A cross-tokamak neural network disruption predictor for the JET and ASDEX Upgrade tokamaks

Abstract

First results are reported on the prediction of disruptions in one tokamak, based on neural networks trained on another tokamak. The studies use data from the JET and ASDEX Upgrade devices, with a neural network trained on just 7 normalised plasma parameters. In this way, a simple single layer perceptron network trained solely on JET correctly anticipated 67% of disruptions on ASDEX Upgrade in advance of 0.01 seconds before the disruption. The converse test led to a 69% success rate in advance of 0.04 seconds before the disruption in JET. Only one overall time scaling parameter is allowed between the devices, which can be introduced from theoretical arguments. Disruption prediction performance based on such networks trained and tested on the same device shows even higher success rates (JET: 86%; ASDEX Upgrade 90%), despite the small number of inputs used and simplicity of the network. It is found that while performance for networks trained and tested on the same device can be improved with more complex networks and many adjustable weights, for cross machine testing the best approach is a simple single layer perceptron. This offers the basis of a potentially useful technique for large future devices such as ITER, which with further development, might help to reduce disruption frequency and minimise the need for a large disruption campaign to train disruption avoidance systems.

Detection of Disruptions in the High-β Spherical Torus NSTX

“…On the other hand, the neural network may be better able to discern couplings in the data that lead to improved prediction, compared to the present case, where various pieces of data are evaluated independently. Also, while the present system produces only an "alarm" trigger, neutral networks can be configured to produce estimates of the time until the disruption [Windsor 2005], a piece of information that can prove quite useful in deciding which mitigation strategy to employ.…”

“…Neural networks are typically trained, in the sense that a predetermined sample of input and output data are used to determine the optimal values of the coefficients of the network. Neural networks have been used to predict various forms of disruptions on ASDEX Upgrade [84][85][86], DIII-D [87], ADITYA [88,89], TEXT [90,91], JET [85,92,93], and JT-60 [94,95].…”

“…Typical input data to the network include a measure of the plasma β (the normalized β (β N ) [96,97] or poloidal β (β P )), edge safety factor, plasma density (or Greenwald fraction [98,99]), locked mode amplitude, input power, radiated power, shape parameters, internal inductance, confinement time, and neutron emission. Some early work also used soft X-ray emission [87][88][89]91], high(er) frequency magnetic probes [87][88][89][90], or Dα monitors [87][88][89] though these diagnostics have typically not been used in more recent studies [85,86,92,93,95].…”

“…For instance, Ref. [85] shows that a neural network trained on JET data can anticipate 67% of ASDEX Upgrade disruptions within 10 ms, while a neutral network trained on ASDEX Upgrade can anticipate 69% of disruptions on JET within 40 ms. While this research represents excellent progress towards the formulations of a disruption detector for ITER, more work is required to understand the extrapolation to that larger device.…”

Desenvolvimento e utilização de um injetor de pastilhas de impurezas no estudo da mitigação de disrupções e atenuação de raios-X de altas energias

“…Uma técnica interessante que já foi empregada utiliza redes neurais para a previsão do momento da disrupção [34][35][36][37].…”