Danny Felton, Northrop Grumman Corporation; Mary Ellen Craddock, Northrop Grumman Corporation; Heather Kiley, Northrop Grumman Corporation; Randall J. Alliss, Northrop Grumman Corporation; Eric Page, Northrop Grumman Corporation; Duane Apling, Northrop Grumman Corporation
Keywords: atmospheric characterization and prediction, AI, Deep Learning, SDA
Abstract:
Space based laser and surveillance applications are often impacted by atmospheric effects. Atmospheric attenuation and distortion caused by aerosols, clouds and optical turbulence can produce harmful effects thereby negatively impacting mission outcomes. A paper briefed at the 2019 AMOS conference described ground-based instrumentation installed at the Haleakala summit in 2017. Still actively collecting data, these instruments are providing unprecedented real-time characterization of the space environment, including minute to minute atmospheric transmission losses. Although real-time measurements are a first step for understanding and characterizing the space environment, they are not alone sufficient. Accurate, short-term atmospheric predictions of the space environment are necessary for many applications in order to optimize mission planning. Although atmospheric predictions are nothing new, their skill has recently been greatly improved with the use of 21st century Artificial Intelligence (AI) technologies. These technologies are a union between high performance computing (HPC) and Deep Learning (DL). The ability to train prediction models with terabytes of data from ground-based atmospheric collection systems and accelerate both their training and inferencing with the use of Graphical Processing Units (GPUs) is the subject of this presentation.
This study focuses on three time scales of prediction. These timescales include short-range (0 to 60 minutes), mid-range (1 hour to 3 hours), and long-range (3 to 48 hours). These time scales represent various decision points for laser and/or surveillance applications and missions. In the short-term prediction case, several DL techniques are applied to local data collected from an optical ground station (OGS). These DL techniques include the use of a U-Net convolutional neural network and an ensemble of multi-layer perceptron (MLP) and Random Forest (RF) models. The MLP is used for point data collected from instruments like a laser ceilometer and Infrared Cloud Imager (ICI). For the intermediate time scale, both a convolutional Long Short Term Memory (LSTM) network and a U-Net are trained with imagery from a collection of NOAA geostationary satellite images of clouds. Finally, a combined U-Net and an autoencoder neural network are used to train atmospheric predictors simulated from an HPC Numerical Weather Prediction (NWP) model to make long-range predictions. The NWP produces many terabytes of data and, therefore, the use of these neural networks is ideal to optimize its predictive ability. Several HPC resources are utilized for this study. These include an in-house GPU node consisting of four NVIDIA Tesla V100 GPUs as well as resources at the Maui High Performance Computing Center (MHPCC). Results indicate that in nearly all cases these prediction technologies are outperforming persistence with very little bias. The ability to make predictions in real-time using HPC and DL inferencing is now the focus moving forward and will be reported on at the conference.
Date of Conference: September 14-17, 2021
Track: Atmospherics/Space Weather