RTCAKENN: Predicting 7 kinetic profiles in real-time on DIII-D with enhanced robustness to diagnostic limitations and approaching offline level accuracy for FPP advancement

This work introduces RTCAKENN[1], a machine learning-based model designed for the real-time prediction of kinetic plasma profiles in tokamaks with significant demonstrated robustness against data absence. RTCAKENN leverages deep neural networks to predict seven profiles: pressure, safety factor, current density, electron/ion temperatures, electron density, and ion rotation. RTCAKENN's architecture is tailored for real-time environments, enabling profile predictions within an 8ms timeframe. The model is trained using a dataset of 696 discharges, incorporating a dropout technique during training to enhance its robustness against incomplete diagnostic data.

Experiments on DIII-D confirm RTCAKENN's capability to produce accurate plasma profiles in real-time, closely matching the precision of offline CAKE-level (offline state-of-the-art [2]) outputs, exceeding existing real-time alternatives. Notably, the model demonstrates impressive resilience by maintaining reasonable predictions even when key diagnostic inputs, such as TS or CER data, are missing. RTCAKENN can become a robust tool for real-time plasma control and analysis. Its ability to operate efficiently under diagnostic limitations and its fast execution time align with the operational demands of future FPPs.



Supported by NRF Korea (RS-2023-00255492) and US DOE (DE-FC02-04ER54698, DE-SC0015480, DE-AC02-09CH11466).



[1] R. Shousha et al 2024 Nucl. Fusion 64 026006

[2] Z.A. Xing et al 2021 FED 163 112163