Evolve AI Institute

Case Study 3: DeepMind Rainfall Nowcasting

Lesson 6: AI in Climate Science and Prediction

DeepMind's Rainfall Nowcasting

The Challenge

Predicting where rain will fall in the next 1-2 hours is surprisingly difficult but critically important. Flash floods kill more people than any other weather-related disaster. Farmers need accurate short-term rainfall predictions for planting and harvesting decisions. Energy companies managing hydroelectric dams need to know when water will arrive. Traditional weather radar can show where rain is falling now, but predicting where it will move and how intense it will become remains challenging.

The AI Solution: DeepMind Nowcasting

Technology: Generative AI for precipitation prediction

Developed by: Google DeepMind in partnership with UK Met Office

How it works:

Key Innovation:

Real-World Impact

Flood Prevention and Emergency Response:

Agriculture:

Energy Management:

Transportation:

Validation Results:

The Data Behind the AI

Input Sources:

AI Training:

Machine Learning Approach:

Critical Thinking Questions

For Your Group Discussion:

  1. Short-Term vs. Long-Term: Why is predicting rainfall 90 minutes ahead different from predicting weather 7 days ahead? What makes short-term prediction so valuable despite covering less time?
  1. Uncertainty Communication: The AI provides multiple possible forecast scenarios rather than one "correct" prediction. Why is showing uncertainty important? How should emergency managers use probabilistic forecasts?
  1. Global Applicability: This AI was trained on UK weather radar data. Would it work as well in tropical regions? Deserts? Mountains? What challenges arise in applying AI trained in one location to another?
  1. Physics vs. Data: Traditional weather models simulate atmospheric physics equations. DeepMind's AI learns from patterns in observations. What are the advantages and disadvantages of each approach?
  1. Failure Consequences: What could go wrong if people rely on 90-minute rainfall predictions that turn out to be inaccurate? Are some use cases riskier than others?
  1. Technology Access: This requires high-resolution weather radar networks and significant computing power. How can developing countries without this infrastructure benefit from similar AI?
  1. Climate Change Impact: How might climate change affect the accuracy of AI rainfall predictions? Can an AI trained on historical data predict unprecedented weather patterns?
  1. Decision-Making: A farmer sees a 60% chance of rain in 90 minutes. Should they start harvesting or wait? How do people make decisions based on probabilities?

Your Analysis Task

Complete the following for your case study presentation:

Summary (2-3 sentences):

Key Data & Methods:

Environmental Problem Addressed:

Real-World Impact:

Limitations or Concerns:

Climate Change Connection:

One Question for Class Discussion:

Real-World Scenario

Agricultural Decision-Making - June 2024, UK

Situation:

Traditional Forecast:

DeepMind Nowcasting Prediction at 10:00 AM:

Farmer's Decision:

Alternative Scenario Without AI:

Discussion: How much prediction accuracy is "good enough" for high-stakes decisions?

Additional Resources

Learn More:

Related AI Applications:

Technical Terms:

Interesting Facts

Why 90 minutes specifically?

Computational Efficiency:

Generative AI Application:

Vocabulary

Nowcasting: Very short-term weather forecasting (next 0-6 hours)

Generative Adversarial Network (GAN): Two AI models competing - one generates predictions, one judges their realism

Probabilistic Forecast: Prediction showing multiple possible outcomes with their likelihood

Radar Reflectivity: Measure of precipitation intensity based on radar return signal strength

Extrapolation: Simple prediction method that assumes current patterns will continue unchanged (often inaccurate for rapidly evolving weather)

© 2025 Evolve AI Institute LLC. May be reproduced for educational purposes.

Lesson 6: AI in Climate Science | Case Study 3 of 6