Forecasting the first Edge Localized Mode (ELM) after LH-transition with a neural network trained on Doppler Backscattering data from DIII-D

TL;DR

This paper demonstrates a neural network model trained on Doppler backscattering data that can predict the first ELM crash in DIII-D tokamak plasmas approximately 100 ms before it occurs, aiding in mitigation efforts.

Contribution

The study introduces a neural network approach trained on DBS data to forecast ELMs in tokamak plasmas, achieving early prediction within 100 ms.

Findings

Neural network reliably forecasts first ELM 100 ms before occurrence.

Model trained on DIII-D data shows promising initial results.

Future work will enhance model robustness and expand training data.

Abstract

In H-mode tokamak and stellarator plasmas, edge localized modes (ELMs) lead to the expulsion of heat and particles beyond the edge transport barrier. ELMs cause a loss of energy and have the potential to damage the divertor and other plasma facing components, which motivates efforts to forecast such events to work alongside mitigation systems. In this paper, we use the Doppler backscattering (DBS) diagnostic data as input to train a neural network model, adapted from DeepHit [Lee et al., Deephit, AAAI 2018], to forecast the first ELM crash of H-mode discharges in DIII-D. The model takes 50 ms of DBS spectrogram data and predicts the probability of an ELM crash occurring within set time windows. Training and testing on shots found in the DIII-D database, we find the initial results promising, with the model reliably forecasting the first ELM 100 ms before it occurs. This successful proof-of-concept lays a strong foundation for a predictive tool that can deploy ELM-mitigation techniques before an ELM crash occurs. Future work will expand the training set with carefully selected shots and refine the neural network architecture to improve model robustness to noise and data variation.