The 37th Annual Conference on Neural Information Processing Systems, informally known as NeurIPS and representing one of the premier machine learning conferences, was held in New Orleans on Dec 10-16th, 2023. In addition to mainline machine-learning research, the conference also features a number of workshops on the intersection of machine learning and various domains of science and engineering.

Postdoc Lucas Spangher and PhD student Allen Wang both attended, each representing two workshop papers accepted as poster presentations on applying state-of-the-art ML techniques to the problems of disruption prediction, hybrid physics-ML dynamics modelling, and retrieving information from historical fusion experimental devices text databases (i.e., logbooks).

Several considerations clearly emerged from the conference:

  • Plasma physics and fusion are currently minimally-represented at machine learning conferences, making it a great opportunity to spread awareness of the challenges of fusion to help engage the ML community. There was a significant grassroots interest in applying new machine learning techniques to nuclear fusion datasets, so fusion dataset preparation would be a worthwhile activity.
  • Conversely, it was exciting learning from the work done in other scientific fields that have a longer history and tighter coupling with the ML community! Genomics, astronomy, climate modeling, and chemistry applications all have interesting niches of model development and excited communities.
  • An array of models was presented, but some model classes stood out. Diffusion models may hold promise as denoisers or data augmenters in fusion datasets.
  • Graph NN may be interesting for encoding spatial or static-distance information into multiple sensors present in tokamaks.
  • Multi-modal world models may eventually be interesting for nuclear fusion datasets which may feature words, time series, static metaparameters, and images.
  • Developments in the controls community will continue to prove significant to applications in fusion controls.

For more details on the various contributions, please refer to the table below.

Authors Title Workshop
LJ Spangher, WF Arnold, A Spangher, AD Maris, C Rea Autoregressive Transformers for Disruption Prediction in Nuclear Fusion Plasmas Workshop on Machine Learning and Physical Sciences
WF Arnold, LJ Spangher, C Rea Continuous Convolutional Neural Networks for Disruption Prediction in Nuclear Fusion Plasmas Workshop on Tackling Climate Change with Machine Learning
AM Wang, DT Garnier, C Rea Hybridizing Physics and Neural ODEs for Predicting Plasma Inductance Dynamics in Tokamak Fusion Reactors Workshop on Tackling Climate Change with Machine Learning
V Mehta, J Abbate, AM Wang, A Rothstein, I Char, J Schneider, E Kolemen, C Rea, DT Garnier Towards LLMs as Operational Copilots for Fusion Reactors Workshop on AI for Scientific Discovery: From Theory to Practice