KARINA: An Efficient Deep Learning Model for Global Weather Forecast

TL;DR

This work presents KARINA, an efficient global weather forecaster that achieves competitive accuracy at 2.5° resolution with substantially reduced compute, training on 4 NVIDIA A100 GPUs in under 12 hours. The method combines ConvNext with SENet and a novel Geocyclic Padding to preserve spherical continuity and dynamic channel emphasis, enabling high-quality forecasts across multiple atmospheric variables. KARINA surpasses ECMWF S2S reforecasts up to 7 days and rivals recent high-resolution DL approaches that use much larger pixel grids, demonstrating a favorable balance of accuracy and resource efficiency. The findings highlight the importance of boundary-aware padding and channel recalibration in scalable data-driven weather prediction and suggest a practical path toward broader deployment of efficient global forecasting models.

Abstract

Deep learning-based, data-driven models are gaining prevalence in climate research, particularly for global weather prediction. However, training the global weather data at high resolution requires massive computational resources. Therefore, we present a new model named KARINA to overcome the substantial computational demands typical of this field. This model achieves forecasting accuracy comparable to higher-resolution counterparts with significantly less computational resources, requiring only 4 NVIDIA A100 GPUs and less than 12 hours of training. KARINA combines ConvNext, SENet, and Geocyclic Padding to enhance weather forecasting at a 2.5° resolution, which could filter out high-frequency noise. Geocyclic Padding preserves pixels at the lateral boundary of the input image, thereby maintaining atmospheric flow continuity in the spherical Earth. SENet dynamically improves feature response, advancing atmospheric process modeling, particularly in the vertical column process as numerous channels. In this vein, KARINA sets new benchmarks in weather forecasting accuracy, surpassing existing models like the ECMWF S2S reforecasts at a lead time of up to 7 days. Remarkably, KARINA achieved competitive performance even when compared to the recently developed models (Pangu-Weather, GraphCast, ClimaX, and FourCastNet) trained with high-resolution data having 100 times larger pixels. Conclusively, KARINA significantly advances global weather forecasting by efficiently modeling Earth's atmosphere with improved accuracy and resource efficiency.

Figures (9)

• Figure 1: Overall description of KARINA, which merges ConvNext with SENet and Geocyclic Padding, targeting 2.5° forecasting precision. Features a [3, 3, 9, 3] stage design, with initial blocks using a kernel size of 7 and depth scaling layers employing a kernel size of 3, channel expansion from 96 to 768, preserving spatial consistency.
• Figure 2: Description of GeoCyclic Padding. It minimizes projection distortions by circularly padding the edges and reordering at the poles to ensure longitudinal continuity in geographic data such as the ERA5 dataset.
• Figure 3: Comparative forecast skill evaluations of forecasting techniques over a 7-day horizon. The graphs display the globally averaged latitude-weighted RMSE for T2M, Z500, T850, and MSL over a 7-day forecast period. Three variations of the KARINA model are compared: 'KARINA Plain' without any modifications, 'KARINA Padded' incorporating GeoCyclic Padding, and 'KARINA Padded+SENet' integrating both GeoCyclic Padding and SENet.
• Figure 4: Horizontal distribution of prediction skill improvement during the test period. a) Effect of GeoCyclic Padding: RMSE difference between Padded and Plain. b) Effect of SENet: RMSE difference between Padded+SENet and Padded. c) ACC improvement of geopotential variable at each pressure level from GeoCyclic Padding (solid line) and SENet (dashed line) in the equatorial region (20°S–20°N).
• Figure 5: Regressed field of variables from $-\sigma$ of $z'_{500\text{NA}}$. Shaded colors represent the 2m temperature ($T2M$), contour lines represent mean sea level pressure ($MSL$), and green arrows indicate the 850 hPa wind vectors. The red boxes delineate the North Atlantic region ($40^\circ\text{W} - 10^\circ\text{W};\ 30^\circ\text{N} - 45^\circ\text{N}$).