PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2609.03210v3physics.ao-phcs.LG

TL;DR

Precipitation is difficult to forecast, and AI weather models trained on ERA5 inherit biases in drizzle and extreme rainfall. The paper retrains AIFS-CRPS using IMERG precipitation targets; Laxmi improves global skill over AIFS-CRPS and IFS across the 15-day forecast horizon, while mitigating those biases.

  • Problem

    Precipitation is sparse, intermittent, and localized, while ERA5-trained AI weather models inherit drizzle overprediction and extreme-rainfall underprediction.

    Source: Introduction · p. 2

  • Method

    Laxmi retrains AIFS-CRPS with IMERG precipitation targets while retaining its graph-transformer architecture, CRPS loss, and core ERA5 atmospheric variables.

    Source: Introduction · p. 3

  • Results

    Laxmi is more skillful than AIFS-CRPS and IFS across the 15-day forecast horizon, reducing 1-day accumulated precipitation CRPS by 19.3% and 20.1%, respectively.

    Source: Global medium-range skill · p. 5

  • Takeaways & Limitations

    Training directly on IMERG mitigates drizzle and extreme-precipitation biases inherited from reanalyses.

    Source: Discussion · p. 9

    Training on IMERG introduces retrieval and calibration biases of its own.

    Source: Discussion · p. 9

Abstract

from arXiv · show

Precipitation forecasts shape decision-making across the global economy, particularly in sectors such as agriculture. However, unlike variables such as temperature, precipitation is highly intermittent and localized, making it difficult to forecast. While recent advances in AI weather prediction systems have enabled them to surpass physical models on globally averaged metrics, improvements in mean error rarely translate to actionable forecasts of severe flooding or dry crop fields. Furthermore, most of these models are trained and evaluated against a reanalysis data product, ERA5, which has well-known biases. Here we retrain AIFS, ECMWF's widely-used, open-source operational 0.25{\deg} probabilistic graph-transformer weather model, on satellite-based precipitation observations. Our model, Laxmi, improves global probabilistic accuracy by 19% and resolves systematic distributional biases in ERA5. Specifically, Laxmi reduces drizzle overprediction by 33% for amounts less than 3 mm per day. It also mitigates extreme rainfall underprediction, improving the global 95th percentile Brier skill score by 57%. Across a case study of 10 Indian tropical storms, Laxmi delivered the most accurate forecast of 150 mm event-total precipitation in 7 events, compared to 1 for AIFS and 2 for the leading physical model, IFS. Our results demonstrate that incorporating observation-based precipitation data directly into training can substantially improve forecasts.

Introduction

Precipitation is difficult to forecast because it is sparse, localized, and affected by both model-loss and training-data biases. This paper retrains AIFS using satellite precipitation targets to address drizzle and extreme-rainfall biases while retaining its architecture and core atmospheric inputs.

Introduction

  • AI weather models consistently overpredict drizzle and underpredict extreme precipitation, despite recent gains in forecasting performance.Sparse, intermittent, localized precipitation is especially challenging to evaluate and predict; global observational datasets disagree, and rain gauges poorly represent gridbox averages.Source: Introduction · p. 2
  • ERA5-trained models inherit structural precipitation biases from parameterizations, including drizzle overprediction and high-intensity rainfall underprediction.Deterministic loss functions can also smooth forecast fields, underestimating the highest intensities while overpredicting drizzle.Source: Introduction · p. 2
  • These biases can contribute to missed warnings before extreme events or delayed or suspended harvests, including in sugarcane operations.The paper notes that satellite precipitation datasets offer an alternative to parameterization biases but carry retrieval and calibration biases of their own.Source: Introduction · p. 2
  • Laxmi retrains AIFS-CRPS with IMERG precipitation targets while retaining its graph-transformer architecture, CRPS loss, and core ERA5 atmospheric variables.The authors use this setup to isolate the impact of training data.Source: Introduction · pp. 2–3

Results

Against IMERG, Laxmi improves global precipitation forecast skill and better represents observed rainfall frequencies, including heavy precipitation. In Indian storm evaluations, it leads on detection-oriented metrics, while IFS remains useful at the highest thresholds.

Improved precipitation distribution

  • ERA5 underproduces 99th-percentile precipitation across the Maritime Continent and has pervasive drizzle bias in regions where IMERG records less drizzle.The cited passage notes substantial deficits in India, Brazil, West Africa, and the central and eastern U.S.Source: Improved precipitation distribution · p. 3
  • Laxmi captures 68% of observed events above 50 mm day−1, versus about 30% for AIFS-CRPS and IFS; above 100 mm day−1, it captures 38%, versus 17% for IFS and 4% for AIFS-MSE.Above 300 mm day−1, Laxmi produces no precipitation, while IFS captures about 10% of observed frequency.Source: Figure 1 · p. 4, Improved precipitation distribution · p. 4
  • Laxmi reduces drizzle below 3 mm day−1 and better captures precipitation frequency above 10 mm day−1 relative to the other models.The intensity-ratio evaluation compares model frequency with observed IMERG frequency across intensities.Source: Improved precipitation distribution · pp. 3–4, Figure 1 · p. 4

Global medium-range skill

  • At 1-day lead, 24-hour accumulated precipitation CRPS falls 19.3% vs AIFS-CRPS and 20.1% vs IFS; 95th-percentile Brier skill score improves 57.2% and 54.4%, respectively.Laxmi is more skillful than both models across all reported metrics over the 15-day forecast horizon.Source: Figure 2 · p. 5, Global medium-range skill · p. 5
  • Laxmi’s spread-skill ratio is 0.83 at 1-day lead and exceeds 0.97 by day 15, while AIFS-CRPS and IFS remain below 0.77 and 0.83, respectively.A ratio of 1 denotes a calibrated ensemble; the reported comparison indicates less overconfidence for Laxmi.Source: Figure 2 · p. 5, Global medium-range skill · p. 6

A case study of Indian tropical storms

  • In five U.S. landfalling hurricanes, Laxmi has the best Brier score at the 100 mm day−1 threshold in four events.The paper also presents metrics for all 46 tropical cyclones globally.Source: A case study of Indian tropical storms · p. 7
  • Across 10 Indian storms, Laxmi has the highest probability of detection and critical success index with a competitively low false alarm rate.At 50 and 100 mm thresholds, it produces the lowest Brier score 12 of 20 times, compared with 4 each for AIFS-CRPS and IFS.Source: A case study of Indian tropical storms · pp. 6–7, Figure 4 · p. 8
  • At the 250 mm threshold, IFS and Laxmi each produce the lowest Brier score in 5 of 10 cases, while AIFS-CRPS is not competitive at any threshold.The authors describe Laxmi and IFS as potentially complementary operationally at very high precipitation thresholds.Source: A case study of Indian tropical storms · p. 7, Figure 4 · p. 8

Discussion

Laxmi improves medium-range skill and precipitation-intensity representation, while higher resolution brings further gains. Remaining retrieval and calibration biases and real-time initialization needs define important boundaries for deployment.

Discussion

  • 19.3% at day 1 and 3.4% at day 10 CRPS reductions versus AIFS-CRPS accompany superior medium-range global skill and a better-calibrated ensemble.Laxmi rivals or exceeds the 6.4% maximum improvement achieved by AIFS-CRPS over IFS.Source: Discussion · p. 8
  • Laxmi improves heavy-precipitation representation at 30 to 200 mm day−1 while reducing pervasive light-rain overprediction.Remaining intensity biases still need attention.Source: Discussion · pp. 8–9
  • For tropical storms, Laxmi raises detection rates without inflating false alarms for moderate-to-heavy events, but IFS remains essential for the highest precipitation events.Source: Discussion · p. 8
  • Over 10%: Laxmi (∼25 km) surpasses NeuralGCM-precip (∼280 km), while systematically outperforming its ∼100 km variant and resolving features at ten times finer scale.Higher spatial resolution is described as crucial for AI weather model precipitation accuracy.Source: Discussion · pp. 8–9
  • IMERG training mitigates reanalysis-inherited drizzle and extreme-precipitation biases but introduces retrieval and calibration biases; real-time deployment requires fine-tuning on operational analysis.That fine-tuning is needed for compatibility with real-time initial conditions.Source: Discussion · p. 9
  • Additional satellite-derived fields, including ocean temperatures, soil moisture, and radiation, could enhance learned AI weather dynamics and extend predictability horizons.Source: Discussion · p. 9

Methods

The methods train Laxmi on IMERG precipitation observations while retaining ERA5 atmospheric inputs and the AIFS-CRPS graph-transformer architecture. They define global and localized evaluations using probabilistic scores and operational warning metrics.

Data and pre-processing

  • IMERG precipitation rates were converted to 6-hour totals and regridded for training, while ERA5 atmospheric inputs excluded convective precipitation, runoff, and snowfall.IMERG-trained models used data from June 2000 through December 2023, with 2024 and 2025 held out; global evaluations were limited to 60°S–60°N.Source: Data and pre-processing · p. 9

Model architecture and optimization

  • Laxmi retains AIFS-CRPS’s graph-transformer architecture and uses the almost fair CRPS loss to reduce precipitation-field over-smoothing.Training uses a four-member ensemble and a three-stage autoregressive rollout; all stages use ERA5 except precipitation, which is scored against IMERG.Source: Model architecture and optimization · p. 10

Probabilistic climatology baseline

  • A 50-member IMERG climatology ensemble was constructed by resampling historical observations from nearby calendar dates and matching the target time of day.Source years were drawn uniformly from 2001–2019, with source dates selected within a 15-day window centered on the target date.Source: Probabilistic climatology baseline · p. 10

Indian monsoon case study setup

  • The storm study evaluated hindcasts for 10 qualifying Indian monsoon systems using four-day precipitation accumulations spanning one day before to three days after landfall.Forecasts were initialized at the latest 06:00 or 18:00 UTC cycle providing at least 72 hours’ lead time before estimated landfall.Source: Indian monsoon case study setup · pp. 10–11

Evaluation metrics

  • Global medium-range skill was assessed with CRPS, unbiased ensemble-mean RMSE, and spread-skill ratio, while extreme-event probabilities were scored with fair Brier scores and Brier Skill Scores.The Brier Skill Score compares the fair Brier score with a climatological score based on the resampled IMERG ensemble.Source: Evaluation metrics · p. 11
  • Localized storm scores were restricted to an IBTrACS-based region with a 500 km buffer and grid cells where observations or models assigned nonzero exceedance probability.Operational warnings used a 30% ensemble threshold; hits, misses, and false alarms supported POD, FAR, and CSI calculations.Source: Evaluation metrics · pp. 11–12

Data availability.

The paper lists its data sources, code access conditions, funding-related affiliations, and competing-interest disclosures.

Data availability.

  • ERA5 and IMERG data, IBTrACS storm tracks, and IFS ensemble forecasts were obtained from their respective climate, NASA, NOAA, and ECMWF sources.IFS ensemble forecasts were downloaded in July 2026 using Herbie.Source: Data availability · p. 12

Code availability.

  • Training used the open-source Anemoi framework, evaluations used WeatherBenchX, and code is available upon request for peer review.Source: Code availability · p. 12

Funding.

Competing interests.

  • Three authors own Alphabet stock, and Aditi Sheshadri contributed as a paid consultant outside her Stanford University duties.Source: Competing interests · p. 12

Supplementary information for: Improving global precipitation forecasts with an AI

The supplement defines the model-variant nomenclature and documents precipitation-distribution and forecast-score comparisons across regions, resolutions, and storm cases. It reports that Laxmi exhibits superior medium-range skill past day 6, while noting a slight IFS evaluation advantage.

Note: Model Variants and Nomenclature

  • Laxmi is IMERG-N320-CRPS; the supplement also compares lower-resolution IMERG-O96-CRPS and ERA5-O96-CRPS, and includes a ∼2.8° comparison with NeuralGCM-precip.The variant labels distinguish precipitation training data, model resolution, and loss function.Source: Note: Model Variants and Nomenclature · p. 16, Figure 10 · p. 25
  • ERA5 overestimates light rain below 3 mm day−1 and underestimates heavy precipitation extremes in both the tropics and extratropics.The supplementary figures show ERA5 and IMERG mean precipitation climatology and regional precipitation distributions.Source: Figure 1 · p. 17, Figure 2 · p. 18, Figure 3 · p. 19
  • Past day 6, Laxmi exhibits superior medium-range skill; the global comparison evaluates CRPS, ensemble-mean RMSE, Brier skill score, and spread-skill ratio.IFS evaluation switches forecast initialization times after day 6, which may give it a slight advantage.Source: Figure 4 · p. 20
  • The global extreme-precipitation figures evaluate Brier scores and Brier skill scores at the 90th, 95th, and 99th precipitation quantiles, at 0.25° and 1° resolutions.The 1° comparisons add ERA5-O96-CRPS and IMERG-O96-CRPS.Source: Figure 5 · p. 21, Figure 6 · p. 22
  • Tropical supplementary evaluations cover forecasting scores and extreme-precipitation Brier scores and skill scores at multiple quantiles and resolutions.The forecasting-score evaluation includes CRPS, unbiased ensemble mean RMSE, Brier skill score, and spread-skill ratio.Source: Figure 7 · p. 23, Figure 8 · p. 24, Figure 9 · p. 25
  • Additional decision-oriented evaluations cover 10 Indian storms, five U.S. cyclones, and 46 globally landfalling tropical cyclones.The figures include event-level precipitation forecasts and comparisons across precipitation thresholds.Source: Figure 11 · p. 26, Figure 12 · p. 27, Figure 13 · p. 28
Loading 2609.03210v3…