Mesh-Free Algorithmic Framework

• Mesh-free algorithmic frameworks are numerical methods that solve partial differential equations using scattered nodes instead of traditional meshes.
• They employ local stencils, high-order polynomial reproductions, and neural operator learning to achieve accuracy and adaptivity in complex simulations.
• These frameworks are vital for modeling evolving geometries and fluid dynamics, overcoming limitations of conventional mesh-based methods.

A mesh-free algorithmic framework is a numerical paradigm for solving partial differential equations (PDEs), dynamical systems, and inverse problems without reliance on an underlying mesh or grid. Rather than discretizing the domain via elements or control volumes, these frameworks operate directly on scattered nodes or points, leveraging local stencils, basis expansions, or neural representations. Mesh-free frameworks facilitate computation on irregular domains, evolving geometries, and situations where local adaptivity or flexibility is essential. This article surveys the principles, methodologies, and representative classes of mesh-free algorithmic frameworks, encompassing meshless interpolation/collocation, high-order mesh-free differential operators, Lagrangian and moving-boundary schemes, mesh-free operator learning, and recent advances in deep learning–driven mesh-free solvers.

1. Principles of Mesh-Free Discretization

The core principle of mesh-free discretization is the construction of local or global numerical approximations for differential operators, field values, or observables, using function values at scattered nodes. Two foundational strategies dominate this landscape:

• Radial Basis Function (RBF) Interpolation and Collocation: Functions are interpolated as weighted sums of radial kernels (e.g., polyharmonic splines, Gaussians), possibly augmented with polynomials for high-order consistency. The interpolation coefficients are determined by enforcing exact reproduction at nodes, and differential operator weights are constructed via analytical differentiation of the kernel expansion. This yields exponentially convergent discretizations when appropriate kernels and polynomial degrees are selected (Shahane et al., 2021, Merten, 2014).
• Moving Least Squares, Generalized Finite Differences, and Local Basis Expansions: High-order mesh-free differential operators are obtained by reproducing polynomials up to degree $p$ on a stencil of neighbor nodes. The weights enforce exact action of the operator on all basis monomials, typically via a local system or moment matrix (Feng et al., 26 Nov 2025, King, 2023). Anisotropic or Hermite basis functions may be used for enhanced accuracy and adaptivity.

Table: Core Mesh-Free Discretization Approaches

Approach	Key Feature	Representative Frameworks
RBF Collocation (and RBF-FD)	Global/local kernel interpolation	PHS-RBF, Gaussian RBF, LABFM
Moving Least Squares / GFDM	Polynomial reproduction by stencils	CPHSF, LABFM, mesh-free GFDM
Method of Fundamental Solutions	Boundary-only meshless expansion	ALE-MFS

Both approaches are compatible with unstructured node layouts and complex, evolving geometries, distinguishing them from classical mesh-based methods.

2. High-Order Mesh-Free Operators and Adaptivity

Recent developments have extended mesh-free discretizations to high-order accuracy and spatial adaptivity:

• High-Order Anisotropic Basis Function Methods (LABFM): LABFM locally reconstructs strong-form derivatives by expanding the solution in anisotropic Hermite polynomials modulated by compactly supported radial envelopes. Operator weights ensure exactness on all monomials up to degree $p$, and stability/convergence properties are rigorously established (Feng et al., 26 Nov 2025, King, 2023, King, 2023). The method scales efficiently to high polynomial degrees and supports arbitrary node distributions.
• p-Adaptive Mesh-Free Methods: Polynomial order $p$ used in the local approximation is dynamically adjusted at each node based on a posteriori error indicators (e.g., discrepancies in the Laplacian evaluated at adjacent orders). Threshold criteria determine local refinement or de-refinement. This reduces computational cost while sustaining accuracy, as demonstrated in simulations of fluid/combustion problems in complex geometries, which report up to 50% cost savings at fixed error (Feng et al., 26 Nov 2025).

3. Time Integration and Semi-Implicit Solvers

Temporal discretization in mesh-free frameworks must be compatible with high-order and locally adaptive spatial stencils.

• Hermite Integrators: For Lagrangian particle-based mesh-free methods, fourth-order Hermite schemes allow the same high spatial and temporal order while reducing the number of costly neighbor interactions compared to Runge–Kutta. The integration of acceleration and higher derivatives (jerk, snap) is performed by consistent local polynomial fits. Hermite integrators have demonstrated efficiency and stability for CFL numbers surpassing those of RK schemes, provided the spatial order matches or exceeds the temporal order (Yamamoto et al., 2018).
• Semi-Implicit Pressure-Correction (Projection) Schemes: In the context of incompressible flow, mesh-free frameworks often employ backward-difference time integration (BDF2) for momentum and semi-implicit splitting to enforce divergence-free velocity via a pressure-Poisson solve. Local mesh-free operators are used for all necessary derivatives, and iterative corrector-predictor loops enforce convergence at each time step (Shahane et al., 2021, King, 2023, Suchde, 2020).

4. Mesh-Free Methods for Evolving and Complex Geometries

Mesh-free frameworks excel in handling moving boundaries, large deformations, and topological changes, for which mesh-based approaches frequently struggle.

• Lagrangian Mesh-Free Motion on Surfaces: Fully Lagrangian frameworks represent evolving surfaces or curves by cloud points advected according to the flow or geometric laws. Node distributions are adaptively maintained through insertion/deletion algorithms, supporting point-cloud surface representations free of intrinsic meshing (Suchde et al., 2019, Suchde, 2020).
• Arbitrary Lagrangian-Eulerian (ALE) Mesh-Free Motion: Meshless method-of-fundamental-solutions (ALE-MFS) extends velocity boundary motion harmonically (by solving Laplace's equation with boundary velocities) to the interior in a mesh-free manner, using collocated fundamental solutions. Boundary normals and curvature are extracted adaptively by local B-spline fitting, and high-order accuracy is sustained even for strongly nonconvex or evolving shapes (Ammad et al., 16 Jan 2026).
• Handling Topological Changes & Collisions: Mesh-free Lagrangian frameworks implement collision detection, signed-distance fitting, and node merging/splitting strategies to address interface pinching and merging. These processes are essential for simulating flows and PDEs on manifolds undergoing topology changes (Suchde et al., 2019).

5. Mesh-Free Operator Learning and Neural Methods

A rapidly advancing area involves learning mesh-free operators or discovering governing equations from data, often by leveraging neural architectures:

• Neural Mesh-Free Operators: Graph neural networks trained via polynomial moment constraints can directly predict stencil weights for mesh-free differentiation. The learned operators display polynomial consistency, resolution- and geometry-agnostic performances, and can be deployed as drop-in replacements in Lagrangian or Eulerian mesh-free solvers (Starepravo et al., 25 Mar 2026).
• Mesh-Free Neural Operators (e.g., BelNet): Projection/construction networks (operator learning frameworks) learn both the input-to-latent "coefficient" mapping and the continuous output "basis" functions via neural architectures, thus realizing discretization- and domain-invariant mappings from arbitrary function samples to arbitrary prediction points. These frameworks, such as BelNet, outperform conventional operator learners for high-contrast and multiscale PDE problems (Zhang et al., 2022).
• Mesh-Free PDE Discovery (Sparse Identification): Algorithms such as mesh-free SINDy leverage neural-network surrogates and automatic differentiation to identify governing PDEs from scattered, irregular, and noisy data. The process separates regression and surrogate modeling stages, is highly robust, and avoids mesh entanglement, achieving rapid and accurate equation extraction across canonical examples (Gao et al., 21 May 2025).
• Deep Mesh-Free Variational and PINN Approaches: Physics-informed networks and mesh-free eigenanalysis frameworks (e.g. Eig-PIELM) solve eigenvalue, optimization, or variational-inequality problems on arbitrary domains by enforcing operators and boundary conditions at collocation points, sometimes utilizing proximal operator reformulations or ADMM splitting for non-smooth or constrained problems (Mishra et al., 21 Aug 2025, Gao et al., 20 May 2025, Song et al., 2023).

6. Applications and Empirical Performance

Mesh-free algorithmic frameworks have been extensively validated across a spectrum of complex applications:

• Fluid Dynamics in Complex Domains: High-accuracy mesh-free discretizations (LABFM, PHS-RBF, GFDM) have been demonstrated for steady and unsteady incompressible and compressible flows, including high-fidelity simulation of flames, turbulence, interface-driven motion, and flows on evolving surfaces. Empirical results confirm exponentially convergent rates for sufficiently smooth solutions and exponential synchronization in mesh-free data assimilation schemes (Shahane et al., 2021, King, 2023, Suchde et al., 2019, Suchde, 2020, Biswas et al., 2021).
• Moving and Evolving Interfaces: ALE-MFS enables robust mesh-free ALE motion in domains exhibiting large deformations and nonconvexity, outperforming classic mesh-based FEM strategies in mesh quality and regularity preservation (Ammad et al., 16 Jan 2026). Mesh-free Lagrangian frameworks seamlessly handle surface PDEs with contact, collision, and topology change (Suchde et al., 2019).
• Operator Learning and Inverse Problems: Mesh-free neural operators and physics-informed neural networks provide state-of-the-art performance in operator learning, PDE-constrained optimization, eigen-analysis, and secondary tasks such as shape optimization and inverse obstacle scattering, leveraging mesh-invariance and the flexibility of neural representations (Zhang et al., 2022, Vlašić et al., 2022, Gao et al., 20 May 2025, Mishra et al., 21 Aug 2025).
• Computational Efficiency and Parallelization: Advanced mesh-free frameworks deploy local adaptivity (p-adaptivity), efficient neighbor searching, and hybrid parallelization (MPI/OpenMP), scaling to millions of nodes and outperforming classical mesh-based approaches for complex 2D/3D domains (King, 2023).

7. Advantages, Challenges, and Outlook

Mesh-free algorithmic frameworks provide numerically robust, flexible, and adaptively accurate tools for a wide range of computational problems—especially those involving complex, evolving, or poorly parameterizable geometries. Key advantages include:

• No need for explicit mesh generation or maintenance, simplifying pre-processing and obviating mesh-induced errors.
• Direct compatibility with irregular, time-dependent, or topologically evolving domains.
• Efficient high-order accuracy via local polynomial reproduction and adaptivity.
• Seamless integration with data-driven, neural, or hybrid learning-physics solvers.

Challenges and open issues persist:

• Ensuring conservation properties (e.g., mass, energy) requires careful algorithmic design, especially for strong-form solvers.
• Large stencil sizes and moment-matrix solves can incur high per-stencil computational cost, though learning-based approaches (NeMDO) and adaptivity mitigate this (Starepravo et al., 25 Mar 2026, Feng et al., 26 Nov 2025).
• Parallel load-balancing is more complex in locally adaptive frameworks.
• Extension to truly large-scale or high-frequency problems (e.g., eigenvalue spectra) demands further research on numerical efficiency and robustness (Mishra et al., 21 Aug 2025).

Future directions include fully three-dimensional mesh-free adaptivity, tighter integration of learned and analytic operators, robust mesh-free model discovery from experimental or field data, and broader deployment of mesh-free frameworks in multi-physics and multi-scale applications. The synthesis of high-order accuracy, geometric flexibility, and data-driven adaptivity positions mesh-free algorithmic frameworks as a central tool in advanced scientific computing.