A Novel ML-NWP Coupled Optimization Approach on a Reconstructed WRF Model

Optimizing Numerical Weather Prediction (NWP) or correcting errors in it by data-driven machine learning (ML) methods constitutes a highly challenging problem. To address this, we propose a coupled ML-NWP optimization framework that embeds data-driven model bias calibration directly within NWP integration loops. By reimplementing the Weather Research and Forecasting (WRF) dynamic solver and physical parameterizations in Python, we developed the tensor-based PyWRF model, which fully supports PyTorch’s automatic differentiation algorithm. A single-column neural network (NN) is coupled with PyWRF to enable observation-guided backward gradient propagation, facilitating feedback between observations and integration steps to modulate specific model variables. Online coupled training of the NN with PyWRF during numerical integration is achieved through local gradient masking, truncated backpropagation with a finite number of steps, and skip-connection-based variable transfer. Eight mesoscale precipitation events over East China were selected for training and testing the coupled model. The results demonstrate that the model trained on limited observational data within this framework exhibits robust generalization ability. Comparative experiments indicate that the proposed framework improves forecast accuracy relative to standalone NWP and traditional single-step AI correction methods. This study provides a pathway for physics-data-intelligent next-generation weather forecasting systems.