Using Bayesian Analysis to Study Tokamak Plasma Equilibrium Greg T. von Nessi (joint work with Matthew Hole, Jakob Svensson, Lynton Appel, Boyd Blackwell, John Howard, David Pretty and Jason Bertram)
Outline Motivation Brief overview of Bayes Formula Using Bayesian analysis to study fusion plasmas via the MINERVA framework Current tomography results obtained for discharges on the Mega- Ampere Spherical Tokamak (MAST) Future plans for using Bayesian Analysis to study fusion plasmas
Why use Bayesian Analysis to Study Fusion Plasmas? Inference of localised density, temperature and current profiles in a fusion plasma from diagnostic data constitutes a grossly underdetermined problem that is difficult to handle with current techniques without using strong underlying assumptions Can give reliable statistical results for intrinsically underdetermined problems Excels at allowing for the straight-forward comparison of different physical models Influence of prior assumptions on final inference lessens with the addition of more data
Background: Bayes Formula Is a probabilistic technique. Using P to denote a probability distribution: H = Hypothesis (Model Parameters) D = Data (Observations) P (H D) = P (D H) P (H) P (D) P(H D) is the Posterior, P(D H) is the Likelihood, P(H) is the Prior, P (D) is called the Evidence By using a resulting posterior as a new prior, an iterative process is formed that can incorporate all data and prior knowledge into a final posterior
The MINERVA Framework Written by Jakob Svensson (IPP Greifswald) A framework written in Java to perform Bayesian Inference on data coming from complex systems like the ones encountered in fusion research.
Current Tomography: Plasma beam model 2.5 2 Model the plasma current as a cluster of rectangular, axis-symmetric current beams; it is the current distribution for each of these plasma beams that we wish to infer. 1.5 1 0.5 0 Prior Assumptions: 0.5 Current distributions for beams are Gaussian 1 The peak of each distribution is correlated to that of its neighbours. The strength of this correlation is controlled via a hyper-parameter τ 1.5 2 2.5 0.5 1 1.5 2
Bayesian Equivalent of Feynman Diagrams
Putting it all together Magnetic Predictions: Pred j (I) =B (or A) =M ij I i + C j MSE Predictions: Pred j (I) =tanγ = A 0 B Z (I)+A 1 B R (I) A 5 B φ (I) Bayes Formula: Posterior(I) j N a Obs j Pred j (I); σ j Prior(I)
Current Tomography: Expectation of Beam Currents Give Flux Surfaces 2.5 2.5 LCFS comparison with EFIT data over current 2 1.91676 2 tomography and Flux Surface Plots 1.5 1.32792 1.5 1 1 Results inferred from Magnetics & MSE Data 0.5 0.73908 0.5 Poloidal flux surfaces 0 0 plotted for currents 0.15024 corresponding to the 0.5 0.5 expectation of the 1 0.438599 1 posterior 1.5 1.5 Pulse = 22254 Time = 350ms 2 1.02744 2 τ = 200A 2.5 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2
Current Tomography: Variance of Current Distributions Show Flux Surface Uncertainties 2.5 2.5 2 1.91676 2 1.5 1.32792 1.5 Plots are of Flux Surfaces 1 1 corresponding to 200 beam current samples 0.5 0.73908 0.5 from the posterior 0 0 0.15024 0.5 0.5 1 0.438599 1 Pulse = 22254 1.5 1.5 Time = 350ms τ = 200A 2 1.02744 2 2.5 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2
q-profile Expectation and Distribution 11 Loosely, q is a measure of the pitch of magnetic field lines Important for plasma stability Plot is of q-profiles corresponding to 200 beam current samples from the posterior 10 9 8 7 6 5 4 3 2 Pulse = 22254 Time = 350ms τ = 200A 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Normalized Flux
Where we are going Developing Bayesian analytic techniques to validate, compare, and modify various force balance models. High-energy particle population corrections to Grad-Shrafranov Bayesian version of EFIT: EFIT + flux surface uncertainties Developing Bayesian inference framework for mode structure on H-1. Ultimately, this could be used to guide the development of new physics of 3D MHD wave analysis. Development of Bayesian inference techniques for diagnostic optimisation. Development of real-time control codes via linearised, analytic models or more-advanced inversion methods. Priors trained on historical data Advanced sampling algorithms
Wider Possibilities for Experiments Combining Different diagnostics to infer hard-to-measure quantities (e.g. electric field). Deployment of MINERVA to other experiments beyond JET, MAST and H-1 (e.g. ITER).
Conclusions MAST Magnetics and MSE have been successfully used to infer current tomography for MAST discharges using Bayesian analysis techniques via the MINERVA framework. Thompson Scattering has been integrated into the current tomography analysis to infer electron temperature and density profiles. Spin-off: Bayesian Analysis methods (e. g. Tikhonov crossvalidation) have been developed to quantitatively identify problematic magnetic diagnostics on MAST and effectively handle outlier data.
References Model Data Fusion: Developing Bayesian Inversion to Constrain Equilibrium and Mode Structure, Hole, von Nessi, Bertram, Svensson, Appel, Journal of Plasma Fusion Research (2010) Current Tomography in Axisymmetric Plasmas, Svensson and Werner, Plasma Physics and Controlled Fusion (2008) Large Scale Bayesian Data Analysis for Nuclear Fusion Experiments, Svensson and Werner, Intelligent Signal Processing (2007)