Mapping methane risk to rank wells for review

Not every orphan well deserves equal urgency. A venting wellhead near homes is an emergency; a quiet capped well in open country is a work order. Our review-priority model turns that judgment into a repeatable score. Here is what goes into it, and where it falls short.

The inputs

Three signal families, each normalized before combining:

• Exposure — distance to occupied structures, water wells, and schools, decayed with distance rather than thresholded. A well 100 feet from a home scores dramatically higher than one 1,000 feet out, with no cliff at an arbitrary line.
• Emissions evidence — direct gas readings where available, aerial plume flags where not, and wellhead condition as a weak prior. Measured beats modeled beats guessed, and the model knows which is which.
• Well factors — age, depth, last-known operator class, and plug status. Older, shallower, never-plugged wells with defunct operators score higher.

Each family produces a 0–100 subscore. The composite is a weighted blend, currently 45/35/20 across exposure, emissions, and well factors.

Why weights and not machine learning

We tried. With a few thousand labeled outcomes — confirmed high-emitters versus confirmed quiet wells — a learned model overfit happily and explained nothing. The weighted blend performs within a few points on held-out data and, crucially, an analyst can read the three subscores and tell you why a well ranks where it does. For a score that drives field-crew dispatch, explainability is a requirement, not a nicety.

Calibration and limits

We calibrate quarterly against new field confirmations, and the weights above are the third revision. Known limits, stated plainly:

1. No reading, no certainty — wells without any emissions evidence lean entirely on exposure and age. The model flags these as low-confidence rather than pretending.
2. Aerial flags are noisy — roughly one in three plume flags confirms on the ground. Useful for discovery, weak for ranking.
3. Water risk is understated — our groundwater proximity data is the thinnest input, and we are actively working to improve it.

How reviewers use it

The queue sorts by composite score within each program, with subscores visible on every row. Reviewers consistently tell us they triage on the composite and investigate on the subscores — exactly the workflow the design intended. The model does not replace judgment; it rations attention.