Processing Graph
Latest high-level configuration, including default self-supervised real VLF pretraining and the supervised real fine-tune gate.
flowchart TD
subgraph Synthetic["Synthetic simulation path"]
Sim["sandpile simulation<br/>summary/sensors/piezo/avalanche CSVs"]
SimEvents["direct avalanche event extraction<br/>*.avalanche_events.csv"]
SimWindows["hourly synthetic seismic windows<br/>synthetic_seismic_*"]
SimSeq["sequence manifests<br/>synthetic_direct_avalanche<br/>synthetic_piezo_vlf<br/>synthetic_summary"]
SimAligned["aligned synthetic rows<br/>gt0 future-lookahead target"]
Regimes["post-burn-in regimes<br/>regime-balanced split"]
DeepTrain["deep patch Transformer synthetic pretrain<br/>d_model=64, 3 layers, 4 heads"]
Checkpoint["synthetic checkpoint<br/>deep_patch_transformer_synthetic.pt"]
Sim --> SimEvents --> SimWindows
Sim --> SimSeq
SimWindows --> SimAligned
SimSeq --> SimAligned
SimAligned --> Regimes --> DeepTrain --> Checkpoint
end
subgraph Real["Real Italy-scoped data path"]
INGV["INGV seismic events<br/>Italy / Central Italy"]
VLF["Cumiana VLF image captures<br/>systemd service"]
Astro["astronomy / space-weather features"]
Norm["normalization and UTC provenance"]
Prospective["prospective VLF-anchored windows"]
RealSeq["real VLF image sequence manifest<br/>cumiana_vlf_image_sequence"]
SelfSup["default self-supervised VLF pretrain<br/>masked sequence autoencoder"]
VLFEmbedding["real VLF embeddings<br/>self_supervised/*.csv"]
DomainDiag["synthetic-to-real embedding diagnostic<br/>compare-vlf-embedding-domains.sh"]
TransferDiag["synthetic-inlier transfer diagnostic<br/>evaluate-vlf-synthetic-inlier-transfer.sh"]
MixedAlign["mixed-domain alignment diagnostic<br/>evaluate-vlf-mixed-domain-alignment.sh"]
RealAligned["real VLF-aligned rows<br/>all_italy / central_italy"]
Readiness{"real label readiness<br/>both classes?"}
INGV --> Norm
VLF --> Norm
Astro --> Norm
Norm --> Prospective
VLF --> RealSeq
RealSeq --> SelfSup --> VLFEmbedding
RealSeq --> DomainDiag
DomainDiag --> TransferDiag
TransferDiag --> MixedAlign
Prospective --> RealAligned
RealSeq --> RealAligned
RealAligned --> Readiness
end
subgraph Modeling["Model evaluation and transfer"]
Baselines["baselines and ablations<br/>seismic-only, VLF-only, multimodal"]
Probe["real VLF sequence probe<br/>sequence_real_vlf_image_only"]
FineTune["future real fine-tune<br/>load VLF encoder / synthetic checkpoint"]
Blocked["blocked now<br/>all-Italy: 55 positive / 0 negative<br/>central Italy: 0 positive / 55 negative"]
Reports["reports<br/>model summaries, readiness, docs/report.md"]
SimAligned --> Baselines
SimSeq --> DomainDiag
SimSeq --> TransferDiag
SimSeq --> MixedAlign
DeepTrain --> Baselines
VLFEmbedding --> Reports
DomainDiag --> Reports
TransferDiag --> Reports
MixedAlign --> Reports
RealAligned --> Probe
Readiness -- no --> Blocked --> Reports
Readiness -- yes --> FineTune
Checkpoint --> FineTune
VLFEmbedding --> FineTune
RealAligned --> FineTune
FineTune --> Reports
Baselines --> Reports
Probe --> Reports
end
Default path: self-supervised real VLF pretraining can run now without target labels.
Key constraint: supervised real fine-tuning is intentionally blocked until real VLF-aligned rows contain both positive and negative labels.