Use of Neural Networks to Guide the Capture of NEOs into Low Earth Orbit

The awareness of special comets and asteroids called Near Earth Objects (NEO's) has been growing. What is the possibility of intercepting, capturing and placing such objects into low earth orbit, and subsequently using them as sources of energy, materials, and water for orbiting platforms? To this end we evaluated the ability of a neural network to accurately learn the dynamics of these objects, and from this data, adjust firing sequences and intensities to efficiently capture NEO's. We simulated the dynamics of NEO's using energy models, and caused neural networks to learn these dynamics. This would lead to a control strategy that possibly applies firing bursts to drop an NEO from sun orbit into low earth orbit. Conceptually, the neural network will control apparatus that would be set up on an NEO passing by, and the firing sequence initiated on the next pass, perhaps one to several years later. The neural network would learn the NEO's dynamics, and fine tune itself by sensing the response to each firing burst. The main objective was to determine how best to train the neural network by developing training-sample-size criteria. Our procedure was to develop an equation for the minimum training-sample-size, then train the neural network on this set, and finally, test the neural network's ability to predict future NEO trajectories. For the assumed conditions of our test problem, the results were that less than 700 observations of NEO trajectory were necessary to train the neural network. This network then predicted the next 2000 trajectory observations with only gradual loss in accuracy. By comparison, using the same training criteria on linear state variable models gave equal or better results in our tests on industrial mineral processing circuits. Thus we concluded that, in general, the non-linear nature of neural networks was not as important as our stringent training criteria.