Introduction
Simulating the internal environment of a pressurized water reactor (PWR) is a formidable multiphysics challenge. It requires modeling turbulent fluid flow, heat transfer between fuel rods and coolant, and neutronics—all of which are deeply coupled. Researchers at Sandia National Laboratories have developed a robust framework to address these complexities, focusing on improving predictive reliability and reducing the computational burden of uncertainty quantification (UQ).
The Problem: Complexity and Cost
Engineers face two primary hurdles in reactor modeling. First, the computational cost of high-fidelity simulations makes it nearly impossible to perform the thousands of runs required for traditional UQ. Second, distinguishing between physical phenomena (like localized hot spots) and numerical artifacts caused by mesh resolution is difficult. Standard "black-box" simulation tools often fail to provide the necessary sensitivity data to overcome these issues efficiently.
Methods: Stabilized Finite Elements and Integrated Adjoints
The proposed solution utilizes the Drekar::CFD framework, built on the Trilinos library. The core of the approach is a stabilized finite element (FE) method using Streamline-Upwind/Petrov-Galerkin (SUPG) and Pressure-Stabilized/Petrov-Galerkin (PSPG) formulations. These allow for stable, equal-order interpolation of velocity and pressure even in high Reynolds number turbulent flows.
The physical behavior is governed by the Reynolds-Averaged Navier-Stokes (RANS) equations, closed with the Spalart-Allmaras (SA) turbulence model.
A key technical innovation is the integration of Automatic Differentiation (AD) and Adjoint-Based Analysis. By solving a "dual" or adjoint version of the flow equations, the solver can simultaneously calculate the sensitivity of a specific Quantity of Interest (QoI)—such as peak temperature—to every input parameter. This integrated approach also facilitates rigorous numerical error estimation.
Validation and Key Insights
The framework was validated against several benchmarks, including laminar pipe and duct flows, and the CASL Test Case #2 (swirling turbulent flow).
- Accuracy: The solver matched analytical Nusselt numbers for laminar flow within 0.6% to 1.5%.
- Error Estimation: Adjoint-based error estimates showed an "effectivity ratio" near 1.01 in duct flow tests, indicating that the estimated error almost perfectly matches the true numerical error.
- Gradient-Enhanced UQ: The most significant impact was seen in UQ. By using adjoint-based gradients to train Gaussian Process surrogate models, the researchers reduced the average relative error from 29% (standard model) to just 10% for a 6-dimensional turbulent pipe flow study.
Conclusion
This integrated adjoint-based framework represents a significant step forward in reactor simulation. By embedding sensitivity and error analysis directly into the solver, it enables more accurate and efficient predictive modeling. While current work focuses on steady-state RANS in prototype geometries, the scalability of the underlying Newton-Krylov solvers (tested up to $10^5$ cores) paves the way for future applications in full-scale, unsteady reactor core simulations.
Summary updated