NeurFrame: Learning Continuous Frame Fields for Structured Mesh Generation
Abstract: Structured meshes, composed of quadrilateral elements in 2D and hexahedral elements in 3D, are widely used in industrial applications and engineering simulations due to their regularity and superior accuracy in finite element analysis. Generating high-quality structured meshes, however, remains challenging, especially for complex geometries and singularities. Field-guided approaches, which construct cross fields in 2D and frame fields in 3D to encode element orientation, are promising but are typically defined on discrete meshes, limiting continuity and computational efficiency. To address these challenges, we introduce \emph{NeurFrame}, a neural framework that represents frame fields continuously over the domain, supporting infinite-resolution evaluation. Trained in a self-supervised manner on discrete mesh samples, NeurFrame produces smooth, high-quality frame fields without relying on dense tetrahedral discretizations. The resulting fields simultaneously guide high-quality quadrilateral surface meshes and hexahedral volumetric meshes, with fewer and better-distributed singularities. By using a single network, NeurFrame also achieves lower computational cost compared to prior self-supervised neural methods that jointly optimize multiple fields.
Summary
- The paper introduces a continuous neural function using spherical harmonics to encode frame fields for structured mesh generation.
- It employs a self-supervised loss combining smoothness, boundary, and feature alignment to ensure high-quality meshing with fewer singularities.
- Experimental results demonstrate improved mesh regularity, enhanced surface conformity, and reduced computational cost compared to traditional methods.
NeurFrame: Self-Supervised Learning of Continuous Frame Fields for Structured Mesh Generation
Introduction
The generation of structured meshes—specifically, quadrilateral meshes in 2D and hexahedral meshes in 3D—is central to engineering analysis, geometric modeling, and simulations that require high accuracy and regularity in discretization. A persistent bottleneck in the field is the construction of smooth, feature-aligned frame fields that successfully guide the production of high-quality structured meshes on complex geometries, often containing singularities. Traditional field-guided approaches are mostly formulated on discrete domains, which couples field fidelity to the density of the underlying tetrahedral mesh, often resulting in a computational trade-off between efficiency and quality. The paper "NeurFrame: Learning Continuous Frame Fields for Structured Mesh Generation" (2603.12820) introduces a paradigm shift by representing the frame field as a continuous neural function, thereby supporting infinite-resolution evaluation and efficient optimization on coarse discretizations while circumventing the limitations and inefficiencies of previous discrete and neural methods.
Methodology
Continuous Neural Representation with Spherical Harmonics
NeurFrame parameterizes the frame field using a multilayer perceptron (MLP) that maps any 3D coordinate in the domain to a vector of spherical harmonic (SH) coefficients. SHs provide a rotation-aware basis in which each frame—three mutually orthogonal vectors—can be encoded as a $9$-dimensional coefficient vector with respect to bands $0$ and $4$.
Figure 1: Spherical harmonic (SH) representation of a frame, with basis projections highlighting the correspondence between spatial axes and SH minima.
This continuous functional representation uniquely empowers NeurFrame to evaluate frame direction at arbitrary spatial locations, decoupling field quality from input mesh discretization. The SH encoding naturally supports rotation-invariant comparison and smooth interpolation, facilitating the definition of loss functions for optimization.
Self-Supervised Loss Formulation
The training of NeurFrame is entirely self-supervised and proceeds on sparse volumetric samples. The principal loss components are:
- Smoothness: A discretized Dirichlet energy is computed over dual edges connecting neighboring tetrahedral centroids, weighted by edge length. This enforces globally smooth frame transitions and precludes the formation of spurious, high-frequency singularities.
- Boundary Alignment: The network aligns one principal axis of the frame at each boundary tetrahedron centroid to the outward surface normal by leveraging SH rotation properties. This is instrumental for surface conformity and regular parameterization.
- Feature Alignment: Explicit alignment along sharp feature curves is implemented through a locally decaying weight, which confines feature adherence to neighborhoods of geometric saliency to avoid over-constraining the field.
These terms are linearly combined with tunable weights, making the formalism adaptable to prioritization of global smoothness, boundary shape fidelity, or internal feature propagation.
Experimental Results
Quadrilateral and Hexahedral Mesh Quality
NeurFrame demonstrates state-of-the-art performance in both 2D and 3D structured meshing tasks. The neural frame field is used to derive a cross field for quadrilateral meshing on surfaces via mixed-integer quadrangulation and a parameterization-driven pipeline for hexahedralization in the bulk. Comparative experiments reveal the following:
- Mesh Regularity and Singularity Distribution: NeurFrame yields meshes with markedly fewer and more evenly distributed singularities, improving both numerical properties for simulation and aesthetic mesh alignment.
Figure 2: The NeurFrame pipeline progressively smooths an initial volumetric frame field, achieving alignment with surface normals and features, guiding the generation of high-quality structured meshes.
- Feature Alignment and Surface Fidelity: Even without explicit feature constraints, the cubic symmetry of the frame field coupled with the boundary alignment substantially enhances the mesh's adherence to complex feature lines and principal curvature directions.
(Figure 3)
Figure 3: Comparison with competing field-guided quadrilateral meshing methods, highlighting singularity count and alignment with geometric features.
Frame Field Quality and Hexahedralization
NeurFrame's continuous SH-encoded field produces highly smooth, well-aligned volumetric parameterizations, resulting in fewer degenerate hexahedral elements:
(Figure 4)
Figure 4: Frame field and singularity graph comparison with prior methods. NeurFrame's field achieves superior smoothness and geometric awareness.
Notably, in benchmarks against the recent Neural Octahedral Field (NeurOcta) [Zheng-2025] and traditional discrete methods such as [Ray-2016], NeurFrame delivers better boundary conformity, avoids noisy topologies, and leads to successful hexahedralization in challenging scenarios:
(Figure 5)
Figure 5: On a complex tetrahedral input, NeurFrame maintains boundary alignment and mesh validity; NeurOcta fails to produce valid hexahedral elements.
(Figure 6)
Figure 6: Hexahedral mesh, cross-section, and singularity structures for various field generation methods. NeurFrame outperforms on mesh quality and topological organization.
Computational Efficiency
NeurFrame achieves a per-iteration runtime reduction compared to NeurCross [Dong-2025b] and NeurOcta, attributed to its single-field approach and architectural efficiency, as evidenced by reduced parameter counts and faster convergence.
Ablation and Limitations
Systematic ablation studies confirm that the chosen weightings for smoothness, boundary, and feature alignment yield optimal mesh quality. Reducing boundary alignment, for instance, increases singularities, while excessive feature weight can overfit local structures (Figure 7).
Nevertheless, there remain topological and algorithmic limits: some models with non-globally meshable singularity graphs still thwart structured hex-mesh extraction even with a smooth and simple field. Theoretical guarantees for global meshability under frame field optimization remain elusive.
Implications and Outlook
NeurFrame sets a new standard in neural field-guided structured meshing, combining infinite-resolution evaluation with practical computational efficiency and broad applicability. The invariance and analyticity of the SH representation, coupled with continuous field learning, inform both future neural geometric processing pipelines and theoretical investigations into global meshability.
Pragmatically, NeurFrame's mesh outputs have immediate benefits for finite element analysis, computer graphics, and scientific computing, where regularity, low-valence singularities, and alignment to geometric features are critical. Its continuous parametrization framework paves the way for tightly coupled shape optimization and simulation-in-the-loop strategies where mesh adaptivity and field smoothness must co-evolve.
Remaining open challenges include the integration of topological correction (e.g., locally meshable singularity graphs [Liu-2023]) within the learning loop and prescribing field singularities directly via explicit constraints or generative priors, which would further enhance the automation and reliability of structured meshing in complex volumes.
Conclusion
NeurFrame introduces a continuous, self-supervised neural framework for frame field learning and structured mesh generation. By leveraging a compact SH encoding and an efficient MLP, it enables smooth, boundary- and feature-aligned fields that support high-quality quadrilateral and hexahedral mesh extraction with low computational cost. The approach surmounts discretization-bound limitations of prior field-guided methods and neural joint-optimization pipelines. Future work will address the theoretical guarantees of global meshability and integrate topological correction directly into the learning process, with implications for robust automation of high-quality mesh generation in simulation science and geometric modeling.
References
- "NeurFrame: Learning Continuous Frame Fields for Structured Mesh Generation" (2603.12820)
Paper Prompts
Sign up for free to create and run prompts on this paper using GPT-5.
Top Community Prompts
Explain it Like I'm 14
What is this paper about?
This paper is about making neat, grid‑like “digital Lego” for shapes on a computer. In 2D, that grid is made of squares; in 3D, it’s made of cubes. Engineers and scientists use these grids (called “structured meshes”) to simulate things like airflow around a car or stress inside a bridge. The problem: fitting nice squares/cubes onto a curved, detailed shape is hard.
The authors introduce NeurFrame, a new way to tell the computer how to place and orient those squares/cubes smoothly across a shape. Instead of deciding directions only at a few fixed points, NeurFrame uses a neural network to describe directions everywhere, continuously, like a smooth “map of tiny compasses.”
What questions did the researchers ask?
They focused on three simple questions:
- Can we describe the “direction map” across a 3D shape as a smooth, continuous function instead of only on a coarse grid?
- Can this map line up with the shape’s surface and sharp edges so squares/cubes follow features nicely?
- Can we do this faster and with fewer mistakes (“singularities,” places where the grid pattern must change) than previous methods?
How did they do it?
Think of covering a statue with perfectly aligned tiles (2D) or stacking cubes (3D) so they follow curves and edges. To do this well, you need, at every point, a small “frame”—three tiny arrows pointing along the best local directions (like the x, y, z axes of a tiny coordinate system). The collection of all these tiny frames across the shape is called a “frame field.”
Here’s the approach in everyday terms:
- A neural network acts like a smart rulebook. You give it a location in space, and it returns a compact 9‑number “code” that represents the three tiny arrows (the local directions). This 9‑number code is a math trick (called spherical harmonics) that stores orientations in a stable, rotation‑friendly way. You can think of it as a tidy ID tag for the local directions.
- The network is trained without ground‑truth examples (self‑supervised). Instead, it follows three simple rules:
- Smoothness: neighboring points should have similar directions, like combing hair so it lies flat and even.
- Boundary alignment: on the surface, one tiny arrow should match the surface normal (the “straight out” direction), so tiles/cubes sit properly on the outer skin.
- Feature alignment: near sharp edges or creases, one arrow should follow the edge, so grid lines trace the feature instead of cutting across it.
- After training, the network can be asked for the directions at any point—on the surface (to guide square grids) or inside the volume (to guide cube grids). Standard meshing tools then turn this direction map into actual quadrilateral (squares) or hexahedral (cubes) meshes.
Why this is helpful:
- Continuous: Because the network describes directions everywhere, you can evaluate it at “infinite resolution” without relying on super‑dense 3D subdivisions.
- Compact and efficient: A single neural network handles the whole job, instead of juggling multiple models.
What did they find, and why does it matter?
Their main results show clear improvements:
- Smoother, cleaner direction maps with fewer and better‑placed singularities (special points where the grid must bend or split). Fewer singularities usually means neater, more regular meshes.
- Good alignment with surfaces and sharp features, so squares and cubes follow shape details instead of fighting them.
- Works for both 2D quadrilateral meshes and 3D hexahedral meshes using the same continuous field.
- Faster training than other neural methods that try to learn similar fields, while using a smaller network.
- Less dependence on very fine 3D subdivisions, which saves time and memory.
Why this matters:
- Better meshes mean more accurate and reliable simulations (like predicting how something will deform or how fluids will flow).
- Cleaner grids are also helpful for texture mapping, modeling, and manufacturing (e.g., 3D printing) where regular structure is beneficial.
- Lower computational cost means this can be run on more models, more quickly.
What does this mean for the future?
NeurFrame shows that teaching a neural network to produce a smooth, continuous “direction map” can make high‑quality square and cube meshes easier and faster to generate. This can help engineers and artists create better models and run better simulations.
There’s still a challenge: sometimes, even with a nice direction map, building a perfect cube mesh everywhere fails because the special points and lines (the singularity graph) don’t match the strict global rules cubes must follow. The authors plan to teach the network to aim for singularity patterns that are not just clean but also guaranteed to be “meshable,” so valid cube meshes can be produced more reliably.
Knowledge Gaps
Knowledge gaps, limitations, and open questions
Below is a focused list of what remains missing, uncertain, or unexplored in the paper, phrased to guide actionable future research.
- Lack of theoretical guarantees that the learned continuous SH coefficients correspond to valid frames everywhere (normalizing q to unit norm is only a necessary, not sufficient, condition); no constraint ensuring validity throughout training rather than only via post-hoc projection.
- No analysis of continuity after projecting SH coefficients to a rotation matrix; the projection step could introduce discontinuities or “snapping” that affect IGM construction and mesh quality.
- Absence of integrability/meshability constraints during training (e.g., local/global hex-meshability conditions per Liu et al.); the method can still produce non-meshable singularity graphs and degenerate IGMs.
- No mechanism to control or penalize singularity creation explicitly; the method relies on smoothness and alignment only, with no topological regularization to steer singularity count, type, or placement.
- No formal analysis of singularity stability with respect to initialization, discretization resolution, or model orientation; reproducibility across random seeds and rotations is unreported.
- The MLP is not rotation- or symmetry-equivariant; robustness of the learned field to arbitrary global rotations or reflections of the input geometry is neither enforced nor evaluated.
- Smoothness uses an isotropic Euclidean Dirichlet term on dual edges; there is no metric-aware or anisotropic smoothness (e.g., conformal or application-driven metrics) that could better reflect geometric or physical anisotropy.
- The boundary alignment enforces only normal alignment; the in-plane (tangential) degree of freedom is left unconstrained unless feature edges exist, leaving surface-aligned directions underdetermined on smooth regions.
- Feature alignment relies on externally detected feature edges and a fixed exponential distance decay; sensitivity to feature detection errors, parameter σ, and model scale is not analyzed, and geodesic vs Euclidean distance is not explored.
- Training constraints are applied at tetrahedral centroids and dual edges only; there is no study of how coarse discretization and sparse sampling affect off-mesh behavior of the continuous field (potential high-frequency artifacts between samples).
- The claim of “avoiding dense tetrahedral discretizations” still requires a volumetric tetrahedralization; dependence on tet quality (slivers, anisotropy) and on target edge length for accuracy and convergence is not quantified.
- No adaptive sampling strategy is used to concentrate supervision near high-curvature regions, features, or evolving singularities; potential benefits of active sampling remain unexplored.
- No joint optimization with parameterization (MIQ/CubeCover) or a differentiable IGM; the decoupled pipeline precludes end-to-end objectives that directly optimize for low-distortion parameterizations or mesh quality.
- No quantitative evaluation of field smoothness (e.g., integrated Dirichlet energy), boundary alignment error distributions, or cross-field integrability on surfaces; most evidence is qualitative.
- Hex-mesh quality metrics (e.g., scaled Jacobian, orthogonality, skewness, element inversion rates) and success rates across a dataset are not reported; it is unclear how field improvements translate to mesh quality at scale.
- Limited runtime reporting (per-iteration only); total convergence time, iteration counts to reach a target error, memory footprint, and scaling with mesh size (number of tetrahedra/edges) are not analyzed.
- Hyperparameter sensitivity (λS, λB, λF, σ) is only partially explored; no automated tuning or principled schedules are proposed, and cross-model robustness is untested.
- No ablation or comparison of neural architectures (e.g., ReLU MLPs, Fourier features, positional encodings), alternative SH/tensor representations (odeco, symmetric 4th-order tensors), or SO(3)-equivariant networks that may better respect rotational symmetries.
- The approach does not leverage curvature directions or principal frames explicitly; integrating differential-geometric cues (e.g., curvature-aligned priors) to constrain the in-plane angle remains unexplored.
- Robustness to real-world noise (e.g., scanned data with normal errors, holes, non-manifold artifacts) and to imperfect tetrahedralization (self-intersections, poor quality) is not evaluated.
- Handling of complex boundaries (thin shells, multiple adjacent boundary faces, small features) is heuristic (tet subdivision) with no analysis of failure modes or alternative constraint discretizations.
- No uncertainty estimation or confidence measures for the predicted field; strategies like ensembles, Bayesian layers, or calibration could help assess reliability prior to meshing.
- The method is demonstrated on closed volumes; applicability to open surfaces, internal boundaries, multi-material domains, and models with internal voids or embedded networks is not addressed.
- The surface cross-field is obtained by discarding the axis most aligned with normals; continuity and coherence of the induced cross field across triangles (sign flips, jumps) and its impact on MIQ are not studied.
- Fairness and completeness of comparisons are limited: dataset size, diversity, parameter parity, and standardized metrics are not detailed; no statistical tests or confidence intervals are provided.
- Code, pretrained models, and detailed experimental protocols are not declared as available; reproducibility and ease of adoption by practitioners are uncertain.
- The proposed future direction—using a “corrected” singularity graph as training target—remains unspecified: how to obtain such targets robustly, enforce them differentiably, and preserve global topological consistency during learning is an open technical challenge.
Practical Applications
Practical Applications of NeurFrame (Continuous Frame Fields for Structured Mesh Generation)
NeurFrame introduces a continuous, self-supervised neural representation of 3D frame fields (and derived 2D cross fields) that guides both quadrilateral and hexahedral mesh generation. It reduces singularities, improves feature/boundary alignment, and lowers computational cost by avoiding dense tetrahedral discretizations. Below are actionable applications, grouped by immediacy, with sector links, prospective tools/workflows, and feasibility notes.
Immediate Applications
- High-quality quadrilateral meshing for digital content creation and CAD
- Sectors: software (DCC/VFX/games), manufacturing, education
- What: Generate smooth, feature-aligned cross fields from the volumetric frame field, then extract quad meshes via MIQ + libQEx for UV mapping, retopology, and texture alignment.
- Tools/workflows: Blender/Maya/Houdini plug-ins calling NeurFrame for cross-field prediction; MIQ/libQEx-based quad extraction; consistent LOD retopology using infinite-resolution queries.
- Assumptions/dependencies: Requires surface normals and a tetrahedralization (e.g., fTetWild). Training on a GPU (10k iterations). Quality depends on MIQ parameterization and feature detection.
- Faster, smoother frame-field generation for hexahedral meshing in CAE
- Sectors: aerospace, automotive, energy, civil/structural engineering (CFD/FEA), isogeometric analysis
- What: Use NeurFrame to produce boundary- and feature-aligned frame fields on coarse tetrahedral meshes, then run CubeCover + libHexEx to extract hex meshes with fewer/better-distributed singularities.
- Tools/workflows: CAD/CAE plug-ins for Gmsh, Abaqus/CAE, ANSYS Meshing, Altair HyperMesh; replace or augment existing frame-field steps with NeurFrame to reduce meshing iterations.
- Assumptions/dependencies: Hex mesh extraction may fail on certain topologies (global meshability not guaranteed); depends on quality of input tet mesh and parameterization; requires GPU; cube-cover pipeline integration.
- Feature-preserving meshing near creases and sharp edges
- Sectors: turbomachinery/CFD, heat exchangers, mechanical design
- What: Leverage boundary and localized feature alignment losses to align cells with creases/edges, improving element orientation and reducing distortion near sharp features.
- Tools/workflows: Feature detection during preprocessing; NeurFrame field optimization with distance-weighted feature constraints; downstream extraction using existing parameterization tools.
- Assumptions/dependencies: Feature detection quality is critical; over-constraining may require tuning λF; noise in surface normals can degrade alignment.
- Robust quadrilateral meshing for scientific visualization and surface parameterization
- Sectors: academia/research, scientific computing, geosciences (surface data), medical imaging (surfaces)
- What: Produce smooth cross fields for rectangular parameterizations and structured quad grids for visualization pipelines.
- Tools/workflows: Integration with surface parameterization libraries (e.g., Directional, MIQ/RSP); interactive visualization tools consuming the continuous field for consistent sampling.
- Assumptions/dependencies: Surface-only workflows can query NeurFrame on triangle centroids; relies on accurate normals and good surface quality.
- Teaching and research demonstrators in geometry processing
- Sectors: academia, education
- What: Use NeurFrame to illustrate SH-based frame representations, singularity graphs, field smoothness, and boundary/feature alignment in courses and labs.
- Tools/workflows: Jupyter notebooks and lightweight GUIs that query the continuous neural field at arbitrary points; comparative studies against discrete methods.
- Assumptions/dependencies: Modest GPU required for training; visualization modules for singularity graphs.
- Meshing pre-processing for topology optimization and iterative design loops
- Sectors: automotive/aerospace structures, mechanical design
- What: Stable, smooth fields reduce singularity churn across iterations, lowering remeshing overhead and numerical artifacts during optimization.
- Tools/workflows: Integrate NeurFrame into topology optimization pipelines for repeated meshing; maintain consistent element orientation across iterations.
- Assumptions/dependencies: Training+inference time must fit design loop cadence; parameter tuning to balance smoothness and feature alignment per iteration.
- Mesh generation as-a-service (internal platform/tooling)
- Sectors: software (platforms), CAE/CFD consultancies
- What: Deploy a GPU-backed service that takes a surface (and optionally, tet mesh) and returns cross/hex/quad meshes; continuous field enables on-demand, resolution-independent queries.
- Tools/workflows: REST/RPC API wrapping NeurFrame + MIQ/CubeCover pipelines; batch meshing for design variants.
- Assumptions/dependencies: GPU orchestration; SLA considerations; dependence on open-source libraries (fTetWild, libQEx, CubeCover, libHexEx) and their licenses.
Long-Term Applications
- Industrial-grade, globally meshable hexahedral meshing
- Sectors: aerospace/automotive certification, nuclear/energy, medical devices
- What: Extend NeurFrame training to incorporate meshability constraints and corrected singularity graphs (e.g., Liu et al.) so that generated fields consistently yield valid, high-quality hex meshes.
- Tools/workflows: End-to-end hex meshing pipeline with embedded global constraints; QA/verification tooling for meshability; integration into commercial solvers.
- Assumptions/dependencies: Research needed on differentiable meshability constraints and global singularity control; certification requires rigorous validation.
- Interactive co-design: real-time field-guided meshing during CAD modeling
- Sectors: CAD/CAE software, product design
- What: Live updates of continuous frame fields and structured meshes as geometry changes; designers inspect singularities and field lines while editing.
- Tools/workflows: Incremental/online learning (meta-learning, warm-starts) for sub-second updates; GPU inference of trained networks integrated into CAD kernels.
- Assumptions/dependencies: Significant optimization of training/inference; robust incremental tet meshing; UI/UX for field visualization.
- Field-guided isogeometric analysis (IGA) and spline patch layout automation
- Sectors: CAE/analysis, academic research
- What: Use the continuous frame field to define integrable parameterizations with fewer singularities for generating tensor-product spline patches (NURBS/T-splines).
- Tools/workflows: IGA toolchains that consume NeurFrame fields to infer patch layouts and parameter lines; automatic conversion to spline solids.
- Assumptions/dependencies: Need advances in integrability enforcement and patch layout extraction; handling of complex topologies.
- Additive manufacturing: anisotropic lattice and toolpath orientation
- Sectors: advanced manufacturing, medical implants, aerospace lightweighting
- What: Guide microstructure orientation (lattices/gyroids) or toolpaths using volumetric frame fields to improve stiffness, strength, and fatigue performance.
- Tools/workflows: Plug-ins for slicers (Cura/PrusaSlicer) or lattice design platforms (nTopology) that consume continuous fields; coupling with material models and print constraints.
- Assumptions/dependencies: Integration between field orientation and lattice/toolpath generators; validation with mechanical testing; printability constraints.
- Digital twins and adaptive remeshing at runtime
- Sectors: industrial IoT, predictive maintenance, process engineering
- What: Continuous frame fields enable on-the-fly structured remeshing as geometries or operating conditions evolve, maintaining simulation fidelity in digital twins.
- Tools/workflows: Cloud/HPC deployments that retrain/refine fields from updated geometry; adaptive simulation steering with field-aware refinement.
- Assumptions/dependencies: Efficient retraining and robust update strategies; data assimilation pipelines; computational budget.
- Multi-physics and biomedical simulation pipelines
- Sectors: healthcare (biomechanics, surgical planning), soft robotics
- What: Use hex meshes guided by NeurFrame for stable, low-distortion FEM in soft tissues, implants, and soft robots, improving convergence and accuracy in coupled simulations.
- Tools/workflows: Integration with FEBio, SOFA, SimVascular, or custom FEM stacks; automated meshing from segmented scans and CAD.
- Assumptions/dependencies: Meshability guarantees and regulatory-grade validation; robust preprocessing of noisy medical data.
- Standardization and benchmarking for mesh quality and singularity control
- Sectors: policy/standards bodies, academia, software vendors
- What: Define benchmarks and metrics for field smoothness, singularity distribution, and feature alignment under continuous representations; guide procurement/specification of meshing tools.
- Tools/workflows: Open datasets and leaderboards comparing discrete vs. continuous-field methods; standardized QA protocols for safety-critical domains.
- Assumptions/dependencies: Community and vendor adoption; agreement on metrics and data formats; transparent reporting.
- Reconstruction-to-simulation pipelines (scan/NeRF/SDF → mesh → solve)
- Sectors: reverse engineering, AR/VR, robotics
- What: After geometric reconstruction, use NeurFrame to produce simulation-ready structured meshes directly from scanned or learned surfaces.
- Tools/workflows: Bind NeurFrame to NeRF/SDF pipelines; automatic tet generation and field optimization; downstream CAE post-processing.
- Assumptions/dependencies: Robust handling of reconstruction artifacts; coupling with meshable parameterizations; compute overhead management.
- Domain-aware PDE solvers on field-aligned coordinates
- Sectors: scientific computing, metamaterials
- What: Employ continuous frame fields to define computational coordinates for anisotropic PDEs (e.g., diffusion, elasticity), potentially improving solver conditioning and accuracy.
- Tools/workflows: Field-aligned discretizations; operator splitting in aligned frames; custom solvers exploiting orientation coherence.
- Assumptions/dependencies: Research into consistent discretizations and stability; integration with existing solver stacks.
Notes across applications:
- Performance and cost: NeurFrame uses a single SIREN MLP and is shown to be faster per iteration than comparable neural methods, but still requires GPU resources and training time.
- Dependencies: Existing third-party tools (fTetWild, MIQ, CubeCover, libQEx, libHexEx) are integral to the current pipeline; their robustness and licensing affect deployment.
- Limitations: Hex meshing is not guaranteed—global meshability remains an open challenge; boundary/feature detection and tet quality influence outcomes.
- Data/normalization: Inputs must be normalized; noisy normals/features require preprocessing for best results.
- Validation: For regulated sectors (e.g., medical, nuclear), extensive verification/validation and uncertainty quantification are prerequisites for adoption.
Glossary
- 4-RoSy: A 2D directional field with fourfold rotational symmetry where directions are equivalent up to 90-degree rotations. "At each point, a cross contains two orthogonal directions and exhibits 4-rotational symmetry (4-RoSy)"
- atlas of charts: A collection of coordinate charts covering a manifold, with transition functions defining how charts relate. "The IGM is an atlas of charts whose maps and transition functions satisfy boundary alignment, local injectivity, and conformity."
- boundary alignment: A constraint requiring field directions (or mesh elements) to align with the surface normals along the boundary. "satisfy boundary alignment, local injectivity, and conformity"
- Cartan’s method of moving frames: A differential geometry technique that represents frames moving along a manifold to analyze curvature and invariants. "\citet{Corman-2019} used Cartanâs method of moving frames with the Darboux derivative as the main representation."
- cross field: A 2D field assigning, at each point on a surface, a pair of orthogonal directions up to 90-degree symmetry to guide quad meshing. "field-based methods that construct cross fields in 2D"
- cubic-symmetry frame fields: 3D frame fields invariant under the symmetry group of a cube (24 rotations), used to represent orientations with octahedral symmetry. "enforcing local integrability and generating cubic-symmetry frame fields guided by Riemannian metrics and alignment constraints"
- CubeCover: An algorithm that computes integer-grid maps aligned with a frame field to enable hexahedral meshing. "we compute a frame-aligned IGM using CubeCover~\cite{Nieser-2011}"
- curl-free fields: Directional fields with zero curl, implying local integrability into potentials or parameterizations. "later extended to curl-free fields in \cite{Diamanti-2015}"
- Darboux derivative: A differential form associated with moving frames that captures how a frame changes along a surface or volume. "with the Darboux derivative as the main representation"
- Dirichlet energy: A measure of smoothness of a function/field based on the integral of squared gradients, often minimized to smooth fields. "we adopt a discretization of the Dirichlet energy similar to prior work~\cite{Liu-2018}"
- dual edge: In a dual mesh (e.g., between tetrahedra), an edge connecting cell centers of adjacent primal elements, used to discretize differential operators. "let E denote the set of dual edge connecting centroids between adjacent tetrahedra"
- fTetWild: A robust tetrahedralization tool that generates volumetric meshes from surfaces. "we employ fTetWild~\cite{Hu-2020} to tetrahedralize the enclosed volume."
- frame field: A 3D field assigning an orthonormal triad (three orthogonal unit vectors) at each point to encode local orientations for hex meshing. "We represent the spatially varying frame field using an MLP"
- Ginzburg--Landau energy: An energy functional (from phase transition models) used to encourage smooth, well-structured fields with controlled singularities. "\citet{Viertel-2019} compute boundary-aligned cross fields by minimizing a Ginzburg--Landau energy using MBO threshold dynamics."
- hexahedral mesh: A volumetric mesh composed of six-faced (cube-like) elements used in simulation and CAD. "resulting in a high-quality hexahedral mesh"
- hex-meshability: The property that a frame field can yield a valid hexahedral mesh under topological and compatibility conditions. "\citet{Liu-2018} derived necessary local and global conditions for frame-field hex-meshability"
- integrable fields: Directional fields that can be integrated into globally consistent parameterizations without rotational inconsistency. "including power fields~\cite{Knoppel-2013}, polyvector fields~\cite{Diamanti-2014}, integrable fields~\cite{Diamanti-2015}, and GinzburgâLandau fields~\cite{Viertel-2019}"
- integer-grid map (IGM): A volumetric parameterization whose coordinates align to integer grid lines, enabling extraction of structured hex meshes. "the frame field guides the construction of an integer-grid map (IGM)"
- isogeometric analysis: A simulation paradigm that integrates CAD geometry (e.g., splines) directly into finite element analysis. "isogeometric analysis \cite{Hughes-2005}"
- Karcher mean: The Fréchet mean on a Riemannian manifold, used to average rotations/frames by minimizing geodesic distances. "compute each tetrahedron's frame as the Karcher mean of the frames at its four vertices"
- Laplace–Beltrami eigenfunctions: Eigenfunctions of the Laplace–Beltrami operator on a manifold, forming an orthogonal basis for functions on surfaces (e.g., the sphere). "an orthogonal basis for spherical functions using LaplaceâBeltrami eigenfunctions"
- libHexEx: A library for extracting hexahedral meshes from integer-grid maps. "The final hexahedral mesh is then extracted from the IGM using libHexEx~\cite{Lyon-2016}."
- libQEx: A library for extracting quadrilateral meshes from surface parameterizations. "The final quadrilateral mesh is extracted from the parameterization using libQEx~\cite{Ebke-2013}."
- local injectivity: A mapping property ensuring no fold-overs or local inversions (one-to-one locally), crucial for valid parameterizations. "satisfy boundary alignment, local injectivity, and conformity"
- MBO threshold dynamics: An iterative diffusion-thresholding scheme (Merriman–Bence–Osher) for minimizing variational energies, used here to compute fields. "using MBO threshold dynamics"
- mixed-integer quadrangulation (MIQ): A method that computes globally consistent quad-aligned parameterizations by solving mixed-integer problems. "using mixed-integer quadrangulation (MIQ)~\cite{Bommes-2009}"
- NeurCross: A self-supervised neural method that learns cross fields without ground-truth by coupling with an SDF. "the self-supervised method NeurCross \cite{Dong-2025b}"
- Neural Octahedral Field (NeurOcta): A neural representation method for continuous octahedral/frame fields learned from data. "The most closely related competing method is Neural Octahedral Field (NeurOcta)~\cite{Zheng-2025}"
- N-PolyVector fields: A representation of N-RoSy fields as roots of complex polynomials enabling efficient optimization. "\citet{Diamanti-2014} introduce N-PolyVector fields, representing N-RoSy fields as roots of complex polynomials"
- N-RoSy: N-rotational symmetry fields where directions are equivalent modulo rotation by 360/N degrees. "representing N-RoSy fields as roots of complex polynomials"
- octahedral field: Another name for a 3D frame field emphasizing the 24-fold octahedral rotational symmetry. "a frame field (also referred to as an octahedral field) in 3D"
- odeco fields: Orthogonally decomposable tensor fields allowing independent scaling along frame axes to better capture singularities. "\citet{Palmer-2020} proposed odeco fields, allowing independent scaling along three orthogonal directions"
- polyvector fields: Direction fields represented by polynomial vectors, enabling robust and smooth design of cross fields. "polyvector fields~\cite{Diamanti-2014}"
- power fields: A quadratic energy-based method for smooth cross field design on surfaces. "power fields~\cite{Knoppel-2013}"
- quadrilateral mesh: A surface mesh composed entirely of four-sided faces, favored for parameterization and simulation. "This produces a structured quadrilateral mesh."
- Riemannian metrics: Metrics defined on manifolds that measure lengths/angles smoothly and anisotropically, guiding field smoothness and alignment. "guided by Riemannian metrics and alignment constraints"
- rectangular surface parameterization (RSP): A technique producing parameterizations aligned to axis-aligned rectangles for quad extraction. "rectangular surface parameterization (RSP)~\cite{Corman-2025}"
- SDF (signed distance field): A scalar field giving the signed distance to a surface; positive outside, negative inside, used for geometry processing. "a neural signed distance field (SDF)"
- SIREN: A neural network architecture using sinusoidal activations for modeling high-frequency continuous signals. "We adopt a SIREN~\cite{Sitzmann-2020} architecture"
- singularity graph: The network of curves/points where a field’s direction is undefined or discontinuous, governing mesh topology. "compare frame fields and their singularity graphs"
- SO(3): The Lie group of 3D rotations (special orthogonal group), the natural space for averaging and measuring distances between rotations. "distance on "
- spherical harmonics (SH): An orthonormal basis of functions on the sphere used to compactly represent rotation-aware fields. "Spherical harmonic (SH) representation of a frame."
- tetrahedral discretizations: Volumetric meshes of tetrahedra used to discretize 3D domains for numerical optimization. "In 3D, tetrahedral discretizations further introduce a trade-off between computational cost and field quality."
- Thingi10K: A large dataset of 3D models commonly used for geometric processing benchmarks. "Our experimental datasets are drawn from Thingi10K~\cite{Zhou-2016}"
- universal approximation theorem: A result stating that sufficiently wide neural networks can approximate continuous functions on compact domains. "a capability supported by the universal approximation theorem \cite{Kim-2003, Lecun-2015}"
- neural radiance fields: Neural representations that model volumetric scene appearance and density for view synthesis. "including neural radiance fields \cite{Mildenhall-2021}"
Collections
Sign up for free to add this paper to one or more collections.