Neural network based fast prediction of double tearing modes in advanced tokamak plasmas

Abstract

An advanced tokamak operation featuring reversed magnetic shear configuration can trigger double tearing modes, a deleterious resistive magnetohydrodynamic (MHD) instability that plays a crucial role in causing major disruptions. Generally, the properties of the double tearing mode can be investigated through solving MHD equations. However, solving MHD equations is widely recognized as a challenging task due to the complex physical model, massive CPU requirements, and lengthy computation time. A surrogate model based on MHD equations can quickly provide physics-based instability information, such as the mode growth rate, serving as references for future control needs. In this work, a reliable surrogate prediction model, LTM-NN, is established using a neural network (NN) based on a numerical database. This work employs a linear tearing mode (LTM) code with a large aspect ratio tokamak configuration to construct a supervised learning dataset of equilibria and growth rates. It is found that the trained LTM-NN predicts growth rates with a high accuracy up to 97%, and the classification accuracy for the stability of the double tearing mode reaches 100%. Furthermore, LTM-NN is applied to systematically investigate the dependence of the growth rate on the inter-resonance distance, the magnetic shear, and the magnetic Reynolds number. The results are consistent with existing theories. Thus, high accuracy, fast computing speed, and correct parameter dependencies demonstrate that LTM-NN is capable of serving as a rapid predictor of double tearing modes to provide instability information for the integrated model or the disruption warning model in the future.