Tearing Instability Prediction for MAST Shots Using a Transformer-Based Deep Learning Model

This article introduces a deep learning model which uses Transformer-based architectural components for next time-step prediction of plasma state parameters and tearing instability probability. The developed model features modern activation functions and utilizes the self-attention mechanism to capture long-range spatial dependencies within multivariate diagnostic signals. The model provides real-time uncertainty estimates by predicting probabilistic output distributions via the use of a negative log-likelihood loss function. Training and evaluation of the model were conducted using shot data from the Mega Ampere Spherical Tokamak (MAST) experimental program. The model offers the capability for the development of uncertainty-aware real-time plasma control for instability avoidance.