Preprint: GPU-native flood model Inunda claims faster, higher-resolution simulations and better Hurricane Harvey hindcast than NOAA model

A newly posted arXiv preprint says a GPU-native flood model called Inunda can simulate large, high-resolution flood events far faster than typical approaches — and, in one Hurricane Harvey hindcast, outperformed NOAA’s operational National Water Model on several accuracy measures.

The paper is a preprint, not a peer-reviewed study. Posted to arXiv as arXiv:2607.09614v1 and submitted July 10, it lists Zhi Li as the author. That means its results have not yet been vetted through peer review or independently verified, and its model comparisons should be treated as provisional.

If the claims hold up, they could matter for flood forecasting and emergency response. High-resolution inundation modeling — simulating where water goes across streets, neighborhoods and floodplains — is computationally expensive. A faster physics-based model could make it more practical to run real-time forecasts, test multiple rainfall scenarios and produce updated hazard maps as a storm unfolds.

The preprint, “Inunda: A GPU-Native, Agent-enabled, Differentiable Solver for High-Resolution Flood Inundation Modeling,” describes Inunda as a flood inundation model built to run on graphics processors rather than conventional CPU-based workflows. According to the arXiv abstract, it solves the two-dimensional shallow water equations, a standard set of equations for moving surface water, using a mass-conservative local-inertial scheme. The abstract says the model “runs multi-day events over millions of cells in minutes on a single GPU.”

The paper also argues that Inunda is differentiable by construction. In plain terms, that means every operator is compatible with automatic differentiation, allowing model parameters to be adjusted through gradient descent by backpropagating through the full simulation. The preprint presents that as a way to calibrate flood models more directly, rather than relying only on more traditional trial-and-error parameter tuning.

Its strongest reported evidence comes from Hurricane Harvey in Harris County, Texas, a widely used test case because of the extensive surveyed flood observations collected after the storm. In the abstract, the author reports that Inunda matched surveyed high-water marks with a mean absolute error of 0.67 meters. For water levels at gauges, the abstract reports a median Nash-Sutcliffe efficiency of positive 0.72, which it says was “more than double the +0.31 of the operational National Water Model v3.0.”

That comparison gives readers a practical benchmark because the National Water Model is NOAA’s operational continental hydrologic model. But the abstract alone does not establish whether the models were run with identical inputs, gauge sets and evaluation methods, all of which can affect head-to-head results.

The preprint also includes two other real-world applications. One examines the July 2025 Central Texas flash flood, where the abstract says Inunda used an 18-member, 1-kilometer convection-allowing precipitation ensemble to generate probabilistic flood forecasts, with skill improving as lead time to crest shortened. The other focuses on flash flooding in the Rio Ruidoso burn scar, where the abstract says differentiable calibration recovered saturated hydraulic conductivity — a measure of how readily water moves through soil — as a spatially explicit field and tracked its recovery over multiple years after fire.

The abstract describes Inunda as an “open, end-to-end pipeline” for real-event flood modeling. Public CV material cited in the research identifies a hydrology and hydraulics researcher named Zhi Li as an assistant professor at the University of Connecticut as of an August 2025 CV update, matching the author name and field.

For now, though, the headline claims are still just that: claims in a newly posted preprint. Before Inunda should be treated as an established operational advance in flood forecasting, its speed, accuracy and calibration results will need independent verification and peer review.