Spatiotemporal Dual-Stream Transformers for Cloud Microphysical Parameterization

Abstract. Accurate precipitation forecasting is essential for mitigating weather-related disasters. Numerical Weather Prediction (NWP) precipitation forecasting accuracy is largely constrained by microphysical parameterization schemes, which rely on simplifying assumptions that introduce uncertainties. Deep learning provides a promising approach for data-driven modeling of complex microphysical relationships. We propose to model the cloud microphysical process via the Learned Microphysics Transformer (LMP-Tr). LMP-Tr employs a hybrid Convolutional Neural Network (CNN)–Transformer architecture that alternately integrates multi-scale convolutional modules and dual-pathway attention modules to capture both local cloud-scale features and long-range atmospheric dependencies. The key innovation lies in the systematic alternation of multi-scale convolutional modules for local feature extraction and dual-pathway attention modules for global dependency modeling. The proposed model enables progressive refinement of atmospheric representations through height-variable attention pathways and cross-module attention mechanisms. Extensive evaluation on a WRF simulation dataset demonstrates superior performance of the proposed method. LMP-Tr provides a practical and effective solution for enhancing cloud microphysics representation in operational NWP systems, offering improved accuracy and physical consistency compared to other Artificial Intelligence (AI)-based parameterization approaches.