Connecting physics models and diagnostic data using bayesian graphical models

With increasingly detailed physics questions to ask, and with more advanced diagnostics available, there is a strong case for trying to generalise the way analysis of diagnostic data, and connection to underlying physics models, is done in today’s experiments. With current analysis chains, it is difficult, verging on impossible, to fully grasp the exact assumptions, hidden in different legacy codes, that goes into a full analysis of the main physics parameters in an experiment. We show that by using Bayesian probability theory as the underlying inference method, it is possible to generalise scientific analysis itself, and therefore build an effective and modular scientific inference software infrastructure. The Minerva framework [1,2] uses the concept of Bayesian graphical models [3] to model the full set of dependencies, functional and probabilistic, between physics assumptions and diagnostic raw data. Using a graph structure, large scale inference systems can be modularly built that optimally and automatically use data from multiple sensors. The framework, used at the JET, MAST, H1 and W7-X experiments, is exemplified by a number of JET applications, ranging from inference on the flux surface topology to profile inversions from multiple diagnostic systems. Introduction Probability theory in the Bayesian interpretation is often used as an alternative to standard least squares, or heuristic analysis methods to increase the understanding of how different types of uncertainties influence the inference of underlying physics parameters. What is not always emphasised, is the fully generic way in which this is actually done: a scientific inference problem, whether the width of a spectral line is measured, or a tomographic x-ray inversion is done, is always fully defined by the specification of two quantities: an assumption of the range and a priori likely values of the parameters of the model, and a 37 EPS Conference on Plasma Physics O4.117