Machine Learning-Based Global Ensemble Weather Prediction for Medium-Range Forecasting

Accurate medium-range global weather forecasts are crucial for decision-making across various societal and economic sectors. Recently, machine learning-based weather prediction (MLWP) models have significantly transformed weather forecasting, often surpassing traditional numerical weather prediction (NWP) methods in efficiency. These MLWP models leverage autoregressive neural network architectures and are trained using reanalysis data from operational centers, demonstrating proficient forecasting abilities while requiring significantly fewer computational resources compared to traditional NWP models. The Machine Learning Global Ensemble Forecast System (MLGEFSv1.0) model, developed at the NCEP Environment Modeling Center (EMC), exemplifies these advancements. MLGEFSv1.0 integrates the GraphCast model developed by Google DeepMind as its core architecture and is trained with Global Forecast System (GFS) analysis data and fine-tuned with recent ERA5 data to produce forecasts using operational Global Ensemble Forecast System (GEFSv12) initial states. The model aims to deliver 16-day forecasts with skill levels comparable to the operational GEFSv12. To enhance the accuracy and reliability of MLGEFSv1.0, we quantify both epistemic and aleatoric uncertainties by perturbing model weights and initial conditions, respectively. Perturbed initial conditions are sourced from GEFS ensemble initial conditions, while perturbed model weights are derived from different stages of GraphCast model training (ranging from 1 to 14 autoregressive steps) to account for short (0-5 day), medium (5-10 day), and longer (10-16 day) lead times. The ensemble approach, featuring 31 members from combined ensemble models and data, has shown comparable performance in prediction compared to GEFS ensemble mean and GraphCast deterministic forecasts while maintaining reasonable spread (variance). In future iterations of this study, we plan to enhance the model by leveraging a hybrid physical and ML approach combining ensemble members from GEFS as well as GraphCastGFS and FourCastNetGFS models.