BRENNEN

Industries / Frontier science / Cloud seeding

Measuring results

Working out whether seeding added any rain or snow, or took any hail away, against weather that varies a great deal on its own.

7

Milestones achieved, from 1986 to 2024.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1986

    Swiss hail test finds nothing

    Grossversuch IV tested the Soviet method of seeding hail clouds with rockets in central Switzerland from 1977 to 1981. A network of 330 hail pads over 1,300 km² and a calibrated 10 cm radar measured 216 hail cells, and the authors found no statistically significant difference between seeded and unseeded cells.

    Lessona method in wide use can fail its first randomized test, so a test with an unseeded comparison has to come before the claims.

    journals.ametsoc.org

    Credit

    Federer and research groups from France, Italy and Switzerland

  2. 1995

    Israel's results reexamined

    Israel 1 (1961 to 1967, 364 days) reported a 15% rise in rain with p = 0.009, and Israel 2 (1969 to 1975, 388 days) a 13% rise with p = 0.028, by the experimenters' own analyses. Re-reading both, Rangno and Hobbs concluded the first was a "lucky draw" and that the second's effect came from control stations chosen after the experiment began, so neither showed a significant effect.

    Lessonpick the control area before the experiment starts, and repeat a positive result before building a program on it.

    journals.ametsoc.orgjournals.ametsoc.org

    Credit

    Arthur Rangno and Peter Hobbs, University of Washington

  3. 2011

    Silver traced into snow

    The Snowy Precipitation Enhancement Research Project in Australia (2005 to 2009) seeded 107 randomized five-hour units from ground generators and released indium oxide, a second tracer, every time. Silver in snow from the target area was significantly higher after seeded units, which proved the material arrived. The gain in precipitation was positive but not statistically significant; restricted to units with more than 45 generator hours, it was 14%, at the 8% significance level in the primary target.

    Lessonchemistry can prove where seeding material went even when rain gauges cannot prove what it did.

    journals.ametsoc.org

    Credit

    Michael Manton and Lisa Warren, Monash University, for Snowy Hydro

  4. 2018

    Seeding seen start to finish

    The SNOWIE campaign in Idaho (7 January to 17 March 2017) seeded winter clouds over the Payette River basin from an aircraft and watched with radars and a research aircraft. On days with little natural snow, radar showed zigzag lines of new snow following the aircraft's path, and probes inside them caught the ice crystals forming and growing. French and colleagues wrote that, despite decades of experiments, no direct observations of this chain of events had existed before. A 2020 follow-up measured 100 to 275 acre-feet (120,000 to 340,000 m³) of water from single seeding runs of 20 to 86 minutes.

    Lessonthe lines showed only on three days with weak natural snow, so the method does not yet give a season's total.

    journals.ametsoc.orgpar.nsf.govebi.ac.uk

    Credit

    Sarah Tessendorf (NCAR), Jeffrey French and Bart Geerts (University of Wyoming), Katja Friedrich (University of Colorado Boulder), and Idaho Power

  5. 2018

    Wyoming test stays inconclusive

    The Wyoming Weather Modification Pilot Project ran a randomized crossover experiment over two mountain ranges from 2008 to 2013. The statistical test on 118 units gave p = 0.28, too weak to reject "no effect". Computer models of the same storms estimated a 5% gain in seeded storms (middle half of runs 3% to 7%), which the authors put at about 1.5% of yearly precipitation.

    Lessona six-year randomized test can still be too small to detect a gain of a few percent.

    journals.ametsoc.org

    Credit

    Roy Rasmussen and colleagues, NCAR

  6. 2023

    Israel 4 ends operational seeding

    After 38 years of operational seeding in northern Israel, a randomized experiment from 2013 to 2020 found a 1.8% rise in rain, p = 0.4, with a 95% confidence interval of −11% to +16%. The experiment stopped a year early, and the Israel Water Authority stopped operational seeding. The authors note that seawater desalination, begun in Israel in 2006, gave the authority another source to weigh it against.

    Lessona program built on old experiments needs re-testing, because clouds, pollution and alternatives change.

    journals.ametsoc.org

    Credit

    Yoav Benjamini, David Steinberg and colleagues, with the Israel Water Authority

  7. 2024

    US auditors call benefits unproven

    The US Government Accountability Office reviewed the evidence in December 2024. In the studies it reviewed, added precipitation ranged from 0 to 20%, and it found reliable information lacking on how well seeding works. It also found that federal reporting rules may not capture enough to monitor seeding.

    legistorm.com

    Credit

    US Government Accountability Office, report GAO-25-107328

Watch for

  1. Synthetic control across the West

    A study uses synthetic control (a statistical method that builds a "twin" of each seeded area from unseeded ones) on 832 reported US seeding periods from 2000 to 2025, extracted from NOAA reports by a language model, to test whether seeding changes snowpack and precipitation at regional scale. The abstract calls this the first such evaluation across the US but gives no results.

    Achieved whenthe estimates appear in a peer-reviewed paper with the code and data public.

    meetingorganizer.copernicus.org

    Claimed by

    Kara Lamb and Jared Donohue, EGU General Assembly, May 2026

  2. Drone seeding's claimed water

    Rainmaker Technology says its drone seeding in Utah and Oregon produced 143 million gallons (about 540,000 m³) of water in the 2025 to 2026 winter, backed by 82 radar and satellite "seeding signatures" matched to distinctive drone flight patterns. In August 2026 it reported that 10 drone flights over Alaska's Kenai Peninsula on 23 August made about 57 acre-feet (70,000 m³) of rain in three hours; it found seven signatures on the public PAHG NEXRAD radar, tracked with polygons drawn by hand, and estimated the water with a method adapted from SNOWIE.

    Achieved whenthe validation study with the National Center for Atmospheric Research and two Utah universities, or another outside team, reproduces the signatures and water totals from the same public data.

    deseret.comrainmaker.comrainmaker.com

    Claimed by

    Rainmaker Technology, April and August 2026

Open

  1. 01

    Find seeding lines automatically

    An open detector that scans public weather radar data for the lines of extra snow that seeding leaves downwind of an aircraft or drone, and reports each line's start, path and size.

    The brief
    Credit · unclaimed
  2. 02

    Fix the controls first

    A rerun of the yearly target-and-control evaluations that operational programs publish, with control stations chosen in advance by a written rule, from public snow and precipitation gauges, and a side-by-side table of each program's claimed gain and the rerun gain.

    The brief
    Credit · unclaimed

Open 01 · the brief

Find seeding lines automatically

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An open detector that scans public weather radar data for the lines of extra snow that seeding leaves downwind of an aircraft or drone, and reports each line's start, path and size.

Why it is still open

The SNOWIE lines and the 2026 Rainmaker signatures were picked out by people; Rainmaker's report says each signature is tracked with polygons drawn by hand. Zaremba and colleagues (2024) showed a seeding signal must change radar reflectivity by about 5 dB to stand out, so it is lost in heavy natural snow. Mehdizadeh, McDonough and Hosseinpour of the Desert Research Institute (2025) described seeding signatures in GOES satellite and NEXRAD radar data for ground-generator events around Lake Tahoe and in two Nevada ranges, but as a study of selected events, with no automatic detector. No public code for this task turned up on GitHub; the public SNOWIE repository holds plotting scripts only.

par.nsf.govrainmaker.comdoi.org

Built with what exists

NEXRAD Level II data on AWS Open Data, NOAA's MRMS radar mosaics, the Py-ART radar toolkit, PyTorch for a small image model, and the flight times and tracks published in the SNOWIE papers and the Rainmaker report.

  1. First step

    Download NEXRAD data for the SNOWIE days and check whether the lines show at all on the public radars, since SNOWIE used mobile research radars on the mountains.

  2. Done looks like

    The detector finds the three SNOWIE days with traceable lines and the seven Kenai signatures, without being told where to look, and flags lines in fewer than 1 in 20 of 200 unseeded storm hours from the same radars. Code and test set are public.

Open 02 · the brief

Fix the controls first

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A rerun of the yearly target-and-control evaluations that operational programs publish, with control stations chosen in advance by a written rule, from public snow and precipitation gauges, and a side-by-side table of each program's claimed gain and the rerun gain.

Why it is still open

Operators' contractors evaluate their own programs by "historical target/control" comparisons and choose the control sites; Utah's 2022–23 report for its southern and central program claims long-term gains of 11% to 13%, and a Colorado report gives 2% to 5%. Rangno and Hobbs showed that control sites chosen after the fact can manufacture an effect. Independent reruns have been done before, but with the same weakness: Bernard Silverman's 2009 evaluation of the Vail, Colorado, program found significant gains of 6.3% to 28.8% in 5 of 8 sub-basins using "the controls that give the most precise evaluation results possible", which is a choice made after seeing the data. The Lamb and Donohue study works at regional scale, not program by program.

water.utah.govcoloradoriverdistrict.orgjournalofweathermodification.scholasticahq.com

Built with what exists

NRCS SNOTEL data through its public web service, Python with statsmodels, the Donohue and Lamb dataset of reported US seeding periods to mark seeded seasons, and a pre-registration on the Open Science Framework. An AI model reading the program reports' tables, checked by a person (see the Note "Seven more flagged").

  1. First step

    Write and register the control-selection rule (distance, elevation and no seeding upwind) before downloading any gauge data.

  2. Done looks like

    Eight named programs, each with its claimed gain, the rerun gain and a confidence interval, using controls fixed before any result was seen, with code that reproduces every number.