Deep learning models for ADITYA-U MHD equilibrium

Abstract This work presents an integrated deep learning framework for rapid magnetohydrodynamic equilibrium estimation in the ADITYA-U tokamak. A free-boundary equilibrium database containing 100,760 cases was generated using the pyIPREQ Grad-Shafranov solver from experimentally derived operating conditions spanning 766 ADITYA-U plasma discharges, with constraints representative of circular limiter plasmas near the flat-top phase. The database was used to develop a cascaded framework for predicting scalar equilibrium quantities, 1d safety factor profiles, and 2d poloidal flux profiles. Dense neural networks, principal component analysis, convolutional neural networks and physics-informed neural networks were investigated within this framework. The error propagation, cross-models consistency and ablation study of Grad-Shafranov residual were performed for the proposed framework. Across the test dataset, the models achieve median mean absolute errors of 1e-2 for the safety factor profile and 2e-3 for the normalized poloidal flux profile. The largest model requires approximately 1 ms per equilibrium inference on a multi-core CPU. The results demonstrate the potential of the framework as a fast surrogate for equilibrium analysis, experiment planning and real-time plasma control applications for ADITYA-U experiments.