Optimized Learning for Yielding Meteorological Predictions Using Surrogates

NASA’s current physics-based models are the gold standard for simulating atmospheric and heliophysics variables. These models, however, require very large amounts of compute power and several hours to produce a forecast. We propose to build the Optimized Learning for Yielding Meteorological Predictions Using Surrogates (OLYMPUS), comprising machine learning-based surrogate models to dramatically reduce the computational time and resources needed to make a prediction. We do not anticipate that these models will replace the physics-based ones, as they do not offer the same level of transparency and explainability. Rather, these models will be used to produce forecasts quickly and cheaply at a fraction of the cost. At the heart of our approach is a novel approach to dramatically reducing the size of gridded atmospheric data, making it feasible to use a modern deep learning architecture to predict sequences of this data. This methodology entails computing the Spherical Harmonic Transform (SHT) of atmospheric variables. The SHT will give us a global representation of atmospheric data on a sphere, efficiently compressing spatial information to a more compact, multi-scale level representation that captures both small-scale, high-frequency and large-scale, low-frequency variations. It will also reduce aliasing artifacts as the spherical nature of the transform allows for an accurate representation of features across different latitudes, mitigating distortions that arise with traditional grids. Performing principal component analysis across the resulting SHT coefficients from several different time periods will allow us to identify the principal components that are able to explain the bulk of the variance within the system and reduce the number of variables form millions to thousands or even hundreds. This reduction will allow us to use state-of-the-art deep learning model architectures that would otherwise not scale to the size of this problem.