Machine Learning for Data Constrained Planetary Mission Instruments

Many planetary exploration missions aim at evaluating the habitability and the existence of potential life on the target bodies. Missions going further away in our solar system (i.e., Titan, Europa, Enceladus, etc.) or environmental-constrained missions (i.e., Venus, Mercury) will face communication constraints due to limited transfer rates and short communication windows. Current operations processes are centered around ground-in-the-loop activities: data is sent back to scientists on Earth, who must rapidly analyze them and infer about the chemistry to decide which operations to do next and send these instructions to the spacecraft.

Our research leverages machine learning (ML) and data science techniques to improve methods for analyzing mass spectrometry data to support scientists' decision-making process. The goal is to make science operations faster and more efficient, and ultimately maximize missions' scientific returns. While space missions are extremely data-limited, scientists use commercial and flight analog instruments during missions' development. We set up an open-science ML challenge focusing on building models to automatically analyze mass spectrometry data for Mars exploration. ML challenges provide an excellent way to engage a diverse set of experts with benchmark training data, exploring a wide range of approaches, and identifying promising models based on empirical results. This challenge was a proof-of-concept to analyze the feasibility of combining data collected from different instruments in a single ML application. We selected data from 1) commercial instruments and 2) the Sample Analysis at Mars (SAM mass spectrometers onboard Curiosity on Mars) testbed. This challenge organized with DrivenData gathered more than 700 participants, obtained more than 600 solutions contributing to powerful models to the analysis of rock and soil samples using mass spectrometry.

We present the creative - and sometimes surprising - solutions delivered by the top three participants, none of whom had backgrounds in mass spectrometry. We will present the potential of these solutions for ML application in future planetary missions. Our longer-term goal is to deploy these powerful methods onboard the spacecraft to autonomously guide space operations and reduce ground-in-the-loop reliance.