Abstract:
|
As part of its mission, the Mars rover Curiosity is equipped with ChemCam, an instrument used to determine the elemental composition of solid objects using laser-induced breakdown spectroscopy (LIBS). Traditionally, linear regression models are trained to disaggregate the sample (i.e., predict elemental compositions) using a library of real LIBS spectra collected under known conditions. However, the libraries tend to be small and focused around particular types of materials, limiting the ability to evaluate modern machine learning techniques for disaggregation. To quantify empirical performance of disaggregation techniques in larger space of materials, we leverage a physics simulation code to generate synthetic LIBS spectra. To augment the set of simulations, we build accurate emulators of the physics code based on low-dimensional representations. Using the emulator to generate additional training data, we compare modern supervised machine learning techniques to traditional approaches for disaggregation and find that machine learning methods may provide a boost in accuracy and generalization compared to traditional techniques.
|