The NEA held a one-day workshop for Radiation Transport Simulation Developers (RTS 2024) on four cutting-edge topics: graphics processing unit (GPU) acceleration; code management, including extensive discussions on continuous testing and improvement; challenges in geometry descriptions and computer-aided design (CAD) modelling; and extended reality applications. Organised under the auspices of the NEA Expert Group on Physics of Reactor Systems (EGPRS), the event was hosted by the Italian National Institute for Nuclear Physics (INFN) at the National Laboratory of Frascati, Italy, and gathered 79 experts joining in person and online.
The discussion on GPU techniques contrasted the virtues and advances in current modelling and simulation (M&S) capabilities applying massive parallelisation techniques with opportunities to apply artificial intelligence (AI) and machine learning (ML) techniques to solve differential equations, for example based on convolutional neural networks. While M&S techniques are still superior in performance, AI techniques offer interesting opportunities to provide sensitivity information and thus might even add to the explainability of the simulations. The migration of simulation tools to GPUs is costly but designing an isolation of the numerical routines from the hardware by specific hardware libraries yields low future migration efforts when the GPU hardware evolves. The workforce with expertise in both nuclear physics and AI&ML is still sparse and international efforts are required to develop it.
Speakers during the code management-related session emphasised the need for strict version control and continuous automatic as well as manual testing cycles. Benchmarks from international benchmark databases like the NEA Shielding Integral Benchmark Archive and Database (SINBAD) and the NEA International Criticality Safety Benchmark Evaluation Project (ICSBEP) are widely integrated in the automatic testing. Code developers, however, identified a lack of opportunities to share simulation results and models associated with the existing benchmark databases. Participants highlighted that numerical benchmarks provide comprehensive and quick verification of daily changes, but international machine-readable databases do not exist yet. An international standardisation of output formats would also greatly simplify code intercomparisons. Data assimilation-driven error correction, for example with Bayesian techniques, provides the potential for improvements but requires high-quality machine-readable experimental benchmark information. It was noted that there is sufficient workforce in the nuclear sector with expertise in compiled and interpreted languages. However participants agreed that it is challenging to train students in writing efficient codes which combine sufficient accuracy and low resource requirements. Writing such codes typically require a broad knowledge of nuclear physics, the engineering requirements and computer science.
The presentations on challenges in geometry descriptions included a discussion related to the modelling of the International Thermonuclear Experimental Reactor (ITER), which has seen large progress in quality and complexity. Homogenisation techniques still do not yield sufficiently accurate results, and thus the strategy is to model with high complexity and little approximations, which requires extremely complicated geometry models with currently approximately 1.3 million surfaces in the associated Monte Carlo N-Particle (MCNP) model for the 360° geometry. The translation from the CAD model to the transport model geometry is not yet fully automatised, and configuration management remains challenging. It was noted that geometry converters translating geometry models between different code packages provide efficient ways to perform code-to-code intercomparisons. Examples of graphical user interfaces (GUIs) with capabilities to couple different codes exist, such as the FLUKA Advanced Interface (FLAIR), but an international platform for the development of comprehensive geometry converters does not yet exist. It was discussed that direct implementation of particle tracking in CAD geometries can be used as an alternative, though it has not yet been tested for complex models like ITER. The geometry modelling involved various stakeholders with different needs and perspectives (for example, component-based versus material-based modelling), and it was pointed out that the smooth integration of radiation transport simulations in the overall design process remains challenging. It was proposed to compartmentalise the CAD models and RTS geometry models to simplify configuration management. Speakers stressed that time-dependent modelling of deformations is limited by current constructive solid geometry (CSG) implementations and polygon mesh modelling provides opportunities. Participants also agreed that optimisations of geometry descriptions should be aligned with anticipated floating-point accuracies in future hardware architectures, which become partially optimised for AI&ML applications with single-precision accuracy.
A presentation on extended reality applications focused on providing information on dose rate fields and radioactive sources in the workers’ range of vision. A presentation of a Microsoft HoloLens-based application showed a proof-of-concept application for real-time 3D imaging of radiation sources and fields. Several groups participating in the workshop noted that they work on similar applications to support and optimise service work. The optimisation of the visualisation of radiation-related information in the holographic displays with respect to human factors was considered challenging.
The EGPRS will next discuss the feedback of the RTS 2024 pilot workshop, evaluate resulting recommendations, and consider implications on its future programme of work.
Photo: INFN