Advanced Deep Learning Methods for Asteroid Mining Target Identification

With a rise in demand for rare earth metals, largely due to an increase in semiconductor usage, asteroid mining has emerged as an alternative to access many of these critical elements. However, large-scale operations have not yet been undertaken, largely due to challenges in accurately identifying mining targets. Currently, expensive sampling methods are needed to determine the mineral compounds in an asteroid. Therefore, the present study develops a novel Asteroid Deep Convolutional Neural Network (ADCNN) to provide a low-cost, robust method for classifying asteroid types based on their chemical compositions. Near-infrared spectroscopy data from 2000 to 2022 were acquired from the MIT-Hawaii Near-Earth Object Spectroscopic Survey (MITHNEOS) database and classified according to the Bus-DeMeo taxonomy system. For each asteroid, a graphical representation of the wavelengths was created after initial data processing. Unlike traditional models, rather than analyzing numerical data files for each asteroid, the ADCNN only requires a singular image per asteroid. The ADCNN was trained on nine different asteroid classes, mainly in the carbonaceous (C-type), siliceous (S-type), and EMP (X-type) groups. After testing, the model was able to predict unique asteroid classes with a mean accuracy of 98.2%. These results demonstrate the ability of the ADCNN model to be scaled and widely used for asteroid classification, requiring low-cost and little computational power while achieving high accuracy. Additionally, the ADCNN's high performance when trained on graphical spectra displays the broader potential of visual spectroscopy for chemical analysis. Further application and development of this model can allow us to efficiently identify asteroid targets, leading to industrial mining operations in the near future.