This dataset contains modified versions of ECMWF archive datasets (source: www.ecmwf.int) and ERA5, which can be downloaded through the Copernicus Climate Data Store ECMWF archive data sets The ECMWF archive data is published under a Creative Commons Attribution Non-Commercial 4.0 International (CC BY NC 4.0). https://creativecommons.org/licenses/by-nc/4.0/legalcode. License notes are also summarized here. Disclaimer: ECMWF does not accept any liability whatsoever for any error or omission in the data, their availability, or for any loss or damage arising from their use. ERA5 We modified ERA5 data which is a product of the Copernicus Climate Change Service. Neither the European Commission nor ECMWF is responsible for any use that may be made of the Copernicus information or data it contains.

Data sources The inputs to the dataset preprocessing were obtained from the following sources: Deep learning forecasts were produced by re-running the deep-learning weather models ourselves. The code to produce forecasts is available on the respective model repositories(GraphCast, Pangu-Weather, and FourCastNet). The ERA5 and HRES-fc0 ground truth data sets as well as the HRES forecasts were retrieved from the following sources: WeatherBench 2 - used when data was available for our case studies, TIGGE data retrieval portal to retrieve 00:00 / 12:00 UTC HRES forecasts, ECMWF's Operational Archive to retrieve 06:00 / 18:00 UTC HRES forecasts. This required obtaining a research license. Copernicus Climate Data Store to retrieve ERA5 data.

For each case study, we have two files _gt.nc and _fc.nc in which we combine the ground truth and forecast data sets respectively. For the Pacific Northwest heatwave case study, we additionally include a climatology for ERA5, which we adapted from WeatherBench 2, and two files for each of the two years 2020 and 2022, which we used as baselines in some analyses. 2021_NA_winter_storm_college_station_ts.csv contains time-series of wind speed, temperatures, and wind chill during the 2021 North American Winter storm in College Station, Texas. For details, see the paper.

This dataset contains the data created and analysed in "Validating Deep-Learning Weather Forecast Models on Recent High-Impact Extreme Events" by Olivier C. Pasche, Jonathan Wider, Zhongwei Zhang, Jakob Zscheischler, and Sebastian Engelke (DOI: 10.1175/AIES-D-24-0033.1). It contains ground-truth data, numerical forecasts (HRES) and deep learning forecasts (GraphCast, Pangu-Weather, FourCastNet) of the weather during the three extreme events studied in the paper. The corresponding analysis code is available at https://github.com/jonathanwider/DLWP-eval-extremes.

The preprocessing mainly consists of two steps: restricting data to a region relevant to the respective case study & unifying metadata, variable names, dimensions, etc between the different data sets merging HRES forecasts. After downloading, the HRES data sets for initialization times 00:00 / 12:00 UTC and 06:00 / 18:00 UTC differ in their length. For some analyses, we extend the 06:00 / 18:00 UTC forecasts. The details are described in section "2b - Initialization Times" in the paper.