Abstract
In this paper, we introduce the isoGeometric Residual Minimization (iGRM) method for solving stationary advection-dominated diffusion problems. The method achieves stability via residual minimization, discretizing the problem using B-spline basis functions on a tensor-product mesh. We seek to minimize the isogeometric residual over a spline space, constructing the solution over a smooth subspace. The Gramm matrix for the residual minimization is approximated by a weighted H1 norm, expressed as Kronecker products to exploit the mesh structure. This allows for a direction-splitting preconditioner with a computational cost proportional to the number of degrees of freedom in 2D and 3D. Numerical tests against Discontinuous Petrov-Galerkin (DPG) and Streamline Upwind Petrov-Galerkin (SUPG) methods demonstrate that iGRM delivers high-quality solutions on smaller grids without requiring problem-specific stabilization parameters.
Introduction
Numerical simulations of fluid flow frequently encounter advection-dominated diffusion scenarios, where a substance or heat is transported primarily by flow (advection) rather than spreading naturally (diffusion). Standard numerical approaches, such as the Galerkin method, often fail in these regimes, producing non-physical oscillations or "wiggles" near sharp gradients and boundary layers unless the mesh is excessively fine.
Existing solutions like the Streamline Upwind Petrov-Galerkin (SUPG) method require carefully tuned, problem-specific stabilization parameters to maintain stability. Conversely, advanced frameworks like the Discontinuous Petrov-Galerkin (DPG) method offer excellent stability but come with significant computational overhead. The iGRM method aims to bridge this gap by providing a framework that is as stable as DPG but computationally efficient by exploiting the mathematical properties of B-splines.
The iGRM Framework
The iGRM method is built upon three primary pillars:
- Isogeometric Analysis (IGA): By utilizing B-spline basis functions rather than standard finite elements, the method benefits from higher continuity (smoothness) across element boundaries. This typically results in higher accuracy per degree of freedom.
- Residual Minimization: Instead of solving the standard weak form of the equations, iGRM minimizes the residual in a specific dual norm. Stability is guaranteed by selecting a residual space that is slightly larger or more flexible than the solution space (e.g., by increasing polynomial order or continuity).
- Direction Splitting Preconditioner: To ensure speed, the authors exploit the fact that B-splines are defined on tensor-product meshes. The large system matrices are approximated using Kronecker products, allowing the decomposition of complex 2D or 3D problems into a series of simple 1D operations. This results in an iterative solver with linear computational cost ($O(N)$).
Numerical Validation
The method was evaluated using three classic numerical benchmarks to test convergence, stability, and handling of complex patterns:
- Manufactured Solution: Used to verify convergence rates, iGRM achieved less than 1% error on a $32 \times 32$ mesh, whereas the SUPG method required a $64 \times 64$ mesh to reach comparable accuracy.
- Eriksson-Johnson Model: This test involves a sharp boundary layer. iGRM successfully captured the steep gradients with minimal overshoot or undershoot compared to standard methods.
- Vortical Wind Problem: This simulation of rotating flow demonstrated the method's robustness in handling complex advection patterns.
Key Insights
A central finding of this research is that the tensor-product structure of IGA is a powerful computational shortcut. By approximating the Gramm matrix as a Kronecker product, the authors transform a massive matrix inversion into a sequence of fast operations. Furthermore, because iGRM is inherently stable due to the residual minimization framework, it eliminates the need for the "magic" stabilization parameters required by SUPG, which are often difficult to calculate for complex flows.
Limitations and Future Work
While highly efficient, the current iteration of iGRM is limited to tensor-product meshes (rectangular or box-shaped domains). Extending the method to complex, unstructured geometries remains a primary challenge. Additionally, while the cost per iteration is linear, the total number of iterations required for the Conjugate Gradient solver grows with problem size. Future research will focus on developing improved preconditioners for the Schur complement and extending the iGRM framework to non-stationary (time-dependent) problems and other physics equations, such as the Stokes and Maxwell equations.
Summary updated