Bulletin of the American Physical Society
60th Annual Meeting of the APS Division of Plasma Physics
Volume 63, Number 11
Monday–Friday, November 5–9, 2018; Portland, Oregon
Session CM10: Mini-Conference on Machine Learning, Data Science, and Artificial Intelligence in Plasma Research II |
Hide Abstracts |
Chair: Cristina Rea, Massachusetts Institute of Technology Room: OCC C124 |
Monday, November 5, 2018 2:00PM - 2:25PM |
CM10.00001: Disruption prediction: from shallow to deep learning and interpretability techniques Cristina Rea, Robert S Granetz, Kevin J Montes, Roy Alexander Tinguely Disruption prediction has the same appeal as solving the autonomous vehicle challenge: its solution could be a drastic leap forward. So far, machine learning-based disruption predictors have shown different and device-dependent performances [Rea PPCF 2018]. Nevertheless, the realization of a universal predictor, portable across existing and future tokamaks, working without extensive empirical tuning, is extremely important. Such a predictor needs to warn of an impending disruption hundreds of milliseconds in advance to inform the plasma control system (PCS) of the offending feature(s) and thus steer the plasma away from the disruptive operational space. We have recently embedded a Random Forest model in the real-time PCS of a fusion device; a baseline disruption predictor was obtained that, by running continuously over four months of operations, is showing encouraging results. We will discuss this shallow learning approach as well as the description of deeper architectures. To better understand disruption dynamics and optimize strategies for disruption avoidance, explainable predictions need to be provided: we will indeed present several interpretability strategies and their implications for disruption prediction. |
Monday, November 5, 2018 2:25PM - 2:45PM |
CM10.00002: Quantifying and Propagating Uncertainties to Enhance Real-time Disruption Prediction with Machine Learning Craig Michoski, Julian Kates-Harbeck, Gabriele Merlo, Max Bremer, Akash Shukla, Nikolas C Logan, David R Hatch, Cristina Rea, Todd A. Oliver, Jani Salomon Janhunen Having the capability to predict disruptions in tokamak reactors has the potential to dramatically improve/optimize both the performance and operating/repair costs of these reactors. Practical disruption prediction in tokamaks has recently improved by utilizing advanced analytics via data-driven machine learning algorithms that utilize real-time machine diagnostics to predict the onset of major disruptions. In this talk we discuss various ways to enhance these predictive capabilities by incorporating and propagating experimental uncertainties from diagnostic signals into the more conventional machine learning algorithms. |
Monday, November 5, 2018 2:45PM - 3:05PM |
CM10.00003: Statistical Distance-Based Validation Metrics for Probabilistic Plasma Turbulence Validation Studies Payam Vaezi, Chris Holland, Brian A Grierson, Gary M Staebler, Orso Meneghini, Sterling P Smith We investigate how uncertainties in experimentally-derived transport model inputs impact model predictions, using the quasilinear TGLF transport model. We use the rapidly converging and computationally inexpensive probabilistic collocation method (PCM) to propagate probability distribution functions (PDFs) of experimentally-derived input parameters through TGLF, yielding PDFs of predicted transport fluxes and local flux-matching gradient scale lengths. The experimental fidelity of the TGLF predictions is then assessed by comparing the predicted PDFs to PDFs of the corresponding experimental measurements and power balance calculations, using a new validation metric based upon the Wasserstein distance [1]. The utility of these metrics are illustrated via transport modeling of a DIII-D ITER baseline plasma in which the mix of neutral beam injection (NBI) and electron cyclotron heating (ECH) is varied. We find the ion temperature gradient is the dominant uncertainty driver in the NBI-only phase, whereas multiple parameters are important in plasmas with NBI+ECH heating. |
Monday, November 5, 2018 3:05PM - 3:25PM |
CM10.00004: Machine learning and algorithmic approaches in ICF Capsule Design Peter William Hatfield, Steven Rose, Robbie Scott In this talk we will discuss recent work in the UK on developing machine learning approaches to modelling and predicting the yield from NIF-like ICF implosions. We present several new ensembles of 10^3-10^4 simulations, showing that the uncertainty on predictions can be accurately decomposed into uncertainty from lack of data, and uncertainty on input parameters. We also show new approaches to finding novel classes of design with comparatively little human intervention. Finally we will briefly discuss how modern data science techniques are being used to support and maximise the utility of other types of HEDP experiments undertaken at the Central Laser Facility at the Rutherford-Appleton Laboratory. |
Monday, November 5, 2018 3:25PM - 3:50PM |
CM10.00005: A machine learning approach to interpreting complex high-dimensional spaces in Fusion Research Alessandro Pau The ever-increasing volume of multidimensional data generated in the framework of fusion experiments, combined with the exponential growth of computing power and cloud computing technologies, has motivated researchers to explore advanced Statistical and Machine Learning (ML) techniques to solve complex problems. One aspect of particular interest in fusion research is the interpretability of ML models, that is, the capability to explain the connections between the knowledge extracted from data and the output of what has always been considered as a black-box. This contribution summarizes the efforts in the development of an innovative machine learning approach for the investigation of complex high-dimensional spaces and its exploitation as pattern recognition model to predict and classify disruptions on Tokamaks. The tool implemented for the analysis is based on the Generative Topographic Mapping (GTM), a generative model belonging to the class of manifold learning techniques. The advanced visualization capabilities of the tool together with its potential real-time application allow exploiting GTM algorithm for plasma monitoring, identifying patterns which reflect physics mechanisms leading to disruptions. |
Monday, November 5, 2018 3:50PM - 4:10PM |
CM10.00006: Manifold learning to detect the transition from kinetics to hydrodynamics Charles Leland Ellison, Frank R Graziani, Jeff Haack, Elizabeth Munch, Michael Sean Murillo, Liam G. Stanton In certain regimes, hydrodynamic models provide reduced descriptions of kinetic models. These reduced models are expected to be accurate when the dynamics are sufficiently near collisional equilibrium. Precise characterization of this ``nearness'' has been an active area of research, appearing for instance in David Hilbert's influential list of open mathematical problems posed at the beginning of the 20th century. In this work, we use manifold learning to identify transitions from the kinetic regime to the hydrodynamic regime. To employ manifold learning in this context, an ensemble of initial conditions is evolved using a BGK kinetic equation. As collisions increase entropy, the ensemble of initial conditions collapses onto a space of reduced dimensionality --- one that could in principle be parameterized by hydrodynamic fields. Manifold learning provides a novel means of characterizing this reduced dimensionality and therefore informing the realm of validity of hydrodynamic approximations. We will demonstrate the technique in representative problems and discuss sensitivity with respect to variations in the study parameters. LLNL-ABS-753699 |
Monday, November 5, 2018 4:10PM - 4:30PM |
CM10.00007: Nested cross-validation loop for performance optimization in imbalanced problems Kevin Montes, Cristina Rea, Robert S Granetz, Roy Alexander Tinguely A disruption prediction algorithm based on the Random Forests Machine Learning method has been developed using large databases of both disruptive and non-disruptive discharges from EAST and Alcator C-Mod. The algorithm was trained on time samples of several physics parameters during the flattop current phase, which were cast into a binary classification scheme based on their proximity to the time of the current quench. Roughly 80% of each database is composed of non-disruptive discharges, and only a fraction of the time samples from disruptive discharges are designated as the positive (close to disruption) class. Therefore, the preponderance of negative class samples results in an imbalanced classification problem regardless of whether it is framed on a time-sample or discharge-by-discharge basis, and care must be taken to accurately measure prediction performance. This presentation describes a nested K-fold cross-validation procedure to determine an optimal mapping from the individual time sample predictions of the random forest to an alarm trigger of an impending disruption. An exploration of sampling methods and performance metrics to address the imbalance between positive and negative classes is also discussed. |
Monday, November 5, 2018 4:30PM - 4:50PM |
CM10.00008: Hazard function exploration of tokamak tearing mode stability boundaries K. Erik J. Olofsson, Brian S. Sammuli, David Humphreys It is possible to model an event prediction problem by a hazard function (event rate) and a phase space trajectory. Compared with typical event prediction approaches, the hazard function has two significant advantages. First, it has a time localized and quantitative interpretation (events per time). Second, event rate models can be used to generate event predictions for arbitrary look-ahead horizons conditioned on future controls with respect to a dynamical systems model. To be able to apply these advantages, we are first required to develop the necessary tools to effectively learn (correct) hazard function models from data. We report scaled-up hazard function analysis applied to tokamak data for tearing mode onset characterization. A particular tearing delta-prime proxy does not significantly increase the likelihood of the hazard model. Shot-database searching and feature extraction tool-chains developed at the DIII-D National Fusion Facility are exploited. Multiple methods for understanding and visualizing the properties of the estimated hazard function(s) are applied, including partial dependence plots. |
Follow Us |
Engage
Become an APS Member |
My APS
Renew Membership |
Information for |
About APSThe American Physical Society (APS) is a non-profit membership organization working to advance the knowledge of physics. |
© 2024 American Physical Society
| All rights reserved | Terms of Use
| Contact Us
Headquarters
1 Physics Ellipse, College Park, MD 20740-3844
(301) 209-3200
Editorial Office
100 Motor Pkwy, Suite 110, Hauppauge, NY 11788
(631) 591-4000
Office of Public Affairs
529 14th St NW, Suite 1050, Washington, D.C. 20045-2001
(202) 662-8700