AI summary by TwinMind
Can AI help us predict extreme weather?
Vox
8m 11s
Hurricane Lee Case Study (September 2023)
- September 11, 2023: Hurricane Lee strengthened to Category 3, northeast US tracking
- September 16, 2023: Downgraded storm made landfall Nova Scotia, Canada
- Flooding roads, downing trees, power outages tens of thousands
- Traditional weather forecasts predicted trajectory 5 days before landfall
- Google AI model beat traditional forecasts by 3 days (8 days total advance prediction)
Traditional Weather Forecasting Process
- Data collection from satellites, weather stations, buoys worldwide
- Measuring clouds, temperature, pressure, wind speed, humidity
- Computer processing creates 3D atmospheric grid (vertical/horizontal)
- Physics calculations determine condition interactions
- ~150 trillion calculations per forecast
- Requires world's most powerful supercomputers
- European Center for Medium-Range Weather Forecasts
- National Weather Service (US)
- Local meteorologists refine global models with regional expertise
- Adjust for terrain (hills, flat areas, lakes)
Forecast Limitations & Ensemble Methods
- Initial 3D grid never exactly replicates reality (data gaps)
- Accuracy decreases with time distance
- Ensemble forecasting: up to 50 forecast variations
- Tweaking initial data parameters
- Measuring uncertainty levels
- Similar forecasts = higher certainty
- Varied forecasts = lower certainty
- Current release schedule: every 6 hours (computing power limitation)
AI Weather Models Development
- 2020: ERA5 dataset released
- 40 years hourly Earth weather data
- Terabytes of complete, smooth data (no missing values)
- Free access for AI training
- AI learning method: historical pattern recognition vs physics equations
- Training process: snapshot → prediction → accuracy scoring against actual outcomes
AI Model Performance (2023)
- Companies: Google, Huawei, Nvidia developed competitive models
- Performance rivals traditional forecasting for:
- Surface temperature, humidity, wind speed
- Extreme weather: tropical cyclones, atmospheric rivers, extreme temperatures
- Speed advantage: Huawei Pangu Weather model
- Week-long forecast in 1.4 seconds
- Runs on modest laptop vs supercomputers
- Google GraphCast: Hurricane Lee path prediction faster than traditional models
AI Model Limitations
- Hurricane intensity prediction still challenging (vs path prediction)
- Conservative bias in scoring system
- Penalized heavily for large errors
- May miss rare outlier weather events
- Training data limitation: 40 years historical data
- Fewer extreme events than current/future climate change reality
AI Advantages: Enhanced Ensemble Forecasting
- Traditional: 50 ensemble forecasts maximum
- AI capability: 1000+ ensemble forecasts
- Freed from computing/time constraints
- Better uncertainty measurement
- Useful for emergency managers (evacuation decisions)
- Potential to catch rare weather events missed by smaller ensembles
- Extended forecast periods beyond current 10-day limit
- Improved subseasonal forecasting (multiple months out)
Future Implementation & Human Role
- Meteorologist role remains essential for TV/public communication
- Communicating uncertainty effectively
- Walking through "what-ifs" scenarios
- Human decision-making support
- European Center started publishing AI forecasts alongside traditional ones
- Future weather apps may use AI instead of/combined with physics-based models
- Still experimental phase but moving toward public implementation
