AI weather models ace light rain but physics still wins when floods loom

Artificial intelligence has stormed the world of weather forecasting, with deep-learning systems from Google DeepMind, ECMWF and Chinese research groups now producing global forecasts in seconds on a single graphics card. But a new head-to-head trial suggests that when the rain turns heavy enough to kill people and flood towns, the old physics-based machines still hold the upper hand. In a study published in Geoscientific Model Development, Haseeb ur Rehman and Félicia Norma Rebecca Teferle of the University of Luxembourg pitted three leading AI weather models against a carefully tuned regional physics model across three devastating flood events in Luxembourg and its neighbouring countries.

The three AI contenders were GraphCast, Google DeepMind’s graph neural network trained on four decades of reanalysis data; AIFS Single 1.0, the European Centre for Medium-Range Weather Forecasts’ own data-driven system; and FuXi, a cascaded transformer developed at Fudan University and the Shanghai AI Laboratory. Their opponent was the Weather Research and Forecasting model, or WRF, a community physics-based system that solves the differential equations governing atmospheric motion. The WRF configuration ran on a 12-kilometre grid over Luxembourg and the Greater Region, and came in two versions: one started cold from the ERA5 reanalysis, and one continuously corrected by a three-dimensional variational data assimilation scheme that ingested conventional surface, radiosonde, aircraft and satellite observations along with GNSS-derived atmospheric water measurements every six hours.

The test bed was deliberately demanding. The team simulated three high-impact flood episodes: the flash flood of 21 to 23 July 2016 in north-eastern Luxembourg, the Mullerthal flash flood of 31 May to 2 June 2018, and the catastrophic European flood of 14 to 15 July 2021, which claimed more than 200 lives across western Europe. Each event was simulated over a window of roughly one month, giving around 90 simulated days per configuration. Verification relied on an independent network of 259 to 280 rain gauges spread across Luxembourg, Germany, France and Belgium, screened to exclude any station within 12 kilometres of an observation assimilated into the physics model, yielding 92,119 matched forecast-observation pairs for precipitation and 45,925 for temperature.

The results split cleanly along the intensity of the rain. At a modest threshold of 1 millimetre per six hours, the AI systems shone. AIFS detected rainfall better than anything else, with a probability of detection of 0.835 against 0.611 for the assimilated WRF, while GraphCast achieved the highest overall categorical skill, a critical success index of 0.424. Both AI leaders also crushed the physics model on temperature, cutting the root-mean-square error by 19.0 and 17.6 percent respectively, the largest and most statistically secure differences in the entire study. For ordinary weather, the machines trained on historical data are simply better.

Then the thresholds rose, and the picture inverted. At 5 and 10 millimetres per six hours, AIFS was the strongest system, but at 20 millimetres per six hours, the kind of deluge that actually triggers flash floods, WRF took the lead with a critical success index of 0.160, ahead of AIFS at 0.141 and GraphCast at 0.114. FuXi fared worst of all: across the entire verification set it did not forecast a single exceedance of the 20-millimetre threshold. The same reversal appeared in space as well as in counts. Verified against a Belgian weather-radar composite for the July 2021 flood using the Fractions Skill Score, AIFS led at light-rain thresholds at every neighbourhood width, but at 20 millimetres both WRF configurations outperformed all three AI models.

The reason lies in how the AI systems were trained. Models optimised with pixel-wise loss functions are rewarded for predicting the conditional mean of the atmosphere, which suppresses the extremes of the distribution. The resulting precipitation fields are visibly smooth: in the radar comparison for the July 2021 event, GraphCast captured the correct envelope of the intense rain band stretching from Belgium across Luxembourg into western Germany but strongly underestimated its peak intensities, while FuXi produced almost no precipitation signal at all during the flood’s peak hours. The physics model, by contrast, reproduced the banded mesoscale structure and its sharp gradients, and during the flood days it achieved the lowest false-alarm ratios recorded anywhere in the study, meaning that when it did predict rain during such an episode, it was almost always right.

The study also delivered a nuanced verdict on data assimilation. Feeding observations into WRF every six hours improved the detection of rainfall occurrence at the 1-millimetre threshold and brought the near-surface temperature bias essentially to zero, but the benefit did not extend to heavier rain: at 10 millimetres the assimilated run actually scored below the unassimilated one, a degradation that appeared in all three events and, intriguingly, was not ordered by how densely each event was observed. The authors are careful not to attribute a cause, noting that with only three events, observation density, observation type and synoptic character cannot be separated, and they recommend systematic observation-denial experiments as the natural next step.

One further finding will reassure sceptics of machine-learning forecasting. GraphCast was trained on data ending in 2017, yet its categorical skill was highest on the July 2021 flood, an event entirely outside its training period, with a critical success index of 0.493 compared with 0.430 for the in-training 2016 event. That is evidence of genuine generalisation rather than memorisation of training samples, although the authors caution that 2018, also after the training cutoff, was GraphCast’s weakest event, so the pattern is not simply a matter of seen versus unseen data. The same test could not be applied to AIFS, whose fine-tuning window of 2016 to 2022 encompasses all three case studies, meaning its scores should be read as potentially optimistic.

The computational contrast is staggering. One six-hour WRF cycle takes roughly 37 minutes on 52 CPU cores, while an AI forecast for the same step takes about 1.5 minutes on a single NVIDIA A100 GPU, a factor of roughly 140, albeit after a one-off training cost borne by the developers. Yet the authors resist declaring a winner. The AI systems ran on global grids of 25 to 28 kilometres against WRF’s 12 kilometres, so part of their smoothing reflects resolution rather than formulation, and refining the WRF grid to 1.33 kilometres raised the simulated peak accumulation to match the radar-observed maximum without improving neighbourhood scores. Their suggested strategy is complementary rather than competitive: use AI models for rapid temperature fields and rainfall detection, and keep observation-corrected physics models where the heaviest rates, the ones that turn steep valleys into deadly torrents, actually matter.

Subject of Research: Comparative evaluation of AI and physics-based weather models for flood-related rainfall forecasting

Article Title: Physics vs. AI in weather prediction: evaluating GraphCast, AIFS, and FuXi against an observation-corrected WRF model for flash floods

Article References: Rehman, H. U., & Teferle, F. N. R. (2026). Physics vs. AI in weather prediction: evaluating GraphCast, AIFS, and FuXi against an observation-corrected WRF model for flash floods. Geoscientific Model Development, 19(19), 9831-9855. https://doi.org/10.5194/gmd-19-9831-2026

Image Credits: AI Generated

DOI: 10.5194/gmd-19-9831-2026

Keywords: weather forecasting, AI weather models, GraphCast, AIFS, FuXi, WRF, data assimilation, flash floods, precipitation verification, Luxembourg, ERA5, extreme rainfall