Neural network applications for NBI performance studies using synthetic data from BTR code

Abstract Neutral Beam Injectors (NBIs) play a critical role in plasma heating and current drive in fusion devices. Accurate diagnostics of the beam's thermal loads and gas density distribution within the injector chamber are essential for optimizing performance and minimizing power losses due to reionization. This study explores the application of neural networks for analyzing thermal loads on NBI surfaces using synthetic data generated by the BTR (Beam Transmission and Re-ionization) code. We frame the problem as two distinct tasks: (1) classification of power load sources (neutral particles, ions and their mix), and (2) regression-based prediction of gas density profiles within the neutral beam injector. Several machine learning architectures, including multilayer perceptrons (MLPs), convolutional neural networks (CNNs), including AlexNet, and state-of-the-art models such as YOLOv8, are evaluated. Experimental results demonstrate that even relatively simple neural networks can effectively classify thermal load distributions achieving accuracy over 80% and predict gas density peaks achieving R 2 score over 0.96, suggesting their potential for real-time diagnostics. Future work will focus on expanding the dataset and validating the models with experimental data to improve generalizability.