Machine Learning-Accelerated Grid Environment

NASA satellites are generating over 4TB of data each day. Analyzing this data in-orbit is becoming increasingly important for the purposes of accelerating scientific discovery and enabling opportunistic science. State-of-the-art artificial intelligence (AI) and machine learning (ML) data science applications require significant resources to run computationally intensive algorithms and models. To facilitate intensive data analysis in a resource constrained environment such as space, we need to utilize resources efficiently and at scale. Current solutions to this problem require downlinking full datasets to perform ground-based processing or running low computational footprint algorithms that are less effective than state-of-the-art solutions. In this proposal, we explore the capabilities and benefits of developing MAGE (ML Accelerated Grid Environment). MAGE is a software framework and API that facilitates ML training and inference distributed across a networked constellation or swarm of satellites to enable resource intensive ML models to run at the extreme edge. This solution makes complex data processing at the edge possible by running on AI accelerated hardware and distributing ML processing and storage across a grid of compute and storage nodes. Collectively, these nodes comprise a grid computing environment that can be tasked by spacecraft to run resource intensive applications. MAGE reduces the need to downlink full data sets, allows prioritization of data downlinking, enables proliferation of complex autonomous space-based systems, and provides a mission agnostic environment for processing and storage. Utilizing a system such as MAGE would allow NASA to perform efficient, scalable, mission agnostic AI and ML processing at the edge for any scientific mission.