Real-Time Feedback Control of Radiation Front-Based Detachment Enabled by Machine Learning on DIII-D and KSTAR

We demonstrate real-time feedback control of the radiation front position for detachment in H-mode plasmas on DIII-D and KSTAR using machine learning (ML). Despite differing diagnostics, ML enables a shared detachment metric across devices, highlighting the flexibility and universality of radiation front-based control. On DIII-D, the radiation front was extracted from 2D C-III images measured by the RT-Tang TV system [1]. On KSTAR, the radiation front was defined from 2D radiated power measured by the IRVB [2] and reproduced in real time using ML models [3] trained on RT-AXUV and RT-Langmuir probe signals. Both systems used balance-point radiated power and PID control. DIII-D showed dynamic detachment and reattachment control within a single discharge and maintained control even during evolving equilibria, including strike point and X-point shifts, a challenging regime where conventional probe-based detachment control is often limited. On KSTAR, control succeeded in both full and partial detachment, mitigating performance degradation from 11% to 7%, which is the primary goal of all detachment controllers. These results show that radiation front-based feedback control with ML offers a robust and generalizable approach for future reactors.

[1] M. E. Fenstermacher et al. 1997 Rev. Sci. Instrum. 68 974-977

[2] S. Oh et al. 2024 Rev. Sci. Instrum. 95(9) 093528

3] A. Jalalvand et al. 2021 IEEE Trans. Neural Netw. Learn. Syst. 33(6) 2630-2641