NEWS · MODELS · #831
Hapi: a U-Net Swin Transformer for continental-scale 24–72h hydrological forecasts
Hapi is a multivariable U-Net Swin Transformer that forecasts discharge, surface runoff, snow water equivalent, and soil wetness across the contiguous United States at 0.05° resolution for 24–72 hour lead times. On 2024 test data using reconstructed ERA5‑Land inputs it outperformed an operational physics-based model and a state-of-the-art AI model in flood detection, was validated against 3,881 USGS gauges and a Hurricane Helene case study, and runs a four-variable 72-hour forecast in an average 0.11 s on a single A100 GPU.
KEY POINTS
- Hapi is a multivariable U-Net Swin Transformer that forecasts discharge, surface runoff, snow water equivalent, and soil wetness across the contiguous United States at 0.05° resolution for 24–72 hour lead times.
- On 2024 test data using reconstructed ERA5‑Land inputs it outperformed an operational physics-based model and a state-of-the-art AI model in flood detection, was validated against 3,881 USGS gauges and a Hurricane Helene case study, and runs a four-variable 72-hour forecast in an average 0.11 s on a single A100 GPU.
- A transformer that improves medium-range, continental-scale flood detection while running fast enough for operational use could materially improve water-resource management and emergency response.
DECISION BRIEF
CONFIDENCE
WHAT CHANGED
A new research model, Hapi, a multivariable U‑Net Swin Transformer, was introduced and shown (on 2024 test data using reconstructed ERA5‑Land inputs) to (a) forecast discharge, surface runoff, snow water equivalent, and soil wetness across the contiguous United States at 0.05° resolution for 24–72 hour lead times; (b) outperform an operational physics‑based model and a state‑of‑the‑art AI model in flood detection; (c) validate against 3,881 USGS gauges and a Hurricane Helene case study; and (d) produce a four‑variable 72‑hour forecast for the contiguous U.S. with average inference time 0.11 s on a single A100 GPU.
WHY NOW
Accurate flood forecasts several days in advance are essential for flood control, water‑resource management, and emergency response. A transformer that improves medium‑range, continental‑scale flood detection while running fast enough for operational use could materially improve these activities.
WHO IS AFFECTED
Directly relevant parties include operational hydrological and flood‑forecasting centers, water‑resource managers, emergency‑response planners, and hydrology/ML research groups developing continental‑scale forecasting methods.
CONFIRMED
According to the supplied source (ID 984): - Hapi is a U‑Net Swin Transformer that forecasts discharge, surface runoff, snow water equivalent, and soil wetness across the contiguous United States. - Forecast resolution is 0.05° and lead times are 24–72 hours. - Training uses learned Laplacian task weights that adjust each variable's contribution. - On 2024 test data using reconstructed ERA5‑Land inputs, Hapi outperformed an operational physics‑based model and a state‑of‑the‑art AI model in flood detection. - Independent validation was performed against 3,881 U.S. Geological Survey gauges and a Hurricane Helene case study. - Controlled experiments showed learned task weighting strengthens rare‑flood detection, which is sensitive to precipitation inputs. - Hapi produced a four‑variable, 72‑hour forecast across the contiguous United States with average inference time 0.11 seconds on a single A100 GPU.
UNCERTAIN
Missing or unresolved items in the supplied excerpt (ID 984): - Missing evidence on performance using real‑time operational inputs (the reported tests used reconstructed ERA5‑Land inputs). - Missing details on exact metrics, statistical significance, and geographic/seasonal breakdown of the 'outperformed' claim (which metrics and where it outperformed are not specified in the excerpt). - Unclear operational readiness: data assimilation, input latency, end‑to‑end pipeline integration, and resource requirements beyond the single‑GPU inference time are not provided. - Unclear robustness to input errors, sensor outages, or different reanalysis/forecast input sources. - Missing evidence on independent replication or peer‑review beyond the arXiv report in the excerpt.
WHAT TO WATCH
Concrete observable signals to monitor (can be checked against publications, code releases, or operational trials): - Availability of code, pretrained weights, and input preprocessing scripts for Hapi (publication or GitHub release). - Peer‑reviewed publication or conference presentation expanding on the arXiv report with full metrics and significance tests. - Independent evaluations comparing Hapi to operational models using live operational inputs (not reconstructed ERA5‑Land) and reporting metrics by region and season. - Operational pilot deployments or trial integrations by forecasting centers, and measured end‑to‑end latency and resource use in those pilots. - Further validation reports covering more events beyond the Hurricane Helene case study and breakdowns of false positives/negatives in flood detection. - Reproduction studies by other research groups using different input datasets to test robustness.
WHY IT MATTERS
A transformer that improves medium-range, continental-scale flood detection while running fast enough for operational use could materially improve water-resource management and emergency response.
EVIDENCE MAP
4Editorial claims linked to specific sources, with support, contradiction and context shown separately.
Hapi is a multivariable U‑Net Swin Transformer model.
SUPPORTEDChecked 2026-09-24 · 1 supporting
Hapi forecasts discharge, surface runoff, snow water equivalent, and soil wetness over the contiguous United States at 0.05° resolution for 24–72 hour lead times.
SUPPORTEDChecked 2026-09-24 · 1 supporting
On 2024 test data using reconstructed ERA5‑Land inputs, Hapi outperformed an operational physics‑based model and a state‑of‑the‑art AI model in flood detection.
SUPPORTEDChecked 2026-09-24 · 1 supporting
Independent validation against 3,881 USGS gauges and a Hurricane Helene case study supported Hapi's advantage over the physics‑based model in reproducing daily discharge.
SUPPORTEDChecked 2026-09-24 · 1 supporting
SOURCES & TIMELINE
1arXiv:2609.22702v1 Announce Type: new Abstract: Accurate flood forecasts several days in advance are essential for flood control, water-resource management, and emergency response. Producing them at high resolution over a continental domain calls for local hydrological detail together with spatial context extending from river basins to synoptic weather systems. We developed Hapi, a U-Net Swin Transformer that uses fi…
arXiv:2609.22512v1 Announce Type: new Abstract: Consensus among LLM judges is often taken as strong evidence that a decision is correct. This assumes that judges make their errors independently. In practice, LLM judges are often trained and evaluated in similar ways, so they can make the same mistakes. We study how this dependency affects the reliability of consensus. We find substantial error correlation across both…
arXiv:2609.27276v1 Announce Type: new Abstract: Long-horizon language-model agents accumulate reasoning traces, tool exchanges, and observations whose relevance changes with the current decision. Existing compression strategies often score historical units independently, but the safety of deleting several units is generally not determined by their singleton scores: redundant evidence, accumulated small effects, and t…