NextGen In A Box (NGIAB) and CIROH Cyberinfrastructure Power Flash Flood Research at the 2026 Summer Institute
On July 4, 2025, a flash flood on the Guadalupe River caused 119 fatalities in Kerr County, Texas. It was the deadliest flash flood in the United States since 1976, and a reminder that flash floods remain among the most dangerous weather hazards in the country. Forecasting them means working at the speed and scale of the storm: small basins, minutes to hours, and rainfall estimates that can make or break the prediction.
That challenge set the agenda for our theme at the 2026 Water Prediction Innovators Summer Institute: Advancing the Use of Quantitative Precipitation Inputs into the NextGen Framework. Ten graduate students in three teams (Don't Runoff!, FlashCast, and The Post-Processors) had seven weeks to bring sub-hourly radar rainfall, precipitation uncertainty, and ensemble forecasting methods into the NextGen Water Resources Modeling Framework. Seven weeks is not much, and the teams could not spend half of it compiling a modeling framework or downloading terabytes of radar data to laptops. Fortunately, they did not have to. The CIROH cyberinfrastructure team carried that load.
The framework fits in the Box
NextGen is powerful but challenging to run. NextGen In A Box (NGIAB), CIROH's containerized distribution with more than 41,000 Docker pulls, turns that setup problem into a single pull command. This year's Summer Institute members were able to hit the ground running after just a single one-hour NGIAB workshop, demonstrating its sheer accessibility.
However, the teams did more than just run it. NGIAB is also an open box, and they took full advantage by extending it:
- FlashCast built a modified NGIAB Docker image that runs CFE and T-Route on a 15 minute timestep and forced it with 15 minute MRMS radar rainfall, expanding the framework's standard hourly configuration to the sub-hourly scales flash floods demand.
- The Post-Processors modified the NGIAB data preprocess library so NextGen could ingest MRMS radar estimates and HRRR forecast rainfall in place of the default AORC forcing. They then simulated 8,474 catchments around Ellicott City, Maryland.
- Don't Runoff! built a volume-preserving workflow that maps 2 minute, 1 km MRMS rainfall onto NextGen Hydrofabric catchments and used it to drive dHBV2-Flash, a new sub-hourly differentiable model designed for the framework.
Every modification lives in a container or a public repository. Any researcher can pull the same image, rerun the same experiment, and build on it. That is the point of the box: NGIAB enables maximal portability and reproducibility to maximize research's impact. For the complete set of shared resources, see each team's final report.
R2OHC Platform: NSF Jetstream2 and NGIAB did the heavy lifting
None of this fits on a laptop. FlashCast alone ran 2,280 simulation scenarios per basin across three flash flood case studies (Asheville NC, Pittsburgh PA, and Kerrville TX) — 6,840 NextGen runs in total. The Post-Processors pushed 500 synthetic storms through NGIAB to train a neural network surrogate, on top of storm displacement experiments across thousands of catchments. Don't Runoff! processed radar rainfall for over 5,000 flash flood events across the country and trained a differentiable model on years of 15-minute data.
The computing resources came from CIROH cyberinfrastructure, which provided allocations on NSF Jetstream2, Indiana University's NSF-funded academic cloud, through CIROH's ACCESS allocation (EES240087). JetStream2 hosted virtual machines sized for each team's workload, from batch NGIAB runs to model training. The rainfall came from NOAA open data: 2-minute MRMS radar estimates, hourly AORC forcings, and HRRR forecasts, all pulled straight from public cloud archives. CUAHSI HydroShare closed the loop as the publication home for datasets and supplementary materials, and the NWC-CUAHSI-Summer-Institute GitHub organization hosts the code.
What the teams built
Team Don't Runoff! delivered RUNOFF v1.0, a CONUS-scale catalog of over 5,000 flash flood events from the NOAA Storm Events Database, each mapped to hydrofabric catchments and paired with a USGS gauge and quality-controlled MRMS rainfall. RUNOFF is designed to conduct flash flood modeling experiments using high-quality precipitation estimates. They also delivered dHBV2-Flash, a 15-minute differentiable rainfall-runoff model that captured event peaks in the Upper Neuse basin with a median peak flow bias of 16% across 1,132 events. Supplementary materials are on HydroShare.
Team FlashCast addressed a question at the heart of flash flood prediction: how fine, and how accurate, does the precipitation input need to be? Their sensitivity pipeline converts thousands of NGIAB runs into critical uncertainty matrices, the precipitation uncertainty a forecast can tolerate before predictions stop being actionable. Early results from the three case studies point to a scale-dependent answer: for these events, resolution finer than about 7 km and 1 hour added little on its own, as long as storm volume was preserved. The pipeline is built to put that threshold to the test across many more basins and storms. Methods and figures are on HydroShare.
The Post-Processors built NCET, the Neighboring Catchment Ensemble Technique, which turns a single deterministic NextGen simulation into an ensemble streamflow forecast by borrowing hydrographs from neighboring catchments. Applied to the May 2018 Ellicott City flash flood with a real HRRR forecast, NCET recovered a flood the deterministic forecast missed, with the largest gains in probabilistic skill when the ensemble weights were aimed along the storm's displacement. That early result suggests storm location, more than catchment similarity, should steer the ensemble, and it points to a computationally inexpensive path toward flash flood forecasting within NextGen. The team also prototyped Replace and Route, a BMI component that lets NextGen borrow discharge from high-fidelity external models at hydrofabric nexus points. Evaluation materials are on HydroShare.
From asking research questions to prototyping workflows, in seven weeks
Each product is hydrofabric-native, open, and packaged into containers, BMI modules, and reproducible pipelines — the way that the Community NextGen ecosystem expects. This ease of access and reuse is research to operations in miniature, and it is what CIROH cyberinfrastructure is for. NGIAB collapsed setup time from weeks to minutes. NSF's Jetstream2 scaled experiments beyond what any student laptop could run. Open data archives fed the models, and HydroShare and GitHub made every result citable and reusable. The students arrived with research questions. They left behind open prototypes, datasets, and code for the community to build on.
The 2026 Water Prediction Innovators Summer Institute took place June 8 — July 22, 2026. 24 graduate student fellows from 21 unique U.S. institutions formed research teams under the guidance of seven Theme Leaders, spanning topics from precipitation forecasting to flood risk communication. Research outcomes are summarized in the 2026 WPI SI final report.
The teams thank the CIROH Cyberinfrastructure & DevOps team and Hydroinformatics team at the Alabama Water Institute (Arpita Patel, Josh Cunningham, Quinn Lee, Nia Minor, and colleagues) for cloud provisioning and NextGen support throughout the program, Dr. Fred Ogden for guidance on developments with the NextGen framework, and the Summer Institute coordinators, Megan Vardaman and Nana Oye Djan. This work used Jetstream2 through allocation EES240087 from the NSF ACCESS program. This research was supported by the Cooperative Institute for Research to Operations in Hydrology (CIROH) under award NA22NWS4320003 from the NOAA Cooperative Institute Program.
