Simulations of large-scale plasma systems are typically based on fluid approximations. However, these methods do not capture the small-scale physical processes available to fully kinetic models. Traditionally, empirical closure terms are used to express high order moments of the Boltzmann equation, e.g. the pressure tensor and heat flux. In this paper, we propose different closure terms extracted using machine learning techniques as an alternative. We show in this work how two different machine learning models, a multi-layer perceptron and a gradient boosting regressor, can synthesize higher-order moments extracted from a fully kinetic simulation. The accuracy of the models and their ability to generalize are evaluated and compared to a baseline model. When trained from more extreme simulations, the models showed better extrapolation in comparison to traditional simulations, indicating the importance of outliers. We learn that both models can capture heat flux and pressure tensor very well, with the gradient boosting regressor being the most stable of the two models in terms of the accuracy. The performance of the tested models in the regression task opens the way for new experiments in multi-scale modelling.