OMG-HD: A High-Resolution AI Weather Model for End-to-End Forecasts from Observations

TL;DR

OMG-HD tackles the problem of short-term, high-resolution weather forecasting by removing the data assimilation step and training directly on observational data. The authors propose a two-block architecture—Assimilating Block (Swin Transformer) and Forecasting Block (AFNO)—trained end-to-end on CONUS data with RTMA labels to predict up to 12 hours ahead. Results show significant RMSE improvements over HRRR and IFS-HRES for temperature, wind, humidity, and surface pressure, and demonstrate robustness to missing data and good generalization to unseen stations. Case studies on a March 2023 winter storm and a August 2023 Chicago heatwave illustrate improved biases and humidity/heat-index forecasts. The work highlights the potential of end-to-end, observation-driven forecasting to accelerate forecasts and reduce reliance on NWP-derived inputs, while noting limitations and directions toward global coverage and longer lead times.

Abstract

In recent years, Artificial Intelligence Weather Prediction (AIWP) models have achieved performance comparable to, or even surpassing, traditional Numerical Weather Prediction (NWP) models by leveraging reanalysis data. However, a less-explored approach involves training AIWP models directly on observational data, enhancing computational efficiency and improving forecast accuracy by reducing the uncertainties introduced through data assimilation processes. In this study, we propose OMG-HD, a novel AI-based regional high-resolution weather forecasting model designed to make predictions directly from observational data sources, including surface stations, radar, and satellite, thereby removing the need for operational data assimilation. Our evaluation shows that OMG-HD outperforms both the European Centre for Medium-Range Weather Forecasts (ECMWF)'s high-resolution operational forecasting system, IFS-HRES, and the High-Resolution Rapid Refresh (HRRR) model at lead times of up to 12 hours across the contiguous United States (CONUS) region. We achieve up to a 13% improvement on RMSE for 2-meter temperature, 17% on 10-meter wind speed, 48% on 2-meter specific humidity, and 32% on surface pressure compared to HRRR. Our method shows that it is possible to use AI-driven approaches for rapid weather predictions without relying on NWP-derived weather fields as model input. This is a promising step towards using observational data directly to make operational forecasts with AIWP models.

Figures (14)

• Figure 1: The framework of OMG-HD.a, The forecasting pipeline with an annotated timeline. At the initial time $T_0$, observations (satellite, radar, and station measurements) and the topography from the preceding 6 hours ([$T_0-5$,$T_0$]) are passed through the Assimilating Block to produce an initial state, which is subsequently used by the Forecasting Block to predict the following 6 hours ([$T_0+1$,$T_0+6$]). The forecast can be fed back into the Forecasting Block in an autoregressive fashion to produce temperature, humidity, pressure, and wind predictions for the next 6 hours ([$T_0+7$,$T_0+12$]), and so on. b, Detailed architecture of the Assimilating Block and the Forecasting Block.
• Figure 2: OMG-HD achieves lower forecasting error than baselines across varying lead hours.a, RMSE for temperature, wind speed, specific humidity, and surface pressure verified against the RTMA dataset. b, RMSE for temperature, wind speed, and specific humidity verified against station observations. c, As in b but instead evaluated only on the hold-out set of observations (see Section \ref{['ablation']}). The absence of certain curves indicates some variables are unavailable in certain model evaluations.
• Figure 3: OMG-HD provides accurate, consistent temperature forecasts throughout the CONUS region. Spatial distribution of temperature RMSE evaluated against the RTMA dataset for OMG-HD and the baseline HRRR, ECMWF, and GFS models, as labeled. Average metrics are in the top row, while metrics for specific lead times of 1, 3, 6, 9, and 12 hours are shown in subsequent rows.
• Figure 4: OMG-HD provides accurate, consistent wind speed forecasts throughout the CONUS region. As in Fig. \ref{['fig:temp_vis']} but for forecasts of wind speed.
• Figure 5: The initial state produced by OMG-HD is more accurate than that of the baselines. Shown are box-and-whisker plots of RMSE values for all forecast grid points compared to RTMA. OMG-HD has the lowest average and inter-quartile range of RMSE across all variables, and the fewest outliers (points).