Author: Mark Potts
EPIC Team: Solutions Architect
The Earth Prediction Innovation Center (EPIC), in collaboration with the NOAA Physical Sciences Laboratory (PSL), NOAA Global Systems Laboratory (GSL), and others, is working to establish the capability to accurately predict the timing, location, and overall precipitation of atmospheric river events on the west coast of North America. Atmospheric rivers (AR) are large streams of atmospheric moisture that form over the ocean and flow to the coast where they account for 30 to 50 percent of annual precipitation along the west coast and can cause significant damage due to the intensity and concentration of precipitation in localized areas. These events form through complex interactions between the Pacific Ocean and the atmosphere. Emerging needs for management of water resources and emergency preparedness for severe AR events require increased fidelity of forecast models The Water in the West project aims to address these shortcomings and has been making steady progress towards a more robust and accurate forecast system.
To improve AR forecasts there is a need to accurately capture the detailed and interconnected physical processes occurring in the atmosphere, in the ocean, and over land. These processes include winds and moisture transport, formation of clouds, rain, and snow, and precipitation over the mountainous terrain along the west coast of North America. Scientists at NOAA PSL and NOAA GSL have been refining a new physics suite that will enhance weather prediction capabilities. They have been testing this suite with support from EPIC across several supercomputing platforms, including cloud resources from Amazon Web Services (AWS), Google Cloud Platform (GCP), and Microsoft Azure. The AR forecast system incorporates a high-resolution nest, or grid that layers over the global forecast model’s grid, specifically covering the eastern Pacific Ocean and most of the continental United States (CONUS). This nested grid approach allows for more detailed and localized weather predictions within these areas.
Having the capability to run a model forecast with high resolution is an essential part of improving weather prediction, but it is only the first step. Without accurate initial conditions that reflect the actual environmental state of the globe, no model can correctly predict the future state of the environment. To get the best possible initial conditions for the model forecast, NOAA uses a variety of data sources, including satellite observations, weather stations, radiosondes, aircraft measurements, weather buoys, etc., and assimilates them into a refined starting point for the model. This sort of cycling—where the model is run for a short time to produce an initial forecast, data is collected to compare to the forecast, and data assimilation techniques are employed to mathematically combine the two to provide a refined starting point for the model—is done on a regular basis as part of NOAA’s operational weather forecasts. Starting from refined initial conditions greatly improves the forecast skill of the model stretching out five to seven days. While the capability to do this sort of cycling exists, it has not yet been done in combination with a high-resolution nest over part of the global domain. This new capability is currently under development for the Water in the West project.
Running a single-cycled forecast with a high-resolution nest coupled to data assimilation (DA) will be a big advance in our ability to predict AR events, but the plan is for that to be just the start. Once the singular cycled forecast capability is in place, NOAA will begin testing the ability to run an ensemble of cycled forecasts with DA in order to get a statistically-based forecast for AR events. Similar to charts that show multiple hurricane paths to help enhance forecasts, using ensemble forecasts and DA improves overall forecast skill and provides better statistical possibilities for events. The Water in the West project aims to run 80 ensemble members in a cycled manner with DA in order to achieve the goal of accurately predicting AR events.
Running 80 ensemble members with DA is both incredibly complicated and expensive, and the workflow that will run this system is the next step in development for the project. The time it will take to run such a system is likely to become an issue since DA is computationally expensive and currently requires many large files to be both written out and then read back in by each of the ensemble members. These disk read and write operations are hobbling efforts to increase the cadence and resolution of the DA cycle. In this project, we endeavor to link model forecasts with the DA system, allowing for faster data transfer using “in-core” memory (RAM), eliminating the need to move data between different storage areas. This will greatly simplify the overall workflow as well as reduce the time required to run a full ensemble forecast.
As ambitious as this proposed system is, there are further plans to improve the forecast skill. As mentioned earlier, the interactions between ocean and atmosphere are a key component driving AR events. In order to best capture this, a coupled forecast model that includes not just the atmosphere, but also the ocean and sea ice, is required. Once an 80-ensemble cycled atmosphere-only forecast system has been completed, the project will focus on extending the system to run a coupled model with atmosphere, ocean, and sea ice. The additional complexity in the model is also matched by added complexity in the DA and the workflow where more data and files will be required. Similarly, directly linking the coupled forecast system to the DA system for in-core memory transfer is complicated but also has the potential to provide even greater computational improvements.
With in-core DA, the proposed coupled system is expected to be very computationally expensive, perhaps even prohibitively so. There are now plans to replace the atmospheric model, the most computationally expensive component in the coupled system, with an Artificial Intelligence/Machine Learning (AI/ML) model. Such models have been developed by a number of groups recently and have proven both accurate and fast, but they are also dependent on the data they are trained upon. NOAA will be undertaking the task of training an AI/ML model using new nested models developed in this project. As such, the trained model should provide a direct replacement for the atmospheric component in the coupled model system and significantly reduce the computational expense of the overall AR forecast and DA systems.
https://www.climatehubs.usda.gov/hubs/northwest/topic/atmospheric-rivers-northwest-0
