Development of a Neural Network-Based Surrogate Model for Neutral Beam Injection in KSTAR

A neural network-based surrogate model for neutral beam injection (NBI) has been developed for the KSTAR tokamak. The model is trained and validated using high-fidelity simulation data generated by the Monte Carlo code NUBEAM. The training dataset is constructed from assumed kinetic profiles of expected KSTAR operating scenarios, including various magnetic equilibrium configurations. In addition to conventional one-dimensional outputs—such as beam heating and current drive profiles—the model is also trained on two-dimensional data, including beam birth profiles in the (R, Z) plane and slowing-down distributions in (E, μ) space, which are particularly useful for energetic particle physics studies. To ensure efficient and stable learning, one-dimensional profiles are preprocessed using principal component analysis, while two-dimensional data are encoded and decoded using autoencoders. The model also supports uncertainty quantification via ensemble averaging of multiple networks. The trained surrogate model drastically reduces inference time—from several minutes for a full NUBEAM run to just a few milliseconds for 1D outputs and a few tens of milliseconds for 2D profiles—making it well-suited for real-time control applications and between-shot analysis. The model is now ready for integration into plasma control systems, scenario development frameworks, and fast-ion diagnostics for KSTAR operations.