| MCNP(Monte Carlo N-Particle Transport Code) is a general-purpose Monte Carlo N-Particle Transport code, which is used to calculate the transport progress of photon, neutron, electron, coupling photon, coupling neutron and coupling electron in the complex 3-dimensional geometric structures. MCNP also has the ability of eigenvalue calculation for the nuclear criticality system(namely the supercritical system and subcritical system).However, the input file of MCNP is extremely complex by artificial handwriting, which has the characteristics of complicated structure and being error-prone. With the increase of complexity of the geometrical model, the error rate will also rise. Aiming at the artificial input error-prone of MCNP input file, lots of researches on automatic generation of MCNP input file by computers are presented in the computer physics research field, and a majority of them are based on CAD(Computer Aided Design) modeling software. The complex mathematical calculation is applied to implement the conversion from CAD model to MCNP model.This thesis is based on the conversion algorithm from CAD model to MCNP geometry model. It redesigns the algorithm from the angle of information extraction and closed cell location evaluation, and presents a conversion algorithm based on Convolutional Neural Networks(CNNs). The number matching replaces the string matching to redesign STEP file information extraction method, whilst the utilization of Multilayer Perceptron(MLP), Support Vector Machines(SVM) and CNNs carries out image classification to replace the complex mathematical calculation and implement location relation judgement for closed cells.After lots of experiments, we find that CNNs based on 60×60 images obtained the highest accuracy. At last, the corresponding CNNs model is adopted to replace the complex mathematical calculation, which realizes the closed cell location relation judgement and reduces the error rate of complex mathematical calculation. Furthermore, the STEP file information extraction part obtains advanced acceleration effect. The experiments results demonstrate that the algorithm in this paper owns a better scalability and performance than our former algorithm. Also, it provides a new approach for feature recognition about CAD model as well as a solution for MCNP input file automatic generation. |
