Mass-conserved hierarchical multi-resolution deep learning for long-term global atmospheric pollutant transport prediction

Abstract

Long-term predictive capability of deep learning models has always been a bottleneck, as errors gradually accumulate along with autoregressive iterations, ultimately deviating significantly from the real data distribution. However, the long-term predictive capability of deep learning is crucial in practical applications, especially for complex real-world physical systems, where modeling requires forecasting the system’s long-term evolution. A representative system is atmospheric pollutant transport. Simulating the spatiotemporal evolution of long-lived atmospheric pollutants under realistic meteorology is a basic requirement in Earth system analysis. Such simulations span from the minute scale to the monthly scale. However, traditional high-resolution numerical models are too costly, and existing deep learning models suffer from the long-term prediction bottleneck. To overcome these issues, this study proposes a mass-conserved Hierarchical Multi-Resolution Transport Model (HMR-Trans) for global atmospheric methane (CH 4) pollutant transport simulation based on deep learning. The model is inspired by the multi-scale coupling characteristics of atmospheric transport to design a hierarchical multi-resolution architecture with a coarse-resolution layer capturing macro-scale evolution and a fine-resolution layer recovering high-frequency details. Moreover, HMR-Trans decouples transport from the whole process and independently models the physical transport process to improve the physical interpretability and provide a necessary precondition for imposing strict mass conservation constraints. Then, the conservation constraint is derived through the mass conservation law using the Lagrange multiplier method. The study achieves, for the first time, long-sequence and mass-conserved global atmospheric pollutant transport simulation at high temporal resolution based on deep learning, with monthly-scale average relative error as low as 0.0198 and mass change controlled at the hardware precision, demonstrating its potential to replace high-computation-load transport operators in traditional numerical methods.