Physics-constrained neural networks for real-time reconstruction of radiated power in DIII-D plasma control system

A physics-constrained neural network is developed to predict radiated power across multiple plasma regions from foil bolometer measurements using a limited set of real-time equilibrium features. The surrogate model enforces a linear mapping between measured brightness and radiated power, enabling efficient inference with microsecond latency, suitable for real-time deployment. Model training is performed on a combination of large synthetic and real experimental datasets, allowing the neural network to learn the mapping between the measured signal and radiated power across varying plasma shapes while accounting for experimental noise, thereby improving its robustness. The model is optimized to estimate near-axis radiated power, enabling early detection of ongoing impurity accumulation. The approach has been implemented in the DIII-D plasma control system, demonstrating its capability for real-time operation.