The NEMS GFS Aerosol Component Version 2.0 (NGACv2) for global multispecies aerosol forecast has been developed at the National Centers of Environment Prediction (NCEP) in collaboration with the NESDIS Center for Satellite Applications and Research (STAR), the NASA Goddard Space Flight Center (GSFC), and the University at Albany, State University of New York (SUNYA). This paper describes the continuous development of the NGAC system at NCEP after the initial global dust-only forecast implementation (NGAC version 1.0, NGACv1). With NGACv2, additional sea salt, sulfate, organic carbon, and black carbon aerosol species were included. The smoke emissions are from the NESDIS STAR's Global Biomass Burning Product (GBBEPx), blended from the global biomass burning emission product from a constellation of geostationary satellites (GBBEP-Geo) and GSFC's Quick Fire Emission Data Version 2 from a polar-orbiting sensor (QFED2). This implementation advanced the global aerosol forecast capability and made a step forward toward developing a global aerosol data assimilation system. The aerosol products from this system have been used by many applications such as for regional air quality model lateral boundary conditions, satellite sea surface temperature (SST) physical retrievals, and the global solar insolation estimation. Positive impacts have been seen in these applications.
Aerosol radiative effects have been studied extensively by climate and weather research communities. However, aerosol impacts on radiance in the context of data assimilation (DA) have received little research attention. In this study, we investigated the aerosol impacts on the assimilation of satellite radiances by incorporating time-varying three-dimensional aerosol distributions into the radiance observation operator. A series of DA experiments was conducted for August 2017. We assessed the aerosol impacts on the simulated brightness temperatures (BTs), bias correction and quality control (QC) algorithms for the assimilated infrared sensors, and analyzed temperature fields. We found that taking the aerosols into account reduces simulated BT in thermal window channels (8 to 13 μm) by up to 4 K over dust-dominant regions. The cooler simulated BTs result in more positive first-guess departures, produce more negative biases, and alter the QC checks about 20%/40% of total/assimilated observations at the wavelength of 10.39 μm. As a result, assimilating aerosol-affected BTs produces a warmer analyzed lower atmosphere and sea surface temperature which have better agreement with measurements over the trans-Atlantic region.
Data driven machine learning-based weather prediction (MLWP) models have been under rapid development in recent years. These models leverage autoregressive neural network architectures and are trained using reanalysis data generated by operational centers, demonstrating proficient forecasting abilities. One remarkable advantage of these MLWP models is that, once trained, they take significantly less amount of computational resources to produce forecasts compared to traditional numerical weather prediction (NWP) models while maintaining or surpassing the NWP performance. NCEP has started machine learning development collaborating with the research community for several years. This presentation will provide an overview of the development of MLWP models for the global ensemble forecast system at NCEP Environment Modeling Center (EMC). The model adopts state-of-the-art MLWP models such as GraphCast and leverages the methodologies from FuXi global ensemble system. The development includes developing cascade MLWP models, training the model with GEFSv12 reanalysis data and producing forecasts with the operational GEFSv12 initial states. The model will be validated using two years of GEFSv12 operational forecast data. The ultimate objective is to deliver 15-day forecasts with skill levels comparable to the operational GEFSv12.
At NOAA/NWS/National Centers for Environmental Prediction (NCEP), a global ensemble forecast system based on GraphCast—a machine-learning weather model developed by Google DeepMind—has been tested. This study compares the machine-learning-based global ensemble forecast system (MLGEFS) with the NCEP operational physics-based global ensemble forecast system (GEFS). The goal is not to determine which system performs better or worse, but rather to understand the similarities and differences between them. We aim to gain insights into how to improve or better design AI models, particularly ensemble prediction systems based on AI, as well as AI-driven forecast post-processing, product generation, and verification methods. To capture both warm and cold season dynamics, one winter case and one summer case were carefully examined across various aspects of ensemble performance. Three similarities—forecast accuracy, ensemble spread, and inter-variable correlation—and four differences—ensemble-averaging effectiveness, nonlinearity, chaos, and blurriness—were identified. Additionally, one assessment of forecast information was conducted. As a side result (the “fifth difference”), the less uniform performance of ensemble members observed in MLGEFS compared to GEFS, may indicate that ensemble post-processing could be more important in AI-based ensemble systems than in physics-based NWP ensembles. Given the early exploratory stage of AI-based weather forecasting, this study may raise more questions than it answers, which is expected to stimulate further research in this emerging field.
Accurate forecasting of Convective Available Potential Energy (CAPE) is critical for severe weather prediction. However, the operational GFS model exhibits a persistent low-CAPE bias. In this work, we apply a two-step regression–diffusion model (NVIDIA CorrDiff) to address this issue. Our results indicate that while a standard U-Net can successfully reduce the bulk systematic bias, the generated output remains overly smoothed. This occurs because, for data with long-tailed statistical distributions such as CAPE, standard models trained on mean squared error fail to capture rare, high-magnitude events. In contrast, generative diffusion models can reproduce realistic, small-scale features similar to the ground truth by learning to reverse a noise-corruption process through a series of iterative denoising steps. Our study begins with bias correction for the 24 h forecast. We then extend this by applying the model—trained solely on 24 h data—to correct forecasts of up to 120 h. This strategy leverages our finding that the GFS forecast bias is highly persistent over time. Furthermore, our examination of CAPE’s joint Probability Density Functions emphasizes the necessity of matching machine learning models to the target variable’s statistical properties. Ultimately, the effectiveness of CorrDiff highlights its potential for other challenging applications involving small-scale phenomena with long-tailed distributions.