Skip to content

Cross-Region Generative Smoke Test

Purpose

Exercise the intended transfer path end to end:

  1. Pretrain the shared Transformer representation on substantial synthetic multimodal sequences.
  2. Continue representation training on Japan ISEE VLF and seismic sequences.
  3. Fine-tune the event head on Italy data, holding out the latest mature Italy week.
  4. Emit selected probabilities over fixed Italy cells and render them beside the actual held-out events.

The first CPU run completed with 45 Japan self-supervised training windows, 3 Japan held-out reconstruction windows, 4,240 Italy training rows, 1,023 Italy held-out rows, 3 actual M>=2.5 events, and 19 generated cells. These counts confirm pipeline wiring only; they are not evidence of useful earthquake prediction.

This is a pipeline and visualization test, not a prediction experiment. It must not be presented as evidence of earthquake forecasting ability.

Data Policy

The core target remains Italy-scoped. Japan ISEE data and derived features are used only for scientific research, with source provenance and the research-use restriction preserved. Japan is an intermediate representation-learning domain, not an additional Italy label source.

Training Stages

  • Synthetic: masked sequence reconstruction plus optional synthetic event-list supervision using the existing seismic, direct-avalanche, piezo/VLF, and summary branches.
  • Japan: self-supervised continuation on native Moshiri CDF sequence features. Do not use Japan target labels while the observed VLF coverage remains sparse.
  • Italy: supervised fine-tuning on fixed 1.5-degree Italy cells with seismic history, VLF image features, astronomy features, and explicit missing-data masks. Every qualifying INGV event in each target cell/horizon is retained in a two-slot target tensor; each slot contains latitude, longitude, magnitude, and presence. The occurrence head selects a small event budget; the coordinate/magnitude head emits continuous event-list rows.

The encoder and modality adapters are retained between stages. The Italy event head is reinitialized or fine-tuned only after the Italy feature schema is checked against the checkpoint schema.

Holdout

Use the latest mature Italy VLF-anchored week. Exclude its target horizon and all later rows from training and calibration. The current 2026-07-29 capture is pending and must not be used as a held-out label until its seven-day horizon matures. Select thresholds using earlier Italy rows only.

Output

Write a stable event-list CSV containing event time, continuous latitude/longitude, magnitude proxy, probability, and provenance. The fixed cells remain an internal training scaffold; the rendered red crosses are coordinate-head outputs and must not be described as cell centres.

Report the historical-rate and always-negative controls, train/test class counts, selected threshold, and missing-modality rates beside the image. Do not report the map as a forecast.

Gates

The smoke test is successful when all stages run without schema leakage, Japan data remains provenance-labelled, the Italy holdout is chronological, and the map contains both actual and model-derived layers. It fails as a model-evidence test if the holdout is one-class, if Japan labels leak into Italy targets, if thresholds are selected on the held-out week, or if coordinate outputs collapse to a single location or leave Italy before boundary projection.