SAM Capability Enhancement for Thermal Mixing and Stratification in LMFR Transient Simulations

Authors:
  • A(Argonne Argonne National Laboratory (Anl)
    Argonne Argonne National Laboratory (Anl)
Open Access
Publisher website
Google Scholar
Abstract
Accurate and computationally efficient representation of key thermal-fluid phenomena is essential for improving the transient simulation capabilities of liquid-metal fast reactors (LMFRs). This report addresses an important modeling gap associated with thermal mixing and stratification in large plena through the development and integration of advanced data-driven turbulence models in the System Analysis Module (SAM). The proposed approach retains the conventional numerical solution of the mass, momentum, and energy equations while using neural networks to provide the turbulent-viscosity closure required by the multidimensional fluid solver. Artificial intelligence (AI) and machine learning (ML) thereby offer a means to improve the fidelity of reactor system simulations while retaining the computational efficiency required for engineering-scale transient analysis. Two neural network architectures, Convolutional Neural Networks (CNNs) and transformers, are investigated and compared using CFD-generated training data. The results demonstrate that the data-driven model can reproduce the turbulent-viscosity field with substantially improved accuracy compared with the simplified Prandtl mixing-length model currently used in SAM. The transformer architecture provides further improvement over CNNs, achieving a 32% lower RMSE than the CNN for the same benchmark problem. To address the computational and memory requirements associated with the quadratic scaling of conventional self-attention, a compressed-attention transformer is subsequently developed. The approach compresses the full mesh representation into a fixed number of latent tokens, reducing the attention complexity from O(N 2 ) to O(NL + L 2 ), where L ≪ N. The resulting model enables application to mesh sizes that are compatible with the typical requirements of coarse-mesh reactor simulations. Finally, the AI-driven turbulence model is integrated into a system-level simulation of an advanced sodium fast reactor. For comparison, a protected loss-of-flow transient is simulated using three different modeling approaches for the sodium hot pool: a 0D perfectly mixed volume, a 1D mixing and stratification model, and a 2D axisymmetric multi-dimensional model with AI-driven turbulence closure. The results demonstrate that the spatial distribution of temperature within the hot pool can produce non-negligible differences in natural circulation and reactor temperatures, highlighting the need for detailed simulation capabilities for thermal mixing and stratification, such as those developed in this work. Overall, the work demonstrates the feasibility of integrating AI-based physical models into SAM and combining system-level and multi-dimensional simulations to enable AI-enhanced multiscale modeling of nuclear reactor systems.
Funding Information
  • USDOE Office of Nuclear Energy (NE), Nuclear Energy Advanced Modeling and Simulation (NEAMS) (AC02-06CH11357)