Japan's Meteorological Agency (JMA) announced on June 26 that it will develop a new "AI weather model" that learns from past meteorological data to predict future weather, aiming for practical deployment around 2030.
June 26, 2026 · Japan Meteorological Agency
JMA to Build an AI Weather Model for the 2030s
The agency will develop a data-driven AI model that learns from past meteorological data to forecast the future — aiming for practical use around 2030, working in tandem with conventional physics-based numerical weather prediction.
~2030
Target for practical, in-service use
~5 min
AI model run time vs tens of minutes for physics models
1970s
Machine learning first used in JMA forecast guidance
Compute time: physics model vs AI model
Speed is the AI model's headline advantage — a single run can be roughly 6× faster.
Physics model (NWP)
How: Simulation solving atmospheric equations.
Strengths: Physical consistency, intensity forecasting.
Challenge: Compute cost.
AI weather model
How: Learns past data, predicts by pattern.
Strengths: Speed, recognizing specific patterns.
Challenge: Black-box nature, GPU resources.
The realistic direction
A hybrid: AI's speed + physics' consistency
AI model
fast pattern recognition
+
Physics model
consistent intensity forecasts
→
Better forecasts
typhoons, rain bands
Overseas models like Microsoft's Aurora — trained on over one million hours of data — have surpassed major centers on tropical-cyclone track prediction, yet physics still wins on storm intensity. As pre-deployment work, future accuracy verification will determine success.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…