episode simulation scaffolding to reduce one-run lifecycle bias; the first aggressive 3000-step profile fixes target-rate drift in a small probe.
- A larger warmed synthetic batch wrapper for CPU-only Transformer pretraining data expansion.
- A reproducible synthetic-to-real transfer suite with real-only and synthetic-pretrained ablations, rolling time holdouts, training-only grid selection, and final evaluation against a historical spatial-rate baseline.
- A 20,000-step CPU simulation episode at seed
4300, expanding the dense synthetic transfer corpus to 79,976 records and 190 weekly spatial samples across four long episodes. - Sandpile simulation with separate seismic-like avalanche outputs and piezo/VLF analogue outputs.
The complete command sequence for acquisition, preprocessing, simulation, model training, evaluation, and service operation is maintained in docs/steps.md.
The simulation is an artificial mountain-like grid where broad background loading is combined with repeated localized stress at fixed point sources. As slopes become unstable, small avalanches redistribute height to neighbouring cells. The aim is to generate synthetic sequences that are close enough in shape to real-world seismic data to be useful as training data for a deep learning system, especially before enough matched seismic, VLF, and astronomical data is available.
It also includes piezo-like sensors that watch quartz-like susceptible regions near failure and produce the VLF/WAV analogue channel. Direct seismic-like event data is kept separate and is derived from avalanche/toppling behavior.
Default outputs use data/derived/sim/mountain_256x256_seed42_10000 as the prefix. The normal piezo image is *.piezo_vlf_summary.png from *.piezo.csv; the direct seismic event analogue is *.avalanche_events.csv. The older FFT diagnostic is opt-in with RUN_FFT=1.
The event-map demo projects avalanche-derived locations over an Apennine-style Italy belt and uses point size for synthetic magnitude.
This is a simplified stress-and-release analogy, not a geological model. Its value depends on whether the generated data has useful structural similarity to real observations. Good performance on simulated avalanches would only show that the tooling can learn synthetic patterns; real claims still require held-out seismic, VLF, and astronomical data.
- Overview
- Documentation Index
- Processing Graph
- Next Actions
- Command Steps
- Forecast Interface
- Success Criteria
- Development Environment
- Source Inventory
- Multimodal Feasibility
- Systemd Service
- Initial Model Trials
- Model Comparison
- Model Scaling Requirements
- Sandpile Simulation
- Japan ISEE VLF
- Japan Earthquake/VLF Window