Data-Driven Plasma Properties

Abstract

We currently enjoy a wide range of theoretical and computational models for predicting plasma properties, ranging from equations of state to thermal conductivities to ionic structure. When available, experimental data is used to validate these models. As we accumulate ever more data, it is natural to reverse this situation and ask the question: can we build predictive models of plasma properties using data alone? We can formally cast this question in terms of function approximation, which seeks an approximate functional relation trained with data to approximate an unknown function. Using electrical conductivity as an example, we imagine that there exists a functional relationship σ(Z,ρ,T) that we wish to learn purely from data. Here, radial basis function neural networks are used to express this function. Novel aspects of the present approach are the use of a theoretical model (modified Lee-More) for detrending, unsupervised center selection with silhouette scores and anisotropic basis functions with Mahalanobis "radial" distances. This model will be described and predictions made from a dataset of electrical conductivities. Improvements to this approach, mainly in the area of data needs and extensions to other properties, will be discussed.