Observation Driven Multi-Scale Medium Range ML Weather Forecasts

Weather forecasting is undergoing a paradigm shift, moving from physics-based numerical modeling to data-driven solutions using artificial intelligence. Physics-based numerical weather prediction (NWP) systems have been used operationally for decades, providing valuable forecasts across spatial and temporal scales for countless applications. While undeniably useful, the technical capabilities of physics-based NWP systems have plateaued due to a combination of factors, including algorithmic limitations, parameter sensitivity, computational resources, and scientific understanding of complex processes. These technical limitations directly lead to lower-resolution forecasts with less than optimal predictive accuracy. In contrast, recent advances in machine learning weather prediction (MLWP) have highlighted new capabilities for more efficient forecasting systems by emulating physics-based systems. The next step in this transition is replacing the dependence on physics-based data assimilation with observation-to-observation AI forecasting. While the advances in MLWP cannot be ignored, the current generation is dependent on reanalysis and physics-based data assimilation systems. The availability of ERA5 and other reanalysis datasets has enabled the machine learning community to develop powerful forecasting models with novel methodologies. This has created a dependence on analysis data from physics-based data assimilation systems, which contain their own uncertainties and are computationally expensive, causing infrequent update cycles.ÝThis project extends our development of EarthNet to generate medium-range forecasts from a pure AI observation-driven system. EarthNet is a multi-modal foundation model that ingests heterogeneous and sparse Earth data from all types of sensing systems. The model learns through masked modeling, where data is randomly removed with a reconstruction training objective. This approach natively handles sparse and missing data, making masked modeling ideal for gap-filling remote sensing data. We have demonstrated EarthNet's ability to reconstruct atmospheric temperature and humidity from geostationary satellites and low-Earth orbit microwave soundings. Here, we will extend our data pipelines with surface observations, radiosondes, and more satellite data to perform forecasting. Once trained on a larger dataset, EarthNet is capable of performing zero-shot forecasting autoregressively, along with fine-tuning for further performance. We will also experiment with initializing open-source MLWP models. EarthNet also has strong scaling capabilities, which indicate that reconstruction and forecast performance will improve with data, compute, and model size.