Over the years, enhancing the accuracy of the gradient operator, which is regarded as a key component of differential operators, has remained a fundamental research objective in mathematics, science, and engineering. For instance, it finds wide-rangin...
Over the years, enhancing the accuracy of the gradient operator, which is regarded as a key component of differential operators, has remained a fundamental research objective in mathematics, science, and engineering. For instance, it finds wide-ranging applications in research areas such as inverse problems, optimal design, Partial Differential Equation (PDE) analysis, regression & interpolation, signal & image processing, manifold analysis, and Artificial Intelligence (AI). In particular, the Least Squares Method (LSM), a representative approach for gradient estimation, has long been utilized in the field of computational fluid dynamics (CFD). In recent years, its utility has extended beyond gradient reconstruction to the LSM-based spatial discretization in meshless method.
However, in terms of accuracy, LSM still faces limitations in its application to spatial discretization and gradient reconstruction, particularly for boundary layer problems involving complex geometries with high Aspect Ratios (AR), where it becomes difficult to accurately estimate the gradients of physical quantities. In addition, when combining upwind schemes for compressible flow analysis with LSM-based gradient operators for spatial discretization, issues related to robustness and stability are observed in the local point clouds of specific geometric configurations. These problems can lead to inaccurate prediction of physical quantities or even numerical divergence in localized regions, which ultimately propagate and deteriorate the accuracy of the global numerical solution. Finally, in terms of efficiency, high computational cost of matrix inversion, which is an essential procedure of LSM-based gradient operators for spatial discretization, remains an unresolved challenge. Therefore, the successful implementation of meshless framework in CFD fundamentally requires an accurate LSM-based gradient operator, numerical strategies for robustness and stability, and computationally efficient algorithms for fast inversion.
First, mathematical correlation between numerical accuracy of LSM-based gradient operator and geometric characteristics of two-dimensional local point clouds, including AR, curvature, and skewness, is theoretically analyzed. Based on this theoretical analysis, two gradient correction approaches are proposed to reduce numerical errors in gradient estimation across multi-dimensional spaces: (1) a multi-stage optimization procedure using the new LSM based on the principle of superposition, and (2) a Lagrange Multiplier (LM)-based LSM for improved gradient estimation in skewed point clouds. To ensure robust operation of the newly developed method-based algorithms, limiting strategies are introduced from the perspectives of boundedness and condition number. The new
gradient operator is evaluated using various test functions, local point clouds, and varying levels of point connectivity in accuracy studies. This operator is evaluated by comparison with conventional approaches, including LSM, Geometric Conservation Least Squares Method (GC-LSM), Green-Gauss theorem (GG), and Green-Gauss theorem with centroidal values using volume-weighted averaging (SGG). In addition, numerical simulations are conducted to investigate the influence of gradient operators on spatial discretization and gradient reconstruction in high-order flux schemes. Furthermore, the time convergence characteristics of the proposed operator are also examined.
Second, the robustness and stability issues arising from the combination LSM-based spatial discretization in meshless method and upwind scheme under specific local point cloud configurations are analyzed in one-dimensional space to gain numerical and physical insights. In addition, because numerical results from the two-dimensional blast wave problem reveal that the findings derived from one-dimensional analysis are insufficient to fully capture the complex behavior observed in multidimensional problems, the one-dimensional analysis was extended to multi-dimensional space, leading to the derivation of new mathematical conditions: one to ensure robustness, and another to guarantee conditional stability based on von Neumann stability analysis. To satisfy these conditions, a new algorithm is developed in this work to generate aligned meshless coefficients. Furthermore, in order to mitigate the potential loss of numerical accuracy resulting from enforcing robustness and stability, an additional indicator is introduced. These proposed methods are validated on point clouds including sharp features, demonstrating their effectiveness in terms of robustness and stability.
Third, a fast matrix inversion algorithm is developed to alleviate the high computational cost associated with local approximation in LSM-based gradient operators with LM for multi-dimensional space. To reduce the time complexity of matrix inversion in the LM-based LSM, tensor product techniques are employed to decompose high-dimensional block matrices into lower-dimensional sub-blocks and repetitive patterns. Singular Value Decomposition (SVD) is integrated into the algorithm to further improve efficiency in underdetermined systems, and block matrix inversion is adopted to enable the complete inversion process using only low-scale computations. The performance of the proposed algorithm is quantitatively evaluated against the conventional LU decomposition benchmark, demonstrating a significant improvement in computational efficiency.
All proposed methods are integrated into an in-house CFD solver implemented in Fortran 90. To evaluate the meshless framework’s applicability across a broad range of validation benchmarks, the following validation cases are considered: (1) inviscid flow (e.g., Sod shock tube), (2) viscous flow (e.g., shock wave boundary layer interaction), (3) turbulent flows (e.g., RAE2822 transonic airfoil, ONERA M6 wing), (4) hypersonic flows (e.g., frozen flow over a circular cylinder, non-equilibrium flow over a Hypersonic Glide Vehicle (HGV)), and (5) supersonic turbulent combustion flows (e.g., Burrows and Kurkov supersonic mixing/combustion). For performance benchmarking, FVM and GC-LSM are employed as reference spatial discretization approaches. The numerical results demonstrate that the developed meshless framework achieves significant improvements in accuracy, robustness, stability, and computational efficiency across a wide range of flow regimes.