Original title: Hourly U.S.-wide flood simulation beyond the limits of traditional and data-driven models
Wencong Yang, Leo Lonzarich, Yalan Song, Haoyu Ji +3 · September 2026
Floods can hit just hours after a storm, even along rivers without water sensors. Current prediction tools struggle to forecast these fast, rare events across whole nations. To fix this, researchers built a new model that blends physical river science with machine learning. It tracks long-term moisture, sudden storms, and water that cannot soak into the ground, letting it spot extreme hourly spikes that daily tools miss. Across the conterminous United States, it outperforms the government's current operational system. It cuts flood-peak timing errors down to 4.5-6 hours from 7-8 hours. It also catches 33% more >=50-year floods than the national system, while beating other artificial intelligence tools at estimating peak water sizes during rare, major floods.
Imagine cooking a meal while a guest is about to arrive. The sauce can go from fine to ruined in under a minute, so checking once at the start or even once every hour tells you almost nothing about the moment the pot boils over. Here, a flood model watches every river segment every hour, so a sudden rainstorm that creates a fast spike in water level is seen right away instead of being smoothed into a flat daily average.
Floods can strike within hours, often in places with no river gauge. This model simulates water levels hourly across more than 800,000 U.S. river reaches, catching extreme flood peaks that daily forecasts miss. That gives communities a real chance to act before the water arrives.
The abstract doesn't state limitations directly; it focuses on reported improvements over NWM3.0 and LSTM baselines. It notes that overall hourly skill is only comparable to recent AI models, so gains concentrate on rare floods. It does not mention computational cost, forecast lead time, or behavior outside the conterminous U.S.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Hourly U.S.-wide flood simulation beyond the limits of traditional and data-driven models — Wencong Yang, Leo Lonzarich, Yalan Song et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.