Bulletin of the American Physical Society
63rd Annual Meeting of the APS Division of Plasma Physics
Volume 66, Number 13
Monday–Friday, November 8–12, 2021; Pittsburgh, PA
Session TM10: Mini-Conference: Machine Learning in Plasma Sciences IIOn Demand
|
Hide Abstracts |
Chair: Michael Churchill, Princeton Plasma Physics Laboratory Room: Room 406 |
Thursday, November 11, 2021 9:30AM - 10:00AM |
TM10.00001: Machine learning surrogates of Bayesian models of nuclear fusion experiments Andrea Pavone The diagnosis of large nuclear fusion experiments involves several different measurement devices relying on the observation of different plasma physics processes and leading to the collection of a vast amount of heterogeneous measurements. Bayesian inference provides a framework to integrate such data sources to infer a relatively small number of plasma parameters together with their uncertainties. It is performed by positing a predictive model of the observable plasma physics processes. The complexity of the models makes Bayesian inference computationally demanding: the inference of plasma profiles from a single measurement record can take tens of minutes. Here we show that artificial neural networks (ANN) can be trained as fast surrogates of Bayesian inference. The training data are generated with Bayesian models implemented within the Bayesian modeling framework Minerva: the data sampled from the joint probability distribution of the models can be used to teach ANN to reconstruct the model free parameters and joint probability values within hundreds of microseconds, making real time application possible. Such combination of Bayesian inference, modeling and ANN allows us to tackle different challenges at once: integrated modeling, uncertainty propagation and scalable inference. Our implementation is general and not bound to a specific experimental device: we show promising results with tests on different kinds of measurements from the W7-X and JET experiments. Future works can lead to an entirely automatic procedure for training machine-learning models as approximate inference algorithms for any Bayesian model implemented within a common framework. |
Thursday, November 11, 2021 10:00AM - 10:30AM |
TM10.00002: Overview and status of the FES Scientific Machine Learning project, "Accelerating radio frequency modeling using machine learning" John C Wright, Gregory M Wallace, Wes Bethel, Zhe Bai, T. Perciano, Robert Sadre, Nicola Bertelli, Syun'ichi Shiraiwa Our machine learning (ML) project focuses on accelerating radio frequency (RF) computational models to enable real time feedback control such as needed in advanced tokamak (AT) scenarios. Lower hybrid and high harmonic fast fasts are two types of waves used for current profile control in tokamaks. The former typically uses geometric options (ray tracing) coupled with a Fokker-Planck code to simulate the currect deposition profile. The later with its longer wavelengths uses physical options (full wave) also coupled with Fokker-Planck. The project aims to general surrogate models with machine learning to reduce simulation time to the order of milli-seconds (ms) needed for real time control during experimental discharges. The first part will develop a fast surrogate modelfor predicting RF heating and current drive with regression analysis using training data from raytracing calculations of LH waves in tokamaks. The second part will provide a ML generated solution to the inverse problem of relating line-integrated experimental bremsstrahlung measurements back to the associatedpower deposition and current drive profiles. This inverse problem does not have an analytic solution and the ML solution to it will provide real time information on the current profile. The third part involves accelerating full wave solvers by accelerating the inverse of the large matrices generated by that method. ML techniques will be used to find preconditioners to iteratively solver the matrices rather than use the more computationally intensive direct inversion. Once an efficient preconditioner is found, its is feasible to apply the first two methods to HHFW full wave models. Taken together, these three efforts result in ms times for surrogate model predictions suitable for real time control. |
Thursday, November 11, 2021 10:30AM - 10:45AM |
TM10.00003: Novel aggregate machine learning and transport modeling profile predictions Joseph A Abbate, Rory Conlin, Egemen Kolemen Progress is presented on a first-of-its-kind "aggregate" model that forecasts the state of a tokamak ~an energy confinement time into the future, to be used in tokamak control and as a "black-box" representation of devices for offline analysis. This aggregate model has been trained on DIII-D+ASDEX experimental data alongside outputs from TRANSP and ASTRA predictive transport simulations. The result is a model more portable and adaptable (i.e. applicable to new regimes and devices) than ML or empirical models trained exclusively on experimental data, yet more accurate than transport simulations. We show the aggregate model's superior performance over our previous fully data-driven forecaster for predicting unseen regimes. Results from a June 2021 DIII-D experiment implementing a similar neural network forecaster for model-predictive control (of temperature and pressure profiles via neutral beam power) are also presented. We discuss how our new "aggregate" methodology can be incorporated into the realtime DIII-D algorithm for a follow-up model-predictive control experiment in the 2022 campaign. |
Thursday, November 11, 2021 10:45AM - 11:00AM |
TM10.00004: Practical Techniques for Machine Learning Control of Fusion Plasmas Rory Conlin, Joseph A Abbate, Laura Fang, Azmaine Iqtidar, Yunona Iwasaki, Aaron Wu, Egemen Kolemen Achieving controlled nuclear fusion requires complex and robust control systems, and significant interest has been directed towards machine learning and artificial intelligence as techniques for developing these controllers. We present three components that will be necessary for any successful machine learning based control system. The first is control oriented models that allow efficient and robust calculation of actuator signals in real time via Linear Recurrent Autoencoder Networks and quadratic programming. We also demonstrate methods for uncertainty quantification for machine learning models and methods to estimate the uncertainty and confidence intervals for predictive control models. Finally, we discuss the need for accurate estimates of the plasma state which can be obtained using machine learning models for approximate Bayesian inference combining both irregularly sampled measurements and dynamical models of plasma behavior. |
Thursday, November 11, 2021 11:00AM - 11:15AM |
TM10.00005: Subgrid Modeling of Gyrokinetic Turbulence using Machine Learning Nathaniel Barbour, William D Dorland Turbulence in magnetic confinement fusion (MCF) experiments acts as a significant catalyst for driving the radial transport of heat. Future MCF experiments must operate in regimes that suppress turbulence in order to efficiently sustain core temperatures that facilitate fusion events. Modeling radial transport is an inherently multiscale objective, as small-scale instabilities influence macroscale heat and particle diffusion. In the interest of decreasing the computational costs of turbulence models, we will present a machine-learned subgrid model for gyrokinetic turbulence. We have reproduced the recent reservoir computing results of J. Pathak and collaborators for the Kuramoto-Sivashinsky system in the framework of GX [1]. We will present a successful extension of that work on a spectral domain. We will discuss ways in which this methodology can interface with coarsely-resolved numerical solutions. |
Thursday, November 11, 2021 11:15AM - 11:30AM |
TM10.00006: Data-driven models for Alfvén eigenmode classification based on high resolution ECE diagnostics at DIII-D Azarakhsh Jalalvand, Alan Kaptanoglu, Alvin V Garcia, Andrew O Nelson, Joseph A Abbate, Max E Austin, Geert Verdoolaege, Steven L Brunton, William W Heidbrink, Egemen Kolemen Modern day tokamaks have made significant advances in fusion understanding, but further progress towards steady-state operation is challenged by a host of kinetic and MHD instabilities. Alfvén eigenmodes (AE) are a class of mixed kinetic and MHD instabilities that are important to identify and control because they can reduce confinement and potentially damage tokamak components. In the present work, we utilize an expert-labeled database of DIII-D discharges and use (deep) recurrent neural networks such as a Reservoir Computing Network (RCN) to classify five AE modes, namely, BAE, EAE, LFM, RSAE and TAE. To deploy the model on a high-throughput FPGA accelerator for integration in the real-time plasma control system, we consider a data processing pipeline with minimum complexity. We trained a simple yet effective RCN on 40 raw ECE diagnostics down-sampled from 500kHz to only 2kHz. Our preliminary results show that such a model achieves a True Positive Rate of 91% with only 7% False Positive Rate, indicating promise for future investigation of AE modes such as detecting shape and location of these instabilities inside plasma and consolidation of the model into a real-time control strategy . |
Thursday, November 11, 2021 11:30AM - 11:45AM |
TM10.00007: Data-enabled Fusion Technology (DeFT): Machine Learning Tools in the Ousai Platform Craig Michoski, David R Hatch, Todd Oliver, Dongyang Kuang, Steph-Y. Louis, Siwei Luo, Matthieu Vitse This talk provides an overview of the Ousai framework—a framework facilitating cutting edge application of machine learning, data analytics, and AI to problems in fusion and beyond. The three categories of machine learning/AI tool types utilized in the Ousai framework are: analysis tools, optimization tools, and anomaly detectors. Our suite of analysis tools includes system identification and regression tools, model discovery and extraction tools, and model enhancement tools. Our suite of optimization tools include spectroscopy tools, configuration performance tools, and machine performance enhancement tools. Our suite of anomaly detectors include performance evaluation tools and contaminant detectors. We will discuss some highlights from the Ousai platform of machine learning tools for data-enabled fusion technology in this talk. |
Thursday, November 11, 2021 11:45AM - 12:00PM |
TM10.00008: A deterministic Gaussian-Mixtures Coulomb-collision algorithm for particle-in-cell Truong Nguyen, Luis Chacon, Guangye Chen, William T Taitano Coulomb-collision modules in PIC simulations are typically Monte-Carlo-based. Monte Carlo (MC) is attractive for its simplicity, efficiency in high dimensions, and conservation properties. However, it is noisy, of low temporal order (typically O(√△t), and has to resolve the collision frequency for accuracy.1 In this study, we explore a machine-learning- based, multiscale alternative to MC for PIC. The approach is based on the reconstruction of the particles’ velocity distribution function (VDF) using a Gaussian Mixtures Model (GMM) via the Maximum Likelihood Estimation principle.2,3 A key element of our algorithm is to decompose each Gaussian in the GMM into a convex linear combination of isotropic Maxwellians for which an exact set of evolution equations can be de- rived according to the Landau-Fokker-Planck collision operator.4 The proposed method is deterministic, free of instability, positivity-preserving, and strictly conservative, and is orders of magnitude faster than either MC or Eulerian Fokker-Planck solvers. We will illustrate the accuracy and performance of the proposed method with several examples of varying complexity. |
Thursday, November 11, 2021 12:00PM - 12:15PM |
TM10.00009: Data augmentation for disruption prediction via robust surrogate models Katharina Rath, Christopher G Albert, Bernd Bischl, Udo von Toussaint This work presents methodological developments regarding disruption prediction using machine learning techniques. When working with neural networks, a comprehensive training data base is important to achieve satisfying and reliable results. Here, we aim for a robust augmentation of the training data base using surrogate models. |
Thursday, November 11, 2021 12:15PM - 12:30PM |
TM10.00010: Upgrading LAPD diagnostic pipelines for training generative ML models Phil Travis, Steve Vincena Machine learning may transform the way science is conducted. We seek to update the Large Plasma Device (LAPD) data acquisition system to better capture machine state information (MSI) and global diagnostics that are important for ML-based analysis pipelines. Plasma processes in LAPD are typically studied by gathering high-spatial-resolution data using probes, combining measurements over many discharges at a 1 Hz shot rate. ML can instead be used to infer behavior over larger spatial regions from a few localized probe measurements and global or spatially-averaged diagnostics. An auxiliary system was appended to the current labview data acquisition routines to record LAPD MSI and auxiliary diagnostics including interferometers, visible light diodes, a diamagnetic loop, and a fast framing camera. An implicitly-generative, neural network-based energy based model (EBM) constructed in pytorch was trained on these auxiliary diagnostics, MSI, and probe measurements. Artificial discharges are sampled from the learned model via Langevin dynamics to predict time evolution of ion saturation current profiles. The EBM is able to reproduce general trends in profile evolution for a variety of magnetic field configurations. In addition, the EBM can be conditionally sampled to find the machine state required for a desired profile. Results from this model and comments on its accuracy will be presented, and the scientific prospects of this type of generative model will be discussed. |
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