A graph neural network-based method for reusing Monte Carlo CSG models and its application

BackgroundAccurate and efficient construction of geometries is crucial for Monte Carlo neutronics calculations. The Constructive Solid Geometry (CSG) descriptions generated by traditional Boundary Representation (B-Rep) to CSG conversion methods are complex and difficult to understand, and they are not always successful because of unstable Boolean operations. Since topologically similar B-Rep models tend to share CSG structures, leveraging geometry reuse can greatly improve modeling efficiency.PurposeThis study aims to develop an intelligent auxiliary modeling method that enables automatically convert B-Rep models into CSG geometry by reusing the CSG cell expressions, and is applied to fusion reactor similarity modules and seeks to reduce manual effort and improve stability while ensuring physical accuracy in subsequent neutronics calculations.MethodsFirstly, a graph-based geometry representation was proposed to transform each B-Rep model into an attributed adjacency graph (AAG) with topological features, and a Graph Attention Network (GAT) was trained to perform graph-level classification, enabling efficient retrieval and reuse of pre-constructed CSG cell expressions from a manually constructed geometry database. Meanwhile, a preprocessing pipeline based on the principal axes of inertia was introduced to ensure consistency across similar models, and standardize node ordering and spatial alignment. Then, the retrieved CSG expressions were transformed into the target geometry using rigid transformation and automatically integrated into Monte Carlo programs' input files via embedded TRCL cards. Finally,the method was tested on two representative blanket modules (#3 inner and #9 outer) and other two complex structures from the China Fusion Engineering Test Reactor (CFETR).ResultsThe average retrieval time is 0.405 ms, demonstrating superior time efficiency in CSG construction for the test cases, and the generated CSG geometry preserves structural consistency with the original Computer Aided Design (CAD) model. Neutronics simulation results showed a maximum deviation of less than 0.5% in the Tritium Breeding Ratio (TBR) compared to traditionally converted models. Additionally, the method demonstrated superior time efficiency in CSG construction for the test cases.ConclusionsThe proposed approach effectively reuses CSG expressions for topologically similar B-Rep models, while maintaining fidelity in Monte Carlo simulations, offering potential for similarity geometries' auxiliary modeling in nuclear fusion engineering applications.