AI+Science: AI for Earth

The panel provides a high-quality overview of cutting-edge AI applications in Earth science, featuring leading researchers with strong credentials. The technical depth is substantial, with detailed explanations of neural operator architectures, training objectives (CRPS), and evaluation against traditional numerical models. The claims are supported by references to specific datasets (ERA5) and models (FourCastNet, Spherical Fourier Neural Operator), and the presenters acknowledge limitations, such as the need for better long-term stability and the challenges of adapting vision techniques. The discussion is well-structured, with each speaker providing concrete examples and results. However, the panel is primarily a presentation of ongoing research rather than a critical review, and some claims could benefit from more rigorous peer-reviewed citations. The adéquation between title and content is excellent, as the panel directly addresses AI for Earth. Overall, the information is reliable and valuable for an audience with some technical background, but it may not be accessible to complete novices. The lack of audience questions in the transcript limits the depth of critical engagement, but the panelists’ expertise and the relevance of the topic make this a strong resource.

The title accurately reflects the content: a panel discussion on AI applications for Earth science, focusing on climate and weather modeling.

8/10

The panel features leading researchers from top institutions (Stanford, NVIDIA, Boston University, NYU) presenting peer-reviewed and published work. Claims are supported by references to specific datasets (ERA5), models (FourCastNet, Spherical Fourier Neural Operator), and metrics (CRPS). The discussion is technical and grounded in established scientific methods, though it represents expert opinion and ongoing research rather than a formal review.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction by moderator David Lobell and panelists.
• 02:00Jean Kossaifi introduces AI-aided engineering and the need for speed in weather forecasting.
• 05:00Kossaifi explains neural operators and the spherical Fourier neural operator for probabilistic forecasting.
• 10:00Kossaifi discusses challenges in adapting vision techniques and the importance of CRPS loss.
• 15:00Kossaifi presents results on Storm Dennis and the speed advantage of AI models.
• 20:00Laure Zanna introduces ocean-atmosphere climate emulators and their ability to run 4,000 years per day.
• 30:00Zanna discusses using emulators for large ensembles and studying climate variability.
• 40:00Elizabeth Barnes presents her work on AI for climate model interpretability and causal discovery.
• 50:00Panel discussion on open science, limitations, and future directions.

• ERA5 reanalysis dataset — Mentioned by Jean Kossaifi as the training data for the neural operator.
• FourCastNet — Referenced as the first neural operator for global weather forecasting.
• Spherical Fourier Neural Operator — Discussed by Kossaifi as the architecture for probabilistic forecasting.
• Continuously Ranked Probability Score (CRPS) — Mentioned as the loss function for probabilistic forecasts.

• FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators — Supports the claim that neural operators can match or exceed traditional models.
• ERA5 reanalysis dataset — Standard dataset used in weather forecasting research.

• Limitations of AI weather models — Some studies have pointed out that AI models may struggle with extreme events and long-term stability, which is acknowledged in the panel.