Deep learning-based 12-hour global dust distribution forecasting on Martian

Martian dust storms have a profound impact on atmospheric structure, pose multiple risks to Mars landers, and greatly affect the accuracy of sounders. This makes the accurate short-term prediction of dust storms extremely important for future Mars exploration missions. However, traditional statistical analyses fail to accurately capture the variation patterns of dust. Here, we show that the ConvGRU-Seq2Seq model can fully utilize the data to achieve a 12-h forecast of global dust. We found that considering multiple interconnected meteorological elements, particularly the wind field, and accounting for seasonal variations can enhance forecast accuracy. The addition of the Seq2Seq structure reduced the mean squared error (MSE) by 85.3% and the mean absolute error (MAE) by 75.07%, compared with the original ConvGRU model. Among the six models compared, the ConvGRU-Seq2Seq model exhibited the best test performance, with MSE, MAE, and R2 values of 8.73×10−4, 13.48×10−3, and 98.12×10−2, respectively. The model exhibited stable and reliable prediction performance and a more concentrated and accurate spatial distribution of errors. We achieved a rapidly changing dust activity forecast within 12 h with