Catalyst with Shayle Kann

Episode Summary: "How AI is Changing Weather Forecasting"

Release Date: January 2, 2026
Host: Shayle Kann
Guest: Peter Battaglia, Senior Director at Google DeepMind

Overview

In this episode, Shayle Kann explores the intersection of cutting-edge artificial intelligence and weather forecasting with guest Peter Battaglia from Google DeepMind. The discussion delves into how traditional weather forecasting works, the transformative role new AI approaches can play, key technical differences between old and new models, data challenges, and what the future might hold if AI lives up to its promise in this field.

Key Discussion Points & Insights

1. The Foundations of Weather Forecasting

Timestamps: 03:27–10:05

• Brief History:

  • Forecasting moved from informal observations to systematic scientific collection (approx. 100–150 years ago) and then to major government bureaus like NOAA (US) and ECMWF (Europe) in the 1970s.
  • Weather forecasts are considered a public good, funded by taxpayers, and offer strong economic returns.

• Current Process:

  • The physical basis is fluid dynamics—the atmosphere is a fluid modeled by the Navier-Stokes equations, approximated with enormous supercomputing resources (a practice called "numerical weather prediction").
  • These models are run at global scale (often at coarse resolution), then specialized through downstream post-processing for specific uses (e.g., your phone's weather app).

• Data Collection as an Ongoing Challenge:

  • Estimating the current state of the weather (from satellites, ground stations, balloons) is as difficult and important as predicting the future.

Peter Battaglia [03:53]: "The first step in weather forecasting isn't actually predicting. It's taking all the satellite data and all the stations and all the different observations and estimating the current state of the weather across the earth."

2. Limits and Progress of Traditional Approaches

Timestamps: 10:05–18:09

• Linear but Significant Progress:

  • Forecast accuracy has gradually improved, due to better data, more compute, and smarter model tweaks.
  • Forecasts now span up to two weeks ahead with reasonable confidence—unimaginable 100 years ago.

• Hard Problems Remain:

  • Predicting precipitation (because it changes at small scales), and fine spatial details (like thunderstorms) is still very hard, even with today's best approaches.
  • Temperature trends smoothly, but storms and wind patterns shift rapidly at local scales.

Peter Battaglia [18:42]: "What doesn't vary smoothly is precipitation... It's everything happening at a finer scale, finer scale than a lot of our models even capture."

3. How 'AI' Enters the Forecast: Old and New

Timestamps: 21:12–31:08

• First Steps:

  • Early "AI" meant using statistical techniques or machine learning (ML) for post-processing applications (tweaking NOAA/ECMWF forecasts for local use).

• Modern AI/ML in Forecasts:

  • Now, researchers train large, supervised learning models (neural networks) to directly predict atmospheric states from data.
  • Transformers and graph neural networks, popular in language and vision, are reshaping weather forecasting too.

• What's New with Transformers:

  • Traditional models (like convolutional neural networks) handle local, hierarchical information transfer.
  • Transformers enable models to "connect" data from distant points directly, mapping complex long-range spatial interactions—essential in weather, where, for instance, a hurricane's influence spans hundreds of kilometers.

Peter Battaglia [27:33]: "We use transformers and graph neural networks to capture the short and long range spatial dependencies...those interactions between what's nearby and what's about to happen next are what determine weather."

4. Comparing AI for Weather vs. Language Models

Timestamps: 31:08–34:04

• Key Differences:

  • Language models (LLMs) predict "next word" based on textual sequence; weather forecasting is about predicting physics-based, Markovian systems where the immediate past is most relevant.
  • Weather data is spatially and temporally distributed, and predictions must be grounded in physical reality (not just probability distributions of words).

• Uncertainty and Structure:

  • Physics is deterministic, but lack of precise information (e.g., not knowing where a "butterfly flapped its wings") makes accurate prediction uncertain—a case of "randomness from ignorance".
  • The challenge: Weather models must predict millions of variables simultaneously—much more complex than predicting a single next word.

Shayle Kann [31:08]: "On one hand, a much harder problem than an LLM because you've got the entire physical world and the data is sparse...On the other hand, it's determinative in a way that LLMs are not."

5. Data: The Lifeblood and Limitation

Timestamps: 34:04–40:16

• Data Abundance—and Scarcity:
  • Weather’s historical datasets (like ECMWF's ERA5 reanalysis) are highly valuable but limited by changes in data collection quality over time.
  • There's a ceiling; "new weather" only arrives one day at a time, limiting the amount of training data.
  • Researchers are turning to unconventional data sources: IoT devices, video doorbells, car sensors, even social media reports—hoping for more granular or novel input.

Peter Battaglia [38:37]: "Cars, your car...it's got a thermometer in it. It's got your windshield wipers...I get very excited about the possibility of using all these kinds of things."

• Universal Truth in AI:
  • "You're always data poor," says Battaglia: More data always enables better models, and this lesson has repeatedly proven itself after initial skepticism.

6. The Future: What Improved AI Weather Forecasts Could Unlock

Timestamps: 40:16–43:53

• Potential Impact Areas:
  • Consumer Guidance: More precise, tailored forecasts could inform individual choices—what you wear, what you buy, when you travel.
  • Energy Sector: From solar and wind forecasting for grid stability, to better demand management as extreme events (like heatwaves and storms) occur with increased frequency.
  • Supply Chains and Logistics: Improved planning and operations.
  • Disaster Preparedness: Better and earlier warnings for cyclones, wildfires, floods—potentially enabling earlier, more targeted intervention.
  • Overall: Technology-driven progress could leave the next generation with a safer, more predictable world.

Peter Battaglia [42:30]: "I have a feeling that there's a lot of headroom...in how we, you know, plan how to operate our electrical grids, how we predict what the electrical demand is going to be...I don't think we've even begun to get into this."

Notable Quotes & Memorable Moments

• On Fluid Dynamics and Complexity:

  "It's hard to even imagine, like, the scale of that type of predictive accuracy was...unimaginable 100 years ago."
  — Peter Battaglia [09:45]

• On the Data Bottleneck:

  "You're always data poor, right? Like, that's sort of the...story of modern AI is you're basically always kind of data poor because...just how well it scales with data. More data just means better models."
  — Peter Battaglia [39:42]

• On Why AI Can Outperform Traditional Models:

  "AI models...are capable of treating the hurricane as almost like a large macroscopic scale object that is moving because they can see all the structure of the hurricane..."
  — Peter Battaglia [27:33]

• On the Promise and Optimism of AI for Weather:

  "My team and I really believe in the power of technology that can have a lot of positive benefits. So we often are trying to look for ways that we can put technology to best use and leave our kids with a world that's better than the one that we grew up in."
  — Peter Battaglia [43:06]

Structure & Flow

• 0:00–01:54 — Introduction & Sponsors (Skipped)
• 01:54–03:27 — Shayle frames the challenge and introduces Peter Battaglia
• 03:27–13:18 — Walkthrough of traditional weather forecasting: history, physics, and current challenges
• 13:18–21:12 — Discussion of improvement drivers and ongoing ML tweaks
• 21:12–31:08 — How AI/ML is now being applied, technical deep-dive into model architectures (transformers, GNNs)
• 31:08–34:04 — Comparing weather models to LLMs, physical determinism vs statistical prediction
• 34:04–40:16 — Data richness, its constraints, and the search for better input sources
• 40:16–43:53 — Forward-looking speculation: what improved forecasts could mean for society
• 43:53–End — Closing remarks

Takeaways for the Uninitiated

• AI is not replacing fundamental physics, but augmenting and optimizing it—unlocking spatial awareness and efficiency unimaginable for traditional models.
• Weather forecasting will never be perfect, due to both inherent chaos and data collection limits—but each leap forward in model structure, compute, and (especially) data translates into societal benefits: resilience, safety, efficiency.
• The field is highly interdisciplinary, drawing in methods and lessons from language, vision, and even unexpected data sources; collaboration and data-sharing will underlie the next big breakthroughs.
• Optimism is high that AI will yield not only better forecasts but also a wide array of societal benefits—energy savings, improved disaster response, and even daily life conveniences.

For more on this topic, explore Latitude Media’s coverage and follow the developing work on AI-driven weather forecasting from agencies and tech leaders like Google DeepMind.