| name | elfquake-synthetic-modeling |
| description | Use when training, comparing, diagnosing, or selecting ELFQuake synthetic PyTorch models, including multimodal sequence models and missing-modality ablation checks. |
ELFQuake Synthetic Modeling
Run from the repository root with the project virtual environment active. Keep models CPU-compatible and preserve swappable model backends.
Usual Order
- Train current synthetic baselines:
./scripts/train-synthetic-torch-model.sh
./scripts/train-synthetic-sequence-model.sh
- Run sequence comparisons and selection checks:
./scripts/sweep-synthetic-sequence-model.sh
./scripts/matched-sequence-comparison.sh
./scripts/repeat-sequence-training-seeds.sh
./scripts/train-sequence-full-regime.sh
./scripts/train-sequence-full-balanced.sh
./scripts/train-tiny-patch-transformer.sh
- Summarize and diagnose model outputs:
./scripts/compare-model-runs.sh
./scripts/compare-real-synthetic-models.sh
./scripts/estimate-model-scale.sh
./scripts/diagnose-sequence-models.sh
./scripts/summarize-sequence-selection.sh
- Check missing-modality behavior:
./scripts/test-sequence-missing-modalities.sh
Use an 80 percent training, 20 percent held-out test split unless a script has a documented time-based split. Interpret synthetic runs as engineering validation only; do not imply real earthquake prediction skill.
After material model changes, update docs/report.md, docs/model-candidates.md, and docs/next-actions.md as appropriate.
