Research On Beam Offset Calibration Based On Deep Reinforcement Learning

Nuclear physics technology has promoted the rapid development of contemporary national defense and energy technologies,but improper disposal of nuclear waste has led to leakage of radioactive elements,causing environmental pollution and cancer induction.Proton linear accelerator is seen as a new modern approach to nuclear waste disposal and for cancer treatment.However,the proton linear accelerator is susceptible to the influence of the surrounding environment during beam acceleration,and the beam offset occurs,which causes the accelerated beam to generate particle loss and affect the accuracy of the physical experiment.Therefore,effective beam offset calibration is an extremely important scientific issue.At present,most offset calibration methods are built on physical derivation and numerical calculation.However,in the complex environment with manufacturing precision,installation accuracy and complex magnetic field changing between multiple magnets interaction,it is difficult to directly derive a certain function by physical derivation to express the relationship between the offset of the beam and the current value of the calibration magnet.Therefore,the use of innovative artificial intelligence methods to solve beam offset calibration has become a novel research direction.In this thesis,the current popular deep reinforcement learning technology is used to study the beam offset calibration problem on the medium energy transmission line of proton linear accelerator.The feasibility and advantages of deep reinforcement learning in beam offset calibration are analyzed in principle.A proton beam offset calibration model is proposed based on three methods as Deep Deterministic Policy Gradient,Proximal Policy Optimization,and Asynchronous Advantage Actor-Critic three methods.On this basis,this thesis further studies the application of six kinds of reward functions such as segmentation reward function and arctangent reward function in beam offset calibration.The experimental results show that the average offset value of the beam in the X and Y directions after offset calibration is reduced to 0.1515 mm and 0.1526 mm,which is better than the traditional physical method based calibration result(x=0.434 mm,y=0.158mm).Therefore,the beam offset calibration method based on deep reinforcement learning studied in this thesis has great research potential and application value.