Development of Emulators: Process Emulators

Numerical models of the atmosphere are typically a collection of sub-models that represent phenomena that are smaller than the spatial resolution that is feasible with available computing capability. For example, weather models typically predict weather as a single value for temperature, precipitation etc on a box of size 3 km x 3 km. (i..e every 3.5 square miles) and climate models on a box of size 4000 square miles. To describe all the physics that are smaller than these box, we need to use an approximation of that phenomena as a sub-model (e.g. clouds are typically smaller than these grid box sizes). These are called parameterization, in which they model the relationships at a scale greater than that of the actual phenomena. Argonne environmental scientists and computational scientists are collaborating to use deep neural networks to replace the parameterizations of certain physical schemes in the Weather Research and Forecasting (WRF) model — a comprehensive model that simulates the evolution of many aspects of the physical world.

Recently, Argonne scientists focused on the planetary boundary layer (PBL) — the lowest part of the atmosphere that is most affected by human activity. PLB dynamics (e.g., wind velocity, temperature and humidity profiles) are critical in determining many of the physical processes in the rest of the atmosphere and on Earth. The team used 20 years of computer-generated data from the WRF model to train the neural networks and two years of data to evaluate whether they could provide an accurate alternative to the physics-based parameterizations, feeding the network more than 10,000 data points and discovering that a trained deep neural network can successfully simulate wind velocities, temperature, and water vapor over time and predict behavior across nearby locations with correlations higher than 90 percent compared with the test data. The result was an algorithm that could replace the PBL parameterization in the WRF model. Such less-expensive models will allow researchers to achieve higher-resolution simulations to predict how short-term and long-term changes in weather patterns affect the local scale — down to neighborhoods or specific critical infrastructure.

Similarly, we have developed a very accurate ML model for one of the most computationally expensive components of the weather and climate models, radiative transfer. Radiative transfer describes the amount of energy reaching the earths’ surface at any given moment, passing through clouds, atmospheric gases and particles as short-wave radiation and going back into space reflected by the ground clouds and atmosphere as long wave radiation.  Describing this process accurately is critical; for building a model of weather and climate and  radiation calculations are computationally expensive as it requires consideration of a number of factors for each layer of the atmosphere. We have built a ML emulator for the radiative transfer that is approximately 1000 times faster than the models that are currently used in weather/climate numerical models.