Development of a neural network-based surrogate model for neutral beam injection in KSTAR

A neural network-based surrogate model for the plasma heating and current-drive response to neutral beam injection is developed for KSTAR using high-fidelity NUBEAM simulations. The model is trained on a synthetic dataset of prescribed kinetic profiles and magnetic equilibria representing typical KSTAR H-mode scenarios. Principal component analysis suppresses Monte Carlo noise inherent in NUBEAM profile outputs, preventing overfitting. A time-filtered beam power, which encodes the finite slowing-down time of injected fast ions, is introduced as an additional input, enabling the model to better capture their delayed thermalization dynamics. Predictive uncertainty is quantified by training an ensemble of neural networks and evaluating the variance across their predictions. The surrogate reproduces key NUBEAM outputs with high accuracy within a fraction of a second, achieving a speed-up of several orders of magnitude over conventional NUBEAM simulations. As a first demonstration, the model is integrated into the V-KSTAR post-analysis system, with an out-of-distribution handling strategy ensuring robust operation for experimental KSTAR inputs.