Optimization of Performance Prediction Model for Radio Frequency–Based Negative Ion Source

To accelerate the development of the radio frequency (RF)–based negative ion source (NIS), an initial performance prediction model was constructed using a back propagation neural network to predict the values of the H⁻ ion current and the extraction electron current by the extraction grid under set parameters. However, as experimental progress continued, the predictive capability of the model was no longer sufficient to meet practical requirements.

To enhance the model’s predictive accuracy, this paper proposes a comprehensive upgrade and optimization of the RF-NIS performance prediction model based on deep learning. The new model integrates the operational principles of the RF-NIS with neural networks and is structurally implemented through a combination of residual networks and deep neural networks. Additionally, a physical supervision and correction algorithm is introduced to correct prediction errors during the feature fusion phase.

Experimental evaluations show that the optimized model demonstrates superior performance. It will be implemented in experiments, replacing the original model, to support kilo-second-scale long-pulse experiments for negative ion-based neutral beam injection.