Methodology · Continuous learning · Verification

How DeepMap AI actually works.

No black boxes. This page walks through the full pipeline — from raw geophysical data ingestion, through the 16,000+-dimensional Earth State Tensor, to the classifiers that fire live alerts and the ledger that pins every prediction to a tamper-evident hash chain.

The pipeline, end to end

Five stages, each built from public data + peer-reviewed physics.

STAGE 1
Ingest — ~568 data sources
Every 15 minutes we pull from USGS FDSN (12,000+ seismic stations), IRIS waveforms, NOAA SWPC (solar wind, Kp, ACE/DSCOVR), NASA FIRMS (wildfire IR), INTERMAGNET (geomagnetic), NMDB (cosmic ray), IGETS (superconducting gravimeter SFTP), GeyserTimes v5, USGS NWIS (groundwater), NDBC (ocean buoys), NGL (GNSS positions), Tomsk observatory (Schumann resonance) — and dozens more. All free, all public, all archival.
STAGE 2
28 Virtual Quantum Instruments
Raw feeds are fused into 28 virtual sensors (the first 22 are the original instrument set; later additions include the traffic-cam strainmeter, infrasound array, GNSS fleet, and muon shadow imager). Each represents a physical quantity you can't buy off the shelf: a virtual gravimeter from ocean tide + GPS fleet, a virtual magnetometer from Schumann + telluric, a virtual muon imager from seismic-array decorrelation. Each instrument outputs 10–40 features per location per snapshot.
STAGE 3
The Earth State Tensor
All 28 instruments' features concatenate into a single 16,000+-dimensional state vector, snapshotted every 15 minutes. We run PCA to extract eigenmodes (the dominant ~50 components explain most of the variance). Anomaly scores come from projecting the current snapshot against the 30-day historical distribution in eigenmode space.
STAGE 4
Predict & gate
Two parallel classifier families: (a) per-hazard tensor classifiers trained on real USGS / SWPC / GeyserTimes outcomes, and (b) the multi-physics convergence gate that only fires when ≥3 independent physics channels (gravity, thermal, strain, cosmic-ray, magnetic, seismic, hydro, tilt) simultaneously show |z|≥2 at the same location. Both write to the ledger.
STAGE 5
Auto-match & learn
Every 10 minutes, 22 auto-matchers poll USGS / EMSC / NOAA / GeyserTimes / NWIS for new events and link them to predictions via bounds-first containment + magnitude-scaled distance. Matched outcomes feed back into the training corpus, and models retrain weekly (or nightly on staleness).

What the accuracy numbers mean

AUC translated for humans.

You'll see AUC 0.92 (cross-validated) on the tensor earthquake classifier and AUC 0.999 on the solar flare classifier. That's Area Under the ROC Curve — a standard machine-learning metric. It is not an accuracy figure, and we don't write it as one.

Unpacked: if we pick a random time when a significant event happened and a random time when it didn't, the model ranks the event-time higher 92% of the time for earthquakes and 99.9% for solar flares. Both figures come from cross-validation, and both carry caveats we set out below. The solar figure in particular should be treated as unvalidated: every positive label in its training window falls inside a single 7.8-hour flare episode, so a forward-in-time test of it is mathematically undefined rather than merely uncertain. The live-ledger hit rate is the honest number for that product.

The validation method is LOOCV — Leave-One-Out Cross-Validation, and in July 2026 we established that it flatters this particular model. We re-tested the per-region earthquake classifier with a temporal holdout — train only on earlier data, score only on later data — and the mean AUC fell from 0.992 to 0.493. That is chance. The cause is that our tensor snapshots are 15 minutes apart while carrying a 7-day forward label, so a randomly held-out sample has near-identical neighbours sitting in the training set with the same label. Random folds cannot separate them; a temporal split can. We are publishing this because it is exactly the kind of thing a reader should not have to discover for us.

Important caveat we don't hide: AUC is an ordering metric, not a pointwise hit rate. High AUC means the model is excellent at ranking — not that every alert will fire correctly. We pair AUC with operating-point tables (precision at threshold, recall at threshold) to pick the right precision/recall trade-off for each product. High-confidence alerts get tight thresholds (fewer, more certain); broad-coverage products get looser ones (more, less certain).

Rule of thumb

0.50Coin flip
0.70"It's doing something"
0.80Operationally useful
0.90Strong — publishable
0.917Tensor EQ (ours)
0.95Excellent
0.999*Tensor Solar — *LOOCV, n≈70; we report the live-ledger hit rate as the operational number

The flywheel

Why accuracy compounds every week.

Most ML systems ship once and decay. DeepMap AI is built as a flywheel — every confirmed outcome becomes a new labeled training sample, and the whole stack retrains on it.

Every 15 min · Tensor snapshot
Fresh 16,000+-dim state vector captured. Stored as compressed float32 blob in the tensor DB. Snapshots are captured every 15 minutes since April 14, 2026 (~96/day) and accumulate indefinitely.
Every 10 min · Auto-match
22 matchers poll USGS, EMSC, NOAA, GeyserTimes, NWIS, FIRMS, GOES, and more. Confirmed matches link back to the prediction that forecasted them. Every match becomes a labeled training example.
Nightly · Staleness-guarded retrain
Models older than 24 h retrain incrementally on the last 500 samples. Adaptive weights bias toward recent confirmed outcomes. Warm-start from previous weights — no cold retraining unless necessary.
Sunday 04:30 UTC · Full retrain
Every classifier retrains from scratch on 2,000 samples with fresh hyper-parameter search. Per-region earthquake classifier re-evaluates all 89 regions. LOOCV AUC is recomputed. New meta file anchors the week.
Every prediction · SHA-256 chain
Every row in the ledger hashes against its predecessor. Tamper-evident. Currently 3,700+ entries, hash chain verified valid end-to-end. Bitcoin-anchored via OpenTimestamps so a researcher can reproduce any claim against a public timestamp.
Continuous learning
88 online learners update weights in real time on a persistent buffer. Temporal decay + class balancing (2× positive boost) keeps the model responsive to regime changes. A/B shadow model framework compares the current production model against candidate replacements before promotion.

Verification

Every claim on this site, reproducible.

Three ways to independently verify any number we publish:

1. Hash-chain verify
Every prediction shows its sha256_hash. The chain itself is queryable at /api/v1/ledger/integrity. Any row rewrite would break every subsequent hash. Try it →
2. Public-data replay
The 2020 Magna M5.7 precursor claim (+92.8 % RMS at UU.BMUT) rebuilds from IRIS FDSN waveforms in ~3 minutes. All input data is open. Methodology is posted. Replay recipe →
3. Daubert-grade audit trail
Every alert stores its full feature snapshot. A forensic reviewer can pull the exact 16,000+-dim tensor slice that fired any prediction and recompute the classifier's output. No "trust me"; show the math.

Scientific advisory panel

The people who keep us honest.

Independent domain experts who review our methods, challenge our claims, and hold the platform to discipline-standard rigor. Active recruitment of three founding advisors: seismology, isotope hydrology, and reinsurance quantitative risk.

Inquire →
Open
Founding advisor · Seismology
Seeking: broadband-array specialist
Reviews the dense-array Chandler retrieval, per-region tensor EQ classifier, and dv/v stretching methodology. Ideal: published on Chandler wobble, IGETS, or large-N seismic array processing. Time commitment ~4 hours/quarter.
Open
Founding advisor · Hydrogeology
Seeking: isotope hydrology specialist
Reviews the deep-water scoring framework, NWIS isotope fusion, and Rush Valley / Utah Corridor validation. Ideal: tritium/³H-³He/¹⁴C field experience in the Basin & Range or Colorado Plateau.
Open
Founding advisor · Reinsurance quant
Seeking: parametric risk modeler
Reviews GeoRisk scoring, cat-bond parametric triggers, and the historical earthquake backtest loss model. Ideal: actuarial or cat-modeling background at a reinsurer or specialty MGA.

Advisory compensation: modest equity grant (DeepMap AI Inc., Wyoming). Typical vest: 18-24 months with a 6-month cliff. No publication restrictions; we encourage advisors to publish their own reviews.

For scientists & researchers

Bring your own event, test the pipeline.

Create a free account to get programmatic access to the ledger, the tensor eigenmode stream, and the convergence-gate feed. Query historical predictions by region, by confidence, by hazard type. Reproduce our backtests against your own event list.

Free researcher account   API docs