Climate

Improving precipitation forecasts in an AI weather model using observational data

Julian F. Schmitt, Bertrand Delorme, Robert C. King, Yashica Patodia, Tapio Schneider, Aditi Sheshadri, Ravi Jain

Featured September 6, 2026

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

By teaching an AI weather model using real satellite rain data instead of older simulated data, the authors made it much better at predicting both light drizzles and heavy storms, giving more trustworthy forecasts.

In depth

The paper addresses persistent biases in AI weather prediction (AIWP) models, specifically the overprediction of drizzle and underprediction of extreme rainfall, which stem from training on reanalysis datasets like ERA5. The authors demonstrate that fine-tuning a graph-transformer AIWP model (dubbed Laxmi) with high-resolution, observation-based IMERG precipitation data significantly mitigates these biases, leading to improved medium-range forecasts and better calibrated probabilistic predictions.

Key Takeaways

1
AI weather models trained on reanalysis data like ERA5 inherit systematic biases, particularly over-forecasting drizzle and under-forecasting extreme precipitation.
2
Fine-tuning a graph-transformer architecture with high-resolution, observation-based IMERG precipitation data (creating the Laxmi model) substantially reduces these biases.
3
The Laxmi model achieves up to 19% improvement in Continuous Ranked Probability Score (CRPS) and 57% improvement in Brier Skill Score for extreme rainfall, demonstrating superior skill for tropical storms and drizzle events.

Conceptual Flow

HIGH LEVEL

Methodology: Training with Real Rain Data

The scientists taught a smart weather computer model by showing it actual rain measurements from satellites instead of just computer simulations.

Old Weather Data

Satellite Rain Data

Learn from Both

Smarter Weather Model

Results: Better Rain Predictions

This new way of teaching made the model much better at predicting when and where it would rain, especially for big storms and light drizzles.

Old Rain Forecasts

New Rain Forecasts

Compare Accuracy

More Accurate Rain Maps

Fewer Mistakes

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.