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 BM10: Mini-Conference on Machine Learning, Data Science, and Artificial Intelligence in Plasma Research I |
Hide Abstracts |
Chair: J. Luc Peterson, Lawrence Livermore National Laboratory Room: OCC C124 |
Monday, November 5, 2018 9:25AM - 9:30AM |
BM10.00001: WELCOME REMARKS J. Luc Peterson . |
Monday, November 5, 2018 9:30AM - 9:55AM |
BM10.00002: Reconstruction of fusion plasma state with a Plasma Debugger M. Dikovsky, E. A. Baltz, Y. Carmon, R. Koningstein, I. Langmore, T. Madams, P. C. Norgaard, J. Platt, J. Romero, M.C. Thompson, E. Trask, E. Granstedt, H. Gota, R. Mendoza We built a "Plasma Debugger", a tool to reconstruct the state of the FRC plasma in the TAE Technologies' experimental machine C-2W. This generalized Bayesian inference approach combines data from magnetic sensors, fast cameras, FIR interferometer, Thomson Scattering system, Bremsstrahlung measurements and neutral beams shine-through SEE detectors. It then reconstructs electron density, temperature and bulk plasma currents, with confidence intervals. Computation takes hundreds of CPUs and is performed in the cloud, with results showing up in the plasma machine control room within several minutes of the experiment. The display shows time evolution of the basic plasma properties, giving machine operators additional insight into the plasma behavior. |
Monday, November 5, 2018 9:55AM - 10:15AM |
BM10.00003: Unsupervised Reinforcement Learning of ALE Mesh Management Strategies for Hohlraum Simulations in HYDRA Jay David Salmonson, Han Truong, Joseph M Koning, Jayson Dean Lucius Peterson We report on our implementation of reinforcement learning (RL) algorithms in PyTorch to learn and automate mesh management strategies of hohlraum simulations in HYDRA. We define regions of the simulation mesh and extract a set of features for each that contain information about the quantity and degree of distortion and irregularity of nodes within the region. This feature set (and its time history) comprise the state of the system at a given time step. Based on a reward function defined to favor minimal intervention but enable the simulation to continue (i.e. not crash), the RL algorithm predicts an action on that region, e.g. relax the mesh, freeze it, or do nothing. The simulation is advanced another time step with this action implemented, and the process is repeated. In this way training episodes are run and recorded for later replay and training. The trained net can then be used as an inference engine to make mesh relaxation decisions in the regions as a simulation runs. |
Monday, November 5, 2018 10:15AM - 10:35AM |
BM10.00004: Development of plasma control algorithm design via machine learning Brian Scott Sammuli, Erik Olofsson, David Humphreys, Martin Margo, Mark Kostuk In this work we explore the use of machine learning for vertical stability control of the DIII-D tokamak. We describe the application of a search and machine learning computing toolchain for system identification of highly non-linear coil/vessel/plasma interactions as a function of equilibrium state. Extraction of the training data used in this process is achieved at rates over two orders of magnitude greater than previously attainable. The identified system model is then integrated with a closed loop controller and tested in simulation. Additionally, we investigate creation of a purely data-driven vertical control algorithm (using, for example, reinforcement learning). Development toward integration of these algorithms for real time use in the DIII-D plasma control system is discussed. The use of large-scale machine learning techniques for plasma control is still novel within the fusion community, and this work provides a template for future data- driven approaches. |
Monday, November 5, 2018 10:35AM - 10:55AM |
BM10.00005: Accelerated predictive models based on TRANSP for scenario optimization and control of NSTX-U Dan Boyer, Keith Erickson, Stanley Kaye, Vaish Gajaraj, Justin Kunimune, Michael Zarnstorff Model-based control and scenario development for fusion devices rely on a hierarchy of models of varying fidelity and speed. Integrated modeling codes, like TRANSP, can provide high-fidelity simulation capability, but are not well-suited for real-time implementation. Data-driven reduced modeling based on higher fidelity models provides a path for developing accelerated models for these tasks. Several such models have been developed for NSTX-U including a real-time capable neural network beam model, NubeamNet, that calculates heating, torque, and current drive profiles from equilibrium parameters and measured profiles. Models have also been developed for real-time evaluation of plasma conductivity, bootstrap current, and flux surface averaged geometric quantities for use in current profile control algorithms. Approaches to hyper parameter tuning have been studied to enable optimization of generalization and complexity. Hardware-in-the-loop simulations in the NSTX-U plasma control system show suitability of the models for real-time applications. Initial applications, including estimation of anomalous fast ion diffusivity to match measured neutron rates, will be presented. |
Monday, November 5, 2018 10:55AM - 11:20AM |
BM10.00006: Scientific Data Services Framework for Plasma Physics Kesheng Wu, Bin Dong, Surendra Byna
Plasma physics experiment and simulations are producing petabytes of data. Hundreds of diagnostic tools are being used with thousands of different analysis tasks on these datasets to generate scientific insight. Often I/O operations are the bottleneck in these analysis operations. This work address the I/O efficiency issue by developing techniques for common data access patterns, for deep storage hierarchies, and for massive parallelism. Additionally, we present a thorough theoretical analysis of the data access cost to exploit the structural locality, and select the best array partitioning strategy for a given operation. In a series of performance tests on large scientific datasets, we have observed that our framework outperforms Spark by as much as 2070X on the same tasks. |
Monday, November 5, 2018 11:20AM - 11:40AM |
BM10.00007: Learning-based predictive models: a new approach to integrating large-scale simulations and experiments Brian K. Spears, Jayson Dean Lucius Peterson, Timo Bremer, Brian Van Essen, John E Field, Peter Robinson, Jessica Semler, Bogdan Kustowski, Jim A Gaffney We will describe a large research effort at Lawrence Livermore National Laboratory aimed at using recent advances in deep learning, computational workflows, and computer architectures to develop an improved predictive model – the learned predictive model. Our goal is to first train these new models, typically cyclic generative adversarial networks, on simulation data to capture the theory implemented in advanced simulation codes. Later, we improve, or elevate, the trained models by incorporating experimental data. We will present work using inertial confinement fusion as a testbed for development. We will describe advances in machine learning architectures and methods necessary to handle the challenges of ICF science, including rich, multimodal data and strong nonlinearities. We will also cover state-of-the-art tools that we developed to manage our computational workflow. The tools manage a wide range of tasks, including developing enormous simulated training data sets, driving the training of learned models on simulation data, and elevating learned models through exposure to experiment. We will end by drawing ties to other scientific applications both within LLNL and in the broader computation and science community. |
Monday, November 5, 2018 11:40AM - 12:00PM |
BM10.00008: Transfer learning for the calibration of the inertial confinement fusion simulations Bogdan Kustowski, Jim A Gaffney, Brian K. Spears, Gemma J. Anderson, Jayaraman Jayaraman Thiagarajan, Rushil Anirudh Transfer learning refers to exploiting the knowledge gained from solving one problem and applying it to solve a different but related problem. A well-known example is reusing publicly available neural network models that have been trained on large sets of images, and partially retraining them to solve a new task, for which fewer data are available. Transfer learning has not been extensively applied in physical sciences yet. In this presentation, we discuss numerical tests that have been carried out to investigate the applicability of transfer learning to calibrate the Inertial Confinement Fusion (ICF) computer simulations against the experimental data, which will be obtained at the National Ignition Facility (NIF). A neural network is initially trained to predict the simulation results and then retrained to match the sparse experimental data. A validation data set is then used to investigate the calibration accuracy as a function of the experimental data volume, retrained model capacity, and the size of discrepancy between simulations and experiments. Preliminary results are encouraging and motivate further investigation of transfer-learning-based calibration using larger data volumes. |
Monday, November 5, 2018 12:00PM - 12:20PM |
BM10.00009: Finding structure in large datasets of particle distribution functions using unsupervised machine learning Randy Michael Churchill, Choong Seock Chang, Seung Hoe Ku The raw data generated by simulation codes on supercomputers can be so large that it requires data reduction methods to allow scientists to understand it. Physics based reductions are often used, for example taking moments of particle distribution functions. It must be realized, however, that there will be a loss of information in these reductions. Here we explore the use of unsupervised machine learning algorithms, to see if patterns and structure in the data itself can be learned and discovered, to give researchers further insight into areas of interest. This has the potential benefit of discovering kinetic structure which would be lost by some physics based reductions. We utilize the 5D, gyrokinetic distribution function in simulations from the full-f code XGC1. We find that in spatial regions of “blobby” turbulence in the edge, the electron distribution function has a very distinct signature, with higher energy regions varying across space separately from the lower energy component, and higher energy regions showing a distinction near passed/trapped boundaries. |
Monday, November 5, 2018 12:20PM - 12:40PM |
BM10.00010: Neural-Network accelerated fusion transport simulations for ITER scenario modeling Chieko Sarah Imai, Orso Meneghini, Joseph McClenaghan, Sterling P Smith, Gary M Staebler, Alberto Loarte Preparation for ITER operation relies on our ability to efficiently predict the plasma confinement with high physics fidelity. High-fidelity turbulent transport models such as TGLF remain one of the major bottlenecks in this process. To accelerate prediction of the turbulent transport, the neural network (NN) approach of TGLF (TGLF-NN) described in [Meneghini NF 2017] has been generalized to support predictions for ITER both during its commissioning phase (H only) and nuclear phase (D+T and He ash). In this talk we describe the techniques used to sample the 20+ dimensions input parameters space and assemble a simulation database that is suitable for training robust machine learning based models. Tuning of the NN model hyper-parameters was carried out on GPU enabled clusters, both using Gaussian process based optimization, and random sampling of the configuration space. Dimensional reduction with auto-encoders, and training on a latent-space data-set was also investigated. Results of coupled core-pedestal ITER simulations leveraging the latest TGLF-NN model will be presented. |
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