Automating the Inference of Asteroid Physical Properties and Motion

Abstract:

Camera networks and Earth-monitoring satellites frequently observe bright impactors in the atmosphere through measurements of brightness as a function of time, known as light curves. These optical observations provide a key source of information to infer small Near-Earth Objects’ size, bulk density, and strength, which are essential physical properties to making reliable asteroid impact risk assessment. In order to infer these initial physical properties from light curves, scientists simulate the impactor’s entry and breakup. However, the rate at which light curve data is collected is exceedingly greater than the rate at which scientists can manually model these events. Manually modeling these events requires extensive trial-and-error variations of physical and modeling parameters, and solutions can be non-unique, with different combinations of parameters producing similar fits. Consequently, the process is laborious, lacks systematic evaluation of the fit, and requires separate measurements of motion to decrease the number of unknowns.

In this work, we aim to increase the speed and reliability of analyzing energy deposition curves, which are derived from light curves, by automating the inference of asteroids’ physical properties using genetic algorithms, deep neural networks, convolutional neural networks, and random forest regressors. We remedy the lack of observational data with known initial conditions needed to train our machine learning algorithms by modeling meteoroid entry and breakup using the fragment-cloud model (FCM). In our methods, we harness FCM that uses asteroid pre-entry properties to generate energy deposition profiles. We combine FCM with a genetic algorithm to infer pre-entry properties by finding the optimal match between the observed and modeled energy deposition curves. FCM also enables the creation of a diverse, realistic, and labeled synthetic dataset to train the supervised learning methods. We compare their accuracy with well-studied events, such as Chelyabinsk. We also show that supervised learning methods do not require initial velocity and entry angle and use fewer computational resources than other methods. The genetic algorithm, however, facilitates the interpretability of results and relations between variables.