Advancing ECMWF’s AIFS Using IMERG and MSWEP Precipitation Products Authors: Gabriel Moldovan (European Centre For Medium Range Weather Forecasts); Ewan Pinnington (ECMWF); Matthew Chantry (ECMWF); Richard Forbes (ECMWF); Ana Prieto Nemesio (ECMWF) Corresponding author email: gabriel.moldovan@ecmwf.int Abstract: Forecasting precipitation remains a persistent challenge for both numerical weather prediction (NWP) and AI-based models. The AIFS v1.0, ECMWF’s first operational AI forecasting system, has demonstrated strong performance—often surpassing the physically-based IFS in key precipitation forecast metrics—but also inherits the biases present in its training data: ERA5 and IFS operational analyses. To explore further improvements, we investigate the use of alternative precipitation datasets for fine-tuning the AIFS-ERA5/IFS analysis trained model, specifically IMERG (Integrated Multi-satellitE Retrievals for GPM) and MSWEP (Multi-Source Weighted-Ensemble Precipitation). These global products combine satellite observations and gauge-based corrections, offering high-resolution (0.1°) estimates and multi-decadal coverage. Early experiments using these datasets for AIFS training show promising results, particularly with IMERG. Evaluation over the summer of 2023 highlights increased skill in tropical regions for IMERG. However, forecasts in the Southern Hemisphere deteriorate, likely due to sparser observational data and lower input quality. This degradation is less pronounced for the MSWEP-trained model. In contrast, in the Northern Hemisphere, the AIFS model trained on only ERA5/IFS analysis continues to perform best, with the MSWEP-trained version performing comparably. These findings suggest that satellite-gauge blended datasets could play a key role in improving AI-based precipitation forecasting, particularly in the tropical regions.

Ground Validation of the Tomorrow.io Pathfinder Radars over the Contiguous U.S. Authors: Daniel Watters (University of Oklahoma); Pierre Kirstetter (University of Oklahoma: NOAA Severe Storms Laboratory); Aimee Matland-Dixon (Rainmaker Technology Corporation); Sarah Ringerud (NASA Goddard Space Flight Center) Corresponding author email: daniel.watters@ou.edu Abstract: Smallsats are supporting earth observations from space though their performance is generally unknown, especially with respect to classical spaceborne missions. The Tomorrow.io Pathfinder radars (R1 & R2) are the first commercial smallsat Ka-band radars to observe precipitation, though their performance in retrieving surface precipitation rates is yet to be quantified by the scientific community. The objective of this work is to conduct the first validation of the Tomorrow.io Pathfinder radars relative to the Ground Validation Multi-Radar Multi-Sensor (GV-MRMS) product, a trusted ground reference for precipitation over the Contiguous U.S. (CONUS) which is used to support spaceborne retrieval validation and development. Specifically, GV-MRMS ingests the MRMS surface precipitation estimates (0.01°, 2-minute resolution), produced from NEXRAD ground-based radars, rain gauges and models, and subjects them to a range of quantity and quality control procedures to tailor the product toward validation of spaceborne precipitation products. Furthermore, we quantify the performance of Tomorrow-R1 & -R2 relative to the Global Precipitation Measurement (GPM) mission’s Ka-band Precipitation Radar (KaPR) by setting the GPM KaPR ground validation metrics relative to GV-MRMS as a benchmark for the Tomorrow.io sensors. Our methodology focuses on validating Tomorrow-R1 & -R2 level-2 surface precipitation footprint retrievals against GV-MRMS over CONUS during October-November 2023, the subset period acquired from NASA’s Commercial Satellite Data Acquisition division, and validating the GPM KaPR both in this Tomorrow.io availability period and over a larger timeframe. Initial key results highlight that the Tomorrow.io radars are skillful in distinguishing between rain and no rain events, including detecting approximately 82% of rain events over CONUS. Furthermore, Tomorrow-R1 & -R2 tend to overestimate light precipitation rates and underestimate heavier precipitation rates (>1-2 mm h-1) over CONUS, as similarly found for the GPM KaPR. Tomorrow-R2 outperforms Tomorrow-R1 in quantifying CONUS rain rates as demonstrated by a smaller underestimating mean relative bias (R1: -6%; R2: -22%) and superior correlation coefficient (R1: 0.73; R2: 0.93). This ground validation work is directly relevant to IPC Session 1.2 ‘Ground Validation of Precipitation Products: From Reference Data to Trustworthy Use’.

Reviewing past strategies and examining new approaches on the development of hybrid observational–modeling datasets Authors: Ana M. B. Nunes (Universidade Federal do Rio de Janeiro) Corresponding author email: ana.nunes@igeo.ufrj.br Abstract: More accurate initial conditions, together with better representation of physical processes, are still among the challenges that numerical weather prediction (NWP) faces. Developed for researchers at Florida State University, the assimilation of precipitation using a global spectral model emerged in the late 80s to provide more truthful NWP’s initial conditions in the tropics. Here, precipitation assimilation was applied in the Regional Spectral Model, which was originally developed at the National Centers for Environmental Prediction for improving hydroclimate simulations. This algorithm was developed at the Federal University of Rio de Janeiro in Brazil for the South American hydroclimate reconstruction. It uses a spectral nudging technique and assimilates precipitation estimates from the Climate Prediction Center (CPC) Morphing Technique (MORPH). Results using CMORPH show improvements in the soil moisture representation in tropical areas, which include the Amazon rainforest. A brief overview of the precipitation assimilation applications spanning from improving initial conditions for short-range precipitation forecasting to hydroclimate reconstructions will be given. The numerous applications of precipitation assimilation in atmospheric modeling will be discussed. Further developments will also be presented on the assimilation of higher-resolution, satellite-based precipitation estimates. Many of the results form part of a review paper in preparation.