Improving global precipitation forecasts with an AI weather model trained on satellite observations
Julian F. Schmitt, Bertrand Delorme, Robert C. King, Yashica Patodia, Tapio Schneider, Aditi Sheshadri, Ravi Jain
physics.ao-ph · cs.LG
Quick verdict
Retraining AIFS on IMERG satellite precipitation (Laxmi) cuts CRPS 19% and drizzle bias 33%, a strong but IMERG-verified result needing independent ground truth.
Summary
The authors retrain ECMWF's operational probabilistic AI weather model AIFS-CRPS, replacing the ERA5 reanalysis precipitation target with NASA's IMERG satellite precipitation observations while keeping the graph-transformer architecture, CRPS loss, and ERA5 atmospheric inputs fixed; the resulting model is called Laxmi. Across 94 fifty-member ensemble forecasts from 2024 evaluated with WeatherBenchX at 0.25° resolution (60°S–60°N), Laxmi cuts day-1 24-hour precipitation CRPS by 19.3% versus AIFS-CRPS and 20.1% versus the physics-based IFS, improves the 95th-percentile Brier skill score by 57%, and raises the spread-skill ratio from 0.50 to 0.83 at day 1. It reduces drizzle overprediction by 33% below 3 mm/day and captures 68% of observed events above 50 mm/day versus roughly 30% for both baselines, though it still fails above 300 mm/day where IFS retains some skill. In case studies of 10 landfalling Indian storms, Laxmi delivered the lowest-Brier forecast in 7 of 10 events at the 150 mm threshold (IFS: 2, AIFS-CRPS: 1), while IFS remained competitive at 250 mm. The main limitation is that all headline verification uses IMERG — the same product used for training — so gains may partly reflect matching the reference dataset's statistics. Code is available only on request.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Retraining AIFS-CRPS with IMERG instead of ERA5 as the precipitation target reduces day-1 global 24-hour precipitation CRPS by 19.3% (vs AIFS-CRPS) and 20.1% (vs IFS) at 0.25°, 60°S–60°N.
- Laxmi resolves ERA5-inherited distributional biases: 33% less drizzle below 3 mm/day and 68% capture of observed events above 50 mm/day versus ~30% for both baselines, though it fails above 300 mm/day.
- Fine-tuning on IMERG improves ensemble calibration, raising the spread-skill ratio from 0.50 (AIFS-CRPS) to 0.83 at day 1 and above 0.97 by day 15.
- In 10 landfalling Indian storms, Laxmi produced the lowest-Brier forecast in 7 of 10 events at the 150 mm threshold (IFS: 2, AIFS-CRPS: 1), while IFS remained essential at the highest totals (best in 5 of 10 at 250 mm).
Strengths
Near single-variable ablation design
Architecture (Anemoi graph-transformer), afCRPS loss (α=0.95), and ERA5 atmospheric inputs are held to match AIFS-CRPS while only the precipitation target swaps ERA5→IMERG (Introduction, Methods), directly attributing the gains to the training data.
Rigorous probabilistic verification stack
Fair Brier score with finite-ensemble correction (Eq. 2), BSS against a 50-member resampled IMERG climatology, spread-skill diagnostics, 90/95/99th-percentile sweeps at 0.25° and 1°, tropics/extratropics splits, and 95% CIs over 94 initializations (Fig. 2, Supplementary Figs. 4–9).
Decision-oriented evaluation beyond averaged metrics
Baseline and resolution coverage
Candor about failure modes
Weaknesses
Verification circularity on the training reference
Ablation is not perfectly single-variable
No inferential statistics for case studies
Accumulation method and physical consistency unexamined
Code not released
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.03210v3, physics.ao-ph/cs.LG) presents Laxmi, a retraining of ECMWF's operational 0.25° probabilistic graph-transformer AIFS-CRPS in which the ERA5 precipitation target is replaced by NASA IMERG v07 Final satellite precipitation, with architecture, afCRPS loss (α=0.95), and ERA5 atmospheric inputs otherwise fixed. Central claims: (1) 19.3% day-1 CRPS reduction versus AIFS-CRPS (20.1% versus IFS) over 60°S–60°N; (2) a 57% improvement in the 95th-percentile Brier skill score; (3) 33% less drizzle below 3 mm/day and far better heavy-rain frequency (68% versus ~30% capture above 50 mm/day); (4) superior decision metrics across 10 landfalling Indian storms (best Brier in 7 of 10 at 150 mm). The claims are stated precisely and mostly supported by the presented evaluation — but every score is computed against IMERG itself.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the Laxmi weather model?
Laxmi is a retrained version of ECMWF's operational AI weather model AIFS-CRPS in which the precipitation the model learns to predict comes from NASA's IMERG satellite dataset instead of the ERA5 reanalysis. It was developed by researchers at X, the Moonshot Factory (Google), with Caltech and Stanford, and is described in arXiv paper 2609.03210 (2026).
How much better is Laxmi than existing AI weather models at forecasting precipitation?
In the paper's evaluation against IMERG satellite observations, Laxmi cuts the CRPS error of 24-hour precipitation by 19.3% at a 1-day lead versus AIFS-CRPS and 20.1% versus the physics-based IFS ensemble, improves the 95th-percentile Brier skill score by about 57%, and reduces drizzle overprediction by 33% below 3 mm/day. Note that the verification used the same IMERG product the model was trained on, which favors it relative to models trained on ERA5.
Why train a weather model on satellite observations instead of ERA5?
ERA5 inherits biases from the convection parameterization of its underlying physical model: too much light rain (drizzle), too little heavy rain, and wrong timing of precipitation. Training directly on IMERG satellite estimates avoids inheriting those structural biases, at the cost of taking on IMERG's own retrieval and calibration biases.
Does Laxmi replace physical weather models like the IFS?
Not entirely. The paper shows the physics-based IFS still produces the best forecasts for the highest rainfall totals: at the 250 mm threshold IFS was best in 5 of 10 Indian storm events, and above 300 mm/day Laxmi fails to produce precipitation while IFS retains some skill. The authors position the two as operationally complementary.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure panels appear in the extraction only as fragmented axis-label and caption text, so judgments about plot-level details (curve separation vs CI shading) rest on reported numbers and captions rather than visual inspection.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.