Onboard Estimation of Physical Parameters in Active Neutron Spectroscopy Data

Neutron instruments, such as the Dynamic Albedo of Neutrons (DAN) instrument on the Mars Science Laboratory Curiosity rover, measure the abundance and energies of neutrons after they have undergone interactions with the environment around them. Active neutron spectrometers include a pulsed neutron generator to increase the signal of these measurements. Neutron counts across multiple energy ranges are measured and binned after the time of each pulse and integrated into neutron die away curves. Retrievals performed on neutron die away curves estimate geochemical parameters of the soil such as water equivalent hydrogen and bulk neutron absorption cross section. Parameter estimation must be completed quickly for operations teams to assess the physical parameters at a given location and make tactical follow-on decisions. Past work has reduced the time required to compute physical parameters, but still requires hours and supercomputing scale resources. Here we demonstrate a rapid approach for parameter retrieval using machine learning capable of estimating physical parameters with computational resources available in an embedded environment onboard the spacecraft. A genetic algorithm approach was used to jointly optimize model hyperparameters and perform feature selection for six machine learning model types. Models are then evaluated against the results from the traditional Bayesian retrieval process to characterize model variance from the baseline Bayesian approach. Top models achieve 0.2 wt. % mean absolute error retrieving the water equivalent hydrogen parameter, and 0.0005 cm<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>/g mean absolute error retrieving the bulk neutron absorption cross section parameter. We assess feature importance and model selection to reason about the driving characteristics of neutron die away curves that enable this approach. Finally, we benchmark runtime and memory footprint of the top models on a Snapdragon 855 processor to demonstrate their path to flight on a representative compute platform. Rapid parameter estimation can be leveraged in the mission planning cycle to generate ‘quick look’ estimates useful for tactical decision making. Embedded deployment of such models can also be utilized onboard future instruments to make in-situ estimates of parameters. This work demonstrates data treatments to support future onboard science instrument autonomy efforts for neutron spectrometers. This is particularly important for applications where active neutron and gamma-ray spectrometers can be used to make rapid, first order geochemical assessments to prioritize follow-on observations without the need for contact science instruments.