Model & Research · Model card v0.1

How the model works — and how we tested it.

Everything on this page is reproducible from public data. Where we have results, we show them. Where we have plans, we say "target," not "result."

01 — The model today

v0: a yield model trained on every well California has on record.

The current model predicts how much water a new well will produce at a given point, before anything is drilled. It is a gradient-boosted decision-tree model — the same family used for credit scoring and fraud detection — trained on the outcomes of real wells.

Model card — v0Live · validated
Task
Predict well yield (gallons per minute, log scale) at an untested location, pre-drill.
Training data
California DWR well completion reports — 1,107,899 records; 189,571 usable new-construction wells (domestic, agricultural, public, industrial) with yield, coordinates, depth and completion date.
Inputs (features)
Location (lat/lon) · year drilled · well use type · outcomes of the 10 nearest same-type historical wells (median yield) · distance to nearest logged well.
Deliberately excluded
Completed depth and static water level — both only known after drilling. A pre-drill model must not see them; including them inflates accuracy dishonestly.
Algorithm
Histogram gradient-boosted trees (scikit-learn), 400 iterations. Trains on a laptop CPU in minutes. Zero cloud cost.
Output
Expected yield + a ranking score; we only recommend sites in the model's high-confidence band.
02 — The back-test

We hid the future from the model, then asked it to predict it.

A model that has seen the answers proves nothing. So we used a temporal holdout: the model learned only from wells drilled before 2018, then scored every well drilled after — wells it had never seen — and we compared its scores to what actually came out of the ground.

01

Train on the past

165,799 wells completed 1960–2017.

02

Predict the future

23,772 wells completed 2018–2025, fully hidden during training.

03

Compare to baselines

Against what a driller's gut approximates (county average by well type) and against the obvious shortcut (copy the 10 nearest wells).

04

Ask the money question

For agricultural wells: if you drilled only where the model was most confident, how often would you hit strong water (≥200 gpm, the ag threshold)?

Result A — how much of yield the model explains (R², higher is better)

PredictorR² on 2018+ wellsIn sparse areas (no logged well within 2 km)
County + use-type average (the "gut" baseline)0.4590.320
Copy the 10 nearest wells0.555
Clarity v0 model0.5990.494

The sparse-area column matters most: where there are no nearby wells to copy is exactly where dowsers get hired — and where the model's edge is largest.

Result B — the decision test (4,714 agricultural wells, 2018+)

51%
hit ≥200 gpm drilling the usual way (base rate)
95%
hit ≥200 gpm in the model's top 20% of sites
2%
hit ≥200 gpm in the model's bottom 20%
Disappointing wells (<50 gpm), all sites32%
Disappointing wells (<50 gpm), model's top 20%3%

Result C — is the ground really that lumpy?

Among 31,208 pairs of wells less than 1 km apart, the median yield difference between neighbors is 1.9×, and 18% of pairs differ by more than 5×. Your neighbor's well is a weak guide to your ground.

What these numbers mean — and don't. The 95% figure is top-quintile precision on a temporal holdout, not "95% accuracy." v0 is strongest at ranking sites and parcels; meter-level spotting within a field needs the geology and aquifer-scan features below. Yields are driller-reported with varying test methods, so there is noise in the ground truth. All numbers come from public state data and the scripts are reproducible on request.
03 — What comes next

v0 reads where wells were drilled. The next models read what they found.

Every item below uses data that already exists and is public. The order is the order of value to a client standing on their land with a $400,000 quote.

Lithology model — reading a million driller logs

In progress
v1 · core

Every well report contains the driller's layer-by-layer description ("0–40 ft clay, 40–110 ft sand & gravel, water at 95…"). An LLM pipeline parses all 1M+ free-text logs into structured layers: coarse-material fraction, aquifer thickness, depth to first water-bearing zone.

Unlocks: micro-siting — reasoning about the ground itself, not just location. This is the jump from "which parcel" to "which 50 meters."

Airborne aquifer-scan model (AEM)

In progress
v1 · core

California flew statewide airborne electromagnetic surveys and published 3D resistivity grids — effectively an X-ray of the aquifers. We sample these grids at every historical well and at every candidate site, teaching the model what productive ground looks like in the scan.

Unlocks: predictions between wells, in places no one has drilled — the dowser's territory.

Depth & lifetime-cost model

Planned
v1.5

Predict the depth of the water-bearing zone, not just the yield — and translate it into the two bills that matter: drilling cost (per foot) and the lifetime electricity cost of lifting every gallon from that depth.

Unlocks: siting on total cost of water, not just "is there water."

Water-quality risk layer

Planned
v1.5

Salinity, nitrate and arsenic vary layer by layer. Using state groundwater-quality monitoring data, the model flags zones where "wet" does not mean "usable."

Unlocks: avoiding the technically-wet, economically-dry well.

Longevity model — will this well survive 2040?

Planned
v2

Water tables are falling. Using decades of groundwater-level monitoring, the model projects whether a site's water is stable, declining, or heading for the 2014–22 dry-well pattern.

Unlocks: a 30-year answer for a 30-year asset.

Legal extractability layer

Planned
v2

Finding water and being allowed to pump it are different questions. This layer scores each site against its basin's management rules and allocation status, so the map never points to water you can't legally use.

Unlocks: hydrology + water law in one score — hard to copy, impossible to skip.

Well & rig detection from imagery

Planned
v2

A vision model that detects wellheads and active drilling rigs in satellite and aerial imagery — finding unreported wells that hide from state records and watching where drilling activity is moving.

Unlocks: ground truth no database has, plus live market intelligence on where the next wells are going.

Nationwide data ingestion

Planned · client-driven
Expansion

Most states keep public well databases (Texas, Arizona, Nebraska, Kansas, Oregon…). Each new state is a data-ingestion task, not a reinvention: the same pipeline, the same validation, new records.

Unlocks: the map goes wherever the client is.
04 — v1 targets

What v1 has to prove — stated before we build it.

v1 has no results yet; pre-registering its targets keeps us honest. We will report against these, pass or fail, with the same temporal holdout.

Targetv0 todayv1 targetWhy it matters
Overall yield R² (2018+ holdout)0.599≥ 0.70Geology + AEM should explain what location alone can't.
Sparse-area R² (no well within 2 km)0.494≥ 0.60The dowser zone — where we must win.
Top-quintile ag precision (≥200 gpm)95%hold ≥ 93% at 50 m resolutionKeep the hit rate while narrowing from parcel to spot.
Depth-to-water error≤ 15% medianTurns yield into a drilling and pumping bill.
Field validation10 pilot sitings, outcomes publishedReal drills, real water, real numbers.
Honesty clause. These are targets, not results. If v1 misses one, that line will say so here — and a missed target on a public page is worth more to a buyer than a perfect claim nobody can check.