Just out todayAugmented reality: The AR trainees were faster on the model they trained on, and no better anywhere elseVirtual reality: They matched the exercise dose exactly, and the VR advantage fell below the line that mattersVirtual reality: Both groups used the same headset. One watched a video first and picked up 10.8 more blocks

Artificial intelligence/China

Snow or rain, answered in real time instead of four hours late

A hybrid AI system determines whether precipitation is falling as rain or snow in real time, where operational products take at least four hours.

What the paper found

RePPIC-Net integrates 3D atmospheric physics fields from the AI-driven FuXi model with operational geostationary satellite observations to determine surface precipitation phase in real time, against the at-least-4-hour latency of current operational systems. Validated against ground stations in China it reached a Critical Success Index of 0.1574 for snowfall and 0.3147 for rainfall, against 0.1001 and 0.3064 for the operational products.

Grade 12 reading level6 min read

What happened

Whether precipitation arrives as rain or as snow determines whether a road stays open, whether a roof takes a load, whether a river floods now or in the spring. It is one of the most operationally consequential things a weather service can tell you.

It is also hard, because satellites see cloud tops and the answer depends on the temperature profile near the ground. Surface observations that would settle it are scarce.

Current operational systems answer with at least four hours of latency, by which point the storm has moved.

A team at the Key Laboratory of Polar Atmosphere-Ocean-Ice System for Weather and Climate, the College of Future Information Technology at Fudan University and the State Key Laboratory of Severe Weather Meteorological Science and Technology at the Chinese Academy of Meteorological Sciences built RePPIC-Net to answer in real time.

The test

The architecture is hybrid in a specific sense: it fuses real-time 3D atmospheric physics fields from the AI-driven FuXi model with operational geostationary satellite observations, through a hierarchical structure.

Using an AI weather model’s output as an input to a satellite retrieval is the interesting move. The satellite sees the cloud; FuXi supplies the atmospheric profile beneath it; the network combines them to infer what is reaching the ground.

Validation was against ground stations in China, with direct comparison to the operational products currently in use – which is the comparison that matters, rather than against machine learning baselines.

The Critical Success Index was the metric, reported separately for snowfall and rainfall in the 0.1 to 5 millimetre per hour precipitation band.

The numbers

Measure RePPIC-Net Operational products
Latency real time at least 4 hours
Snowfall CSI (0.1-5 mm/h) 0.1574 0.1001
Rainfall CSI (0.1-5 mm/h) 0.3147 0.3064
Inputs FuXi 3D physics fields + geostationary satellite observations
Validation ground stations in China
Target application 1-3 hour global phase transition warnings
Data public FY-4B and GPM satellite datasets

Both CSI columns deserve a moment. A Critical Success Index of 0.1574 means that counting hits, misses and false alarms together, the system gets a small minority of snowfall phase determinations right in this precipitation band.

The result

RePPIC-Net enables real-time monitoring of surface precipitation phase, against the at-least-four-hour latency of current operational systems.

Validated against ground stations in China, it achieved a Critical Success Index for phase and detection of 0.1574 for snowfall and 0.3147 for rainfall in the 0.1 to 5 millimetre per hour band, outperforming the operational products’ 0.1001 and 0.3064.

The real-time discrimination capability allows development of a satellite-based surface precipitation phase nowcasting system, meeting the need for one to three hour global phase transition warnings.

The authors present RePPIC-Net as a replicable blueprint for AI-powered real-time weather monitoring, filling a gap in wintertime weather disaster warnings.

How this compares to what came before

Precipitation phase determination from satellites has used threshold methods on brightness temperature and, more recently, machine learning classifiers trained on reanalysis. Those systems inherit the latency of the reanalysis they depend on, which is where the four hours comes from.

The structural contribution here is using an AI weather model’s forecast fields as the atmospheric profile input instead of waiting for reanalysis. FuXi produces those fields fast enough to be used in real time, which is what collapses the latency.

That is a pattern worth noting beyond this application: AI weather models are being evaluated mainly as forecasting systems competing with physical models, and this paper uses one as a fast source of atmospheric state for a downstream retrieval. That second role may matter as much as the first.

What it means

The latency improvement is the real contribution and it is substantial. Four hours to real time, for a phenomenon that changes within the hour, is the difference between a product that describes what happened and one that supports a decision.

The skill numbers need to be read honestly. A 57 percent relative improvement in snowfall CSI is meaningful and 0.1574 is a low absolute score. Most light snowfall phase determinations in this band are still wrong or missed. For rainfall the improvement is marginal – 0.3147 against 0.3064 – which means the gain is concentrated in the harder case.

That split makes sense physically. Rain is the default; snow requires the specific near-surface conditions that are hardest to infer from space. It also means the system’s value is specifically in the snow case, which is where the operational need is.

Anyone deploying this should treat it as a better-than-nothing real-time indicator rather than a reliable determination, and should say so to whoever acts on it.

Business ideas from this paper

Road and transport winter operations. Real-time phase information, even at modest skill, is better than the four-hour-old product road authorities currently use for gritting and closure decisions.

AI weather model fields as a data product. If FuXi-style outputs are useful as inputs to downstream retrievals, there is a business in serving those fields at low latency to anyone building environmental products.

Phase-aware hydrological modelling. Whether precipitation falls as rain or snow changes runoff timing by months. Water managers in mountain catchments have a direct use for real-time phase.

Skill-honest product design. A system with CSI 0.1574 needs an interface that communicates uncertainty rather than a binary answer. Designing that well is the difference between a useful tool and a misleading one.

What could be wrong with it

The absolute skill is low. CSI 0.1574 for snowfall means most events in this band are missed or misplaced, and the headline should not be read as a solved problem.

The rainfall improvement, 0.3147 against 0.3064, is small enough to be operationally negligible.

The reported figures cover the 0.1 to 5 millimetre per hour band, which is light precipitation. Heavy snowfall – the case that closes roads and collapses roofs – is not covered by these numbers.

Validation is against ground stations in China only. Phase determination depends on regional atmospheric structure and the result may not transfer.

The system depends on FuXi outputs, so its performance is bounded by that model’s accuracy and availability.

No independent replication.

How sure can you be?

We score this 8 out of 10.

It is a peer-reviewed paper in Nature Communications, validated against ground stations with a direct comparison to the operational products it would replace, with named institutions including Fudan University and the Chinese Academy of Meteorological Sciences, and it uses publicly available FY-4B and GPM satellite datasets.

It loses points because there is no independent replication and the evidence is retrospective validation rather than operational deployment.

The comparison against operational products rather than research baselines is what distinguishes this from most remote sensing machine learning papers. It also makes the low absolute scores visible, which is to the authors’ credit.

Do this today

If you make winter operational decisions, ask what the latency of your current precipitation phase product is. Four hours is common and most users have never checked.

If you build environmental retrievals, AI weather model fields are now available fast enough to use as live inputs. That changes what is possible downstream.

If you read remote sensing results, look for the Critical Success Index rather than accuracy. Accuracy is flattered by the fact that most of the time it is not precipitating; CSI is not.

Source: Snow or rain? Hybrid AI deciphers surface precipitation phase from satellite observations, Nature Communications, February 2026. PMID 41698920 · doi.org ().

Just Out Tech explains new research in plain language. This article was drafted with AI assistance and checked by a human against the original source.

What to remember
  • Surface precipitation phase transition drives snowstorms and avalanches and is limited by the scarcity of surface observations.
  • Real-time output against at least 4 hours of latency in current operational systems.
  • Snowfall CSI 0.1574 against 0.1001; rainfall CSI 0.3147 against 0.3064.
  • Uses physics fields from an AI weather model as input, fused with geostationary observations.
Where this really is

RePPIC-Net was validated against ground stations in China and compared directly against the operational products it would replace, using operational geostationary satellite observations.

Questions people ask

What is Critical Success Index?

A forecast verification score combining hits, misses and false alarms. It runs from 0 to 1 and a score of 0.1574 means most events are being missed or wrongly placed.

Why is the phase so hard to determine?

Satellites see cloud tops, not the ground. Whether precipitation arrives as rain or snow depends on the temperature profile in the lowest part of the atmosphere, which satellites infer rather than observe.

What is FuXi?

An AI-driven weather model that produces 3D atmospheric physics fields. Using its output as an input to a satellite retrieval is the hybrid part of this hybrid AI framework.

Is beating 0.1001 with 0.1574 meaningful?

It is a 57 percent relative improvement on a low base. In absolute terms both systems miss most light snowfall events, and that should be stated alongside the improvement.

About the author

Mark Alex

Mark Alex is the founder and Managing Director of Real Biz Digital, a technology company operating out of Nairobi since 2018. He works in agentic AI and the Model Context Protocol, AI governance, enterprise software architecture and cybersecurity. He holds an MSc in Mechatronical Engineering from Obuda University in Budapest and a BSc in IT, Forensic Technology and Cybercrime, from USIU-Africa in Nairobi, and has published IEEE conference research on an AI-powered digital twin for greenhouse systems. He is the author of seven books. Between 2020 and 2024 he mentored more than 200 university students and interns in Nairobi. He writes every Just Out Tech article from the original research paper.