C-HiDeNN: Convolutional Hierarchical Deep NN

• C-HiDeNN is a mesh-based neural approximator that enriches finite-element interpolation with locally supported convolution patches.
• It achieves high-order smoothness and reproducing capability while retaining the local support and Kronecker-delta properties critical for boundary handling.
• Extensions like tensor decomposition, isogeometric analysis, and multilevel variational frameworks enable scalable PDE solutions and efficient surrogate modeling.

Searching arXiv for papers on C-HiDeNN and closely related hierarchical convolutional models. arXiv search: query for "C-HiDeNN HiDeNN convolutional hierarchical deep neural network". Convolutional Hierarchical Deep Neural Network, usually abbreviated C-HiDeNN in the recent numerical-analysis literature, denotes a structured neural approximation framework that combines finite-element-style interpolation, locally supported convolution patch functions, and hierarchical basis construction to solve partial differential equations and build surrogate models for large-scale physical systems (Guo et al., 2024). In this exact usage, C-HiDeNN is not a standard image-classification CNN; it is a mesh-based, finite-element-inspired neural basis method in which approximation functions retain local support and the Kronecker-delta property while gaining higher-order enrichment through patchwise convolution (Liu et al., 2 Sep 2025). The phrase also has a broader descriptive use for hierarchical convolutional architectures in computer vision, but the older vision literature typically employs different names such as HD-CNN rather than the exact acronym C-HiDeNN (Yan et al., 2014).

1. Terminology, scope, and lineage

The term C-HiDeNN expands to Convolutional Hierarchical Deep-learning Neural Network in the recent PDE and computational mechanics literature (Guo et al., 2024). Its immediate lineage is HiDeNN, a hierarchical deep-learning neural-network formulation for interpolation and discretization, and several later variants. In particular, C-IGA is presented as the isogeometric specialization of C-HiDeNN, where the same convolutional-hierarchical idea is embedded into B-spline or NURBS geometry representations so that exact CAD mappings are preserved (Zhang et al., 2024). Ex-HiDeNN is then described as being built upon prior work on HiDeNN and Convolutional HiDeNN, with C-HiDeNN-TD serving as its first-stage structured surrogate before symbolic regression (Batley et al., 7 Jul 2025).

The literature therefore uses the name in a precise and fairly technical way: C-HiDeNN is a neural-network-represented approximation basis rather than a generic deep model trained only from labeled data. The “convolutional” qualifier refers to local patch interactions over neighboring nodes, not to translationally invariant pixel kernels in the conventional computer-vision sense (Guo et al., 2024). The “hierarchical” qualifier refers to a nested construction in which low-order finite-element interpolation is enriched through patch-level and neural layers, rather than to a label taxonomy or tree of semantic classes (Guo et al., 2024).

A broader usage also exists. Earlier computer-vision models such as HD-CNN are directly relevant when “convolutional hierarchical deep neural network” is used descriptively rather than as a formal name, because they organize prediction hierarchically over classes and combine coarse and fine specialists (Yan et al., 2014). That broader usage, however, should be distinguished from the exact C-HiDeNN family used in scientific computing.

2. Core approximation principle

The defining C-HiDeNN construction is a finite-element-like basis enriched by local convolution patch functions. In the space-time PDE formulation, the approximation is written as

$$u^{h,e}(\xi)=\sum_{i\in A^{e}} \mathbf{N}_{i}(\xi)\sum_{j\in A_{s}^{i}}\mathcal{W}^{i}_{s,a,p,j}\!\left(x^{h,e}(\xi)\right)\mathbf{u}_{j} =\sum_{k\in A_{s}^{e}}\widetilde{\mathbf{N}}_{k}(\xi)\mathbf{u}_{k},$$

where $\mathbf{N}_i(\xi)$ is the standard FE/Lagrange basis, $A^e$ is the element node set, $A_s^i$ is a nodal patch, $\mathcal W_{s,a,p,j}^i$ is a convolution patch function, and $\widetilde{\mathbf N}_k$ is the enriched C-HiDeNN basis (Guo et al., 2024). The same idea appears in the nonlinear finite-element/meshfree formulation as

$$u^h(\xi)=\sum_{I\in A^e}N_I^e(\xi)\sum_{J\in A_I^s}W_{a,p,J}(X(\xi))u_J=\sum_{K}\widetilde N_K(\xi)u_K,$$

which makes explicit the two nested levels: element-level FE interpolation and patch-level convolution interpolation over neighboring nodes (Liu et al., 2 Sep 2025).

This construction is designed to combine two properties that are usually separated. From finite elements it retains local support, Gaussian quadrature compatibility, and the Kronecker-delta property, which simplifies the imposition of Dirichlet boundary conditions (Guo et al., 2024). From neural and meshfree enrichment it gains higher-order smoothness, reproducing capability, and local nonpolynomial flexibility (Liu et al., 2 Sep 2025). The numerical-analysis papers repeatedly emphasize that ordinary multilayer perceptrons behave as global basis functions, whereas C-HiDeNN remains explicitly local and mesh-aware (Guo et al., 2024).

The principal design parameters are the polynomial order $p$, dilation parameter $a$, patch size $s$, and, in the nonlinear mechanics formulation, the nodal position $\mathbf{N}_i(\xi)$0 (Liu et al., 2 Sep 2025). These parameters are stated to “function as the weights and biases within the C-HiDeNN patch” (Liu et al., 2 Sep 2025). Their numerical meaning is direct: $\mathbf{N}_i(\xi)$1 controls reproducing order, $\mathbf{N}_i(\xi)$2 controls kernel spread, $\mathbf{N}_i(\xi)$3 controls how many neighboring nodes participate in the convolution patch, and $\mathbf{N}_i(\xi)$4 supports $\mathbf{N}_i(\xi)$5-adaptivity through nodal relocation (Liu et al., 2 Sep 2025).

Because nonlinear finite-element analysis requires material derivatives, C-HiDeNN also provides derivatives of the enriched basis. In the nonlinear total-Lagrangian formulation, the derivative of the approximation splits into a term from the FE basis and a term from the patch basis, reflecting the product structure of the interpolation (Liu et al., 2 Sep 2025). This allows the method to enter standard weak forms and semidiscrete momentum equations without abandoning finite-element assembly logic.

3. Tensor decomposition and scalable space-time variants

A major extension is C-HiDeNN-TD, where the solution is represented as a low-rank separated product of one-dimensional C-HiDeNN factors. For an $\mathbf{N}_i(\xi)$6-variate field,

$\mathbf{N}_i(\xi)$7

or, in compact notation,

$\mathbf{N}_i(\xi)$8

with $\mathbf{N}_i(\xi)$9 the rank and $A^e$0 the Hadamard product (Guo et al., 2024). The stated effect is to turn a high-dimensional problem into a set of coupled one-dimensional subproblems, reducing storage and computation from full-grid growth to low-rank separated growth (Guo et al., 2024).

In the PDE setting, this is not cast as conventional offline training. The surrogate is obtained directly from the PDE weak form, and the resulting systems are solved iteratively until the factor vectors converge (Guo et al., 2024). On the reported 4D space-time benchmark, the method handles an effective grid of

$A^e$1

points while storing only 15.6 MB of solution data, because only the one-dimensional factor vectors are retained (Guo et al., 2024).

The same separated structure becomes the first stage of Ex-HiDeNN, where C-HiDeNN-TD learns one-dimensional representations for each feature, combines them into modes, and supplies a continuously differentiable surrogate for separability checking and symbolic regression (Batley et al., 7 Jul 2025). This suggests a second role for the C-HiDeNN family: beyond direct PDE solving, it can serve as a structured interpolant whose hierarchy and separability make downstream symbolic model discovery more tractable (Batley et al., 7 Jul 2025).

4. Isogeometric and multilevel generalizations

The isogeometric extension, C-IGA, replaces polynomial reproduction by spline or NURBS reproduction so that the geometry mapping remains exact, as in IGA, while the solution variables remain nodal and retain the Kronecker-delta property (Zhang et al., 2024). In multi-patch CAD systems this creates an interface problem: patchwise high-order approximation is not sufficient unless compatibility is enforced across patch boundaries. The paper develops two levels of interface matching, nodal compatibility and $A^e$2 compatibility, and shows that nodal matching alone is generally insufficient for optimal PDE convergence (Zhang et al., 2024). With full $A^e$3 compatibility, the reported interface deviation is about $A^e$4 and high-order convergence is recovered (Zhang et al., 2024).

A second major generalization is the MultiLevel Variational MultiScale (ML-VMS) framework, which uses C-HiDeNN as the approximation basis on coarse and localized fine meshes (Zhang et al., 27 Oct 2025). In this setting C-HiDeNN is described as enabling arbitrary-order approximation on a linear FE mesh without adding higher-order element nodes or increasing nodal degrees of freedom (Zhang et al., 27 Oct 2025). The multilevel method couples coarse and fine regions through weak forms and interface conditions, while level-dependent hyperparameters such as $A^e$5 and $A^e$6 control local accuracy and bandwidth (Zhang et al., 27 Oct 2025).

The reported multilevel results emphasize efficiency. For a large-scale single-track laser powder bed fusion transient heat-transfer problem equivalent to a full-order finite-element model with $A^e$7 spatial DoFs, the 3-level ML-VMS C-HiDeNN-TD formulation achieves an approximately 5,000x speedup on a single CPU over a single-level linear FEM-TD reduced-order model (Zhang et al., 27 Oct 2025). This use of C-HiDeNN as a high-order local basis, rather than as a generic learned black box, is central to the multilevel methodology.

5. Relation to hierarchical convolutional models in computer vision

Outside numerical analysis, closely related ideas appear in the computer-vision literature under different names. HD-CNN embeds deep CNNs into a two-level category hierarchy, using a coarse classifier to separate easy classes and fine classifiers to discriminate confusing subsets (Yan et al., 2014). Its final prediction is a coarse-weighted combination of specialist outputs, and the reported top-1 error reductions of the three HD-CNN variants over their base CNNs are 2.65%, 3.1%, and 1.1% (Yan et al., 2014). Although the paper does not use the acronym C-HiDeNN, it is an exact example of a hierarchical convolutional deep network in the class-taxonomy sense.

Other related architectures broaden the comparison. Tree-CNN organizes multiple CNN modules in a tree and grows the hierarchy incrementally as new classes arrive, reducing training effort relative to broad fine-tuning while maintaining competitive accuracy on CIFAR-10 and CIFAR-100 (Roy et al., 2018). Multiscale Hierarchical Convolutional Networks treat deeper layers as functions over spatial variables and progressively learned attribute coordinates, with multidimensional convolutions over both spaces (Jacobsen et al., 2017). Towards Deep Compositional Networks keeps the outer convolutional form of CNNs but replaces unconstrained filters with sums of Gaussian components, thereby injecting explicit part structure into each unit (Tabernik et al., 2016).

A theoretical bridge is provided by work on hierarchical tensor factorization, which analyzes a class of deep non-linear convolutional networks with tree-structured locality. That analysis shows an implicit bias toward low hierarchical tensor rank, which in the associated convolutional networks translates into an implicit bias toward locality (Razin et al., 2022). This does not define C-HiDeNN, but it clarifies why “hierarchical” and “convolutional” often co-occur in models that fuse local interactions into progressively larger structures.

6. Applications, strengths, and limitations

In scientific computing, the reported applications are concentrated in PDE surrogate modeling and nonlinear mechanics. C-HiDeNN-TD is demonstrated on a steady 2D Poisson equation and a large-scale transient diffusion problem in 3D space plus time, where it is reported to require much less computational time, GPU memory, and storage than full-grid finite-difference methods for enormous tensor-product discretizations (Guo et al., 2024). The nonlinear mechanics framework applies C-HiDeNN to plane-strain notch problems, brain impact, long-notch tensile loading, and 3D hyperelastic structures, repeatedly showing much better stress and strain-energy predictions than same-mesh low-order FEM (Liu et al., 2 Sep 2025). In one notched-plate example at $A^e$8 ms, the reported errors are 23.4%, 37.5%, and 50.7% for FEM versus 2.4%, 2.2%, and 4.0% for C-HiDeNN in $A^e$9, $A_s^i$0, and von Mises stress, respectively (Liu et al., 2 Sep 2025).

The main strengths stated across the literature are consistent. C-HiDeNN preserves FE-style locality and boundary handling, supports multiple adaptivity mechanisms, can raise approximation order without introducing higher-order element nodes, and scales effectively when tensor decomposition is valid (Guo et al., 2024). In symbolic-regression workflows, its continuously differentiable and separable structure also reduces the search burden for closed-form discovery (Batley et al., 7 Jul 2025).

The limitations are likewise explicit. The tensor-decomposition extension assumes that the target field admits a useful low-rank separated representation; if the required rank is large, the computational advantage diminishes (Guo et al., 2024). The nonlinear FEM paper does not provide a rigorous convergence analysis, stability proof, or a posteriori error estimator, and optimization of discrete parameters such as $A_s^i$1 and $A_s^i$2 is asserted conceptually rather than algorithmically detailed (Liu et al., 2 Sep 2025). In the multi-patch isogeometric setting, full $A_s^i$3 compatibility restores convergence but introduces substantial implementation complexity near patch interfaces (Zhang et al., 2024). Ex-HiDeNN, finally, inherits dependence on the quality of the first-stage C-HiDeNN-TD surrogate and is reported to struggle with discontinuities or kinks (Batley et al., 7 Jul 2025).

Taken together, the literature presents C-HiDeNN as a physics-grounded hierarchical neural basis family rather than a single monolithic architecture. Its exact meaning depends on context, but the dominant recent usage is a locally supported, convolution-enriched, hierarchical approximation framework for PDEs, finite elements, and surrogate modeling, with extensions to tensor decomposition, isogeometric analysis, multilevel variational multiscale methods, and symbolic regression (Guo et al., 2024).