首页|A hybrid Monte Carlo and Artificial Intelligence framework for efficient eye lens dose assessment in nuclear glove box facilities
A hybrid Monte Carlo and Artificial Intelligence framework for efficient eye lens dose assessment in nuclear glove box facilities
Chen, Dr. Jialu Ma, Mr. Yixuan Chu, Dr. Jian Yu, Dr. Libing Chen, Dr. Chun Tang, Dr. Wei
A hybrid Monte Carlo and Artificial Intelligence framework for efficient eye lens dose assessment in nuclear glove box facilities
A hybrid Monte Carlo and Artificial Intelligence framework for efficient eye lens dose assessment in nuclear glove box facilities
摘要
The recent reduction of the occupational equivalent dose limit for the eye lens by the International Commission on Radiological Protection has created an urgent need for accurate and rapid dosimetry in complex radiation environments. This study presents a hybrid framework integrating Monte Carlo (MC) simulation with machine learning for real-time eye lens dose assessment in glove box operations. A comprehensive dataset comprising 25,475 irradiation scenarios (9,412 photon and 16,063 neutron) was generated using MC simulations, covering a wide parameter space of source positions, source energies, shielding configurations, and operator head orientations. Two modelling approaches were systematically evaluated: eXtreme Gradient Boosting (XGBoost) and an uncertainty-guided deep neural network (DNN) with attention mechanism. Both models achieved remarkably high predictive accuracy across all test conditions, with R^2 values exceeding 0.997 for DNN and 0.999 for XGBoost, confirming the feasibility of machine learning-based dose estimation in this challenging scenario. XGBoost consistently outperformed the DNN, attaining test-set Mean Absolute Percentage Error (MAPE) of 0.154% (photon) and 0.453% (neutron), compared to 0.498% and 0.657% for the DNN, respectively. Despite the DNN's slightly higher errors, its performance remains well within acceptable bounds for practical applications. Ablation studies further revealed that the uncertainty input branch is essential for the DNN's performance under high-uncertainty neutron scenarios, while the attention mechanism primarily stabilizes training and suppresses outliers. Both models achieved millisecond-level inference times on standard hardware, satisfying the latency requirements for real-time occupational monitoring. Collectively, these results demonstrate that machine learning constitutes a viable and practically deployable approach for organ dose assessment in complex radiation fields, offering a robust alternative to conventional dosimetry while meeting the stringent requirements of the updated regulatory limits.