DeepMap AI
Planetary Early Warning System · Live

Every forecast signed before the event.

We forecast earthquakes from dozens of Earth sensors that have to agree before we say anything. What makes that worth trusting is the part underneath: every forecast is hash-chained and anchored to Bitcoin the day it is issued, so anyone can verify we said it first — and we publish the misses and the disconfirmed hypotheses on the same ledger.

Live Earth digital twin
Explore Live Globe →
6 days early on the M7.8 Philippines (Jun 2026) · 6 of 6 out-of-sample M7.5+ quakes flagged in replay · 4-day lead on the M6.7 Chile (May 2026) · Every warning signed, timestamped, publicly auditable
Earthquake warnings that came true · 14-day window · fully settled — last 14 days our numbers, explained →
Get Free Alerts
View Live Dashboard → API Access →
— critical · — active alerts · — matches / 24h
tensor: —

Why any of this is checkable

The forecast is the claim. This is the receipt.

A hit rate is easy to assert and hard to audit. So the record comes first: every forecast is written to a hash-chained ledger the moment it is issued, the day's root is anchored to Bitcoin, and the misses stay in place beside the catches. You do not have to take the number on trust — you can reconstruct it.

Cumulative forward-only predictions over time
The moat, drawn as a line that only goes up: every prediction logged forward-only since March 2026. Our numbers, explained →
7,600
forecasts on the public ledger, each hash-chained at issue and anchored to Bitcoin daily
Timestamped before the event
The anchor is what a competitor cannot manufacture later. No amount of compute produces a hash that predates an earthquake.
Locked first
hypotheses registered and Bitcoin-anchored BEFORE the analysis runs
Pre-registered, not back-fitted
The pass criterion is fixed on the public record before we see a result, which removes the freedom to reinterpret a number after the fact.
Published
disconfirmed findings, including several of our own headline claims
We publish what failed
In July 2026 an internal audit retired eleven findings, most of them previously ours. They are on the record with the method that killed them. A group that only shows you its wins has not told you its error rate.

Verify the chain → See the pre-registrations and the nulls →

The same record runs in public as it happens. A watch board posts each day naming the regions we have flagged, and seven days later a reply scores it — what arrived and what did not, misses included. @deepmapai

What the engine actually does

Most of what earns a forecast its place is clustering — earthquakes arrive in sequences, and a model that knows the statistics of those sequences beats a long-run average. Ours does, measured on the earthquake field's own yardstick and replicated across independent regions. The underlying method, ETAS, is a published 1998 model. We did not invent it and we do not present it as ours.

On top of that sits one physical precursor that has survived every internal attack we could design. What it is, how it is measured, and the exact pass criterion are locked in a Bitcoin-anchored pre-registration — sealed before its forward test began. If the test passes, we publish everything, and the hash proves we found it first. If it fails, that result goes on the ledger too, in public, like every other number here.

What we do not claim: that fusing thousands of multi-physics signals adds forecasting skill on top of that. We tested it three ways — how many, where, and how large — and all three came back null. Those results are published. Any group can assert a bigger number; the reason to trust a smaller one is that you can check it.

Earthquakes we caught

Real quakes. Real warnings. Issued before they happened.

These are the most recent earthquake warnings our system issued where the earthquake actually happened. Each one is on the public prediction ledger with a signature — you can verify we said it before it happened, not after.

See all wins →
7 days early
M7.1 Venezuela · June 2026
Venezuela M7.1 — caught 7 days out
On June 17 our Caribbean regional model issued a HIGH earthquake warning. On June 24 an M7.1 struck near Morón, Venezuela (confirmed by the USGS and the European EMSC) — inside the window at a full match score. Every alert is on the public ledger with a signature, timestamped before the quake — a regional bounds-first catch, not a pinpoint forecast.
6 days early
M7.8 Philippines · June 2026
Major Philippines earthquake — caught 6 days out
Starting June 1, our Philippines regional model held a CRITICAL warning — renewed each day for six straight days. On June 7 an M7.8 struck off Mindanao (confirmed by the USGS and the European EMSC), inside the window and dead on the forecast. We then credited the M4.4+ aftershock swarm that followed. Every alert is on the public ledger with a signature, timestamped before the quake — not after.
+92.8%
M5.7 Magna · March 2020 · public-data replay
Salt Lake City quake — textbook early signal
One week before the 2020 Magna earthquake, the ground-wave speed at a station 8 km from the epicenter jumped 93% — a classic pre-quake signal. A second station further away showed the opposite. We can reproduce this finding from free public data in about 3 minutes. No private sensors required.
6 of 6
out-of-sample M7.5+ replay · flagged hours ahead
Six major quakes the model never trained on — all flagged
We took six major M7.5+ earthquakes from 2016–2018 — a stretch the model was never trained on — and replayed the real Earth-state data before each one. All six landed in the top ~1% of global risk in the hours before the quake. It's a short-fuse signal (hours, not days), and it reproduces from a committed script.

Live alerts on the board right now

Written clearly. Signed for the record.

Each alert below is currently inside its validation window. Every number came from the live ledger — no templates, no demos.

loading live alerts…
How we build that confidence number

Each sensor family has to agree independently before a warning fires.

  • Seismic: 11,000+ stations worldwide
  • Geodetic: GPS ground-motion networks
  • Atmospheric: pressure, ionosphere, lightning
  • Space weather: solar wind, cosmic rays
  • Gravity: superconducting gravimeters

When at least three disagree, we stay quiet. Being quiet matters as much as being loud.

How the physics works →

Live now

Active predictions on the board

Every alert here is currently inside its validation window. Each has a cryptographic hash logged before the event it's watching.

Full track record →
loading live predictions…
Things no one else has done — without the jargon
2026 · Multi-sensor agreement First system that only warns when 8 sensor types independently agree. Being quiet is a feature, not a bug.
2026 · Real outcomes, not simulations First earthquake classifier trained on real earthquakes, not on data someone made up in a lab.
2026 · Cosmic-ray weather Experimental: reading cosmic-ray flux through the atmosphere as a weather signal — a free channel almost no one uses.
2026 · Tamper-proof record First prediction system where every warning is signed and recorded on Bitcoin — so we can't edit it after the fact.
2026 · Caves without digging Pioneering passive cave & void detection from background seismic noise — no drilling, no active survey. Benchmarked against 19 mapped caves.
2020 · Salt Lake replay First to show a Magna M5.7 earthquake precursor using only free public data. Anyone can reproduce it.
2026 · One model of the whole planet First to fuse 15,000+ live signals from the ocean floor to the inner core into a single model — most systems watch one thing at a time.
2026 · Freshwater under the sea First atlas of offshore freshwater springs mapped from multi-physics signals — satellite, seismic, and gravity — fresh water at a fraction of desalination's cost.
2026 · Verify it yourself First hazard trigger anyone can re-run from public data and check against the signed record — no trust required.
Built on
16,000+
Cross-correlated features
28
Virtual quantum instruments
89
Earthquake regions monitored
778+
Trained models
300K+
Interferometry pairs/day

Narrow, high-skill products

Specialized predictors, each tuned to one well-behaved signal.

These percentages aren't frozen. Every confirmed outcome — every matched eruption, every observed flare, every M5+ earthquake — feeds back into the Earth State Tensor flywheel. Models retrain weekly on real labels. Accuracy compounds.

All products →
9 / 9
predicted eruptions caught (last 30 days)
Daisy Geyser
Regular-interval Upper Geyser Basin feature. Predictions trained against GeyserTimes reports with IRIS tremor + thermal context.
12 / 13
predicted eruptions caught (last 30 days)
Beehive Geyser
Paired with Old Faithful preplay signal; narrow in time by design.
11 / 11
predicted eruptions caught (last 30 days)
Riverside Geyser
~6 hour regular interval. High predictability, confirmed against GeyserTimes.
83%
of predicted eruptions caught (last 30 days)
Steamboat Geyser
World's tallest active geyser. Multi-day-ahead forecasts using Norris basin thermal + seismic coupling.
AUC 0.92
ROC discrimination — separating high-risk from quiet periods (LOOCV, early-2026 run at n=67 · a ranking score, not a hit rate)
Tensor earthquake classifier
Trained on real USGS outcomes — not simulated. Inputs: the full 17,000-dim tensor. Target: any M≥6.0 anywhere in the world within 7 days. Cross-validated, not validated on a temporal holdout — when we ran that test on the sibling per-region classifier in July 2026, its score fell to chance. Treat this as a ranking score pending the same test.
94%
of 24h solar-flare forecasts confirmed on the live ledger (last 30 days, n=69)
Solar flare classifier
Trained on NOAA GOES X-ray outcomes. Target: C-class flare or larger within 24 hours. This is one narrow product's live-ledger hit rate over a small sample (n=69) — not a platform-wide accuracy. We never headline a single "90%+" figure for the whole platform.

Caught in the wild

Real events. Real signals. Reproducible from public data.

Live hits below are pulled directly from the ledger every 15 minutes, ranked by a magnitude × lead-time / distance score. Every card has a SHA-256 prediction hash you can verify against the public chain.

Built for eight industries

One platform. Pick your lens.

The same tensor, the same ledger, the same flywheel — filtered and priced for the way your industry actually works.

Insurance
Parametric triggers
Hash-chained earthquake and volcano triggers settle from public data without claims adjustment — every one on the same auditable ledger, reproducible by anyone.
Learn more →
Emergency management
Pre-positioning alerts
Multi-day lead times on high-confidence convergence events. Route responders before the event, not after. 89 regions, global coverage.
Learn more →
Agriculture
Soil + thermal + hail
SMAP soil moisture, MODIS thermal stress, hail-precursor atmospheric patterns. Plot-level forecasts for operations that can't afford surprises.
Learn more →
Defense
Passive ground awareness
Muon tomography, passive seismic, and Starlink ionospheric mapping fused by multi-physics convergence — passive tunnel- and void-detection, no drilling.
Learn more →
Energy
Geothermal + grid stability
Thermal anomaly targeting for geothermal exploration (Millard County → Fervo). Grid-frequency and space-weather disturbance intelligence for grid operators.
Learn more →
Water
Aquifer + deep-source discovery
77 deep-water targets tracked across the Great Basin. Six discrimination tests, 10 sensing channels per observation. Find water where geologists can't.
Learn more →
Research
Free API + time machine
Query the ledger, replay any historical epoch, pull feature snapshots for any prediction. Reproducibility is the product.
Learn more →
Infrastructure
Rail · bridge · pipeline
Continuous monitoring of critical spans. Interferometry for ground stability, VQ strain for load shift, multi-physics convergence for fault detection.
Learn more →

Why cross-correlation matters

Earth speaks on many channels at once. We listen to all of them.

Multi-physics convergence gate
We require N ≥ 3 independent physics channels (gravity · thermal · strain · cosmic ray · magnetic · seismic · hydro · tilt) to flag the same location with |z|≥2 before a prediction logs. Independent error models mean false-positive rates compound multiplicatively: 30% per-channel FPR → 2.7% triple-gated FPR ≈ 97% specificity if the channels were independent (a design target, not a measured rate).
SHA-256 hash-chained ledger
Every prediction is a row in an append-only ledger hashed against its predecessor. Tamper-evident. Bitcoin-anchored via OpenTimestamps. A forensic reviewer can reproduce the exact feature snapshot that fired any alert. Not a spreadsheet of "trust me" claims.
Learns from every outcome
Every confirmed earthquake, eruption, and solar flare feeds back into the Earth State Tensor. Models retrain weekly on real labels, so accuracy compounds instead of going stale. The methods are public; the feature weights are the trade secret.
2020 Magna M5.7 backtest
Using real FDSN waveforms we replayed the 7 days before the March 2020 M5.7 Magna earthquake. Station UU.BMUT (8 km from epicenter) showed +92.8% seismic-velocity shift in the 0.5–2 Hz band. Station UU.BEI showed −54.6% (stress shadow). Textbook dilatancy-diffusion. Reproducible from public data.

Free tier

See the live alerts for your region.

Sign up in under a minute. Pick the regions you care about. Get live convergence-gate alerts the moment three or more physics channels agree.

Create free account   Compare tiers