Machine learning for optimizing plasma resource utilization on Mars

Abstract: On Mars, the atmosphere plays a pivotal role in the In-Situ Resource Utilization (ISRU) perspective, mainly due to the abundant atmospheric CO2. Plasma technologies, especially the application of low-temperature plasmas (LTPs), have been proposed to decompose the existing CO2, facilitating the recovery of O2 and CO to support life, fuel, and agriculture. However, much remains to be done regarding accurate plasma modeling, which relies on the availability of precise reaction rate coefficients. Determining these constants is a non-trivial task, primarily addressed by scientists with extensive experience in the field.

In this thesis, we investigate the use of machine learning techniques to predict these parameters. This is done with an automated and systematic approach, based on the heavy-species densities of the gas's final state. The training data is generated using the Lisbon Kinetics simulation tool (LoKI).

Both Support Vector Regression (SVR) and Artificial Neural Network (ANN) models were trained using an oxygen (O2) plasma kinetic scheme, encompassing 11 species and 53 heavy-species reaction processes. When evaluated on a test dataset, the top performing SVR and ANN models achieved mean relative error values of 0.10% and 0.22%, respectively. Although both models exhibited significant accuracy, the SVR model emerged as superior, offering both simplicity and heightened performance for our specific low-dimensional dataset. These two machine learning approaches serve as a successful preliminary step towards more comprehensive models to automate the prediction of reaction rate coefficients.