Description
To achieve high-fidelity equilibrium reconstruction, we develop a Long Short-Term Memory (LSTM) framework integrating ex-vessel magnetic and neutron diagnostics. The primary objective of this framework is to bypass the long-term reliability issues of in-vessel diagnostics caused by the harsh radiation environments in Demonstration Fusion Power Reactor (DEMO) class devices. Utilizing KSTAR experimental data from 2023 to 2026, a model will be constructed to resolve data sparsity through virtual-sensor-based data augmentation alongside neutron measurements. By incorporating neutron diagnostics, the project intends to improve the predictive accuracy of plasma beta (𝛽), thereby establishing a DEMO-relevant baseline where precise pressure profiles are critical. The upcoming validation phase aims to show that the integrated model matches the predictive fidelity of off-line EFIT, maintaining high accuracy even through transient operational phases. Additionally, we will employ SHAP(SHapley Additive exPlanations)-based interpretability methods to confirm that the model successfully learns the underlying magnetohydrodynamic (MHD) physics, particularly the Shafranov shift. Expected to operate with a microsecond-level inference speed per time slice, this inverse problem solver is anticipated to prove that combining external and neutron diagnostics offers a highly efficient path toward precise equilibrium analysis for future fusion reactors.