Combining machine-learned regression models with Bayesian inference to interpret remote sensing data

The need to build more comprehensive planetary science theories requires observations from remote sensing and in-situ platforms. Space missions, telescopic surveys and laboratory experiments collect large amounts of data, but their combined interpretation typically requires intensive modeling. This may result in prohibitively high computational costs as the number of parameters increases. Here we describe a case study for asteroid thermophysical analysis [12] to show how neural networks can be trained on a large but sparse data set of model simulations to learn the relationship between the output and input of a physical process. The resulting “surrogate model” can then be efficiently used to infer the properties of the system, facilitating combined observations across instruments and platforms. This approach allows extracting more consistent and accurate information from planetary science data, and can enhance autonomous decision-making onboard robotic spacecraft and landers.