A Scenario-Agnostic Neural Network Model for Identifying NTM Triggers in DIII-D

We present a neural network–based classifier and regressor for determining the onset time of disruptive m/n = 2/1 Neoclassical Tearing Modes (NTMs) with sub-millisecond precision using DIII-D magnetic probe data. The model is trained on synthetic poloidal profiles of the complex-valued Fourier amplitude of the n=1 signal, generated via singular value decomposition (SVD) from tearing- and sawtooth-dominated regimes of 144 ITER Baseline Scenario discharges, without requiring precise human labeling of onset times. Time-dependent linear combinations of the SVD modes provide training and test data with known ground-truth onset across a broad range of sawtooth–tearing mixtures. The method identifies tearing onset in these synthetic mixed-mode scenarios with sub-millisecond mean error and a standard deviation, and generalizes well to unseen experimental data. This fast, precise, and scenario-agnostic framework enables separation of true seeding mechanisms from coincidental correlations, helping resolve longstanding ambiguities in NTM onset interpretation and guiding the development of trigger-resilient, passively stable operating regimes in advanced tokamaks.