Physics vs. AI inWeather Prediction: Evaluating GraphCast, AIFS, and FuXi against an Observation-Corrected WRF Model for Flash Floods

Abstract. Data-driven artificial intelligence (AI) weather models are increasingly positioned as alternatives or complements to physics-based numerical weather prediction (NWP) systems, yet systematic model evaluation studies that compare both paradigms under controlled experimental conditions remain limited. Here we present a structured model evaluation framework applying standardised verification metrics to assess a regional WRF configuration against three publicly available AI models – GraphCast, AIFS Single 1.0, and FuXi – over three high-impact flood events in Luxembourg and the Greater Region (2016, 2018, and 2021), comprising 90 simulation days and over 7300 matched forecast-observation pairs. The WRF model employs a three-dimensional variational (3D-VAR) data assimilation scheme ingesting GNSS Zenith Total Delay and conventional observations on a 6-hourly Rapid Update Cycle. All systems are initialised from ERA5 reanalysis to ensure consistent initial conditions. Categorical precipitation scores at the 1 mm threshold and continuous temperature metrics at surface stations serve as verification targets. Data assimilation measurably improves WRF's categorical precipitation skill (Critical Success Index: 0.306 to 0.341) while leaving near-surface temperature largely unchanged, demonstrating the value of the assimilation scheme. AIFS achieves the highest detection rate (POD 0.765) and net categorical skill (CSI 0.370), GraphCast and AIFS reduce temperature RMSE by ~10 % relative to assimilated WRF, and WRF alone reproduces the observed mesoscale precipitation structure of the catastrophic July 2021 flood. The per-event breakdown reveals no skill degradation for AI models on out-of-sample events, suggesting that meteorological regime rather than training-data overlap governs model performance. These results provide a replicable model evaluation methodology for benchmarking emerging data-driven NWP systems against observation-corrected regional models.