Dirt & Data: Issue #4

Originally published March 19, 2026. Corrected October 8, 2026. The original confused the share of forecasting targets won with the size of an accuracy improvement. We also corrected model timing and removed unsupported farm-level promises. The original issue number is retained.

AI Weather Models Are Here. What Changes for Your Farm?

Better weather models can help you compare possible conditions farther ahead. They do not turn a two-week outlook into a safe spray-window guarantee.

The developments below happened on different dates: Google's GenCast research appeared in December 2024, Microsoft's Aurora paper in May 2025, and the National Oceanic and Atmospheric Administration launched an operational AI model suite in December 2025. Research results, operational models, and the forecast in a particular phone app are separate things.

The Big 3

GenCast: Google DeepMind

In a test using 2019 weather, GenCast beat the European Centre for Medium-Range Weather Forecasts' ensemble on 97.2% of 1,320 variable-and-lead-time targets. For targets beyond 36 hours, it won 99.8%. Those are shares of comparisons won, not 97.2% or 99.8% improvements in accuracy. Read the research explanation.

Google reported roughly eight minutes on one specialized processor for one member of a 15-day forecast ensemble. That is not the time for every scenario combined. Its 0.25-degree grid is about 28 kilometers north-to-south; east-to-west spacing varies by latitude.

Google announced WeatherNext 2 in November 2025, with faster generation and predictions at hourly intervals. Faster computation and finer time steps do not by themselves prove better wind forecasts for a specific field.

Aurora: Microsoft

Microsoft's May 2025 research summary reported that a version of Aurora fine-tuned for medium-range weather at 0.25-degree resolution beat comparison models on 91% of forecasting targets. The same announcement said its code and model weights were available, and that a specialized version was being incorporated into MSN Weather.

The research also tested air-quality, ocean-wave, and tropical-cyclone tasks. These are task-specific results from retrospective tests. They are not a single accuracy score for every forecast or proof that every farm app now uses Aurora.

The National Weather Service's Operational AI Models

The December 17, 2025 launch introduced three applications: a single-forecast AI system, an AI ensemble of possible outcomes, and a hybrid ensemble combining AI with the existing physics-based model. Their short names are AIGFS, AIGEFS, and HGEFS. The agency describes this suite as an outgrowth of Project EAGLE.

It reported that a 16-day AIGFS forecast used about 0.3% of the computing resources of its operational Global Forecast System and took about 40 minutes. It also disclosed a weakness in the first version's tropical-cyclone intensity forecasts. Lower computing cost is not proof of superiority on every weather variable.

The published AIGFS data cycles are 00, 06, 12, and 18 universal time, four daily cycles. The original twice-daily statement was wrong.

What This Means for Your Operation

• Use the longer outlook for options. Identify possible fieldwork windows and backup plans, then revisit them as the date approaches.
• Keep spray decisions local and current. Check the current forecast, actual conditions, and the product label. A global model score does not establish safe wind, rain, or inversion conditions at your field.
• Measure the decision you care about. Save forecasts before an event and compare them with what happened. A useful test might be rain timing before harvest, not an app's general accuracy claim.

The original article's yield-accuracy percentages and insurance predictions had no comparable, independently verified basis. They have been removed.

What You Can Use Today

Start with the National Weather Service's explanation of its model suite and the current forecast for your location at weather.gov. A provider may combine several models with observations and forecaster judgment. Do not assume the home page displays a particular experimental model or that every app uses the same inputs.

For a paid service, ask which weather variables it predicts, at what lead times, and how it performed against observations near your ground. A small grid cell on a map is not proof of equally precise local predictions.

What to Watch

Watch for published local validation, clear uncertainty ranges, and records that let you compare a forecast with the result. Those will tell you more than a market-size projection or a promised date for field-level accuracy.

The original market-growth figures, vendor list, and two-to-three-year claims were not supported well enough to retain. They also mixed global model resolution with separate local forecasting products.

One Thing to Try

For the next two weeks, save the rain and wind forecast from your usual service at a consistent time. Record what actually happened using a suitable local observation. Keep the forecast lead time with each entry so tomorrow's forecast is not compared with next week's. Treat a short trial as a learning exercise, not proof of year-round accuracy.

That's All for Today

AI adds useful tools to weather forecasting. It works alongside an existing system of observations, physics-based models, and human forecasting. The practical test is whether a service helps you make a better decision on your farm at the lead time you need.

The Dirt & Data Team

Sources:

1. Google DeepMind GenCast research, December 4, 2024
2. Google WeatherNext 2 announcement, November 17, 2025
3. Microsoft Aurora research, May 21, 2025
4. National Oceanic and Atmospheric Administration model launch, December 17, 2025
5. National Weather Service AIGFS model data cycles
6. National Weather Service forecast model guide

Dirt & Data is an independent newsletter covering AI in agriculture. Written with AI assistance. Images are illustrations, not documentation of the reported trials. Straight talk for people who work the land.