This paper addresses the stability issue of long-term forecasts using large-scale weather models (LWMs). While existing state-of-the-art models achieve interannual stability by transforming input data into non-standard spatial domains, such as spherical harmonics or HEALPix meshes, this paper demonstrates that similar performance can be achieved on a standard latitude-longitude grid. We propose AtmosMJ, a deep convolutional neural network that directly processes ERA5 data. This network generates stable forecasts for approximately 500 days by preventing error accumulation through a novel Gated Residual Fusion (GRF) mechanism. AtmosMJ achieves 10-day forecast accuracy comparable to models such as Pangu-Weather and GraphCast, while training in a low training cost of 5.7 days on a V100 GPU. This demonstrates that efficient architecture design, rather than non-standard data representations, is key to long-term weather forecasting.
