A physically constrained Energy Spectrum Superposition Method-Machine Learning coupling algorithm for obtaining the neutron spectrum of BNCT rapidly

Abstract: Accurately and rapidly obtaining neutron energy spectrum is one of the significant challenges for Image-Guided Radiotherapy of BNCT (IGRT-BNCT). In this paper, machine learning (ML), including deep neural network (DNN) and Random Forest (RF) algorithm, is integrated into energy spectrum superposition method (ESSM) to predict unknown energy spectrum not present in the database, namely ESSM-ML. The new algorithm can further improve the speed of spectrum acquisition with high accuracy. The predictions of neutron energy spectra of two physical processes by both DNN and RF show high R² values and extremely low RMSE. In the three energy regions of thermal, epithermal, and fast neutrons, the deviations between predicted and true values are low. For the total neutron energy spectrum at the Beam shaping assembly (BSA) emission window, the ratio of the total neutron fluence rate obtained by ESSM-ML to that by the traditional simulation-based method is 95.3%. Moreover, for the time consumption of treatment energy spectrum obtained by ESSM-ML, the required time is only 69s, with a 4500-fold improvement in computational efficiency. The average prediction time for a single energy spectrum by the ML modules is merely 0.0052s. ESSM-ML provides the theoretical and algorithmic foundation for realizing IGRT-BNCT.