Do Fusion Plasma Time-Series Have a Persistent Memory that Machine Learning May Exploit?

Numerous Machine Learning (ML) models [1–4] have been proposed which, with varying degrees of success, attempt to predict the probability of disruption throughout the shot. Here, our work here is two-fold: (1) we give a practical study of model introduction, in which we demonstrate several ML techniques for consideration, and (2) we give a scientific study of model comparison, to interpret why one model may perform better than another. We compare a GPT2-like transformer [5] to several other algorithms, including a random forest and neural nets with limited window convolutions and/or recurrence. All are rigorously tuned to ensure a fair benchmark. One aspect of the transformer sets it apart from the others: masked self-attention, i.e. the ability to explicitly use information across the entire shot (including the ramp-up) when making a decision on the disruption probability. With these model comparisons, we argue for (or against) the persistence of a “memory” throughout the plasma that ML may exploit.

[1] J. Vega et al 2014 Nucl. Fusion 54 123001

[2] J.X. Zhu et al 2021 Nucl. Fusion 61 026007

[3] C. Rea et al 2019 Nucl. Fusion 59 096016

[4] K.J. Montes et al 2019 Nucl. Fusion 59 096015

[5] A. Randord et al 2019, OpenAI blog 1.8