Title : A general AI framework for solving inverse problems in physics with applications to thermonuclear fusion
Abstract:
In the exact sciences and engineering, most analysis tasks require the solution of some form of inversion. Indeed regression, density estimation and causality detection can be all conceptualised as the problem of deriving the origin of some observed effects as in the case of tomography. Leveraging on the enormous amounts of data produced by modern societies, machine learning techniques have become the dominant paradigm to address these tasks. However machine learning does not make explicit use of domain knowledge and their models are typically completely data driven, obtained by the analysis of correlations between different variables. The agnostic nature of these methods render their models poorly interpretable, unrepresentative of the actual dynamics of the processes at play and sometimes utterly wrong. Neglecting completely the available knowledge of the processes to be studied is not an efficient and reliable strategy in many scientific applications. Consequently a new series of deep machine learning tools have been developed for combining data and available physics laws. The synergetic integration of measurements and prior physics allows not only the derivation of more consistent models but also the discovery of new causal relationships and phenomena. Moreover the proposed framework is valid for all the various types of inverse problems and is equally applicable to any form of structured data, from time series to probability distributions. It can also be deployed to identify reduced models for feedback control. The potential of the proposed tools is substantiate by a series of numerical tests and the results of analysing experimental data in one of the most complex laboratory systems ever studied by physicists, high temperature magnetised plasmas for the research on thermonuclear fusion.
Keywords: PINN, autoencoders, tomography, regression, causality detection, dynamical systems, attractors, nuclear fusion.
