Neural Network Gravity Field Model

Neural-inspired computing has significant implications for spacecraft onboard processing. The same innovations that currently allow consumer devices to process large data streams in real-time will, in the future, enable spacecraft to do the same, make intelligent decisions, and achieve mission objectives that are impossible with current ground-in-the-loop systems. However, new algorithms must be developed to reformulate space-related mathematical problems into a form that can take advantage of these computer hardware advances.nbsp;We propose to develop a framework for high-fidelity force fields to be modeled as artificial neural networks (ANNs). Force model evaluation is a fundamental limiting computational step in many astrodynamics algorithms, including mission planning, navigation, maneuver design, and operations planning. Engineers are typically forced to choose between accuracy and speed. Onboard implementationsnbsp;currentlynbsp;require the dynamical models to be greatly simplified to run within limited computational resources. The proposed innovation will be developed for use both on the ground and in space, benefitting space mission design, navigation, and operation. The innovation is relevant and advantageous for current computer systems, and it will become even better over time based on the direction of computer chip research and development.nbsp;nbsp;When used on the ground, the proposed innovation will improve the fidelity and computational performance of standard human-in-the-loop mission design and navigation. When trained on the ground and evaluated onboard a spacecraft, the innovation will enable higher-accuracy onboard operations for lower computational demand than existing capabilities.nbsp;In the future, when neuromorphic processors are available onboard spacecraft, the framework created by the proposed innovation will allow spacecraft to retrain a dynamical model based on data received in-space.