Intelligent Decision Support System for Planetary Defense under Mixed Aleatory/Epistemic Uncertainties

This paper studies the application of Machine Learning techniques in Planetary Defense. To quickly respond to an asteroid impact scenario, an Intelligent Decision Support System is proposed to automatically decide if a deflection mission is necessary, and then select the most effective deflection strategy. This system consists of two sub-systems: the first one is named as Asteroid Impact Scenarios Identifier, and the second one is named as Asteroid Deflection Strategies Selector. The input to the Asteroid Impact Scenarios Identifier is the warning time, the orbital parameters and the diameter of the asteroid and the corresponding uncertainties. According to the Probability of Collision and the corresponding confidence, the output is the decision of action: the deflection is needed, no deflection is needed, or more measurements need to be obtained before making any decision. If the deflection is needed, the Asteroid Deflection Strategies Selector is activated to output the most efficient deflection strategy that offers the highest probability of success. The training dataset is produced by generating thousands of virtual impact scenarios, sampled from the real distribution of Near-Earth Objects. A robust optimization is performed, under mixed aleatory/epistemic uncertainties, with five different deflection strategies (Nuclear Explosion Device, Kinetic Impactor, Laser Ablation, Gravity Tractor and Ion Beam Shepherd). The robust performance indices are considered as the deflection effectiveness, which is quantified by the change of impact probability pre and post deflection. We demonstrate the capabilities of Random Forest, Deep Neural Networks and Convolutional Neural Networks at classifying impact scenarios and deflection strategies. Simulation results suggest that the proposed system can quickly provide decisions to respond to an asteroid impact scenario. Once trained, the Intelligent Decision Support System, does not require re-running expensive simulations and is, therefore, suitable for the rapid prescreening deflection options.