Describing the Upper Atmosphere using Machine Learning to investigate long-term observational data sets

The AtmoSense Background Characterization (ABC) project applies modern machine learning techniques to describe the natural parameters of the atmosphere and its wave population with the goal of identifying transient disturbances seen in observational data.
ABC has completed the first versions of a machine learning system that can both generate realistic atmospheric profiles for a given set of geophysical drivers and discriminate to determine if a given set of observations and driver conditions is sufficiently different from the historically observed atmosphere to represent a true event detection. The system is based on a GAN (Generative Adversarial Network) approach, which simultaneously creates both a Generator and Discriminator, trained from long-term data sets accumulated by the NSF’s Poker Flat Incoherent Scatter Radar (PFISR) and the University of Alaska’s Rayleigh Lidar, both located at the Poker Flat Research Range outside Fairbanks, Alaska. The GAN captures the full distribution of the data, allowing it to be used to not only to determine the average atmospheric profile for a given set of drivers, but also the expected variability. This provides for a deeper understanding of the range of realistic atmospheric parameters, how this range changes with geophysical conditions, and enhances the utility of the data products.
We will discuss the philosophy of our approach, the preparation of training data sets, the machine learning process, and the fidelity of the outputs compared to the observational data. Examples of both Generated atmospheres and Discriminated events will be shown. The tools and tutorials are available publicly on GitHub and as a Python package, allowing scientists to run both the Generator and Discriminator themselves.
The AtmoSense Background Characterization (ABC) project is a component of the Defense Advanced Research Projects Agency (DARPA) AtmoSense program which uses observations of the upper atmosphere to identify transient disturbances as a function of the location, duration, and size of their source events.
This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. HR001121C0026.