This paper addresses the stability issue of long-term forecasts using large-scale weather models (LWMs). Existing models, such as SFNO and DLWP-HPX, achieve stable long-term forecasts by transforming input data into non-standard spatial domains, such as spherical harmonic functions or HEALPix meshes. This has been considered essential for physical consistency and long-term stability. In this paper, we challenge this assumption and investigate whether similar long-term forecast performance can be achieved on a standard latitude-longitude grid. To achieve this, we propose AtmosMJ, a deep convolutional neural network that directly processes ERA5 data. AtmosMJ utilizes a novel Gated Residual Fusion (GRF) mechanism to prevent error accumulation and ensures stability in long-term recursive simulations. Experimental results demonstrate that AtmosMJ generates stable and physically plausible forecasts for approximately 500 days, and is competitive with models such as Pangu-Weather and GraphCast in terms of 10-day forecast accuracy. It is also noteworthy that these results were achieved in a short training time of 5.7 days using a V100 GPU. In conclusion, this paper suggests that efficient architecture design, rather than non-standard data representations, is the key to ensuring the stability and computational efficiency of long-term weather forecasting.
