Conditional Neural Whitney Forms (CNWF)

• Conditional Neural Whitney Forms are reduced finite element models that use Whitney forms and FEEC to enforce exact discrete conservation.
• They integrate transformer-based operator learning to parameterize both the reduced basis and nonlinear flux laws under varying conditions.
• The approach supports real-time digital twins and adaptive sensing while preserving numerical stability via structure-preserving discretizations.

Searching arXiv for the cited CNWF and Whitney-form papers to ground the article in the current literature. Conditional Neural Whitney Forms (CNWF) are a family of data-driven, structure-preserving reduced finite element models built on Whitney forms and finite element exterior calculus (FEEC), whose basis functions and discrete flux operators are parameterized by a transformer conditioned on a latent variable $Z$. In the digital-twin formulation, conditional attention mechanisms learn both a reduced finite element basis and a nonlinear conservation law within FEEC, so that numerical well-posedness and exact preservation of conserved quantities are retained regardless of data sparsity or optimization error (Kinch et al., 9 Aug 2025). In the source-localization setting, the same framework couples the numerical guarantees of FEEC with transformer-based operator learning, using a conditional attention mechanism to identify a reduced Whitney-form basis, reduced integral balance equations, and a source field, each compatible with given sensor measurements (Shaffer et al., 12 Sep 2025).

1. Definition and conceptual scope

CNWF organize learned models around a reduced Whitney complex rather than around an unconstrained function approximator. In the steady formulation emphasized in the digital-twin literature, the model is a parametric family

$$Z \mapsto \big( \mathcal{W}_0^\theta(Z), \mathcal{W}_1^\theta(Z), \mathcal{NN}[\cdot;\theta,Z] \big),$$

together with an FEEC-compatible discrete operator

$$\mathcal{F}(\hat u;\theta,Z) := \delta_0^\top \mathbf{M}_1\,\delta_0 \hat u + \delta_0^\top \mathbf{M}_1\,\mathcal{NN}[\hat u;\theta,Z] = \mathbf{f}_\theta(Z).$$

Here $\hat u$ denotes reduced $0$-form degrees of freedom, $\mathbf{M}_0,\mathbf{M}_1$ are reduced mass matrices, and $\delta_0$ is the reduced coboundary operator (Kinch et al., 9 Aug 2025).

The conditional aspect is central. In one formulation, $Z$ indexes operating conditions or parameters such as material properties, charge location, flow orientation, or battery operating conditions; in another, the conditioning variable is derived from streaming sensor data. In both cases, $Z$ modulates the geometry through the reduced basis and the physics through the nonlinear flux law, yielding a reduced model that is recalibrated by conditioning rather than by retraining (Kinch et al., 9 Aug 2025). In the adaptive source-localization formulation, the conditioned model is used as a real-time digital twin for a coupled hydrodynamic-transport system, specialized to steady advection-diffusion, with a learned reduced basis, learned flux, and learned source operator all driven by sparse sensor measurements (Shaffer et al., 12 Sep 2025).

A defining distinction from generic neural operators is that CNWF do not treat the map from observations to fields as an unconstrained regression. The learned output is always interpreted inside a discrete de Rham complex, and the governing conservation law is enforced as an equality constraint at the reduced level. This yields a physically realizable, regular mapping from sensor data to the source field in the source-localization setting, and a real-time digital twin that interfaces non-invasively with conventional finite element machinery in the digital-twin setting (Shaffer et al., 12 Sep 2025).

2. Whitney-form and FEEC foundations

CNWF inherit their geometric structure from classical Whitney forms. On a simplex $\sigma$, Whitney forms are characterized as the unique $$Z \mapsto \big( \mathcal{W}_0^\theta(Z), \mathcal{W}_1^\theta(Z), \mathcal{NN}[\cdot;\theta,Z] \big),$$0-forms with affine coefficients whose pullbacks to every $$Z \mapsto \big( \mathcal{W}_0^\theta(Z), \mathcal{W}_1^\theta(Z), \mathcal{NN}[\cdot;\theta,Z] \big),$$1-face are constant and whose integrals over those faces reproduce a prescribed simplicial $$Z \mapsto \big( \mathcal{W}_0^\theta(Z), \mathcal{W}_1^\theta(Z), \mathcal{NN}[\cdot;\theta,Z] \big),$$2-cochain. More precisely, for $$Z \mapsto \big( \mathcal{W}_0^\theta(Z), \mathcal{W}_1^\theta(Z), \mathcal{NN}[\cdot;\theta,Z] \big),$$3, $$Z \mapsto \big( \mathcal{W}_0^\theta(Z), \mathcal{W}_1^\theta(Z), \mathcal{NN}[\cdot;\theta,Z] \big),$$4 is the unique $$Z \mapsto \big( \mathcal{W}_0^\theta(Z), \mathcal{W}_1^\theta(Z), \mathcal{NN}[\cdot;\theta,Z] \big),$$5 such that: $$Z \mapsto \big( \mathcal{W}_0^\theta(Z), \mathcal{W}_1^\theta(Z), \mathcal{NN}[\cdot;\theta,Z] \big),$$6 has affine coefficients; $$Z \mapsto \big( \mathcal{W}_0^\theta(Z), \mathcal{W}_1^\theta(Z), \mathcal{NN}[\cdot;\theta,Z] \big),$$7 is constant for every $$Z \mapsto \big( \mathcal{W}_0^\theta(Z), \mathcal{W}_1^\theta(Z), \mathcal{NN}[\cdot;\theta,Z] \big),$$8-dimensional face $$Z \mapsto \big( \mathcal{W}_0^\theta(Z), \mathcal{W}_1^\theta(Z), \mathcal{NN}[\cdot;\theta,Z] \big),$$9; and $$\mathcal{F}(\hat u;\theta,Z) := \delta_0^\top \mathbf{M}_1\,\delta_0 \hat u + \delta_0^\top \mathbf{M}_1\,\mathcal{NN}[\hat u;\theta,Z] = \mathbf{f}_\theta(Z).$$0 for every $$\mathcal{F}(\hat u;\theta,Z) := \delta_0^\top \mathbf{M}_1\,\delta_0 \hat u + \delta_0^\top \mathbf{M}_1\,\mathcal{NN}[\hat u;\theta,Z] = \mathbf{f}_\theta(Z).$$1-face $$\mathcal{F}(\hat u;\theta,Z) := \delta_0^\top \mathbf{M}_1\,\delta_0 \hat u + \delta_0^\top \mathbf{M}_1\,\mathcal{NN}[\hat u;\theta,Z] = \mathbf{f}_\theta(Z).$$2 (Dodziuk, 2022). This local characterization makes explicit that Whitney forms are finite-dimensional, simplexwise, and determined by face integrals.

The de Rham and Whitney maps provide the algebraic backbone. For simplicial complexes, the de Rham map integrates forms over simplices, the Whitney map sends cochains to differential forms, and the classical identities

$$\mathcal{F}(\hat u;\theta,Z) := \delta_0^\top \mathbf{M}_1\,\delta_0 \hat u + \delta_0^\top \mathbf{M}_1\,\mathcal{NN}[\hat u;\theta,Z] = \mathbf{f}_\theta(Z).$$3

express interpolation and compatibility with the exterior derivative (Guzmán et al., 13 May 2025). A related DEC–FEEC framework shows that cochains on primal and dual meshes are equivalent to Whitney and generalized Whitney forms, and that DEC Hodge-star, codifferential, and Hodge-Laplacian operators admit FEEC counterparts with rigorous convergence analysis for Hodge-Laplacian problems on well-centered meshes on contractible domains (Guzmán et al., 13 May 2025).

CNWF use these properties at reduced order. In the reduced FEEC discretization used for source localization, the coarse $$\mathcal{F}(\hat u;\theta,Z) := \delta_0^\top \mathbf{M}_1\,\delta_0 \hat u + \delta_0^\top \mathbf{M}_1\,\mathcal{NN}[\hat u;\theta,Z] = \mathbf{f}_\theta(Z).$$4-form basis $$\mathcal{F}(\hat u;\theta,Z) := \delta_0^\top \mathbf{M}_1\,\delta_0 \hat u + \delta_0^\top \mathbf{M}_1\,\mathcal{NN}[\hat u;\theta,Z] = \mathbf{f}_\theta(Z).$$5 generates coarse $$\mathcal{F}(\hat u;\theta,Z) := \delta_0^\top \mathbf{M}_1\,\delta_0 \hat u + \delta_0^\top \mathbf{M}_1\,\mathcal{NN}[\hat u;\theta,Z] = \mathbf{f}_\theta(Z).$$6-forms via the classical Whitney $$\mathcal{F}(\hat u;\theta,Z) := \delta_0^\top \mathbf{M}_1\,\delta_0 \hat u + \delta_0^\top \mathbf{M}_1\,\mathcal{NN}[\hat u;\theta,Z] = \mathbf{f}_\theta(Z).$$7-form formula

$$\mathcal{F}(\hat u;\theta,Z) := \delta_0^\top \mathbf{M}_1\,\delta_0 \hat u + \delta_0^\top \mathbf{M}_1\,\mathcal{NN}[\hat u;\theta,Z] = \mathbf{f}_\theta(Z).$$8

so the reduced spaces still form a de Rham subcomplex and the reduced coboundary remains surjective (Shaffer et al., 12 Sep 2025). In the digital-twin formulation, the reduced $$\mathcal{F}(\hat u;\theta,Z) := \delta_0^\top \mathbf{M}_1\,\delta_0 \hat u + \delta_0^\top \mathbf{M}_1\,\mathcal{NN}[\hat u;\theta,Z] = \mathbf{f}_\theta(Z).$$9-form space is likewise built from a partition of unity, and higher-degree structure is recovered through standard Whitney constructions (Kinch et al., 9 Aug 2025).

3. Conditional parameterization and exact discrete conservation

The core CNWF parameterization begins with a fine finite element basis $\hat u$0 and constructs a coarse Whitney $\hat u$1-form basis by conditional convex combination: $\hat u$2 If $\hat u$3 has positive entries and unit row sums, then the $\hat u$4 remain a partition of unity (Kinch et al., 9 Aug 2025). In the source-localization formulation, the same construction appears as

$\hat u$5

with $\hat u$6 the sensor-conditioned latent input (Shaffer et al., 12 Sep 2025).

Two neural components then define the reduced model. The first is a conditional basis network, described in one implementation as a ShapeFunctionModel that outputs the convex combination matrix $\hat u$7. The second is a conditional flux network, described as a PairwiseFluxModel or $\hat u$8, mapping reduced $\hat u$9-form coefficients to reduced $0$0-form coefficients (Kinch et al., 9 Aug 2025). In the source-localization architecture, the encoder is a permutation-invariant transformer over sensor tokens, pooled into a latent $0$1 and passed to three heads: a basis head $0$2, a source head $0$3, and a flux head $0$4 (Shaffer et al., 12 Sep 2025).

The resulting reduced conservation law is enforced exactly. In the source-localization formulation,

$0$5

while in the digital-twin formulation the steady problem is

$0$6

(Shaffer et al., 12 Sep 2025, Kinch et al., 9 Aug 2025). The conservation structure is not a soft penalty. It is imposed by construction, either as the defining reduced balance equation or as an equality-constrained optimization with Lagrange multipliers.

This exact structure yields local and global conservation. Because Whitney $0$7-forms are antisymmetric under index exchange and the $0$8-form basis is a partition of unity, internal fluxes cancel pairwise and only boundary fluxes and body forces remain in the global balance. The electrostatics benchmark makes this explicit: even when local reconstruction error is visible near the point charge, the total flux integral equals the source strength to machine precision (Kinch et al., 9 Aug 2025).

Well-posedness is obtained by adding a diffusive term and assuming a Lipschitz nonlinear perturbation. In the digital-twin analysis, the reduced nonlinear system is uniquely solvable if

$0$9

where $\mathbf{M}_0,\mathbf{M}_1$0 is a Poincaré constant and $\mathbf{M}_0,\mathbf{M}_1$1 is the Lipschitz constant of the nonlinear term (Kinch et al., 9 Aug 2025). In the source-localization formulation, the reduced FEEC problem is likewise stated to inherit stability and consistency from earlier FEEC-based theory, with $\mathbf{M}_0,\mathbf{M}_1$2 providing coercivity and $\mathbf{M}_0,\mathbf{M}_1$3 assumed Lipschitz (Shaffer et al., 12 Sep 2025).

4. Digital twins, calibration, and adaptive sensing

CNWF are explicitly formulated as digital twins. The digital-twin framework emphasizes real-time calibration to parametric variables, closed-loop inference, and non-invasive coupling to conventional finite element pipelines (Kinch et al., 9 Aug 2025). In deployment, the model receives current observations, conditions the reduced basis and flux law on those inputs, solves a small reduced nonlinear system, and reconstructs quantities of interest. Because the reduced dimension is the number of coarse partitions, inference is fast: the digital-twin benchmarks report real-time inference $\mathbf{M}_0,\mathbf{M}_1$4s with a speedup of $\mathbf{M}_0,\mathbf{M}_1$5 relative to LES (Kinch et al., 9 Aug 2025).

The adaptive sensor-coverage formulation makes this closed-loop role concrete. Sensor tokens are of the form

$\mathbf{M}_0,\mathbf{M}_1$6

and the transformer-conditioned model outputs a reduced basis, reduced source coefficients, and a reduced flux law (Shaffer et al., 12 Sep 2025). After solving the reduced conservation law, the predicted source field $\mathbf{M}_0,\mathbf{M}_1$7 is normalized into a density

$\mathbf{M}_0,\mathbf{M}_1$8

which is then used as the importance function in a Lloyd-type coverage algorithm (Shaffer et al., 12 Sep 2025).

The coverage functional is

$\mathbf{M}_0,\mathbf{M}_1$9

with $\delta_0$0 in the reported experiments, and the analysis provides conditions for monotone improvement of a coverage functional under a staggered scheme that alternates between evaluating the digital twin and applying Lloyd’s algorithm (Shaffer et al., 12 Sep 2025). One sufficient condition for monotone decrease of the model-relative coverage is

$\delta_0$1

where $\delta_0$2 is a domain-dependent Lipschitz constant for $\delta_0$3 in $\delta_0$4 (Shaffer et al., 12 Sep 2025). The same paper also gives conditions under which true coverage decreases and an upper bound on prediction error decreases, and it proves exponential convergence of continuous Lloyd dynamics to a point source under continuity and contractivity assumptions on the learned density field (Shaffer et al., 12 Sep 2025).

A plausible implication is that CNWF are not merely reduced solvers with neural coefficients; they are feedback-compatible operators whose regularity is part of the sensing and control analysis. That interpretation is consistent with the emphasis on a physically realizable, regular mapping from sensor data to source fields and with the role of regularity as a sufficient condition for localization (Shaffer et al., 12 Sep 2025).

5. Related Whitney-form machinery and generalized formulations

CNWF sit within a broader Whitney-form ecosystem that already supplies several ingredients for higher-order, more stable, or more topologically constrained variants. One important direction is high-order approximation. A least-squares construction of trimmed polynomial Whitney spaces $\delta_0$5 uses weighted integrals over families of small simplices, producing linear systems

$\delta_0$6

for the coefficients of a high-order Whitney approximation (Bruno et al., 2024). That analysis reports Runge-like phenomena for exact interpolatory support choices, and shows that enriched support sets improve conditioning and can mitigate those phenomena in least squares (Bruno et al., 2024). This suggests a natural route to higher-order CNWF: retain the FEEC-compatible trimmed polynomial spaces and replace exact interpolation by differentiable least-squares projection layers.

Another extension concerns homotopy operators and potentials. Discrete Poincaré operators on cochains and Whitney forms satisfy exact identities such as

$\delta_0$7

for $\delta_0$8, while discrete Bogovskiĭ operators preserve homogeneous boundary conditions and satisfy

$\delta_0$9

on Whitney forms with vanishing trace (Guzmán et al., 30 Mar 2026). These constructions furnish explicit right inverses for discrete differential operators, with applications to discrete scalar and vector potentials and to the discrete wedge product of DEC (Guzmán et al., 30 Mar 2026). This suggests homotopy or inverse-divergence layers for CNWF architectures on contractible domains or collapsible complexes.

Generality with respect to mesh topology is also expanding. On cubical meshes, finite element spaces by Whitney $Z$0-forms have been constructed together with compatible discretizations, discrete de Rham complexes, and commutative diagrams (Zhang, 2024). On polyhedral dual meshes, generalized Whitney spaces connect DEC cochains to FEEC form spaces and support convergence analysis for Hodge-Laplacian problems (Guzmán et al., 13 May 2025). A plausible implication is that CNWF need not remain restricted to simplicial meshes, provided the learned representation preserves the relevant discrete complex and commuting structure.

Finally, coordinate-free geometric representations of Whitney forms generalize from Euclidean simplices to flat pseudo-Riemannian manifolds, including Minkowski spacetime, and provide an explicit characterization of Hodge-dual Whitney forms in that setting (Salamon et al., 2014). This opens a principled path to spacetime FEEC and fully covariant discretizations of classical electromagnetism. It also suggests that spacetime or Lorentz-covariant variants of CNWF are mathematically natural, although such architectures are not developed in the CNWF papers themselves.

6. Applications, empirical performance, and limitations

The published CNWF benchmarks span advection-diffusion, shock hydrodynamics, electrostatics, source localization, and battery thermal runaway. In the structure-preserving digital-twin framework, reported examples include one-dimensional and two-dimensional advection-diffusion, a spacetime Sod shock tube, electrostatics of a point charge in a conducting shell, and a battery thermal runaway problem based on LES data (Kinch et al., 9 Aug 2025). In the sensor-coverage framework, experiments are carried out on a circular domain, a Gulf of Mexico geometry using HYCOM surface velocities, and a maze geometry with Stokes flow, all for steady advection-diffusion with localized sources (Shaffer et al., 12 Sep 2025).

The battery digital twin is the most explicit large-scale benchmark. Using 25 LES simulations on a $Z$1 parameter grid in $Z$2, the CNWF model is reported to capture the transition to turbulence and achieve real-time inference $Z$3s with a speedup of $Z$4 relative to LES (Kinch et al., 9 Aug 2025). In the source-localization setting, experimental comparisons with physics-agnostic transformer architectures show improved accuracy in complex geometries when physical constraints are enforced, which the authors interpret as evidence that structure preservation provides an effective inductive bias for source identification (Shaffer et al., 12 Sep 2025).

The reported limitations are correspondingly specific. The digital-twin theory is developed for nonlinear perturbations of a Hodge Laplacian; the discrete Laplacian has a constant nullspace that must be stabilized in practice; training still depends on high-fidelity data; examples mostly use low-dimensional conditioning variables $Z$5; and mesh-quality issues are inherited from standard finite elements (Kinch et al., 9 Aug 2025). The adaptive source-localization analysis assumes two-dimensional conforming simplicial meshes, steady advection-diffusion, smooth compact training sources, divergence-free known velocity, and continuity or contractivity assumptions on the learned density field in the point-source theorem (Shaffer et al., 12 Sep 2025).

Within these limits, CNWF define a distinctive synthesis. They use the rigid local geometry of Whitney forms, the commuting structure of FEEC, and conditional transformer mechanisms to learn reduced bases and nonlinear operators without relinquishing exact discrete conservation. The foundational Whitney literature shows why the underlying representation is so constrained; the FEEC and DEC literature explains why those constraints matter for stability, commuting projections, and cohomological structure; and the CNWF literature shows how those same constraints can be turned into a learnable reduced model for real-time digital twins and adaptive sensing (Dodziuk, 2022, Guzmán et al., 13 May 2025, Kinch et al., 9 Aug 2025, Shaffer et al., 12 Sep 2025).