The Evaluation Of Photonuclear Reaction Data Based On Multi-task Neural Networks

It is essential to evaluate accurately the nuclear data since they are important inputs of nuclear structure and reaction as well as the utilization and engineering development of nuclear energy.Photonuclear reactions' data as the key nuclear data,which help us to study nuclear excitation properties and reveal the interaction between photon and nucleus,are expected to be re-evaluated by IAEA in recent years.Machine learning,as a powerful data processing algorithm,has been widely used in nuclear physics.In this thesis,the neural network algorithm in machine learning will be adopted to develop an effective model,and the characteristic data of photonuclear reaction will be evaluated and predicted.Firstly,the key parameters of nuclear giant dipole resonance(GDR)are described based on neural network algorithm.By correctly classifying the single and double peak features in the excitation spectrum and establishing the multi-task learning network structure to correlate various nuclear data,the description of the resonance energy and width of representative nuclides is optimized,and the accuracy is significantly improved compared with the previous models.Then the key GDR parameters are predicted based on the trained network for 79 nuclei near ?-stability line but lack of experimental data,providing possible suggestions for future experiments and data evaluation.Furthermore,the evaluation of(?,n)photonuclear reaction data is carried out by using the developed neural network method.Due to the limitation of different experimental methods,the data of photonuclear reactions in the nuclear database still keep uncertain for several nuclides,such as Ge and Se isotopes.In order to improve the evaluation accuracy to predict the relevant data,three neural networks are constructed and trained,correspondingly the relations of the cross sections of photonuclear reaction with nuclear structure characteristics and reaction energies are discussed.Then the models are utilized to evaluate(?,n)cross sections for several nuclides which are largely uncertain experimentally.It is found that the model prediction exhibits clear tendency to one of the data from different experimental methods.As an extension of the developed neural network method,the thesis also tries to combine it with the local mass relation to study the nuclear mass data.As summary of the thesis,the multi-task neural network method is applied successfully to the evaluation of nuclear data,with its advantage in analysing complex nuclear data.The links between various nuclear quantities increase the reliability of model prediction,which help us to understand further the physics behind and describe better the nuclear data.