Last updated 2020-01-23.

The Fusion Recurrent Neural Net (FRNN) software is a Python package that implements deep learning models for disruption prediction in tokamak fusion plasmas. Using diagnostic time-series data collected over years of tokamak pulses, or "shots", combined with labels of the disruption time (or "nondisruptive") provided by machine experts, our software aims to predict disruptions at least 30 ms before their onset. Future extensions of the framework will attempt to explain and classify the precursors to disruptions well in-advance of the event.

The main machine learning model used in the module (and behind the project's name) is based on a recurrent neural network (RNN), specifically the long short-term memory (LSTM) architecture. 1D (radial) profile data, if available and enabled by the user, is incorporated through a series of Conv1D layers and concatenated to the 0D scalar signal data at every timestep before being fed to the RNN layers. Future work will involve incorporating 2D (radial x toroidal) data from the electron cyclotron emission imaging (ECEi) advanced diagnostic in operation.

The implementation is built on Keras with a pre-v2.x TensorFlow backend. A second implementation in PyTorch is also available. The software can optionally employ data parallelism with distributed synchronous training through mpi4py. We do not currently support Horovod, but we expect to implement it soon.

Despite the name, the codebase also contains implementations of a variety of "shallow" ML methods for comparison: random forests, SVMs, single hidden layer perceptrons, and gradient-boosted trees. We have added, or are about to add, deep learning methods that are not based on RNNs, e.g. TCN and Transformers.

Refer to Kates-Harbeck, et al. (2019), for additional information.

The public software repository is hosted on GitHub: https://github.com/PPPLDeepLearning/plasma-python

The master host for the raw data used in the project is located on the /tigress GPFS file system hosted by Princeton Research Computing. At the time of writing, it consists of:

• 1.8 TB of SHOT_ID.txt files each containing the data from an individual diagnostic signal data, during a single shot, for the JET, DIII-D, and NSTX tokamaks.

/tigress/FRNN/signal_data/
/tigress/FRNN/signal_data_new_nov2019/

• A few MB of two-column (with 3 space delimiter) .txt files containing ID numbers of shots and the labeled disruption time (or -1.000000 if the shot is nondisruptive).

/tigress/FRNN/shot_lists/

The data is accessible from the ALCF Theta and Cooley systems via the Lustre file system. Periodically, the data from Princeton is copied via Globus to this secondary store:

/lus/theta-fs0/projects/fusiondl_aesp/shot_lists/
/lus/theta-fs0/projects/fusiondl_aesp/signal_data/

Although the subset of raw data that is used to train a single FRNN configuration is typically ~100s of GBs, the raw data must first be preprocessed into individual SHOT_ID.npz files each containing a pickled Shot class object which combines the resampled, cut, clipped, normalized, and transformed diagnostic data that was originally distributed among multiple SHOT_ID.txt files.

An example of such a set of preprocessed shot data files can be found here:

/lus/theta-fs0/projects/fusiondl_aesp/felker/processed_shots/signal_group_250640798211266795112500621861190558178

This pipeline greatly reduces the size of the input data used during training or inference. For the 0D FRNN model applied to our original set of 3449 shots from DIII-D, the preprocessed SHOT_ID.npz files occupy only a few GB, e.g.

A comprehensive tutorial for building and running the main FRNN model on the P100 GPUs of the Tiger cluster can be found in the main repository: https://github.com/PPPLDeepLearning/plasma-python/blob/master/docs/PrincetonUTutorial.md

Similarly, we maintain documentation for a tutorial on ALCF Theta: https://github.com/PPPLDeepLearning/plasma-python/blob/master/docs/ALCF.md

However, it is generally less up-to-date than the GPU documentation.