Practicum Abstracts

XGC, a total-f, electromagnetic gyrokinetic particle-in-cell code has recently had a hybrid spectral - finite difference Poisson equation solver installed in order to accurately treat low toroidal peroidicity (low-n) mode numbers in toroidally confined magnetic confined fusion regime plasmas. This solver requires verification, including unit testing and full simulation runs. The latter will comprise large-scale low-n classical (straight cylinder) and neoclassical (toroidal geometry) tearing mode runs and subsequent comparison of linear growth rates and mode structures against existing MHD and gyrokinetic results in order to demonstrate the validity of the solver's method. Following this verification study, the limits of fluid descriptions of tearing modes will be investigated by scaling the ion gyroradius and ion banana orbit widths.

The multi-scale, multi-physics nature of fusion plasmas makes predicting plasma events challenging. Recently, neural networks based on the transformer architecture have shown great promise in making accurate predictions on sequences which have long-range, multi-scale characteristics. Examples include their use in generative large language models such as ChatGPT and in high-resolution image classification, where vision transformers are the state-of-the-art. In this work, we apply this neural network architecture to the popular problem of disruption prediction in fusion tokamaks, utilizing raw data from multiple multidimensional diagnostics from the DIII-D tokamak. Example diagnostics include the 2D Electron Cyclotron Emission imaging (ECEi) diagnostic, the 1D electron temperature profile, and the scalar q95 safety profile. Previous work has used Recurrent Neural Networks or Temporal Constitutional Networks to solve the same problem. However, these works has not achieved the desired false positive rates necessary to use these algorithms for real-time predictions, and do not make use of the various multidimensional diagnostics considered in this work.

Tokamaks are magnetic confinement devices for burning plasma to yield net positive nuclear fusion energy production. Current experimental tokamaks suffer from plasma disruptions which are impractical to predict from first principles in real time due to the multi-scale and multi-physics nature of tokamaks. Recent applications of machine learning [1, 3] have created predictive models of plasma disruptions in the JET and DIII-D small scale tokamaks and forecasting models for the plasma measurements [2] for use in a real time plasma control system. However, these predictive physics-based models suffer from numerical instabilities when they are used outside of the training domain. Recent work in the deep learning community has shown that when physically relevant symmetries of the system in question are obeyed by the layers of the model, the models generalize from training data [4] to other sets of initial conditions. We propose a deep learning model which uses such physics-based symmetries [5] of the gyrokinetic plasma formulation as a real time propagator of field fluctuations. The training data would be obtained from Gyrokinetic Toroidal Code (GTC) simulations of the DIII-D tokamak. Both the running of GTC and the training of the deep network will require HPC resources. The model will be evaluated by observing the accumulated error in the fields compared to the GTC simulation after an autoregressive simulation. This deep learning based autoregressive simulation could be used in conjunction with other codes to estimate instability growth rates from real time data [2]. Furthermore, this setup could be used in a model predictive control scheme. References: 1. Kates-Harbeck, J., Svyatkovskiy, A. & Tang, W. Predicting disruptive instabilities in controlled fusion plasmas through deep learning. Nature 568, 526–531 (2019). 2. Abbate, J., Conlin, R., & Kolemen, E. (2021). Data-driven profile prediction for DIII-D. Nuclear Fusion, 61(4), 046027. https://doi.org/10.1088/1741-4326/abe08d 3. Ge Dong, Xishuo Wei, Jian Bao, Guillaume Brochard, Zhihong Lin, & William Tang. (2021). Deep learning based surrogate model for first-principles global simulations of fusion plasmas. 4. Wang, R., Walters, R., & Yu, R. (2021). Incorporating Symmetry into Deep Dynamics Models for Improved Generalization. ArXiv:2002.03061 [Cs, Math, Stat]. http://arxiv.org/abs/2002.03061 5. White, R. L., Hazeltine, R. D., & Loureiro, N. F. (2018). Symmetries of a reduced fluid-gyrokinetic system. Journal of Plasma Physics, 84(2), 905840204. https://doi.org/10.1017/S0022377818000247

The goal of this project is to use machine learning to predict plasma disruptions in tokamaks, which is a highly nonlinear data-rich environment. Ultimately, one could create machine learned models to facilitate real time disruption avoidance in operational tokamaks. The lack of simple low order predictive models coupled with the 30 years of experimental data available makes this experimental area a promising area of application for data-driven methods.

Up until this summer, the Fusion Recurrent Neural Network (FRNN) disruption prediction model utilized a combination of various 0D and 1D signals from tokamak fusion plasmas to predict disruptive and potentially harmful reactor events, but it had not yet incorporated 2D signals. As new sequence model deep learning architectures (such as temporal convolutional networks and transformers) which demand more data to be fully utilized are developed, the capability of the FRNN software suite stands to be greatly improved should more complex data signals be added to the input. This summer, my goal was to expand the capability of the FRNN software to train on 2D signals, specifically Electron Cyclotron Emission Imaging (ECEi) data.

My practicum broadly approached the emerging collaboration of machine learning and fusion science at PPPL under Dr. William Tang. In particular, supervised machine learning methods have been used to successfully predict the onset of “disruptions” during the operation of tokamak reactors. Disruptive events must be avoided or mitigated to enable fusion energy production and to prevent damage to the reactor. Conventional first-principles calculations are unable to accurately model the complex dynamics that lead to disruptions in real-time. Using empirical signal data from the Joint European Torus (JET), we seek to apply new modern machine learning methods to the disruption problem, improve the predictive capabilities of the methods via addition of new signal data, and predict disruptions on a tokamak using algorithms trained on a different tokamak’s data. More broadly, we hope to generate promising results to encourage the application of machine learning methods within the fusion science community and establish an open data-sharing collaboration with the major tokamaks.

The prediction and avoidance of disruptions in tokamak fusion plasmas represents a key challenge on the way to stable energy production from nuclear fusion. Using diagnostic data from experimental runs with both disruptive and non-disruptive outcomes, we trained a deep recurrent neural network to predict the onset of disruptions. The algorithm achieves competitive performance on data from the JET tokamak. To deal with very large amounts of data and the need for iterative hyperparameter tuning, we also introduced a distributed training algorithm that runs on MPI clusters of GPU nodes and provides strong linear runtime scaling.

We studied the nonlinear physics of energetic particle-induced geodesic acoustic mode (EGAM) in tokomak plasma confinement. In this phenomenon, fast ions are present that have a velocity distribution far from Maxwellian. To capture the affect of the fast ions, a continuum kinetic approach is used where we keep track of the distribution of fast ions in both position and velocity space. This adds extra dimensionality to the problem and thus more computational complexity. We developed a code to solve this problem that is well suited for many-core machines such as GPUs.

Biological sensory systems are adept at performing a wide array of complex computational tasks related to pattern recognition. In the case of vision, organisms must track objects through space, make comparisons against memory, and learn from past experiences. Understanding how these tasks are completed is important both for gaining insight on natural systems and for developing methods for trend finding. Working toward these goals we explored the behavior of an artificial intelligence system called Thinking Machine 2. We determined that the impact of attention and noise on memory learning varied for different object classes as a function of feature similarity.

Using the Gyrokinetic Toroidal Code (GTC), which is PPPL's flagship code for studying plasma microturbulence in magnetically-confined high-temperature plasmas, running on DOE's Intrepid BG/P system, conduct a systematic study of the long-time evolution of plasma microturbulence including both collisionless dissipation from the ion temperature gradient (ITG) instability and collisonal dissipation from classical Coulomb collisional dynamics.

I assisted in the development of a massively parallel code for simulating Tokamaks.

The TSC code (Tokamak Simulation Code) was developed by scientists at the Plasma Phyiscs Laboratory in 1986 to model the transport and stability of axisymmetric plasmas in tokamak reactors. The code has since been modified by other scientist around the world and has become a useful design tool for new reactors. The code is difficult to use and the existing documentation was incomplete and outdated. My assignment was to learn the physics and numerical methods used in the model and to write a comprehensive users manual.