LN — GraphCast Weather Forecasting (Lam et al., 2023)

Full Reference: Lam, R. R., Sanchez-Gonzalez, A., Willson, M., et al. (2023). Learning skillful medium-range global weather forecasting. Science, 382(6677), eadi2336. https://doi.org/10.1126/science.adi2336


Pass 1 — Bird’s Eye

Main Claim

GraphCast outperforms ECMWF’s HRES system in >90% of 1380 evaluated metrics for 10-day global weather forecasting, at orders of magnitude lower computational cost.

PropertyDetail
TypeResearch paper — Meteorology / AI
Relevance to PUMA⭐⭐ Medium — paradigmatic example of AI as scientific model outperforming traditional mechanistic models; supports PUMA’s positioning of LLM agents as superior to heuristic PM baselines

Pass 2 — Key Content

Model

  • Graph neural network trained on 39 years of ERA5 reanalysis data
  • 10-day global forecast in <1 minute (vs. hours for HRES)
  • Generates implicit physical representations without explicit fluid dynamics equations

Results

  • Outperforms HRES in 90%+ of variables including temperature, wind, precipitation
  • Tracks tropical cyclones with higher accuracy than operational systems
  • Knowledge implication: GraphCast learned physically meaningful representations not explicitly encoded

Level of Knowledge Generation

  • Empirical model (not a new theory): Learned functional map from state to state
  • However, the high accuracy implies the model has captured real physical regularities
  • Can be used to generate physical insights (e.g., sensitivity analysis)

PUMA Connection

PUMA Analogy

GraphCast vs. HRES = analogous to PUMA’s triage agent vs. majority-class baseline:

  • Both AI systems replace rule-based/statistical baselines with learned models
  • Both achieve measurably superior performance on defined evaluation metrics
  • Both operate in domains where ground truth is available (observed weather / labeled Jira issues)

This validates the general approach PUMA uses.


APA7 Citation

Lam, R. R., Sanchez-Gonzalez, A., Willson, M., et al. (2023). Learning skillful medium-range global weather forecasting. Science, 382(6677), eadi2336. https://doi.org/10.1126/science.adi2336


MOCs