betaFeedback

WorksComputer Science

Structure-Preserving Neural Surrogates with Tractable Uncertainty Quantification

Abstract

Recent advances in scientific machine learning provide a means of near-real-time solution to partial differential equations (PDEs), but lack the theoretical underpinnings of conventional simulators that support contemporary verification and validation. In this work, we construct data-driven reduced-order models that serve as structure-preserving, real-time surrogates. Remarkably, the exterior calculus that imposes physical conservation structure also exposes topological structure that we use to build a Gaussian process (GP) representation of uncertainty in state-flux relationships, ultimately yielding a Dirichlet-to-Neumann map for quantities of interest with closed-form expressions for posterior uncertainty. We specifically propose structure-preserving H(div)H(\mathrm{div})–L2L^2 subspaces of conventional Raviart–Thomas and dgP0dgP_0 elements prescribed by a lightweight transformer. Reduced-order dynamics consistent with this subspace are learned by posing a conservation law in which a GP describes the fluxes between volumes. This work hinges on a novel interface between mixed FEM spaces and GP regression; when training is posed as the optimal recovery problem (ORP), the resulting GP regression can be written as an optimization problem with equality constraints that impose a conservation structure, amenable to a fast Schur-complement training strategy. The trained model can then be solved in real time with closed-form estimators for boundary fluxes driven by prescribed Dirichlet data. The paper includes RKHS posterior error bounds for linear functionals to support uncertainty quantification, as well as numerical experiments demonstrating the accuracy of the posterior distribution as a surrogate for error estimation.

Problem overview and relation to the literature

Neural operators and other machine learning surrogates are emerging as practical alternatives to classical simulation due to their computational efficiency and ability to generalize across problem instances. However, the black-box nature of these methods remains a major impediment to adoption in scientific and engineering applications and creates challenges for uncertainty quantification (UQ), where errors can propagate through the learned solution operator and further affect downstream predictions without clear traceability [2, 28]. Moreover, for systems governed by partial differential equations (PDEs), the learned surrogate models should also respect underlying physical constraints, such as conservation laws and initial and boundary conditions.

Motivated by this gap, we propose a surrogate framework for PDE-governed systems that supports fast simulation while providing tractable posterior uncertainty. Consider the following abstract problem, which we later elaborate in § 4. Let u∈Qu \in Q be a field on Ω⊂Rd\Omega\subset\mathbb{R}^{d} (naturally, Q⊂L2(Ω)Q \subset L^{2}(\Omega)), assumed to satisfy an unknown conservation law ∇⋅F(u)=f\nabla\cdot F(u) = f, where F∈VF \in V is a given flux function (naturally, V⊂H(div⁡;Ω)V \subset H(\operatorname{div};\Omega)). We assume access to data consisting of sampled state and flux fields uu and FF. Additionally, we allow for parametric dependence on a conditioning variable ZZ, which may represent constitutive relationships, geometry, or other factors, yielding the training set

DN={(Z(n),udata(n),Fdata(n))}n=1N.\mathcal{D}_{N} = \left\{\left(Z^{(n)}, u_{\mathrm{data}}^{(n)}, F_{\mathrm{data}}^{(n)}\right)\right\}_{n=1}^{N}.

Our goal is to identify a functional form for FF that is consistent with the available data, supports tractable posterior estimation, and generalizes to boundary data not seen during training. This learning problem must be posed in an appropriate discrete setting so that the resulting surrogate preserves the structure of the underlying PDE. We assume that the flux can be decomposed in the form

F=−ϵ∇u+N[u],∇⋅F=f,(1)F = -\epsilon\nabla u + \mathcal{N}[u], \qquad\nabla\cdot F = f, \tag*{(1)}

consisting of a diffusion term with ϵ>0\epsilon> 0 for numerical stability and a nonlinear correction N\mathcal{N} that identifies a state-to-flux map from data. In previous work we demonstrated how neural operators can serve as N\mathcal{N} [21]; here we instead treat N\mathcal{N} as a Gaussian process (GP). Classical GPs are most tractable in small-data, low-dimensional settings, and a primary contribution of this work is that the reduced graph representation developed below brings the otherwise infinite-dimensional state-to-flux learning problem into precisely such a regime.

The proposed method separates the learning task into two parts: a data-driven model mapping the fine-scale space to a coarse-scale space (P1), and a stochastic model for flux correction that supports uncertainty quantification (P2). Because the full construction is technical, we first summarize the salient features of each.

In P1, we work with low-order Whitney forms [4] that provide conforming finite element spaces encoding the topological and cohomological properties tied to conservation structure. Here the Raviart–Thomas and discontinuous-piecewise-constant (RT0/dgP0\mathrm{RT}_0/\mathrm{dgP}_0) pair is naturally conforming, with RT0⊂H(div⁡;Ω)\mathrm{RT}_0 \subset H(\operatorname{div}; \Omega) and dgP0⊂L2(Ω)\mathrm{dgP}_0 \subset L^2(\Omega), and has the surjectivity property ∇⋅:RT0→dgP0\nabla\cdot: \mathrm{RT}_0 \to\mathrm{dgP}_0. Because this space interpolates cell-based scalar degrees of freedom and facet-based flux degrees of freedom, it provides a discrete divergence theorem, allowing the discrete div/grad matrices to be interpreted as adjacency matrices between cells and facets. This dual interpretation of div/grad is the crucial ingredient for Bayesian analysis on fields: in previous work [29, 30] we showed how to use optimal recovery to learn circuit models with tractable posteriors, and this linkage lets us extend the analysis to finite element fields.

The goal is to identify reduced-order spaces Qhc(z;θ)⊆QhfQ_h^c(z;\theta) \subseteq Q_h^f and Vhc(z;θ)⊆VhfV_h^c(z;\theta) \subseteq V_h^f, where the superscripts ⋅f\cdot^f and ⋅c\cdot^c denote fine and coarse spaces and the subscript ⋅h\cdot_h denotes discretization. To preserve exterior calculus structure in the reduced space, we require a construction in which the restrictions rQ:Qhf→Qhcr_Q : Q_h^f \to Q_h^c and rV:Vhf→Vhcr_V : V_h^f \to V_h^c commute with the divergence operator, so that ∇⋅Vhc⊆Qhc\nabla\cdot V_h^c \subseteq Q_h^c. To achieve this, we first design a transformer that evaluates rQr_Q by outputting a coarsening matrix WW, so that Qhc=span⁡{∑aWiaχa}iQ_h^c = \operatorname{span}\{\sum_a W_{ia}\chi_a\}_i, where the χa\chi_a are the indicator functions over the cells spanning QhfQ_h^f. We then design VhcV_h^c by identifying coarsened degrees of freedom associated with coarse facets shared between coarse cells. This novel construction represents a “top-down” coarsening of the (Λd/Λd−1)(\Lambda_d/\Lambda_{d-1}) subcomplex, in contrast to the “bottom-up” coarsening of (Λ0/Λ1)(\Lambda_0/\Lambda_1) developed previously [3]. P1 therefore produces a reduced H(div⁡)H(\operatorname{div})-conforming finite element space that can be conditioned on ZZ.

In P2, we pose the discrete representation of the physics as an optimal recovery problem [29]. We identify F∈VhcF \in V_h^c and u∈Qhcu \in Q_h^c by their basis coefficients F^\hat{F} and u^\hat{u}, where each coarse flux degree of freedom F^ij\hat{F}_{ij} is associated with the coarse oriented boundary shared by the coarse cells u^i\hat{u}_i and u^j\hat{u}_j in the graph induced by P1. We model the nonlinearity edgewise by a Gaussian process, N=GP(uij)\mathcal{N} = GP(u_{ij}), where uiju_{ij} concatenates the states u^i\hat{u}_i, u^j\hat{u}_j with the additional metric features needed to generalize the flux law across the mesh.

In traditional GP regression, the posterior distribution is derived from the joint normal distribution via a Schur complement. Its mean may equivalently be obtained by solving the optimal recovery problem: given a kernel KK with reproducing kernel Hilbert space (RKHS) HK\mathcal{H}_K and NN noisy data pairs (X,Y)={(xi,yi)}i=1N(\mathbf{X},\mathbf{Y) = \{(x_i,y_i)\}_{i=1}^{N}} with noise variance σε2\sigma_{\varepsilon}^{2}, the minimizer of the RKHS norm penalized by data misfit

f^=arg min⁡f∈HK ∥f∥HK2+1σε2∑i=1N(f(xi)−yi)2,(2)\hat{f} = \underset{f\in\mathcal{H}_K}{\operatorname{arg\,min}}\ \lVert f\rVert_{\mathcal{H}_K}^{2} + \frac{1}{\sigma_{\varepsilon}^{2}}\sum_{i=1}^{N}(f(x_i)-y_i)^2, \tag*{(2)}

recovers the conventional GP estimator f^(⋅)=K(⋅,X)(K(X,X)+σε2I)−1Y\hat{f}(\cdot) = K(\cdot,\mathbf{X})\left(K(\mathbf{X},\mathbf{X})+\sigma_{\varepsilon}^{2}I\right)^{-1}\mathbf{Y}.

Recasting the conventional GP problem in this manner allows us to incorporate equality constraints directly. In the graph interpretation of P1, fluxes over VhcV_h^c are edge currents between nodes associated with states in QhcQ_h^c. On each edge we cast an optimal recovery problem mapping state to flux, subject to the equality constraint that the conservation law holds at each node. The connectivity of the coarse bases imposes sparsity on this dense graph through a metric weighting, and after training we are left with a discrete boundary value problem whose boundary fluxes are encoded by GPs. Because this is a constrained optimization problem, its KKT conditions expose a saddle-point structure that we exploit to derive a fast optimizer (Subsection 4.2).

In concert, P1 and P2 amount to a graph discovery problem: by concurrently training the transformer and solving the optimal recovery problem, we interpret the identification of reduced div/grad finite element spaces as the identification of a dense graph, sparsified by the connectivity of the bases, that encodes the flow of conserved quantities through the system (Figure 1). While circuit analogies are commonplace in engineering (e.g., compact models for semiconductor devices [1, 13], hydraulic circuits in hydrodynamics [34, 38], or thermal circuits for heat transfer [39]), they are typically built from simplified analytic solutions and empirical curve-fitting through a slow iterative process. In contrast, our approach can be carried out autonomously to obtain rapid, uncertainty-quantified surrogates that preserve a conservative input/output relationship. We demonstrate this by considering two representative examples. Advection-diffusion on a triangular mesh of Philadelphia’s Liberty Bell serves to illustrate how real-time surrogates with quantified uncertainty can be constructed on complex geometries; in this setting, we condition on the direction of advection as an example of how to construct parametric models. We finally construct a digital twin of a semiconductor device (specifically, a p−np-n diode). Training data can be constructed by TCAD simulations solving the device drift-diffusion equations [27], providing examples of internal device transport and how it defines the voltage-current relationship governing the device. This problem specifically shows how the UQ developed here identifies the range of inputs over which the learned surrogate can be trustworthy.

Image not extracted

Figure 1. Roadmap to structure-preserving surrogates with quantified uncertainty. The transformer learns H(div⁡)H(\operatorname{div})-conforming bases and constructs a conservative coarse graph adapted to the conditioning ZZ (Boxes ①–③). The graph edges carry GP models of the state-to-flux laws while conservation is imposed as an exact divergence constraint (Box ④) and the resulting reduced model serves as a Dirichlet-to-Neumann surrogate with posterior error estimates for boundary fluxes (Box ⑤).

Relation to literature

The construction above brings together three goals usually pursued separately: fast reduced-order simulation, exact enforcement of physical structure, and uncertainty quantification. We review each in turn. Classical discretizations such as finite element methods yield accurate, verifiable predictions. but at substantial cost, motivating a large body of scientific machine learning (SciML) that trades rigor for speed. Physics-informed neural networks (PINNs) embed the governing PDEs in the training loss [20, 43]; neural operators such as DeepONet [24, 40] and Fourier neural operators [23] learn maps between function spaces; and data-driven reduced-order models (ROMs) accelerate parametric simulation by working in a low-dimensional subspace [14, 18, 19]. These surrogates can be fast, but typically forfeit the structural guarantees and error control of the solvers they replace. A related line of structure-preserving model-reduction methods addresses part of this gap by seeking reduced spaces that retain conservation, stability, passivity, or compatibility properties of the full-order discretization [6, 10, 32]. Our construction is close in spirit to this line of work, but learns the compatible reduced H(div⁡)H(\operatorname{div})–L2L^2 complex from data rather than selecting it through a fixed projection, and couples it to a GP optimal-recovery formulation for posterior uncertainty.

Enforcing physical structure reliably is the central difficulty, particularly under limited data and complex geometry. Most physics-informed approaches impose conservation through soft penalties on the PDE residual or boundary conditions [11, 17]; these do not guarantee exact conservation, and the resulting violations can accumulate, amplify errors, and destabilize training on stiff or multiscale problems [8, 22]. Structure-preserving methods instead build the constraints into the function spaces themselves. The mixed finite element theory underlying Raviart-Thomas and discontinuous Galerkin pairs provides the classical foundation for such H(div⁡)H(\operatorname{div})-conforming flux approximations and compatible L2L^2 scalar spaces; see, for example, [7]. Finite element exterior calculus (FEEC) provides a canonical framework for this approach, choosing spaces and operators that form a discrete de Rham complex and so inherit the topological identities underlying conservation [4, 5]. Data-driven exterior calculus extends this to learned, graph-based operators while preserving exact-sequence compatibility [37]. Our coarsened RT0/dgP0\mathrm{RT}_0/\mathrm{dgP}_0 complex is designed to inherit these guarantees by construction.

Finally, trustworthy surrogates require calibrated uncertainty estimates that can be propagated to downstream quantities of interest [2, 31, 41]. UQ for neural PDE surrogates is often pursued through ensembles, Bayesian neural networks, dropout, latent-variable models, or post-hoc calibration; these approaches can be effective but usually do not yield the closed-form functional posterior bounds available in GP/RKHS optimal recovery [36, 42]. Gaussian processes provide a natural alternative because conditioning a GP prior on data yields a posterior mean and covariance in closed-form. The same estimate arises from the optimal recovery problem, which characterizes the minimum-norm RKHS interpolant of noisy data [12, 29, 30] and, as we exploit below, accommodates hard linear constraints. We build most directly on the graph-based optimal recovery of Dirichlet-to-Neumann maps in [30], extending it from fixed graphs to the learned, conditioned finite element complexes produced by P1.

Main contributions

Figure 1 summarizes the resulting framework with structure preservation and uncertainty quantification. Our specific contributions are:

  • a top-down coarsening of the (Λd/Λd−1)(\Lambda_d/\Lambda_{d-1}) subcomplex, in which a lightweight transformer emits a coarsening matrix WW that defines a reduced, H(div⁡)H(\operatorname{div})-conforming RT0/dgP0\mathrm{RT}_0/\mathrm{dgP}_0 pair conditioned on the problem parameters ZZ, in contrast to the bottom-up (Λ0/Λ1)(\Lambda_0/\Lambda_1) coarsening of [3];

  • a structure-preserving optimal recovery formulation in which an edgewise Gaussian process models the state-to-flux law and conservation is imposed

exactly through linear equality constraints, yielding a saddle-point KKT system with a fast Schur-complement solve;

  • a joint “graph discovery” training procedure that learns the reduced complex and the flux Gaussian process concurrently, producing a sparse circuit model of the conserved dynamics that generalizes across geometries and boundary data;

  • a closed-form RKHS posterior error bound for linear functionals of the flux, in particular the boundary fluxes defining the Dirichlet-to-Neumann map, validated on complex geometries and semiconductor device equations.

The remainder of the paper is organized as follows. Section 2 reviews the necessary mathematical background for the proposed framework. Subsection 2.1 reviews the finite element exterior calculus preliminaries for the H(div⁡)H(\operatorname{div})–L2L^2 mixed setting and Subsection 2.2 specifies the finite element space with conforming RT0/dgP0\mathrm{RT}_0/\mathrm{d}gP_0 discretizations. Subsection 2.3 explains the optimal recovery problem for Gaussian processes. Section 3 develops the first component (P1) by introducing the data-driven PoU Whitney construction (Subsection 3.1) and the induced coarse operators (Subsection 3.2) for structure preservation. We show that under this construction, properties such as surjectivity and flux balance still hold in the reduced space (Subsection 3.3). In addition, we provide a graph interpretation of the learned reduced spaces (Subsection 3.4), which supports the subsequent optimal recovery formulation. Section 4 develops the second component (P2). We cast the GP optimal recovery formulation on the coarse spaces subject to equality constraints that guarantee the exact enforcement of conservation laws (Subsection 4.1). In addition to a fast Schur-complement solve for the KKT system, we propose a bilevel training procedure for efficient training with nested coarse variables and models (Subsection 4.2) and incorporate geometric embedding (Subsection 4.3). We also provide the theoretical analysis for the RKHS posterior error bound in Subsection 4.4. Section 5 reports numerical experiments on representative examples, a complex-geometry advection-diffusion problem, and a semiconductor diode example to validate the proposed framework.

Mathematical preliminaries

In this section we gather the necessary background in both finite element exterior calculus (FEEC) and the optimal recovery formalisms. For further background, please see [4, 5, 25] and [37].

Finite element exterior calculus (FEEC)

Exterior calculus generalizes classical calculus to differential forms of higher degree on differentiable manifolds and provides a robust mathematical framework for analyzing PDEs on different geometries. Finite element exterior calculus extends the exterior calculus to discrete finite element meshes using chain complexes and preserves crucial topological and geometric structures underlying the governing PDEs [21, 37].

One advantage of FEEC is that it unifies the main differential operators, i.e., grad, div, and curl, with the de Rham complex. Let Λ(Ω)\Lambda(\Omega) denote the exterior algebra with exterior product ∧\wedge and Ω⊂Rd\Omega\subset\mathbb{R}^{d}. We then have the de Rham complex:

0→Λ0(Ω)→d0Λ1(Ω)→d1⋯→dk−1Λk(Ω)→0,(3)0 \to\Lambda^{0}(\Omega) \xrightarrow{d^{0}} \Lambda^{1}(\Omega) \xrightarrow{d^{1}} \cdots\xrightarrow{d^{k-1}} \Lambda^{k}(\Omega) \to0, \tag*{(3)}

where d:Λk−1(Ω)→Λk(Ω)d:\Lambda^{k-1}(\Omega) \to\Lambda^{k}(\Omega) is the exterior derivative operator. For Ω⊂R3\Omega\subset\mathbb{R}^{3}, the de Rham complex becomes:

0→C∞(Ω)→grad⁡[C∞(Ω)]3→curl⁡[C∞(Ω)]3→div⁡C∞(Ω)→0.(4)0 \to C^{\infty}(\Omega) \xrightarrow{\operatorname{grad}} [C^{\infty}(\Omega)]^{3} \xrightarrow{\operatorname{curl}} [C^{\infty}(\Omega)]^{3} \xrightarrow{\operatorname{div}} C^{\infty}(\Omega) \to0. \tag*{(4)}

If we further consider Sobolev spaces with homogeneous Dirichlet boundary conditions, we can derive the FEEC specialization primal and dual cochain complex:

0⟶H0(grad⁡,Ω)⇄grad⁡−grad⁡H0(curl⁡,Ω)⇄curl⁡curl⁡H0(div⁡,Ω)⇄div⁡−div⁡L2(Ω)⟶0.(5)0 \longrightarrow H_{0}(\operatorname{grad},\Omega) \underset{-\operatorname{grad}}{\overset{\operatorname{grad}}{\rightleftarrows}} H_{0}(\operatorname{curl},\Omega) \underset{\operatorname{curl}}{\overset{\operatorname{curl}}{\rightleftarrows}} H_{0}(\operatorname{div},\Omega) \underset{-\operatorname{div}}{\overset{\operatorname{div}}{\rightleftarrows}} L^{2}(\Omega) \longrightarrow0. \tag*{(5)}

In this work, instead of working on “bottom-up” subcomplexes, we work on the terminal map H(div⁡,Ω)→∇⋅L2(Ω)H(\operatorname{div},\Omega) \xrightarrow{\nabla\cdot} L^{2}(\Omega) because this is a natural algebraic mechanism through which flux balance is imposed. This is illustrated in the commuting diagram in Figure 2.

Image not extracted

Figure 2. Construction of H(div⁡)H(\operatorname{div})-conforming reduced subspace. We design a transformer that outputs a subspace of the de Rham complex appropriate for strongly imposing conservation laws. While previous work builds “bottom-up” dual Λ0/Λ1\Lambda_0/\Lambda_1 reduced subcomplexes [3], this work provides a “top-down” primal subcomplex on Λd/Λd−1\Lambda_d/\Lambda_{d-1}. A transformer prescribes the restriction map rQr_Q conditioned on ZZ, and we provide a compatible construction of rVr_V that defines coarsened RT0−dgP0\mathrm{RT}_0-\mathrm{dgP}_0 subspaces which preserve surjectivity of the divergence.

H(div) space and lowest-order Raviart–Thomas element

Since one core objective of this work is to guarantee the flux conservation law, instead of working on the entire de Rham complex, we focus on its tail, i.e., H(div⁡,Ω)→div⁡L2(Ω)H(\operatorname{div},\Omega) \xrightarrow{\operatorname{div}} L^{2}(\Omega). This segment of the complex supports flux conservation in a natural way through the mixed formulation of elliptic problems. To make this connection explicit, consider the general second-order elliptic equation in a bounded domain Ω⊂Rd\Omega\subset\mathbb{R}^{d}:

−∇⋅(K(x)∇u)=fin Ω,(6)-\nabla\cdot\left(\mathbf{K}(x)\nabla u\right)=f \qquad\text{in }\Omega, \tag*{(6)}

where K(x)∈Rd×d\mathbf{K}(x) \in\mathbb{R}^{d\times d} is a symmetric and uniformly positive definite diffusion tensor. The boundary ∂Ω\partial\Omega is decomposed into disjoint parts ∂Ω=ΓD∪ΓN\partial\Omega=\Gamma_{D}\cup\Gamma_{N} on which Dirichlet and Neumann boundary conditions are defined, respectively. Then by introducing the flux variable F:=−K(x)∇u\mathbf{F}:=-\mathbf{K}(x)\nabla u, the problem can be rewritten as the first-order system:

\begin{cases} \mathbf{K}(x)^{-1}\mathbf{F}+\nabla u=0,\\ \nabla\cdot\mathbf{F}=f, \end{cases} \qquad\text{in }\Omega. \tag*{(7)}\end{cases}

In this setting, the natural space for the flux variable is:

H(div⁡,Ω):={v∈[L2(Ω)]d∣∇⋅v∈L2(Ω)}.(8)H(\operatorname{div},\Omega):=\left\{\mathbf{v}\in[L^{2}(\Omega)]^{d}\mid\nabla\cdot\mathbf{v}\in L^{2}(\Omega)\right\}. \tag*{(8)}

This space ensures that the divergence operator is well-defined in a weak sense and the normal components v⋅n\mathbf{v}\cdot\mathbf{n} are well-defined on element interfaces, thus naturally making it the appropriate space for modeling local flux conservation. The corresponding mixed variational formulation is to find (F,u)∈H(div⁡,Ω)×L2(Ω)(\mathbf{F},u)\in H(\operatorname{div},\Omega)\times L^{2}(\Omega) such that

∫ΩK(x)−1F⋅v dx−∫Ωu(∇⋅v) dx=−∫ΓDuD(v⋅n) ds,∀v∈H(div⁡,Ω),\int_{\Omega}\mathbf{K}(x)^{-1}\mathbf{F}\cdot\mathbf{v}\,dx-\int_{\Omega}u\left(\nabla\cdot\mathbf{v}\right)\,dx=-\int_{\Gamma_{D}}u_{D}\left(\mathbf{v}\cdot\mathbf{n}\right)\,ds,\qquad\forall\mathbf{v}\in H(\operatorname{div},\Omega),
∫Ω(∇⋅F)q dx=∫Ωfq dx,∀q∈L2(Ω).(9)\int_{\Omega}\left(\nabla\cdot\mathbf{F}\right)q\,dx=\int_{\Omega}fq\,dx,\qquad\forall q\in L^{2}(\Omega). \tag*{(9)}

A conforming discretization of this mixed formulation requires a finite element flux space in H(div⁡,Ω)H(\operatorname{div},\Omega) and a compatible scalar space in L2(Ω)L^{2}(\Omega). In this work, we consider the Raviart–Thomas (RT) elements for the flux variable and piecewise constants for the scalar variable. Let Th\mathcal{T}_{h} be a simplicial mesh of Ω⊂Rd\Omega\subset\mathbb{R}^{d}. On each element K∈ThK\in\mathcal{T}_{h}, the Raviart–Thomas space of order k≥0k\geq0 is defined by

RT⁡k(K):=Pk(K)d+xPk(K),(10)\operatorname{RT}_{k}(K):=\mathcal{P}_{k}(K)^{d}+\mathbf{x}\mathcal{P}_{k}(K), \tag*{(10)}

where Pk(K)\mathcal{P}_{k}(K) denotes the space of polynomials of degree at most kk on KK. The global Raviart–Thomas space is defined as

RT⁡k(Th):={v∈H(div⁡,Ω)∣v∣K∈RT⁡k(K), ∀K∈Th},(11)\operatorname{RT}_{k}(\mathcal{T}_{h}):=\left\{\mathbf{v}\in H(\operatorname{div},\Omega)\mid\mathbf{v}|_{K}\in\operatorname{RT}_{k}(K),\ \forall K\in\mathcal{T}_{h}\right\}, \tag*{(11)}

which enforces the continuity of normal components across element interfaces. For k=0k = 0, the lowest-order Raviart–Thomas space on KK is defined as

RT⁡0(K)=P0(K)d+xP0(K).(12)\operatorname{RT}_{0}(K) = P_{0}(K)^{d} + x P_{0}(K). \tag*{(12)}

Equivalently, any v∈RT⁡0(K)\mathbf{v} \in\operatorname{RT}_{0}(K) can be written as v(x)=a+px\mathbf{v}(x) = \mathbf{a} + p x with constant vector a∈Rd\mathbf{a} \in\mathbb{R}^{d} and constant scalar p∈Rp \in\mathbb{R}. The degrees of freedom of this lowest-order RT⁡0\operatorname{RT}_{0} space are defined by the normal fluxes across element faces:

∫e(v⋅n) q ds,∀e⊂∂K.(13)\int_{e} (\mathbf{v} \cdot\mathbf{n}) \, q \, ds, \qquad\forall e \subset\partial K. \tag*{(13)}

Thus the degree of freedom admits a direct interpretation in terms of cochains. In particular, the RT space ensures that the discrete divergence operator is compatible with the underlying differential complex, and the (RT⁡0,P0)(\operatorname{RT}_{0}, P_{0}) can form a meaningful discretization of the H(div⁡)H(\operatorname{div})–L2L^{2} mixed formulation. Within the FEEC framework, this construction corresponds to the final part of the de Rham complex where the scalar field is associated with dd-forms and the flux corresponds to (d−1)(d - 1)-forms. The gradient and divergence operators are unified through the exterior derivative and the codifferential. The resulting compatible structure provides the algebraic foundation for conservation and motivates the learnable coarse space construction below.

Optimal recovery problem for Gaussian processes

Gaussian processes provide a probabilistic model for unknown functions. Formally, a GP is a collection of random variables such that any finite collection has a joint Gaussian distribution. Equivalently, a GP can be viewed as a distribution over functions: each draw from the process is a possible function, and for any finite set of input locations, the corresponding vector of function values is multivariate Gaussian. This makes GPs particularly useful in settings where the uncertainty quantification is important, since conditioning on observed data yields both a posterior mean prediction and a posterior covariance that quantifies uncertainty at new inputs.

Given data D=(X,Y)\mathcal{D} = (\mathbf{X}, \mathbf{Y}) with X=[x1,…,xN]T\mathbf{X} = [\mathbf{x}_{1}, \ldots, \mathbf{x}_{N}]^{\mathsf{T}} and corresponding output Y=[y1,…,yN]T\mathbf{Y} = [y_{1}, \ldots, y_{N}]^{\mathsf{T}} generated by an underlying function yi=f(xi)y_i = f(\mathbf{x}_i), a GP prior on ff is fully specified by its mean function m(x)m(x) and covariance kernel K(x,x′)K(x, x'):

m(x)=E[f(x)](14)m(x) = \mathbb{E}[f(x)] \tag*{(14)}
K(x,x′)=E[(f(x)−m(x))(f(x′)−m(x′))](15)K(x, x') = \mathbb{E}[(f(x) - m(x))(f(x') - m(x'))] \tag*{(15)}
f(x)∼GP(m(x),K(x,x′))(16)f(x) \sim\mathcal{GP}(m(x), K(x, x')) \tag*{(16)}

In the noise-free scenario, for test inputs X∗=(x1∗,…,xN∗∗)T\mathbf{X}_{*} = (\mathbf{x}_{1}^{*}, \ldots, \mathbf{x}_{N_{*}}^{*})^{\mathsf{T}}, the joint distribution of the observed training values Y=f(X)\mathbf{Y} = f(\mathbf{X}) and the unobserved test values f∗=f(X∗)\mathbf{f}_{*} = f(\mathbf{X}_{*}) is

\begin{bmatrix} \mathbf{Y} \\ \mathbf{f}_{*} \end{bmatrix} \sim\mathcal{N}\left( \mathbf{0}, \begin{bmatrix} K(\mathbf{X}, \mathbf{X}) & K(\mathbf{X}, \mathbf{X}_{*}) \\ K(\mathbf{X}_{*}, \mathbf{X}) & K(\mathbf{X}_{*}, \mathbf{X}_{*}) \end{bmatrix} \right). \tag*{(17)}\end{bmatrix}

Conditioning on this joint Gaussian distribution gives

f∗∣X∗,X,Y∼N(μ∗,Σ∗).(18)\mathbf{f}_{*} \mid\mathbf{X}_{*}, \mathbf{X}, \mathbf{Y} \sim\mathcal{N}(\boldsymbol{\mu}_{*}, \boldsymbol{\Sigma}_{*}). \tag*{(18)}

where

μ∗=K(X∗,X)K(X,X)−1Y,(19)\mu_* = K(\mathbf{X}_*,\mathbf{X})K(\mathbf{X},\mathbf{X})^{-1}\mathbf{Y}, \tag*{(19)}
Σ∗=K(X∗,X∗)−K(X∗,X)K(X,X)−1K(X,X∗).(20)\Sigma_* = K(\mathbf{X}_*,\mathbf{X}_*) - K(\mathbf{X}_*,\mathbf{X})K(\mathbf{X},\mathbf{X})^{-1}K(\mathbf{X},\mathbf{X}_*). \tag*{(20)}

In most applications, however, the data are noisy. Suppose Y=f(X)+ε\mathbf{Y}=f(\mathbf{X})+\boldsymbol{\varepsilon}, where ε∼N(0,σε2I)\boldsymbol{\varepsilon}\sim\mathcal{N}(0,\sigma_{\varepsilon}^{2}I) is independent identically distributed (i.i.d.) Gaussian noise. Then the joint distribution in (17)–(18) becomes

\begin{bmatrix}\mathbf{Y}\\ \mathbf{f}_*\end{bmatrix} \sim\mathcal{N}\left(0, \begin{bmatrix} K(\mathbf{X},\mathbf{X})+\sigma_{\varepsilon}^{2}I & K(\mathbf{X},\mathbf{X}_*)\\ K(\mathbf{X}_*,\mathbf{X}) & K(\mathbf{X}_*,\mathbf{X}_*) \end{bmatrix} \right), \tag*{(21)}\end{bmatrix}
f∗∣X∗,X,Y∼N(K(X∗,X)(K(X,X)+σε2I)−1Y,K(X∗,X∗)−K(X∗,X)(K(X,X)+σε2I)−1K(X,X∗)).(22)f_* \mid\mathbf{X}_*,\mathbf{X},\mathbf{Y} \sim\mathcal{N}\left(K(\mathbf{X}_*,\mathbf{X})(K(\mathbf{X},\mathbf{X})+\sigma_{\varepsilon}^{2}I)^{-1}\mathbf{Y}, K(\mathbf{X}_*,\mathbf{X}_*)-K(\mathbf{X}_*,\mathbf{X})(K(\mathbf{X},\mathbf{X})+\sigma_{\varepsilon}^{2}I)^{-1}K(\mathbf{X},\mathbf{X}_*)\right). \tag*{(22)}

The posterior covariance provides a meaningful probabilistic measure of uncertainty at the test points, which is useful for uncertainty-aware scientific machine learning tasks. Indeed, prior work has established that incorporating GPs within SciML frameworks provides rigorous uncertainty quantification for learning complex PDE-governed systems [9, 12, 16, 35].

Given limited noisy data and some prior knowledge of the function space, the optimal recovery problem aims to learn an unknown function by minimizing the worst-case error. Mathematically, given a normed space F\mathcal{F} and an unknown function f∈Ff\in\mathcal{F}, the partial knowledge of ff is defined through point evaluations:

yi=ℓi(f),i=1,…,m,(23)y_i=\ell_i(f),\qquad i=1,\ldots,m, \tag*{(23)}

for some linear functionals ℓ1,…,ℓm∈F∗\ell_1,\ldots,\ell_m\in\mathcal{F}^*, where F∗\mathcal{F}^* is the dual space of F\mathcal{F}. The objective is to approximate ff by some estimate f^∈F\widehat{f}\in\mathcal{F} such that the worst-case error is minimized:

E(f^):=inf⁡f^∈Fℓi(f)=yisup⁡f∈F∥f−f^∥∥f∥.\mathcal{E}(\widehat{f}) := \inf_{\substack{\widehat{f}\in\mathcal{F}\\ \ell_i(f)=y_i}} \sup_{f\in\mathcal{F}} \frac{\lVert f-\widehat{f}\rVert}{\lVert f\rVert}.

Let K:X×X→RK:\mathcal{X}\times\mathcal{X}\to\mathbb{R} be a symmetric positive definite bivariate kernel, and HK\mathcal{H}_K denote its reproducing kernel Hilbert space (RKHS) with accompanying induced RKHS norm ∥⋅∥K\lVert\cdot\rVert_K. Given noisy observations (X,Y)(\mathbf{X},\mathbf{Y}), the corresponding regularized optimal recovery problem is to find the function f^∈HK\widehat{f}\in\mathcal{H}_K that minimizes the sum of the squared RKHS norm and the scaled data-misfit term:

J(X;Y)=min⁡g∈HK∥g∥K2+1σε2∥g(X)−Y∥22.\mathcal{J}(\mathbf{X};\mathbf{Y})=\min_{g\in\mathcal{H}_K}\lVert g\rVert_K^2+\frac{1}{\sigma_{\varepsilon}^{2}}\lVert g(\mathbf{X})-\mathbf{Y}\rVert_2^2.

By the Representer Theorem, the minimizer f^\widehat{f} has a finite-dimensional representation in terms of kernel evaluations at the training inputs. In particular, f^\widehat{f} evaluated at a new point can be written as

f^(⋅)=K(⋅,X)(K(X,X)+σε2I)−1Y.\widehat{f}(\cdot)=K(\cdot,\mathbf{X})(K(\mathbf{X},\mathbf{X})+\sigma_{\varepsilon}^{2}I)^{-1}\mathbf{Y}.

This estimator coincides with the posterior mean of Gaussian process regression with covariance kernel KK and independent Gaussian observation noise of variance σε2\sigma_{\varepsilon}^{2}. The corresponding minimum value of the objective J(X;Y)\mathcal{J}(\mathbf{X};\mathbf{Y}) is

J(X;Y)=YT(K(X,X)+σε2I)−1Y,\mathcal{J}(\mathbf{X};\mathbf{Y})=\mathbf{Y}^{\mathsf{T}}(K(\mathbf{X},\mathbf{X})+\sigma_{\varepsilon}^{2}I)^{-1}\mathbf{Y},

proved in [30].

Method part 1: Data-driven H(div⁡)H(\operatorname{div})-conforming reduced space construction

In this section, we construct the structure-preserving reduced spaces used by the surrogate. The goal is to learn the data-driven Whitney forms from a trainable partition of unity (PoU) conditioned on ZZ. The PoU functions define coarse control volumes, represented by 0-forms, while the associated 1-forms encode generalized fluxes between them (Subsection 3.1). This construction yields reduced scalar and flux spaces compatible with the H(div⁡)H(\operatorname{div})–L2L^2 structure introduced above, allowing the resulting H(div⁡)H(\operatorname{div})-conforming finite element spaces to adapt to the conditioning variable ZZ (Subsections 3.2 and 3.3). Finally, we connect the induced coarse spaces to a graph interpretation in Subsection 3.4 that will support the optimal recovery formulation.

PoU Whitney forms and coarse spaces

We construct the learnable coarse spaces from the fine-scale RT0/dgP0RT_0/dgP_0 discretization. Consider the fine-scale piecewise-constant scalar space QhfQ_h^f and the corresponding fine-scale lowest-order RT0RT_0 flux space VhfV_h^f defined as

Qhf:=span⁡{ϕaP0:a∈Ch},Vhf:=span⁡{ϕeRT:e∈Eh},(24)Q_h^f := \operatorname{span}\{\phi_a^{P_0}: a \in\mathcal{C}_h\}, \qquad V_h^f := \operatorname{span}\{\phi_e^{RT}: e \in\mathcal{E}_h\}, \tag*{(24)}

where Ch={1,…,Ncell}\mathcal{C}_h = \{1,\ldots,N_{\mathrm{cell}}\} denotes the set of fine cell indices and Eh\mathcal{E}_h is the set of edges for fine RT0RT_0 flux degrees of freedom. The basis {ϕaP0}\{\phi_a^{P_0}\} is chosen so that

∑a∈ChϕaP0(x)=1.\sum_{a\in\mathcal{C}_h}\phi_a^{P_0}(x)=1.

In this work, we only use the 0-form and 1-form components needed for the H(div⁡)−L2H(\operatorname{div})-L^2 tail. Furthermore, instead of directly constructing the Whitney forms using the classic barycentric interpolant, we employ neural networks to construct neural Whitney forms that still possess the desired PoU property.

Let N0N_0 denote the number of partitions. For a fixed conditioning variable, let W=[Wia]∈RN0×NcellW=[W_{ia}]\in\mathbb{R}^{N_0\times N_{\mathrm{cell}}} be a learnable PoU weight matrix satisfying:

Wia≥0,∑i=1N0Wia=1for every a∈Ch.(25)W_{ia}\geq0,\qquad\sum_{i=1}^{N_0}W_{ia}=1\quad\text{for every }a\in\mathcal{C}_h. \tag*{(25)}

Then the coarse 0-forms and 1-forms are constructed as follows.

Proposition 3.1 (Coarse 0-forms). For a fixed conditioning variable, let WW be the learned PoU weights as in (25). The coarse 0-form is defined as

ψi0=∑a∈ChWiaϕaP0,i=1,…,N0,(26)\psi_i^0=\sum_{a\in\mathcal{C}_h}W_{ia}\phi_a^{P_0},\qquad i=1,\ldots,N_0, \tag*{(26)}

which forms a partition of unity:

∑i=1N0ψi0=∑a∈Ch(∑i=1N0Wia)ϕaP0=1.(27)\sum_{i=1}^{N_0}\psi_i^0=\sum_{a\in\mathcal{C}_h}\left(\sum_{i=1}^{N_0}W_{ia}\right)\phi_a^{P_0}=1. \tag*{(27)}

The corresponding coarse P0P_0 space is defined as:

Qhc:=W0=span⁡{ψi0}i⊂Qhf.(28)Q_h^c:=\mathcal{W}^0=\operatorname{span}\{\psi_i^0\}_i\subset Q_h^f. \tag*{(28)}

Next, we construct the coarse 1-forms that still follow the same PoU Whitney principle with a special design of separating the interior edges and boundary edges. This separation allows us to maintain the boundary information more efficiently during the coarsening procedure, especially for scenarios where the geometry is complex or where boundary conditions are different for boundary subdomains.

Proposition 3.2 (Interior and boundary coarse 1-forms). For a fixed conditioning variable, let WW be the learned PoU weights as in Equation (25). Let e∈Einte \in\mathcal{E}_{\mathrm{int}} denote the interior fine edges with two adjacent cells KL(e)K_{L}(e), KR(e)K_{R}(e). For each pair of coarse 00-forms (i,j)(i,j) with 1≤i<j≤N01 \le i < j \le N_{0}, the coarse interior 11-form is defined as

ψij1,int(x)=∑e∈Eint(Wi,KL(e)Wj,KR(e)−Wj,KL(e)Wi,KR(e))ϕe1(x).(29)\psi_{ij}^{1,\mathrm{int}}(x)=\sum_{e\in\mathcal{E}_{\mathrm{int}}}\left(W_{i,K_{L}(e)}W_{j,K_{R}(e)}-W_{j,K_{L}(e)}W_{i,K_{R}(e)}\right)\phi_{e}^{1}(x). \tag*{(29)}

Let the boundary fine edges be partitioned into disjoint groups

Ebc=E1∪⋯∪Er,Eγ∩Eγ′=∅(γ≠γ′),\mathcal{E}_{\mathrm{bc}}=\mathcal{E}_{1}\cup\cdots\cup\mathcal{E}_{r},\qquad\mathcal{E}_{\gamma}\cap\mathcal{E}_{\gamma'}=\varnothing\quad(\gamma\ne\gamma'),

where each group may represent a distinct boundary component or boundary type. Let e∈Eγe\in\mathcal{E}_{\gamma} denote the boundary fine edges with one adjacent cell K(e)K(e). For γ=1,…,r\gamma=1,\ldots,r and i=1,…,N0i=1,\ldots,N_{0}, the coarse boundary 11-form is defined as

ψi,γ1,bc(x)=∑e∈EγWi,K(e)ϕe1(x).(30)\psi_{i,\gamma}^{1,\mathrm{bc}}(x)=\sum_{e\in\mathcal{E}_{\gamma}}W_{i,K(e)}\phi_{e}^{1}(x). \tag*{(30)}

Concatenating the interior and boundary 11-forms, we have:

ψ1=[ψ1,int  ψ1,bc],\psi^{1}=[\psi^{1,\mathrm{int}}\ \ \psi^{1,\mathrm{bc}}],

where N1=N1int+N1bc=(N02)+rN0N_{1}=N_{1}^{\mathrm{int}}+N_{1}^{\mathrm{bc}}=\binom{N_{0}}{2}+rN_{0}. The corresponding coarse RT0\mathrm{RT}_{0} space is

Vhc:=W1=span⁡{ψα1}α=1N1⊂Vhf.(31)V_{h}^{c}:=\mathcal{W}^{1}=\operatorname{span}\{\psi_{\alpha}^{1}\}_{\alpha=1}^{N_{1}}\subset V_{h}^{f}. \tag*{(31)}

Remark 3.3. The classic PoU Whitney 11-form between two coarse partitions is

ψij1=λi∇λj−λj∇λi.\psi_{ij}^{1}=\lambda_{i}\nabla\lambda_{j}-\lambda_{j}\nabla\lambda_{i}.

The dense representation applies a double sum over all oriented cell pairs:

ψij1=∑a,bWiaWjbϕabRT0,\psi_{ij}^{1}=\sum_{a,b}W_{ia}W_{jb}\phi_{ab}^{\mathrm{RT}_{0}},

assuming that the oriented flux basis ϕabRT0\phi_{ab}^{\mathrm{RT}_{0}} is defined for each ordered pair (a,b)(a,b). In the actual finite element implementation, however, RT0\mathrm{RT}_{0} basis functions are associated only with fine mesh edges. By the definition, we have

ϕKL(e),KR(e)RT0=ϕeRT0,ϕKR(e),KL(e)RT0=−ϕeRT0.\phi_{K_{L}(e),K_{R}(e)}^{\mathrm{RT}_{0}}=\phi_{e}^{\mathrm{RT}_{0}},\qquad\phi_{K_{R}(e),K_{L}(e)}^{\mathrm{RT}_{0}}=-\phi_{e}^{\mathrm{RT}_{0}}.

Thus Proposition 3.2 is the edge-restricted construction of the same antisymmetric principle, and the coefficient for the interior coarse 11-forms in Equation (29) should be considered the discrete mesh-edge analogue of the classical Whitney representation. For later propositions and proofs, we use the double sum notation.

Coarse operators

For convenience, we hereafter denote by {ψα1}α=1N1\{\psi_{\alpha}^{1}\}_{\alpha=1}^{N_{1}} the full set of interior and boundary coarse 1-form basis functions. With the coarse 0-forms {ψi0}i=1N0\{\psi_{i}^{0}\}_{i=1}^{N_{0}} and 1-forms {ψα1}α=1N1\{\psi_{\alpha}^{1}\}_{\alpha=1}^{N_{1}}, we can further define the corresponding coarse operators.

Proposition 3.4 (Coarse operators). For i,j=1,…,N0i,j=1,\ldots,N_{0}, and α,β=1,…,N1\alpha,\beta=1,\ldots,N_{1}, the coarse mass matrices and divergence matrix are defined by

(M0)ij=(ψi0,ψj0)Ω,(M1)αβ=(ψα1,ψβ1)Ω,Diα=(∇⋅ψα1,ψi0)Ω.(M_{0})_{ij}=(\psi_{i}^{0},\psi_{j}^{0})_{\Omega},\qquad(M_{1})_{\alpha\beta}=(\psi_{\alpha}^{1},\psi_{\beta}^{1})_{\Omega},\qquad D_{i\alpha}=(\nabla\cdot\psi_{\alpha}^{1},\psi_{i}^{0})_{\Omega}.

The above construction for coarse bases and operators is purely algebraic and therefore remains valid for any WW satisfying the property in Equation (25). For a conditioning variable zz, the neural coarsening map gives W=Wθ(z)W=W_{\theta}(z) and the resulting conditional operators in reduced space are Qhc(z;θ)Q_{h}^{c}(z;\theta), Vc(z;θ)V^{c}(z;\theta), Dθ(z)D_{\theta}(z). In practice, for NN training instances, WW can be a tensor of shape W∈RN×N0×NcellW\in\mathbb{R}^{N\times N_{0}\times N_{\mathrm{cell}}}, where W(n)∈RN0×NcellW^{(n)}\in\mathbb{R}^{N_{0}\times N_{\mathrm{cell}}} represents the conditional neural Whitney form for sample nn given a conditioning variable z(n)z^{(n)}. The choice of coarse forms may vary depending on different factors, such as the treatment of boundaries or the choices of finite element spaces, but the fundamental principle is that by introducing the trainable WW, we are able to adapt the PoU flexibly with respect to certain conditioning variables. Moreover, the learned coarse spaces inherit the H(div⁡)H(\operatorname{div})–L2L^{2} structure from the fine space with the coarse divergence operator as in the following diagram:

H(div⁡,Ω)→∇⋅L2(Ω)∪→∇⋅∪Vhc(z;θ)→Dθ(z)Qhc(z;θ).(32)\begin{aligned} H(\operatorname{div},\Omega)&\xrightarrow{\nabla\cdot}L^{2}(\Omega)\\ \mathbin{\cup}\phantom{\xrightarrow{\nabla\cdot}}&\mathbin{\cup}\\ V_{h}^{c}(z;\theta)&\xrightarrow{D_{\theta}(z)}Q_{h}^{c}(z;\theta). \tag*{(32)} \end{aligned}

The following proposition provides a sufficient condition for the full row rank of the coarse divergence matrix. This condition is a nondegeneracy condition of the learned coarse basis and can be checked from the learned PoU weights.

Proposition 3.5 (Rank condition). Let the boundary fine edges be partitioned as in Proposition 3.2. Assume that the fine boundary basis ϕe1\phi_{e}^{1} for boundary edge e∈Eγe\in\mathcal{E}_{\gamma} is oriented so that βeγ:=(∇⋅ϕe1,ϕK(e)P0)Ω>0\beta_{e}^{\gamma}:=(\nabla\cdot\phi_{e}^{1},\phi_{K(e)}^{P_{0}})_{\Omega}>0. Let Dγ∈RN0×N0D_{\gamma}\in\mathbb{R}^{N_{0}\times N_{0}} denote the block of the coarse divergence matrix DD whose columns correspond to the boundary basis functions associated with Eγ\mathcal{E}_{\gamma}. Then

(Dγ)iα=∑e∈EγβeγWα,K(e)Wi,K(e)=(WγBγWγT)iα,(D_{\gamma})_{i\alpha}=\sum_{e\in\mathcal{E}_{\gamma}}\beta_{e}^{\gamma}W_{\alpha,K(e)}W_{i,K(e)}=(W_{\gamma}B_{\gamma}W_{\gamma}^{\mathsf{T}})_{i\alpha},

where

(Wγ)ie=Wi,K(e),Bγ=diag⁡{βeγ:e∈Eγ}.(W_{\gamma})_{ie}=W_{i,K(e)},\qquad B_{\gamma}=\operatorname{diag}\{\beta_{e}^{\gamma}:e\in\mathcal{E}_{\gamma}\}.

If ∑γ=1rWγBγWγT\sum_{\gamma=1}^{r}W_{\gamma}B_{\gamma}W_{\gamma}^{\mathsf{T}} is positive definite, then the coarse divergence matrix has full row rank, rank⁡(D)=N0\operatorname{rank}(D)=N_{0}.

Proof. See Appendix A. □

Surjectivity property under H(div⁡)H(\operatorname{div}) setting and flux conservation

The divergence operator ∇⋅\nabla\cdot is surjective from the Raviart–Thomas space RTk(K)\mathrm{RT}_{k}(K) onto Pk(K)P_{k}(K) in the standard fine scale [7]. In particular, for RT0/dgP0\mathrm{RT}_{0}/dgP_{0}, the divergence represents the cellwise flux conservation. Here, we extend this surjectivity property and conservation structure to coarse operators induced by the Whitney forms construction.

Proposition 3.6. Let W1:=span⁡{ψa1}a=1N1\mathcal{W}^{1}:=\operatorname{span}\{\psi_a^1\}_{a=1}^{N_1}, W0:=span⁡{ψi0}i=1N0\mathcal{W}^{0}:=\operatorname{span}\{\psi_i^0\}_{i=1}^{N_0} be the coarse flux and scalar spaces following Proposition 3.1 and Proposition 3.2. Then:

∇⋅W1⊆W0.(33)\nabla\cdot\mathcal{W}^{1}\subseteq\mathcal{W}^{0}. \tag*{(33)}

Moreover, if the rank condition in Proposition 3.5 holds, then the coarse divergence matrix has full row rank and the coarse divergence is surjective from W1\mathcal{W}^{1} onto W0\mathcal{W}^{0}.

Proof. Recall that we have fine-scale basis functions {ϕ1}⊂RT0\{\phi^1\}\subset RT_0 and {ϕ0}⊂P0\{\phi^0\}\subset P_0. Given the fine-scale divergence identity for (RT0,dgP0)(RT_0,\mathrm{dg}P_0), we have

∇⋅ϕeRT0=ϕKR(e)P0−ϕKL(e)P0.(34)\nabla\cdot\phi_e^{RT_0}=\phi_{K_R(e)}^{P_0}-\phi_{K_L(e)}^{P_0}. \tag*{(34)}

For interior edge e=(i,j)e=(i,j), we can derive the identity in the coarse scale by substituting the definition of coarse basis functions:

∇⋅ψij1=∇⋅(∑a,bWiaWjbϕabRT0)=∑a,bWiaWjbϕbP0−∑a,bWiaWjbϕaP0=(∑aWia)ψj0−(∑bWjb)ψi0=miψj0−mjψi0∈W0.\begin{aligned} \nabla\cdot\psi_{ij}^{1} &=\nabla\cdot\left(\sum_{a,b}W_{ia}W_{jb}\phi_{ab}^{RT_0}\right)\\ &=\sum_{a,b}W_{ia}W_{jb}\phi_b^{P_0}-\sum_{a,b}W_{ia}W_{jb}\phi_a^{P_0}\\ &=\left(\sum_a W_{ia}\right)\psi_j^0-\left(\sum_b W_{jb}\right)\psi_i^0\\ &=m_i\psi_j^0-m_j\psi_i^0\in\mathcal{W}^0. \end{aligned}

Similarly, by linearity, the boundary coarse 1-form is also a linear combination of fine P0P_0 basis functions and belongs to the span of the coarse scalar space. Therefore,

∇⋅W1⊆W0.\nabla\cdot\mathcal{W}^{1}\subseteq\mathcal{W}^{0}.

Next we prove the surjectivity of the coarse divergence operator. Let

v=∑a=1N1F^aψa1∈W1,q=∑i=1N0qiψi0∈W0.(35)v=\sum_{a=1}^{N_1}\widehat{F}_a\psi_a^1\in\mathcal{W}^{1},\qquad q=\sum_{i=1}^{N_0}q_i\psi_i^0\in\mathcal{W}^{0}. \tag*{(35)}

By linearity,

∇⋅v=∑a=1N1F^a∇⋅ψa1∈W0.(36)\nabla\cdot v=\sum_{a=1}^{N_1}\widehat{F}_a\nabla\cdot\psi_a^1\in\mathcal{W}^{0}. \tag*{(36)}

Then testing against ψk0\psi_k^0 for i=1,…,N0i=1,\ldots,N_0 and expanding for coarse variables gives:

(∇⋅v,ψk0)Ω=(∇⋅(∑a=1N1F^aψa1),ψk0)Ω(37)(\nabla\cdot v,\psi_k^0)_\Omega =\left(\nabla\cdot\left(\sum_{a=1}^{N_1}\widehat{F}_a\psi_a^1\right),\psi_k^0\right)_\Omega \tag*{(37)}
=∑a=1N1F^a(∇⋅ψa1,ψk0)Ω=∑a=1N1F^aDia(38)=\sum_{a=1}^{N_1}\widehat{F}_a(\nabla\cdot\psi_a^1,\psi_k^0)_\Omega =\sum_{a=1}^{N_1}\widehat{F}_aD_{ia} \tag*{(38)}
(q,ψk0)Ω=∑i=1N0qi(ψi0,ψk0)Ω=∑i=1N0qi(M0)ik,(39)(q,\psi_k^0)_\Omega =\sum_{i=1}^{N_0}q_i(\psi_i^0,\psi_k^0)_\Omega =\sum_{i=1}^{N_0}q_i(M_0)_{ik}, \tag*{(39)}

where DD and MM are as defined in Proposition 3.4. When rank⁡(D)=N0\operatorname{rank}(D)=N_0, the map D:RN1→RN0D:\mathbb{R}^{N_1}\to\mathbb{R}^{N_0} is onto and thus for arbitrary qq, there exist coefficients F^∈RN1\widehat{F}\in\mathbb{R}^{N_1} such that DF^=M0qD\widehat{F}=M_0q. For this choice of F^\widehat{F}, the corresponding v∈W1v\in W^1 satisfies

(∇⋅v,ψk0)Ω=(q,ψk0)Ω.(40)(\nabla\cdot v,\psi_k^0)_{\Omega}=(q,\psi_k^0)_{\Omega}. \tag*{(40)}

Thus we conclude that the coarse divergence operator is surjective. ∎

Moreover, since the coefficient in (3.6) is the discrete analogue of the classical Whitney exchange form λi∇λj−λj∇λi\lambda_i\nabla\lambda_j-\lambda_j\nabla\lambda_i, the resulting antisymmetry property ∇⋅ψij1=−∇⋅ψji1\nabla\cdot\psi_{ij}^1=-\nabla\cdot\psi_{ji}^1 with respect to the exchange of indices ensures by construction that the 1-forms encode equal and opposite fluxes exchanged between 0-forms. Summing over all indices cancels the internal fluxes, so that only the boundary contributions remain. This recovers a discrete conservation law consistent with the divergence theorem and shows that the learned coarse construction preserves the flux balance.

Graph calculus and circuit interpretation for FEEC

The preceding construction in Subsection 3.1 and Subsection 3.2 produces an H(div⁡)H(\operatorname{div})-conforming reduced space where conservation is preserved by the coarse divergence operator. To formulate the GP optimal recovery problem in the next section, we now reinterpret the reduced subcomplexes as a graph-valued circuit model and connect coarse operators with the graph calculus. This interpretation bridges the structure-preserving FEEC with the graph-based GP model used to learn the nonlinearity and perform uncertainty quantification later. The viewpoint is related to the computational graph completion (CGC) [29] problem, where the unknown functions and unobserved variables can be recovered on a given graph. In contrast, the graph in this work is induced by the coarse finite element subcomplexes, which can be considered an autonomous identification of a dense graph.

Let G=(V,E)\mathcal{G}=(\mathcal{V},\mathcal{E}) be a graph with ∣V∣=N0|\mathcal{V}|=N_0 vertices and ∣E∣=N1|\mathcal{E}|=N_1 edges. The coarse spaces constructed via the partition of unity Whitney forms naturally define a graph where each vertex i∈Vi\in\mathcal{V} corresponds to a coarse 0-form ψi0\psi_i^0 and each edge e∈Ee\in\mathcal{E} corresponds to a coarse 1-form ψe1\psi_e^1. We define the cochain spaces as:

C0(G):=R∣V∣,C1(G):=R∣E∣.C^0(\mathcal{G}):=\mathbb{R}^{|\mathcal{V}|},\qquad C^1(\mathcal{G}):=\mathbb{R}^{|\mathcal{E}|}.

Then the coarse variables admit the interpretation on vertices and edges, respectively:

u^∈C0(G),F^∈C1(G),\widehat{u}\in C^0(\mathcal{G}),\qquad\widehat{F}\in C^1(\mathcal{G}),

where u^\widehat{u} represents node states and F^\widehat{F} represents oriented fluxes or currents.

In order to formulate the mapping between nodal values and edge values, we introduce the edge-vertex signed incidence matrix B0∈RE×VB_0\in\mathbb{R}^{E\times V} defined by

(B0)i,j={−1if edge ei leaves vertex vj,1if edge ei enters vertex vj,0otherwise.(41)(B_0)_{i,j}= \begin{cases} -1 & \text{if edge }e_i\text{ leaves vertex }v_j,\\ 1 & \text{if edge }e_i\text{ enters vertex }v_j,\\ 0 & \text{otherwise.} \end{cases} \tag*{(41)}

Subsequently for an oriented edge e=(i,j)e=(i,j), we can define the discrete graph gradient operator and graph divergence as:

(Graph gradient)(δ0u^)(e)=u^j−u^i,\text{(Graph gradient)}\qquad(\delta_0\widehat{u})(e)=\widehat{u}_j-\widehat{u}_i,
(Graph divergence)(δ0TF^)(v)=∑vi∼vF^(vi,v)−∑v∼vjF^(v,vj).\text{(Graph divergence)}\qquad(\delta_0^{T}\widehat{F})(v)=\sum_{v_i\sim v}\widehat{F}(v_i,v)-\sum_{v\sim v_j}\widehat{F}(v,v_j).

The discrete graph gradient, mapping the nodal values to edge values, captures the difference across edges. Conversely, the graph divergence, mapping edge values back to values at nodes, measures the net flux entering or leaving the node vv over its adjacent edges. Let DG∈R∣V∣×∣E∣D_G \in\mathbb{R}^{|V|\times|E|} denote the graph divergence matrix, defined by

(DGF^)i=∑e∈EDieF^e.(D_G\widehat{F})_i = \sum_{e\in E} D_{ie}\widehat{F}_e.

With the construction of coarse Whitney forms in Subsection 3.2, DD has the same incidence structure as the graph divergence operator, possibly weighted by the mass or basis normalization depending on the specific construction. Thus, this gives the conservation law in the reduced spaces and supplies the finite-dimensional variables on which the GP acts.

Method part 2: Optimal recovery of physics with quantified uncertainty

In this section, we develop a probabilistic representation of physical fields on the graph induced by the coarse spaces. This provides a unified framework where the physical law acts on discrete structures while uncertainty is quantified in a functionally consistent manner.

Structure-preserving Optimal Recovery Problem

We now introduce a GP optimal recovery formulation for learning the nonlinear flux correction in the coarse H(div⁡)−L2H(\operatorname{div})-L^2 space, consistent with the FEEC structure established above. For each coarse edge e=(i,j)∈Ee=(i,j)\in E, we can define the local input with different choices:

u^e:={(u^i,u^j)∈R2,(endpoint representation),δ0u^e=u^j−u^i∈R,(edge gradient representation).\widehat{u}_e := \begin{cases} (\widehat{u}_i,\widehat{u}_j)\in\mathbb{R}^2, & \text{(endpoint representation)},\\ \delta_0\widehat{u}_e=\widehat{u}_j-\widehat{u}_i\in\mathbb{R}, & \text{(edge gradient representation)}. \end{cases}

For training instances n=1,…,Nn=1,\ldots,N, the desired GP input features become

xe(n)=(u^e(n),z(n),s(n)).x_e^{(n)}=\left(\widehat{u}_e^{(n)},z^{(n)},s^{(n)}\right).

Here u^e(n)\widehat{u}_e^{(n)} is the coarse state variable associated with edge ee, z(n)z^{(n)} denotes the conditioning variable, and s(n)s^{(n)} denotes the geometric embedding.

In the independent kernel construction, for a fixed edge ee, we collect the training inputs as Xe={xe(n)}n=1NX_e=\{x_e^{(n)}\}_{n=1}^N and the corresponding global kernel is the block-diagonal matrix defined as

Ke:=Ke(Xe,Xe)∈RN×N,K=diag⁡(K1,…,KN1)∈RN1N×N1N.(42)K_e := K_e(X_e,X_e)\in\mathbb{R}^{N\times N},\qquad K=\operatorname{diag}(K_1,\ldots,K_{N_1})\in\mathbb{R}^{N_1N\times N_1N}. \tag*{(42)}

The GP model on edge ee maps local features to the flux on the edge:

fe(Xe)↦F^egp,F^gp=vec⁡([F^egp,(n)])∈RN1N,(43)f_e(X_e)\mapsto\widehat{F}_e^{\mathrm{gp}},\qquad\widehat{\mathbf{F}}^{\mathrm{gp}}=\operatorname{vec}\left(\left[\widehat{F}_e^{\mathrm{gp},(n)}\right]\right)\in\mathbb{R}^{N_1N}, \tag*{(43)}

where each fef_e belongs to a reproducing kernel Hilbert space HKe\mathcal{H}_{K_e} and F^gp\widehat{\mathbf{F}}^{\mathrm{gp}} is the vectorized global output over all edges.

Instead of assigning independent GPs to each edge, we may also stack all edge samples into a global matrix

X={xe(n):e=1,…,N1, n=1,…,N}(44)X=\left\{x_e^{(n)}:e=1,\ldots,N_1,\ n=1,\ldots,N\right\} \tag*{(44)}

and construct a single shared GP across all cochains:

K=[kθ(xe(n),xe′(m))](e,n),(e′,m)∈RN1N×N1NK=\left[k_\theta\left(x_e^{(n)},x_{e'}^{(m)}\right)\right]_{(e,n),(e',m)}\in\mathbb{R}^{N_1N\times N_1N}

where kθk_\theta denotes the covariance kernel. Since both cases use a large global matrix, for simplicity we denote the global matrix as KK in the following.

Recall the governing equation for the physical law in Equation (1). By restricting the PDE to data-driven Whitney spaces W0(Ω;θ)W^0(\Omega;\theta) and W1(Ω;θ)W^1(\Omega;\theta), the goal is to find (u,F)∈W0(Ω;θ)×W1(Ω;θ)(u,\mathbf{F}) \in W^0(\Omega;\theta) \times W^1(\Omega;\theta) such that

(F,v)Ω=(u,∇⋅v)Ω+(N[u],v)Ω−⟨uD,v⋅n⟩ΓD,v∈W1(Ω;θ)(45)(\mathbf{F},\mathbf{v})_\Omega= (u,\nabla\cdot\mathbf{v})_\Omega+ (\mathcal{N}[u],\mathbf{v})_\Omega- \langle u_D,\mathbf{v}\cdot\mathbf{n}\rangle_{\Gamma_D}, \qquad\mathbf{v} \in W^1(\Omega;\theta) \tag*{(45)}
(∇⋅F,q)Ω=(f,q)Ω,q∈W0(Ω;θ),(46)(\nabla\cdot\mathbf{F},q)_\Omega= (f,q)_\Omega, \qquad q \in W^0(\Omega;\theta), \tag*{(46)}

where ff is the source term and ⟨uD,v⋅n⟩ΓD=∫ΓDuD(v⋅n) ds\langle u_D,\mathbf{v}\cdot\mathbf{n}\rangle_{\Gamma_D} = \int_{\Gamma_D} u_D(\mathbf{v}\cdot\mathbf{n})\,ds is the weak Dirichlet boundary condition. Recall that u^\widehat{\mathbf{u}} denotes the coarse scalar coefficient. Let F^gp\widehat{\mathbf{F}}^{\mathrm{gp}} denote the nonlinear flux correction on the coarse edge and F^\widehat{\mathbf{F}} denote the total coarse flux coefficient. Using the FEEC structure, the problem can be further written as

M1F^=DTu^+M1F^gp−g^D,(47)M_1\widehat{\mathbf{F}} = D^{\mathsf{T}}\widehat{\mathbf{u}} + M_1\widehat{\mathbf{F}}^{\mathrm{gp}} - \widehat{\mathbf{g}}_D, \tag*{(47)}
DF^=f^,(48)D\widehat{\mathbf{F}} = \widehat{\mathbf{f}}, \tag*{(48)}

where f^\widehat{\mathbf{f}} is the coefficient vector of the source term and g^D\widehat{\mathbf{g}}_D represents the Dirichlet boundary contribution with g^D=∫ΓDuDψ1⋅n ds\widehat{\mathbf{g}}_D = \int_{\Gamma_D} u_D\boldsymbol{\psi}^1\cdot\mathbf{n}\,ds. M1M_1 and DD are the discrete mass matrix and divergence operator as defined in Proposition 3.4. For simplicity, we let θ=(θGP,θW)\theta= (\theta_{\mathrm{GP}},\theta_W) denote the set of trainable parameters for the GP model and neural Whitney forms. Eliminating F^\widehat{\mathbf{F}} gives the reduced residual:

Rθ(u^,F^gp;z):=DM1−1DTu^+DF^gp−b^,with b^=f^+DM1−1g^D.(49)R_\theta(\widehat{\mathbf{u}},\widehat{\mathbf{F}}^{\mathrm{gp}};z) := DM_1^{-1}D^{\mathsf{T}}\widehat{\mathbf{u}} + D\widehat{\mathbf{F}}^{\mathrm{gp}} - \widehat{\mathbf{b}}, \qquad\text{with } \widehat{\mathbf{b}} = \widehat{\mathbf{f}} + DM_1^{-1}\widehat{\mathbf{g}}_D. \tag*{(49)}

Finally, we arrive at the full optimization problem, combining the GP recovery and the data-driven Whitney forms on the graph:

min⁡θ,u^min⁡F^gp(F^gp)T(K(X,X)+σε2I)−1F^gp+log⁡det⁡(K(X,X)+σε2I)+∑n=1N∥udata(n)−uapprox(n)∥2+∑n=1N∥FΓ,data(n)−FΓ,approx(n)∥2s.t.Rθ(u^,F^gp;z(n))=0,(50)\begin{aligned} \min_{\theta,\widehat{\mathbf{u}}}\min_{\widehat{\mathbf{F}}^{\mathrm{gp}}}\quad& (\widehat{\mathbf{F}}^{\mathrm{gp}})^{\mathsf{T}}\left(K(X,X)+\sigma_\varepsilon^2 I\right)^{-1}\widehat{\mathbf{F}}^{\mathrm{gp}} + \log\det\left(K(X,X)+\sigma_\varepsilon^2 I\right) \\ &+ \sum_{n=1}^{N}\left\lVert\mathbf{u}_{\mathrm{data}}^{(n)}-\mathbf{u}_{\mathrm{approx}}^{(n)}\right\rVert^2 +\sum_{n=1}^{N}\left\lVert\mathbf{F}_{\Gamma,\mathrm{data}}^{(n)}-\mathbf{F}_{\Gamma,\mathrm{approx}}^{(n)}\right\rVert^2 \\ \text{s.t.}\quad& R_\theta\left(\widehat{\mathbf{u}},\widehat{\mathbf{F}}^{\mathrm{gp}};z^{(n)}\right)=0, \tag*{(50)} \end{aligned}

where uapprox\mathbf{u}_{\mathrm{approx}}, Fapprox\mathbf{F}_{\mathrm{approx}} are the reconstruction of the solution field and flux back to the original fine-scale space. The first two terms promote smoothness in the RKHS norm and penalize overly complex models to avoid overfitting. Moreover, as the recovery problem is now defined for the coarse coefficients u^\widehat{\mathbf{u}} and F^gp\widehat{\mathbf{F}}^{\mathrm{gp}}, there is no guarantee that the solution and flux projected back to the original space can reconstruct the target solution udata\mathbf{u}_{\mathrm{data}} and Fdata\mathbf{F}_{\mathrm{data}} as desired. Thus, we introduce two additional terms consisting of the mean squared error (MSE) reconstruction loss in the fine scale.

The reconstruction of the solution in (50) can be directly obtained by mapping the coarse 0-cochain coefficients back to the fine scale with the neural Whitney bases:

uapprox=u^TΨ0.(51)\mathbf{u}_{\mathrm{approx}} = \widehat{\mathbf{u}}^{\mathsf{T}}\boldsymbol{\Psi}_0. \tag*{(51)}

The total ground-truth flux and reconstructed flux on the target boundary Γ\Gamma from the PoU Whitney form and coarse coefficient are:

FΓ,data=∑e∈ΓFe,FΓ,approx=∑i=1NPoU∑e∈ΓWieF^i(52)\mathbf{F}_{\Gamma,\mathrm{data}}=\sum_{e\in\Gamma}\mathbf{F}_e,\qquad\mathbf{F}_{\Gamma,\mathrm{approx}}=\sum_{i=1}^{N_{\mathrm{PoU}}}\sum_{e\in\Gamma}W_{ie}\widehat{\mathbf{F}}_i \tag*{(52)}

where WieW_{ie} denotes the corresponding weight for the ii-th PoU and fine-scale edge ee and F^=M1−1D⊤u^+F^gp−M1−1gD\widehat{\mathbf{F}}=\mathbf{M}_1^{-1}\mathbf{D}^{\top}\widehat{\mathbf{u}}+\widehat{\mathbf{F}}^{\mathrm{gp}}-\mathbf{M}_1^{-1}\mathbf{g}_D.

Bilevel optimization

The optimization problem is nested and strongly coupled because the GP inputs depend on the unknown coarse potentials u^e\widehat{\mathbf{u}}_e. Therefore, we apply a bilevel optimization procedure to train the transformer for data-driven Whitney forms and the GP model for flux prediction simultaneously. The bilevel training algorithm is summarized in Algorithm 4.1.

Algorithm 4.1 Bilevel training for structure-preserving optimal recovery

Require: D={(z(n),u^data(n),Fdata(n))}n=1N\mathcal{D}=\{(\mathbf{z}^{(n)},\widehat{\mathbf{u}}_{\mathrm{data}}^{(n)},\mathbf{F}_{\mathrm{data}}^{(n)})\}_{n=1}^{N}; fine operators; PoU network WθWW_{\theta_W}; GP kernel KθGPK_{\theta_{\mathrm{GP}}}

Require: Optimizers: Adam for θGP\theta_{\mathrm{GP}}, Shampoo for θW\theta_W

  1. Initialize u^\widehat{\mathbf{u}}, θW\theta_W, and θGP\theta_{\mathrm{GP}}

  2. for epoch =1,…,T=1,\ldots,T do

  3. &nbsp;&nbsp;&nbsp;&nbsp;PoU forward: Compute W←WθW(z)W\leftarrow W_{\theta_W}(\mathbf{z})

  4. &nbsp;&nbsp;&nbsp;&nbsp;Compute coarse bases Ψ0,Ψ1\mathbf{\Psi}^0,\mathbf{\Psi}^1, and operators M1,D,RΓ,b^\mathbf{M}_1,\mathbf{D},\mathbf{R}_{\Gamma},\widehat{\mathbf{b}}

  5. &nbsp;&nbsp;&nbsp;&nbsp;Form coarse GP features X={xe(n)}e,n\mathbf{X}=\{\mathbf{x}_e^{(n)}\}_{e,n}, K~=KθGP(X,X)+σε2I\widetilde{\mathbf{K}}=K_{\theta_{\mathrm{GP}}}(\mathbf{X},\mathbf{X})+\sigma_{\varepsilon}^2\mathbf{I}

  6. &nbsp;&nbsp;&nbsp;&nbsp;Inner problem: Fix u^\widehat{\mathbf{u}}, θW\theta_W, θGP\theta_{\mathrm{GP}}, solve the constrained recovery problem for F^gp,⋆\widehat{\mathbf{F}}^{\mathrm{gp},\star}

  7. &nbsp;&nbsp;&nbsp;&nbsp;Update F^gp,⋆←SolveKKT⁡(K~,D,RΓ,b^,FΓ,data)\widehat{\mathbf{F}}^{\mathrm{gp},\star}\leftarrow\operatorname{SolveKKT}\left(\widetilde{\mathbf{K}},\mathbf{D},\mathbf{R}_{\Gamma},\widehat{\mathbf{b}},\mathbf{F}_{\Gamma,\mathrm{data}}\right)

  8. &nbsp;&nbsp;&nbsp;&nbsp;Update θGP←Adam⁡(θGP,∇θGPLinner)\theta_{\mathrm{GP}}\leftarrow\operatorname{Adam}(\theta_{\mathrm{GP}},\nabla_{\theta_{\mathrm{GP}}}\mathcal{L}_{\mathrm{inner}})

  9. &nbsp;&nbsp;&nbsp;&nbsp;Outer problem: Fix F^gp,⋆\widehat{\mathbf{F}}^{\mathrm{gp},\star}, θGP\theta_{\mathrm{GP}}, solve for u^⋆\widehat{\mathbf{u}}^\star with least squares method

  10. &nbsp;&nbsp;&nbsp;&nbsp;Update u^⋆←lstsq⁡(DM1−1D⊤,b^−DF^gp,⋆)\widehat{\mathbf{u}}^\star\leftarrow\operatorname{lstsq}\left(\mathbf{D}\mathbf{M}_1^{-1}\mathbf{D}^{\top},\widehat{\mathbf{b}}-\mathbf{D}\widehat{\mathbf{F}}^{\mathrm{gp},\star}\right)

  11. &nbsp;&nbsp;&nbsp;&nbsp;Update θW←Shampoo⁡(θW,∇θWLouter)\theta_W\leftarrow\operatorname{Shampoo}(\theta_W,\nabla_{\theta_W}\mathcal{L}_{\mathrm{outer}})

  12. end for 13.

  13. return trained θW⋆,θGP⋆,u^⋆\theta_W^\star,\theta_{\mathrm{GP}}^\star,\widehat{\mathbf{u}}^\star, and F^⋆\widehat{\mathbf{F}}^\star.

For the inner loop, we first fix u^\widehat{\mathbf{u}} and θW\theta_W and solve for the closed-form solution F^gp\widehat{\mathbf{F}}^{\mathrm{gp}} via the Karush–Kuhn–Tucker (KKT) system. Since the flux observations are defined on fine boundary edges, for simplicity, we introduce the coarse-to-fine reconstruction operator RΓ(W):RN1→R\mathbf{R}_{\Gamma}(W):\mathbb{R}^{N_1}\to\mathbb{R} that maps the coarse flux degrees of freedom (F^)(\widehat{\mathbf{F}}) back to the total flux through the boundary region Γ\Gamma. Substituting F^gp\widehat{\mathbf{F}}^{\mathrm{gp}} gives the inner optimization problem as

min⁡F^gp∥FΓ,data−RΓ(W)(M1−1D⊤u^+F^gp−M1−1gD)∥22(53)\min_{\widehat{\mathbf{F}}^{\mathrm{gp}}}\left\|\mathbf{F}_{\Gamma,\mathrm{data}}-\mathbf{R}_{\Gamma}(W)\left(\mathbf{M}_1^{-1}\mathbf{D}^{\top}\widehat{\mathbf{u}}+\widehat{\mathbf{F}}^{\mathrm{gp}}-\mathbf{M}_1^{-1}\mathbf{g}_D\right)\right\|_2^2 \tag*{(53)}
+(F^gp)⊤(K~+σε2I)−1F^gp+log⁡det⁡(K~+σε2I)(54)+\left(\widehat{\mathbf{F}}^{\mathrm{gp}}\right)^{\top}\left(\widetilde{\mathbf{K}}+\sigma_{\varepsilon}^2\mathbf{I}\right)^{-1}\widehat{\mathbf{F}}^{\mathrm{gp}}+\log\det\left(\widetilde{\mathbf{K}}+\sigma_{\varepsilon}^2\mathbf{I}\right) \tag*{(54)}
s.t.DF^gp=b−DM1−1D⊤u^.(55)\text{s.t.}\quad\mathbf{D}\widehat{\mathbf{F}}^{\mathrm{gp}}=\mathbf{b}-\mathbf{D}\mathbf{M}_1^{-1}\mathbf{D}^{\top}\widehat{\mathbf{u}}. \tag*{(55)}

Since the term log⁡det⁡(K~+σε2I)\log\det(\widetilde{\mathbf{K}}+\sigma_{\varepsilon}^2\mathbf{I}) does not depend on F^gp\widehat{\mathbf{F}}^{\mathrm{gp}}, it can be omitted when solving for F^gp\widehat{\mathbf{F}}^{\mathrm{gp}}. Introducing the Lagrange multiplier λ\lambda, the optimality system becomes:

[K~−1+RΓ⊤RΓD⊤D0][F^gpλ]=[RΓ⊤(FΓ,data−RΓ(M1−1D⊤u^−M1−1gD))b−DM1−1D⊤u^].\begin{bmatrix} \widetilde{\mathbf{K}}^{-1}+\mathbf{R}_{\Gamma}^{\top}\mathbf{R}_{\Gamma} & \mathbf{D}^{\top}\\ \mathbf{D} & 0 \end{bmatrix} \begin{bmatrix} \widehat{\mathbf{F}}^{\mathrm{gp}}\\ \lambda \end{bmatrix} = \begin{bmatrix} \mathbf{R}_{\Gamma}^{\top}\left(\mathbf{F}_{\Gamma,\mathrm{data}}-\mathbf{R}_{\Gamma}\left(\mathbf{M}_1^{-1}\mathbf{D}^{\top}\widehat{\mathbf{u}}-\mathbf{M}_1^{-1}\mathbf{g}_D\right)\right)\\ \mathbf{b}-\mathbf{D}\mathbf{M}_1^{-1}\mathbf{D}^{\top}\widehat{\mathbf{u}} \end{bmatrix}.

where K~=K+σ2I\widetilde{K}=K+\sigma^{2}I. Solving by the Schur complement, we obtain the closed-form solution for optimal F^gp\widehat{\mathbf{F}}^{\mathrm{gp}} at the current stage:

λ=(D(K~−1+RΓTRΓ)−1DT)−1(D(K~−1+RΓTRΓ)−1rt−rb)(56)\lambda=\left(D\left(\widetilde{K}^{-1}+R_{\Gamma}^{\mathsf{T}}R_{\Gamma}\right)^{-1}D^{\mathsf{T}}\right)^{-1}\left(D\left(\widetilde{K}^{-1}+R_{\Gamma}^{\mathsf{T}}R_{\Gamma}\right)^{-1}r_t-r_b\right) \tag*{(56)}
F^gp,⋆=(K~−1+RΓTRΓ)−1(rt−DTλ)(57)\widehat{\mathbf{F}}^{\mathrm{gp},\star}=\left(\widetilde{K}^{-1}+R_{\Gamma}^{\mathsf{T}}R_{\Gamma}\right)^{-1}\left(r_t-D^{\mathsf{T}}\lambda\right) \tag*{(57)}

where rt=RΓ(W)T(FΓ,data−RΓ(M1−1DTu^−M1−1gD))r_t=R_{\Gamma}(W)^{\mathsf{T}}\left(\mathbf{F}_{\Gamma,\mathrm{data}}-R_{\Gamma}\left(M_1^{-1}D^{\mathsf{T}}\widehat{u}-M_1^{-1}g_D\right)\right) and rb=b−DM1−1DTu^r_b=b-DM_1^{-1}D^{\mathsf{T}}\widehat{u}.

Then with the optimal F^gp\widehat{\mathbf{F}}^{\mathrm{gp}}, we optimize the parameters for GP kernels and noise with the Adam optimizer. The choice of kernels varies for different cases. For example, for the independent GP, we choose the RBF kernel matrix KijK_{ij} for a given edge eije_{ij}, defined as

Kij(eij)=exp⁡(−∥ui−uj∥22leij2),K_{ij}(e_{ij})=\exp\left(-\frac{\lVert u_i-u_j\rVert^2}{2l_{e_{ij}}^2}\right),

where the length scale ll and σ\sigma are the trainable parameters, i.e., θGP={l,σ}\theta_{GP}=\{l,\sigma\}. Specifically, in numerical experiments, we choose the same initialization on all edges. For the shared GP scenario, we consider the Automatic Relevance Determination (ARD) Matérn kernel [33] with ν=32\nu=\frac{3}{2}:

Kν=3/2(x,x′)=σ2(1+3r)exp⁡(−3r),r2=∑d(xd−xd′)2ℓd2.K_{\nu=3/2}(x,x')=\sigma^2\left(1+\sqrt{3}r\right)\exp\left(-\sqrt{3}r\right),\qquad r^2=\sum_d\frac{(x_d-x'_d)^2}{\ell_d^2}.

Here, ℓd>0\ell_d>0 is a trainable length scale associated with the dd-th input feature, where a small ℓd\ell_d indicates that the predicted flux varies strongly in the dd-th feature direction. This ARD parameterization allows the GP to have anisotropic scaling across different physical or geometric features.

For the outer loop, with F^gp\widehat{\mathbf{F}}^{\mathrm{gp}} and θGP\theta_{GP} fixed, we optimize over u^\widehat{u} and θW\theta_W with the Shampoo optimizer [15] by minimizing:

Louter=∥FΓ,data−RΓ(M1−1DTu^+F^gp,⋆−M1−1gD)∥2+∥udata−u^TΨ0∥2s.t.DM1−1DTu^+DF^gp,⋆=b^.(58)\begin{aligned} \mathcal{L}_{\mathrm{outer}}={}&\left\lVert\mathbf{F}_{\Gamma,\mathrm{data}}-R_{\Gamma}\left(M_1^{-1}D^{\mathsf{T}}\widehat{u}+\widehat{\mathbf{F}}^{\mathrm{gp},\star}-M_1^{-1}g_D\right)\right\rVert^2\\ &+\left\lVert u_{\mathrm{data}}-\widehat{u}^{\mathsf{T}}\Psi_0\right\rVert^2\\ &\text{s.t.}\quad DM_1^{-1}D^{\mathsf{T}}\widehat{u}+D\widehat{\mathbf{F}}^{\mathrm{gp},\star}=\widehat{b}. \tag*{(58)} \end{aligned}

Essentially, we cast the optimal recovery problem on the reduced space subject to the equality constraint that enforces the conservation law exactly. Under this setting, we have a constrained optimization problem where the KKT conditions yield a saddle-point structure that we can use to derive a fast solution with the Schur complement. The bilevel training is not strictly necessary, but can be useful for efficient training when variables are coupled and different model components, here the GP and transformer, are trained jointly.

Geometric embedding for H(div⁡)H(\operatorname{div}) coarse representations

To incorporate geometric information about the coarse degrees of freedom into the GP, we introduce a geometric embedding based on domain integrals of the coarse Whitney basis functions. This embedding is designed to encode both the coarse variables in 0- and 1-forms in a manner consistent with the underlying FEEC structure. Because scalar P0P_0 coefficients represent coarse cellwise quantities while H(div⁡)H(\operatorname{div}) coefficients represent flux degrees of freedom, we use different geometric summaries for 0-forms and 1-forms.

First, we consider the coarse P0P_0 geometric embedding

Proposition 4.1. Let {ψi0}i=1N0\{\psi_i^0\}_{i=1}^{N_0} denote the coarse 00-form basis functions constructed as in Proposition 3.1. For i=1,…,N0i=1,\ldots,N_0, substituting the definition of ψi0\psi_i^0 gives the explicit representation for the geometric embedding on the iith PoU as

si:=∫Ωψi0(x) dx=∑a∈ChWia∫ΩφaP0(x) dx.(59)s_i := \int_{\Omega}\psi_i^0(x)\,\mathrm{d}x = \sum_{a\in\mathcal{C}_h} W_{ia}\int_{\Omega}\varphi_a^{P_0}(x)\,\mathrm{d}x. \tag*{(59)}

This embedding defines a positive coarse area scalar that represents the effective measure of the coarse region associated with the ii-th PoU component and does not encode any directional information. In terms of the GP input, sis_i should be considered the geometric information for coarse node ii.

We next define the coarse RT0RT_0 geometric embedding that accounts for both direction and magnitude.

Proposition 4.2. Let {ψk1}k=1N1\{\psi_k^1\}_{k=1}^{N_1} denote the coarse RT0RT_0 basis functions constructed as in Proposition 3.2. We define the vector-valued geometric embedding associated with the kk-th coarse flux degree of freedom as

sk:=∫Ωψk1(x) dx∈R2.s_k := \int_{\Omega}\psi_k^1(x)\,\mathrm{d}x \in\mathbb{R}^2.

Substituting the expansion of ψk1\psi_k^1, we obtain:

sk=∑eWkK(e)∫ΩφeRT0(x) dx.s_k = \sum_e W_{kK(e)}\int_{\Omega}\varphi_e^{RT_0}(x)\,\mathrm{d}x.

For each fine edge ee shared by adjacent cells KL(e)K_{L(e)} and KR(e)K_{R(e)}, the RT0RT_0 basis function satisfies:

∫KL(e)∪KR(e)φeRT0(x) dx=cen^e∣e∣,\int_{K_{L(e)}\cup K_{R(e)}}\varphi_e^{RT_0}(x)\,\mathrm{d}x = c_e\widehat{n}_e|e|,

where n^e\widehat{n}_e is the unit normal vector associated with the edge orientation, ∣e∣|e| is the edge length, and cec_e depends on the reference element geometry.

From the above expression, sks_k is a weighted sum of edge-normal vectors n^e\widehat{n}_e scaled by edge lengths ∣e∣|e|, and it encodes both the net orientation and the effective magnitude of the coarse flux degree of freedom ψk1\psi_k^1. Specifically, its direction reflects the dominant orientation of the fine-scale flux contributions aggregated into ψk1\psi_k^1, while its magnitude reflects the total weighted edge measure. This provides a geometric characterization of the coarse flux degree of freedom. Therefore, the above geometric embeddings are consistent with the structure of H(div⁡)H(\operatorname{div}) spaces and enable the GP to incorporate geometric and directional information without relying on explicit spatial coordinates.

Posterior error bound for boundary flux functionals

The quantified uncertainty below is conditional on the learned representation from PoU Whitney forms and kernel hyperparameters. Under this conditioning, the boundary flux is a linear functional of the GP output at the coarse level and admits a closed-form posterior error bound. Previous work establishes a pointwise posterior error bound under the assumption that the GP input locations are fixed and known exactly [30]. In this work, we extend this result to weighted linear functionals of the GP posterior, which is the relevant setting for boundary flux quantities.

Proposition 4.3. Let RΓ(W)R_{\Gamma}(W) denote the coarse-to-fine reconstruction operator determined by WW in the proposed method (53). We define the flux functional on a given boundary Γ\Gamma and its estimator by

FΓ:=Lv(f):=vTf(Z):=∑m=1Mvmf(zm),\mathcal{F}_{\Gamma} := L_v(f) := v^{\mathsf{T}}f(Z) := \sum_{m=1}^{M} v_m f(z_m),
F^Γ:=vTf^(Z).(60)\widehat{\mathcal{F}}_{\Gamma} := v^{\mathsf{T}}\widehat{f}(Z). \tag*{(60)}

where the weight vTv^{\mathsf{T}} is the row vector of RΓ(W)R_{\Gamma}(W), i.e., vm=(RΓ(W))m,:v_m = (R_{\Gamma}(W))_{m,:}, and the error is defined by

e:=FΓ−F^Γ.(61)e := \mathcal{F}_{\Gamma} - \widehat{\mathcal{F}}_{\Gamma}. \tag*{(61)}

Theorem 4.4. Let X⊆Rp\mathcal{X} \subseteq\mathbb{R}^{p} be the Gaussian process input space, let K:X×X→RK : \mathcal{X} \times\mathcal{X} \to\mathbb{R} be a positive definite kernel with associated RKHS HK\mathcal{H}_K, and let f∈HKf \in\mathcal{H}_K. Let X=(x1,…,xN)∈XNX = (x_1,\ldots,x_N) \in\mathcal{X}^{N}, Z=(z1,…,zM)∈XMZ = (z_1,\ldots,z_M) \in\mathcal{X}^{M}, where xi,zm∈Rpx_i,z_m \in\mathbb{R}^{p} and assume we have noisy data D=(X,Y)D = (X,Y) with Y=f(X)+εY = f(X) + \varepsilon where ε∼N(0,σε2I)\varepsilon\sim\mathcal{N}(0,\sigma_{\varepsilon}^{2}I). Let f^∈HK\widehat{f} \in\mathcal{H}_K denote the minimizer of the optimal recovery problem:

f^(⋅)=K(⋅,X)(K(X,X)+σε2I)−1Y.\widehat{f}(\cdot) = K(\cdot,X)\left(K(X,X) + \sigma_{\varepsilon}^{2}I\right)^{-1}Y.

Define the posterior covariance on ZZ by

ΣZ:=K(Z,Z)−K(Z,X)(K(X,X)+σε2I)−1K(X,Z).\Sigma_Z := K(Z,Z) - K(Z,X)\left(K(X,X) + \sigma_{\varepsilon}^{2}I\right)^{-1}K(X,Z).

For v∈RMv \in\mathbb{R}^{M}, consider the weighted functionals FΓ,F^Γ\mathcal{F}_{\Gamma},\widehat{\mathcal{F}}_{\Gamma} as defined in Proposition 4.3. Then we have the bounded mean-squared error satisfying:

Eε[e2]≤∥f∥HK2vTΣZv+σε2∥vTKZXA−1∥22.(62)\mathbb{E}_{\varepsilon}[e^2] \le\lVert f\rVert_{\mathcal{H}_K}^{2}v^{\mathsf{T}}\Sigma_Zv + \sigma_{\varepsilon}^{2}\lVert v^{\mathsf{T}}K_{ZX}A^{-1}\rVert_{2}^{2}. \tag*{(62)}

If we consider the diagonal bound, then (62) can be written as

Eε[e2]≤∥f∥HK2(∑m=1M∣vm∣σ(zm))2+σε2∥∑m=1MvmK(zm,X)(KXX+σε2I)−1∥22,(63)\mathbb{E}_{\varepsilon}[e^2] \le\lVert f\rVert_{\mathcal{H}_K}^{2}\left(\sum_{m=1}^{M}|v_m|\sigma(z_m)\right)^2 + \sigma_{\varepsilon}^{2}\left\lVert\sum_{m=1}^{M}v_mK(z_m,X)(K_{XX}+\sigma_{\varepsilon}^{2}I)^{-1}\right\rVert_{2}^{2}, \tag*{(63)}

where σ2(zm):=(ΣZ)mm\sigma^2(z_m):=(\Sigma_Z)_{mm}.

Proof. Let A=K(X,X)+σε2IA = K(X,X) + \sigma_{\varepsilon}^{2}I. For finite ordered sets Z=(zi)i=1MZ=(z_i)_{i=1}^{M} and X=(xj)j=1NX=(x_j)_{j=1}^{N}, define:

KZX:=(K(zi,xj))i,j.K_{ZX}:=(K(z_i,x_j))_{i,j}.

Since Y=f(X)+εY=f(X)+\varepsilon, we can decompose the error as

e:=F−F^=vTf(Z)−vTKZXA−1Y=vTf(Z)−vTKZXA−1f(X)⏟=:edet,v(f)−vTKZXA−1ε⏟=:enoise,v.(64)\begin{aligned} e &:= \mathcal{F}-\widehat{\mathcal{F}} \\ &= v^{\mathsf{T}}f(Z)-v^{\mathsf{T}}K_{ZX}A^{-1}Y \\ &= \underbrace{v^{\mathsf{T}}f(Z)-v^{\mathsf{T}}K_{ZX}A^{-1}f(X)}_{=:e_{\mathrm{det},v}(f)} -\underbrace{v^{\mathsf{T}}K_{ZX}A^{-1}\varepsilon}_{=:e_{\mathrm{noise},v}}. \tag*{(64)} \end{aligned}

We then evaluate the MSE:

MSE⁡(e)=Eε[e2]=Eε[(edet−enoise)2](65)\operatorname{MSE}(e)=\mathbb{E}_{\varepsilon}[e^2]=\mathbb{E}_{\varepsilon}\left[(e_{\mathrm{det}}-e_{\mathrm{noise}})^2\right] \tag*{(65)}
=edet2+Var⁡(enoise).(66)= e_{\mathrm{det}}^2+\operatorname{Var}(e_{\mathrm{noise}}). \tag*{(66)}

For the noise part, since ε∼N(0,σε2IN)\varepsilon\sim\mathcal{N}(0,\sigma_{\varepsilon}^{2}I_N), we have:

Eε[enoise,v]=0,Var⁡[enoise,v]=σε2∥v⊤KZXA−1∥22.(67)\mathbb{E}_{\varepsilon}[e_{\mathrm{noise},v}] = 0,\qquad\operatorname{Var}[e_{\mathrm{noise},v}] = \sigma_{\varepsilon}^{2}\left\|v^{\top}K_{ZX}A^{-1}\right\|_{2}^{2}. \tag*{(67)}

By the reproducing property,

f(zm)=⟨f,K(zm,⋅)⟩HK,f(xi)=⟨f,K(xi,⋅)⟩HK.(68)f(z_m)=\langle f,K(z_m,\cdot)\rangle_{\mathcal{H}_K},\qquad f(x_i)=\langle f,K(x_i,\cdot)\rangle_{\mathcal{H}_K}. \tag*{(68)}

Therefore, we further write the deterministic error in (4.23) as

edet,v(f)=⟨f,∑m=1Mvm[K(zm,⋅)−K(zm,X)A−1K(X,⋅)]⟩HK=⟨f,rv⟩HK.(69)e_{\mathrm{det},v}(f)=\left\langle f,\sum_{m=1}^{M}v_m\left[K(z_m,\cdot)-K(z_m,X)A^{-1}K(X,\cdot)\right]\right\rangle_{\mathcal{H}_K}=\langle f,r_v\rangle_{\mathcal{H}_K}. \tag*{(69)}

Thus

Eε[e2]=⟨f,rv⟩HK2+σε2∥v⊤KZXA−1∥22.(70)\mathbb{E}_{\varepsilon}[e^{2}]=\langle f,r_v\rangle_{\mathcal{H}_K}^{2}+\sigma_{\varepsilon}^{2}\left\|v^{\top}K_{ZX}A^{-1}\right\|_{2}^{2}. \tag*{(70)}

By applying the Cauchy–Schwarz inequality, we have:

∣⟨f,rv⟩HK∣2≤∥f∥HK2∥rv∥HK2.(71)\left|\langle f,r_v\rangle_{\mathcal{H}_K}\right|^{2}\leq\|f\|_{\mathcal{H}_K}^{2}\|r_v\|_{\mathcal{H}_K}^{2}. \tag*{(71)}

We now compute ∥rv∥HK2\|r_v\|_{\mathcal{H}_K}^{2}. Recall the reproducing property that for ordered sets Z=(zi)i=1MZ=(z_i)_{i=1}^{M} and X=(xj)j=1NX=(x_j)_{j=1}^{N}, and vectors a∈RMa\in\mathbb{R}^{M}, b∈RNb\in\mathbb{R}^{N},

⟨a⊤K(Z,⋅),b⊤K(X,⋅)⟩HK=∑i=1M∑j=1Naibj⟨K(zi,⋅),K(xj,⋅)⟩HK(72)\left\langle a^{\top}K(Z,\cdot),b^{\top}K(X,\cdot)\right\rangle_{\mathcal{H}_K}=\sum_{i=1}^{M}\sum_{j=1}^{N}a_i b_j\left\langle K(z_i,\cdot),K(x_j,\cdot)\right\rangle_{\mathcal{H}_K} \tag*{(72)}
=∑i=1M∑j=1NaibjK(zi,xj)=a⊤KZXb.(73)=\sum_{i=1}^{M}\sum_{j=1}^{N}a_i b_jK(z_i,x_j)=a^{\top}K_{ZX}b. \tag*{(73)}

Then rv(⋅)r_v(\cdot) can be equivalently written as

rv(⋅)=v⊤K(Z,⋅)−α⊤K(X,⋅),(74)r_v(\cdot)=v^{\top}K(Z,\cdot)-\alpha^{\top}K(X,\cdot), \tag*{(74)}

where

α⊤=v⊤KZXA−1\alpha^{\top}=v^{\top}K_{ZX}A^{-1}
K(Z,⋅)=(K(z1,⋅),…,K(zM,⋅))⊤,K(Z,\cdot)=\left(K(z_1,\cdot),\ldots,K(z_M,\cdot)\right)^{\top},
K(X,⋅)=(K(x1,⋅),…,K(xN,⋅))⊤.K(X,\cdot)=\left(K(x_1,\cdot),\ldots,K(x_N,\cdot)\right)^{\top}.

Applying (4.32) yields:

∥rv∥HK2=⟨v⊤K(Z,⋅)−α⊤K(X,⋅),v⊤K(Z,⋅)−α⊤K(X,⋅)⟩HK(75)\|r_v\|_{\mathcal{H}_K}^{2}=\left\langle v^{\top}K(Z,\cdot)-\alpha^{\top}K(X,\cdot),v^{\top}K(Z,\cdot)-\alpha^{\top}K(X,\cdot)\right\rangle_{\mathcal{H}_K} \tag*{(75)}
=⟨v⊤K(Z,⋅),v⊤K(Z,⋅)⟩HK−2⟨v⊤K(Z,⋅),α⊤K(X,⋅)⟩HK(76)=\left\langle v^{\top}K(Z,\cdot),v^{\top}K(Z,\cdot)\right\rangle_{\mathcal{H}_K}-2\left\langle v^{\top}K(Z,\cdot),\alpha^{\top}K(X,\cdot)\right\rangle_{\mathcal{H}_K} \tag*{(76)}
+⟨α⊤K(X,⋅),α⊤K(X,⋅)⟩HK(77)+\left\langle\alpha^{\top}K(X,\cdot),\alpha^{\top}K(X,\cdot)\right\rangle_{\mathcal{H}_K} \tag*{(77)}
=v⊤KZZv−2v⊤KZXA−1KXZv+v⊤KZXA−1KXXA−1KXZv(78)=v^{\top}K_{ZZ}v-2v^{\top}K_{ZX}A^{-1}K_{XZ}v+v^{\top}K_{ZX}A^{-1}K_{XX}A^{-1}K_{XZ}v \tag*{(78)}
=v⊤KZZv−v⊤KZXA−1KXZv−σε2v⊤KZXA−2KXZv(79)=v^{\top}K_{ZZ}v-v^{\top}K_{ZX}A^{-1}K_{XZ}v-\sigma_{\varepsilon}^{2}v^{\top}K_{ZX}A^{-2}K_{XZ}v \tag*{(79)}
=v⊤ΣZv−σε2v⊤KZXA−2KXZv(80)=v^{\top}\Sigma_Zv-\sigma_{\varepsilon}^{2}v^{\top}K_{ZX}A^{-2}K_{XZ}v \tag*{(80)}
=v⊤ΣZv−σε2∥A−1KXZv∥22≤v⊤ΣZv.(81)=v^{\top}\Sigma_Zv-\sigma_{\varepsilon}^{2}\left\|A^{-1}K_{XZ}v\right\|_{2}^{2}\leq v^{\top}\Sigma_Zv. \tag*{(81)}

since ∥A−1KXZv∥22≥0\left\|A^{-1}K_{XZ}v\right\|_{2}^{2} \ge0 for KXX⪰0K_{XX} \succeq0 and symmetric positive definite AA. Hence with (4.29) and (4.40), we obtain:

Eε[e2]≤∥f∥HK2 v⊤ΣZv+σε2∥v⊤KZXA−1∥22.(82)\mathbb{E}_{\varepsilon}[e^{2}] \le\|f\|_{\mathcal{H}_{K}}^{2}\,v^{\top}\Sigma_{Z}v+\sigma_{\varepsilon}^{2}\left\|v^{\top}K_{ZX}A^{-1}\right\|_{2}^{2}. \tag*{(82)}

The decomposition of error into noise and deterministic components is readily interpretable. The deterministic component captures how well the training data at XX could reconstruct the functional v⊤f(Z)v^{\top}f(Z) if the observations were noiseless. This error is largely controlled by the RKHS norm (or complexity) of the true function ff and by the posterior covariance at evaluation points ZZ. The noise component, controlled by σε2∥v⊤KZXA−1∥22\sigma_{\varepsilon}^{2}\left\|v^{\top}K_{ZX}A^{-1}\right\|_{2}^{2}, captures how much the observation noise is amplified through the kernel reconstruction and through the linear functional v⊤v^{\top}.

Finally, we derive the diagonal posterior standard deviation from the full covariance form. Recall that ΣZ\Sigma_{Z} is a positive semidefinite block matrix. We therefore have

∣(ΣZ)mn∣≤(ΣZ)mm(ΣZ)nn=σ(zm)σ(zn).(83)\left|(\Sigma_{Z})_{mn}\right| \le\sqrt{(\Sigma_{Z})_{mm}}\sqrt{(\Sigma_{Z})_{nn}}=\sigma(z_m)\sigma(z_n). \tag*{(83)}

Thus, we have:

v⊤ΣZv=∑m,n=1Mvmvn(ΣZ)mn≤∑m,n=1M∣vm∣∣vn∣σ(zm)σ(zn)=(∑m=1M∣vm∣σ(zm))2.(84)\begin{aligned} v^{\top}\Sigma_{Z}v=\sum_{m,n=1}^{M}v_m v_n(\Sigma_{Z})_{mn} \\ &\le\sum_{m,n=1}^{M}|v_m||v_n|\sigma(z_m)\sigma(z_n) \\ &=\left(\sum_{m=1}^{M}|v_m|\sigma(z_m)\right)^{2}. \tag*{(84)} \end{aligned}

Substituting (4.43) into (4.21) gives the desired error bound in the diagonal form (63). This bound can be interpreted as yielding large errors if the functional places a large weight on uncertain locations.

Results

In this section, we demonstrate the performance of our proposed method on various problems, including the 1D Poisson equation with analytic solution (Subsection 5.1), the nonlinear diffusion equation with manufactured solution (Subsection 5.2), and the 2D advection-diffusion equation on a bell-shaped complex geometry (Subsection 5.3). We also present results for a semiconductor p−np-n diode problem (Subsection 5.4) to illustrate the range over which the uncertainty estimates remain reliable for the proposed method. For the training of data-driven Whitney forms, we consider a lightweight cross-attention transformer with dropout.

Toy one-dimensional linear example

Consider a toy example of the one-dimensional Poisson equation on Ω=(0,1)\Omega=(0,1):

F=−∇u,∇⋅F=f,(85)F=-\nabla u,\qquad\nabla\cdot F=f, \tag*{(85)}

with Dirichlet boundary conditions u(0)=αu(0)=\alpha, u(1)=0u(1)=0, and constant source term f≡1f\equiv1. The analytic solution is:

u(x)=−x22+(12−α)x+α,F(x)=x+α−12.(86)u(x)=-\frac{x^{2}}{2}+\left(\frac{1}{2}-\alpha\right)x+\alpha,\qquad F(x)=x+\alpha-\frac{1}{2}. \tag*{(86)}

The goal is to learn the solution uu and the associated flux FF given the left boundary condition α\alpha. In particular, we use this toy example because the flux depends linearly on the left boundary condition.

In the implementation, to construct the fine-scale discretization, we employ the classic Galerkin finite element with P0P_0 elements for solution uu and P1P_1 Lagrange elements for the flux. For a mesh with NN cells, we have Ncell=NN_{\mathrm{cell}}=N cell-averaged values for uu and Nface=N+1N_{\mathrm{face}}=N+1 nodal flux values for FF. Then the sparse basis evaluations in the fine scale ϕ0∈RNcell×Nq\phi^0 \in\mathbb{R}^{N_{\mathrm{cell}}\times N_q} and ϕ1∈RNface×Nq\phi^1 \in\mathbb{R}^{N_{\mathrm{face}}\times N_q} are constructed with the quadrature weight exponent 12\frac{1}{2}. In one dimension, this is analogous to the lowest-order Raviart–Thomas space RT0\mathrm{RT}_0 where each hat function ϕi1\phi_i^1 is supported on the two cells adjacent to node ii, and its derivative dϕi1/dx\mathrm{d}\phi_i^1/\mathrm{d}x plays the role of ∇⋅ϕi1\nabla\cdot\phi_i^1. The resulting P0P_0–P1P_1 pair yields the standard mixed P0P_0–H(div⁡)H(\operatorname{div}) formulation in 1D. In addition, for a prescribed number of PoU, we have N0=NPoUN_0=N_{\mathrm{PoU}} for the coarse 0-forms and N1int=(N02)N_1^{\mathrm{int}}=\binom{N_0}{2} for the interior coarse 1-forms, respectively. For the 1D problem, we explicitly set two boundary 1-forms ψ1left\psi_1^{\mathrm{left}} and ψ1right\psi_1^{\mathrm{right}}, which gives the N1bc=2N_1^{\mathrm{bc}}=2 and total number of coarse 1-forms N1=(N02)+2N_1=\binom{N_0}{2}+2.

Comparison of the ground truth with the prediction is shown in Figure 3 (left). The reconstructed solution field from the coarse 0-forms can approximate the ground-truth solution with the MSE below 10−510^{-5}. Figure 3 (right) illustrates the GP inference of the left boundary flux with respect to different values of α\alpha. For the 1D problem where the target boundary Γ\Gamma is reduced to the left boundary point x=0x=0, the coarse-to-fine reconstruction of flux in Proposition 4.3 is simplified to use the coarse boundary 1-form for the left boundary. Furthermore, as this problem is linear, the error bound is still tight near the training range even in the small extrapolation area, with only a slight increase in the most extreme cases, α=0.5\alpha=0.5 and α=3.5\alpha=3.5.

Image not extracted

Figure 3. Solution reconstruction and quantified uncertainty for the toy example. Left: Solution field uu with respect to different left boundary conditions α\alpha. The black solid line is the analytic solution, and the red dashed line is the reconstructed solution. Right: Inference result for the left boundary flux with respect to left boundary conditions α\alpha with shaded blue area as the error bound from Theorem 4.4.

1D nonlinear example with manufactured solution

We next consider a nonlinear diffusion equation with piecewise nonlinear diffusivity and the same Dirichlet boundary conditions as in Subsection 5.1:

F=−k(u)∇u,∇⋅F=f,(87)F=-k(u)\nabla u,\qquad\nabla\cdot F=f, \tag*{(87)}

where

k(u)={k0if u≤u0,β(u−u0)q+k0if u>u0.(88)k(u)= \begin{cases} k_0 & \text{if } u \le u_0,\\ \beta(u-u_0)^q+k_0 & \text{if } u>u_0. \end{cases} \tag*{(88)}

The analytic solution is u(x)=α(1−x)u(x)=\alpha(1-x), which gives F(x)=αk(α(1−x))F(x)=\alpha k(\alpha(1-x)) and

f(x)=∇⋅F(x)={0if α(1−x)≤u0,−βqα2(α(1−x)−u0)q−1if α(1−x)>u0.f(x)=\nabla\cdot F(x)= \begin{cases} 0 & \text{if } \alpha(1-x)\le u_0,\\ -\beta q\alpha^2\left(\alpha(1-x)-u_0\right)^{q-1} & \text{if } \alpha(1-x)>u_0. \end{cases}

This problem features a nonlinear constitutive law k(u)k(u), a spatially varying source f(x)f(x) that depends on the parameter α\alpha, and a kink in kk at the threshold u=u0u=u_0. Therefore, we have a manufactured problem where the flux is a nonlinear function of the left boundary condition α\alpha with a transition point at x=1−u0αx=1-\frac{u_0}{\alpha}. The visualiza-

tion of coarse shape functions {ψj0}\{\psi_j^0\} in Figure 4 demonstrates that the conditional neural Whitney form allows an adaptive adjustment of the PoU. As α\alpha changes, the coarse basis functions reallocate their support to better accommodate the condition. One coarse basis function becomes much more dominant near the left side at larger values of α\alpha, suggesting that the solution structure there becomes more parameter-sensitive. For this example with both linear and nonlinear regions, the error bound grows more noticeably in the extrapolation range than in the previous linear example (Figure 5). Furthermore, in the left extrapolation region where the flux is still linear in the conditioning variable, the GP-inferred flux deviates less from the ground-truth than in the nonlinear extrapolation region on the right. The error bound grows in both extrapolation directions because the distance in the GP feature space from the training samples increases regardless of whether the extrapolation is in the linear regime or the nonlinear regime.

Image not extracted

Figure 4. Conditional coarse bases for Subsection 5.2. Coarse shape functions {ψi0}\{\psi_i^0\} evaluated on the fine-scale nodes can adapt to different values α=[1,2.1,3]\alpha=[1,2.1,3].

Image not extracted

Figure 5. Quantified uncertainty for the left-boundary flux in the 1D nonlinear problem. The learned surrogate predicts the boundary flux FΓF_\Gamma as a function of the left Dirichlet boundary value α\alpha. The growth of the uncertainty band outside the training range in α\alpha reflects reduced confidence in extrapolation.

Conditioned 2D advection-diffusion equation in complex geometry

We extend the problem to the steady advection-diffusion equation in a complex geometry and an unstructured mesh:

−∇⋅(k∇u)+β⋅∇u=0in Ω,(89)-\nabla\cdot(k \nabla u) + \boldsymbol{\beta} \cdot\nabla u = 0 \quad\text{in } \Omega, \tag*{(89)}

where kk is the diffusion coefficient and β=(cos⁡θ,sin⁡θ)\boldsymbol{\beta} = (\cos\theta, \sin\theta) is the advection direction with θ∈[0,2π)\theta\in[0, 2\pi). The domain is a two-dimensional “bell”-shaped domain Ω\Omega with three distinct Dirichlet boundary conditions as shown in Figure 6. We use an unstructured triangular mesh to discretize the complex bell geometry and parameterize the problem with both θ\theta and ucracku_{\mathrm{crack}}. Specifically, we have the top handle of the bell (Γ1\Gamma_1) with u=1u = 1, the crack boundary (Γ2\Gamma_2) with u=ucracku = u_{\mathrm{crack}}, and the remaining bell-body boundary (Γ3\Gamma_3) with u=0u = 0. The conditioning variable θ\theta controls the advection direction and thus changes the transport pattern inside the domain.

Image not extracted

Figure 6. Complex geometry demonstration. Left figure shows the irregular triangular mesh and right figure shows the nonhomogeneous Dirichlet boundary conditions.

Results are shown in Figure 7. In the implementation, we first train the transformer for the Whitney forms with a dropout rate of 0.1 for 100,000 epochs and then turn off the dropout for another 20,000 epochs. Empirically, this two-stage training strategy alleviates overfitting and prevents the transformer from getting stuck in local minima during training. Figure 8 visualizes the Whitney 0-forms learned by the neural PoU framework for three different values of the conditioning variables, showing that the learned partitions adapt naturally to the geometric structure. As θ\theta and ucracku_{\mathrm{crack}} increase, some 0-forms tend to localize near the crack interface to capture changes near the crack boundary efficiently, while others tend to align with the flow near other boundaries or within the domain.

Image not extracted

Figure 7. Quantified uncertainty on complex geometry. Top: The solution to the advection-diffusion equation, conditioned on the advection direction, is recovered. Bottom: Dropout during training provides a high-quality basis; turning off at 100k epochs allows refinement of the final basis (bottom left). The error in the flux through the crack at the bottom of the Liberty Bell is bounded by the estimator in Equation (63) (bottom right).

Image not extracted

Figure 8. Conditioned basis adapts to advection. The three basis functions are shown conditioned on angles θ∈[0.45π,0.75π,1.05π]\theta\in[0.45\pi, 0.75\pi, 1.05\pi], respectively, highlighting the ability of the basis to adapt itself to the relevant conditioning.

p-n diodes

Finally, we consider the p−np-n diode problem obtained from Charon TCAD simulations [26] on the rectangular device domain Ω=(0,Lx)×(0,Ly)\Omega= (0, L_x) \times(0, L_y) with Lx=1L_x = 1 and Ly=0.5L_y = 0.5. For each applied anode bias VaV_a, the data consist of the electrostatic potential, electron and hole current densities, doping profiles, and the terminal currents at the cathode and anode. The cathode bias is fixed and the anode bias is swept over [−1,1][-1, 1] in thermal-voltage units; multiplication by V0=2.585×10−2 VV_0 = 2.585 \times10^{-2}\,\mathrm{V} gives the corresponding physical voltage scale.

The two-dimensional TCAD fields are reduced to one-dimensional profiles in a way that preserves the terminal-current constraint. Let φ(x,y;Va)\varphi(x,y; V_a) denote the electrostatic potential and let jx(x,y;Va)j_x(x,y; V_a) denote the net conventional current density in the axial direction. On the structured 100 ×\times 10 cell grid, we define:

u(x;Va)=1Ly∫0Lyφ(x,y;Va) dy,J(x;Va)=∫0Lyjx(x,y;Va) dy.(90)u(x; V_a) = \frac{1}{L_y}\int_0^{L_y} \varphi(x,y; V_a)\,\mathrm{d}y,\qquad J(x; V_a) = \int_0^{L_y} j_x(x,y; V_a)\,\mathrm{d}y. \tag*{(90)}

Since the steady drift–diffusion solution satisfies charge conservation (∇⋅j=0\nabla\cdot\mathbf{j} = 0) and the top and bottom boundaries are insulating, integration over the cross section gives:

ddxJ(x;Va)=0,J(x;Va)≃Ia(Va).(91)\frac{d}{dx}J(x;V_a)=0,\qquad J(x;V_a)\simeq I_a(V_a). \tag*{(91)}

We use the same conditional Whitney–GP construction as in previous 1D examples, where the neural Whitney forms are generated by the transformer conditioned on VaV_a. Given that the current values vary over several orders of magnitude, we train the GP on the logarithm of the current magnitude and map the posterior mean back to the original scale by exponentiation, i.e., the posterior mean μ(Va)\mu(V_a) is mapped back to the original current magnitude by exp⁡(μ(Va))\exp(\mu(V_a)). The numerical results are shown in Figure 9 with the inference in original scale on the left and in log scale on the right.

Image not extracted

Figure 9. Quantified uncertainty for the I-V response of a semiconductor component. Top: schematic of the one-dimensional p-n diode and the prescribed anode–cathode orientation. Bottom: Comparison of the ground-truth current and predicted current in original scale (bottom left) and log scale (bottom right). Despite the appearance of a relatively good current reconstruction on a linear scale (bottom left), on a log scale the uncertainty estimator reveals the range of voltages over which the model may be reliably applied.

Despite the fairly good current reconstruction on a linear scale, the uncertainty estimator reveals the range of voltages over which the model may be reliably applied, and flags where the uncertainty becomes large relative to the scale of the quantity of interest. This example should be considered an illustration and guideline for where the proposed model starts to become less trustworthy, i.e., the posterior uncertainty grows rapidly in the exponential regime.

Conclusion

In this work, we developed a structure-preserving neural surrogate framework for PDE-governed systems that incorporates a reduced finite element space and tractable uncertainty quantification with a closed-form posterior error bound. Our method separates the problem into two coupled parts: in P1, the trainable PoU Whitney forms generate the reduced-order H(div⁡)H(\operatorname{div})–L2L^2 spaces that preserve the conservation law with the induced coarse divergence operator and support the graph interpretation; in P2, the GP learns the nonlinear state-to-flux map on the coarse graph by solving the optimal recovery problem with an equality constraint that enforces the conservation exactly. In particular, the conditional Whitney forms provide an adaptive mechanism that enables generalization across different geometries, and the optimal recovery formulation yields a saddle-point KKT condition with a fast Schur-complement solve. Numerical results demonstrate that the proposed method can reconstruct accurate solution fields and fluxes and provide meaningful uncertainty quantification for boundary flux functionals. In future work, we will generalize the framework to more complex problems and real datasets for practical applications, such as large-scale deployments for climate modeling. In addition, the current structurepreserving construction is based on the de Rham complex and is primarily designed for scalar- and vector-valued problems; extending the method to handle tensor-valued variables is an important direction for future work.

Acknowledgments. H. Zhang and Dr. Kinch acknowledge support from the United States Department of Energy under the Advanced Scientific Computing Research (ASCR) program (award number DE-SC0024563). Drs. Kinch, Owhadi, Propp, and Trask acknowledge funding under the ASCR-sponsored Mathematical Multifaceted Integrated Capability Centers program (award number DE-SC0023163).

References

Appendix A. Proof for prop:rank

Proof. For a boundary edge e∈Eγe \in\mathcal{E}_{\gamma}, the divergence ∇⋅ϕe1\nabla\cdot\phi^{1}_{e} is supported entirely on its unique adjacent cell K(e)K(e) and is zero everywhere else. Let βeγ=(∇⋅ϕe1,ϕK(e)P0)Ω\beta^{\gamma}_{e} = (\nabla\cdot\phi^{1}_{e},\phi^{P_{0}}_{K(e)})_{\Omega}. Since the coarse 0-form and boundary 1-form are defined as

ψi0=∑a∈ChWiaϕaP0,ψα,γ1,bc=∑e∈EγWα,K(e)ϕe1,\psi^{0}_{i} = \sum_{a\in C_{h}} W_{ia}\phi^{P_{0}}_{a}, \qquad\psi^{1,\mathrm{bc}}_{\alpha,\gamma} = \sum_{e\in\mathcal{E}_{\gamma}} W_{\alpha,K(e)}\phi^{1}_{e},

we have

(Dγ)iα=(∇⋅ψα,γ1,bc,ψi0)Ω=∑e∈EγWα,K(e)(∇⋅ϕe1,ψi0)Ω=∑e∈EγWα,K(e)Wi,K(e)(∇⋅ϕe1,ϕK(e)P0)Ω=∑e∈EγβeγWα,K(e)Wi,K(e).\begin{aligned} (D_{\gamma})_{i\alpha} &= (\nabla\cdot\psi^{1,\mathrm{bc}}_{\alpha,\gamma},\psi^{0}_{i})_{\Omega} \\ &= \sum_{e\in\mathcal{E}_{\gamma}} W_{\alpha,K(e)}(\nabla\cdot\phi^{1}_{e},\psi^{0}_{i})_{\Omega} \\ &= \sum_{e\in\mathcal{E}_{\gamma}} W_{\alpha,K(e)}W_{i,K(e)}(\nabla\cdot\phi^{1}_{e},\phi^{P_{0}}_{K(e)})_{\Omega} \\ &= \sum_{e\in\mathcal{E}_{\gamma}} \beta^{\gamma}_{e}W_{\alpha,K(e)}W_{i,K(e)}. \end{aligned}

This shows Dγ=WγBγWγ⊤D_{\gamma} = W_{\gamma}B_{\gamma}W_{\gamma}^{\top}, where (Wγ)ie=Wi,K(e)(W_{\gamma})_{ie} = W_{i,K(e)}, Bγ=diag⁡{βeγ:e∈Eγ}B_{\gamma} = \operatorname{diag}\{\beta^{\gamma}_{e}:e\in\mathcal{E}_{\gamma}\}. Then it remains to check the sufficient condition. Let

D∂=[D1 D2 ⋯ Dr]D_{\partial} = [D_{1}\ D_{2}\ \cdots\ D_{r}]

denote the boundary block of the divergence matrix DD, collecting the boundary groups γ=1,…,r\gamma= 1,\ldots,r. Given D∈RN0×N1D \in\mathbb{R}^{N_{0}\times N_{1}}, we have

rank⁡(D)≤N0.\operatorname{rank}(D) \le N_{0}.

If D∂D_{\partial} has full row rank N0N_{0}, since D∂D_{\partial} is the column submatrix of DD corresponding to the boundary basis functions, we have

rank⁡(D)≥rank⁡(D∂)=N0,\operatorname{rank}(D) \ge\operatorname{rank}(D_{\partial}) = N_{0},

and therefore rank⁡(D)=N0\operatorname{rank}(D) = N_{0}. Let G∂=∑γ=1rWγBγWγ⊤G_{\partial} = \sum_{\gamma=1}^{r} W_{\gamma}B_{\gamma}W_{\gamma}^{\top} be positive definite. For any x∈RN0x \in\mathbb{R}^{N_{0}} such that x⊤D∂=0x^{\top}D_{\partial}=0, we have

0=x⊤(∑γ=1rDγ)x=x⊤G∂x⟹x=00 = x^{\top}\left(\sum_{\gamma=1}^{r}D_{\gamma}\right)x = x^{\top}G_{\partial}x \Longrightarrow x=0

since G∂≻0G_{\partial} \succ0. Hence Null⁡(D∂⊤)={0}\operatorname{Null}(D_{\partial}^{\top})=\{0\}, and by the rank–nullity theorem, D∂D_{\partial} has full row rank N0N_{0}.

References

  1. [1]K. Aadithya, P. Kuberry, B. Paskaleva, P. Bochev, K. Leeson, A. Mar, T. Mei, and E. Keiter, Data-driven compact models for circuit design and analysis, in Proceedings of The First Mathematical and Scientific Machine Learning Conference, J. Lu and R. Ward, eds., vol. 107 of Proceedings of Machine Learning Research, PMLR, 20–24 Jul 2020, pp. 555–569, https://proceedings.mlr.press/v107/aadithya20a.html.
  2. [2]M. Abdar, F. Pourpanah, S. Hussain, D. Rezazadegan, L. Liu, M. Ghavamzadeh, P. Fieguth, X. Cao, A. Khosravi, U. R. Acharya, et al., A review of uncertainty quantification in deep learning: Techniques, applications and challenges, Information fusion, 76 (2021), pp. 243–297.DOI
  3. [3]J. A. Actor, X. Hu, A. Huang, S. A. Roberts, and N. Trask, Data-driven whitney forms for structure-preserving control volume analysis, Journal of Computational Physics, 496 (2024), p. 112520.
  4. [4]D. Arnold, R. Falk, and R. Winther, Finite element exterior calculus: from hodge theory to numerical stability, Bulletin of the American mathematical society, 47 (2010), pp. 281–354.DOI
  5. [5]D. N. Arnold, R. S. Falk, and R. Winther, Finite element exterior calculus, homological techniques, and applications, Acta Numerica, 15 (2006), p. 1–155, https://doi.org/10.1017/S0962492906210018.
  6. [6]P. Benner, S. Gugercin, and K. Willcox, A survey of projection-based model reduction methods for parametric dynamical systems, SIAM Review, 57 (2015), pp. 483–531, https://doi.org/10.1137/130932715.
  7. [7]D. Boffi, F. Brezzi, and M. Fortin, Mixed Finite Element Methods and Applications, vol. 44 of Springer Series in Computational Mathematics, Springer, Berlin, Heidelberg, 2013, https://doi.org/10.1007/978-3-642-36519-5.
  8. [8]A. Bonfanti, G. Bruno, and C. Cipriani, The challenges of the nonlinear regime for physics-informed neural networks, Advances in neural information processing systems, 37 (2024), pp. 41852–41881.arxiv.org/abs/2402.03864
  9. [9]S. L. Brunton and J. N. Kutz, Promising directions of machine learning for partial differential equations, Nature Computational Science, 4 (2024), pp. 483–494.DOI
  10. [10]K. Carlberg, Y. Choi, and S. Sargsyan, Conservative model reduction for finite-volume models, Journal of Computational Physics, 371 (2018), pp. 280–314, https://doi.org/10.1016/j.jcp.2018.05.019.
  11. [11]H. Chen, G. E. C. Flores, and C. Li, Physics-informed neural networks with hard linear equality constraints, Computers & Chemical Engineering, 189 (2024), p. 108764.arxiv.org/abs/2402.07251
  12. [12]Y. Chen, B. Hosseini, H. Owhadi, and A. M. Stuart, Solving and learning nonlinear pdes with gaussian processes, Journal of Computational Physics, 447 (2021), p. 110668.arxiv.org/abs/2103.12959
  13. [13]D. Fan, W. Li, H. Qiu, Y. Xu, S. Gao, L. Liu, T. Li, F. Huang, Y. Mao, W. Zhou, et al., Two-dimensional semiconductor integrated circuits operating at gigahertz frequencies, Nature Electronics, 6 (2023), pp. 879–887.
  14. [14]S. Fresca and A. Manzoni, Pod-dl-rom: Enhancing deep learning-based reduced order models for nonlinear parameterized pdes by proper orthogonal decomposition, Computer Methods in Applied Mechanics and Engineering, 388 (2022), p. 114181.DOI
  15. [15]V. Gupta, T. Koren, and Y. Singer, Shampoo: Preconditioned stochastic tensor optimization, in International Conference on Machine Learning, PMLR, 2018, pp. 1842–1850.arxiv.org/abs/1802.09568
  16. [16]M. Harkonen, M. Lange-Hegermann, and B. Raita, Gaussian process priors for systems of linear partial differential equations with constant coefficients, in International conference on machine learning, PMLR, 2023, pp. 12587–12615.arxiv.org/abs/2212.14319
  17. [17]A. Jiao, Q. Yan, J. Harlim, and L. Lu, Solving forward and inverse pde problems on unknown manifolds via physics-informed neural operators, arXiv preprint arXiv:2407.05477, (2024).arxiv.org/abs/2407.05477
  18. [18]K. S. Jung, C. E. Lacey, H. Babaee, and J. H. Chen, Accelerating high-fidelity simulations of chemically reacting flows using reduced-order modeling with time-dependent bases, Computer Methods in Applied Mechanics and Engineering, 437 (2025), p. 117758.DOI
  19. [19]M. G. Kapteyn, D. J. Knezevic, D. Huynh, M. Tran, and K. E. Willcox, Data-driven physics-based digital twins via a library of component-based reduced-order models, International Journal for Numerical Methods in Engineering, 123 (2022), pp. 2986–3003.
  20. [20]G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang, Physics-informed machine learning, Nature Reviews Physics, 3 (2021), pp. 422–440.DOI
  21. [21]B. Kinch, B. Shaffer, E. Armstrong, M. Meehan, J. Hewson, and N. Trask, Structure-preserving digital twins via conditional neural Whitney forms, arXiv preprint arXiv:2508.06981, (2025).arxiv.org/abs/2508.06981
  22. [22]A. Krishnapriyan, A. Gholami, S. Zhe, R. Kirby, and M. Mahoney, Characterizing possible failure modes in physics-informed neural networks, Advances in neural information processing systems, 34 (2021), pp. 26548–26560.arxiv.org/abs/2109.01050
  23. [23]Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anandkumar, Fourier neural operator for parametric partial differential equations, arXiv preprint arXiv:2010.08895, (2020).arxiv.org/abs/2010.08895
  24. [24]L. Lu, P. Jin, G. Pang, H. Zhang, and G. E. Karniadakis, Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators, Nature Machine Intelligence, 3 (2021), pp. 218–229, https://doi.org/10.1038/s42256-021-00302-5.
  25. [25]C. A. Micchelli and T. J. Rivlin, A survey of optimal recovery, Optimal estimation in approximation theory, (1977), pp. 1–54.DOI
  26. [26]L. Musson, X. Gao, M. Negoita, A. Huang, and G. L. Hennigan, Charon: A radiation aware massively parallel tcad modeling code, Sandia National Laboratories (SNL-NM), Albuquerque, NM (United States), 11 2018, https://www.osti.gov/biblio/1575982.
  27. [27]L. Musson, G. Hennigan, X. Gao, R. Humphreys, M. Negoita, and A. Huang, Charon user manual: v. 2.2 (revision1), tech. report, Sandia National Laboratories (SNL-NM), Albuquerque, NM (United States), 2022.DOI
  28. [28]National Academies of Sciences, Engineering, and Medicine, Division on Engineering and Physical Sciences, Board on Mathematical Sciences and Analytics, and Committee on Foundation Models for Scientific Discovery and Innovation, Foundation Models for Scientific Discovery and Innovation: Opportunities Across the Department of Energy and the Scientific Enterprise, The National Academies Press, Washington, DC, 2025, https://doi.org/10.17226/29212, https://nap.nationalacademies.org/catalog/29212/foundation-models-for-scientific-discovery-and-innovation-opportunities-across-the.
  29. [29]H. Owhadi, Computational graph completion, Research in the Mathematical Sciences, 9 (2022), p. 27.
  30. [30]A. M. Propp, J. A. Actor, E. Walker, H. Owhadi, N. Trask, and D. M. Tartakovsky, Discovery of probabilistic dirichlet-to-neumann maps on graphs, SIAM Journal on Scientific Computing, 48 (2026), pp. C191–C215.arxiv.org/abs/2506.02337
  31. [31]A. F. Psaros, X. Meng, Z. Zou, L. Guo, and G. E. Karniadakis, Uncertainty quantification in scientific machine learning: Methods, metrics, and comparisons, Journal of Computational Physics, 477 (2023), p. 111902.DOI
  32. [32]A. Quarteroni, A. Manzoni, and F. Negri, Reduced Basis Methods for Partial Differential Equations: An Introduction, vol. 92 of Unitext, Springer, Cham, 2016, https://doi.org/10.1007/978-3-319-15431-2.
  33. [33]C. E. Rasmussen and C. K. I. Williams, Gaussian Processes for Machine Learning, The MIT Press, Cambridge, Mass., 2006, http://www.gaussianprocess.org/gpml/.DOI
  34. [34]M. Ruderman, S. Kaltenbacher, and M. Horn, Pressure-flow dynamics with semi-stable limit cycles in hydraulic cylinder circuits, in 2021 IEEE International Conference on Mechatronics (ICM), 2021, pp. 1–6, https://doi.org/10.1109/ICM46511.2021.9385620.
  35. [35]Y. Shi, P. Wei, K. Feng, D.-C. Feng, and M. Beer, A survey on machine learning approaches for uncertainty quantification of engineering systems, Machine learning for computational science and engineering, 1 (2025), p. 11.
  36. [36]H. Song, Z. Li, M. Deng, X. Li, D. Pan, Z. Lai, and W. Wang, Structure-aware epistemic uncertainty quantification for neural operator pde surrogates, arXiv preprint arXiv:2603.11052, (2026).arxiv.org/abs/2603.11052
  37. [37]N. Trask, A. Huang, and X. Hu, Enforcing exact physics in scientific machine learning: a data-driven exterior calculus on graphs, Journal of Computational Physics, 456 (2022), p. 110969.arxiv.org/abs/2012.11799
  38. [38]A. Vacca and G. Franzoni, Hydraulic fluid power: Fundamentals, applications, and circuit design, John Wiley & Sons, 2021.DOI
  39. [39]S. WANG, A. L. COTTRILL, Y. KUNAI, A. R. TOLAND, P. LIU, W.-J. WANG, AND M. S. STRANO, Microscale solid-state thermal diodes enabling ambient temperature thermal circuits for energy applications, Physical Chemistry Chemical Physics, 19 (2017), pp. 13172–13181.DOI
  40. [40]S. WANG, H. WANG, AND P. PERDIKARIS, Learning the solution operator of parametric partial differential equations with physics-informed DeepONets, Science Advances, 7 (2021), p. eabi8605, https://doi.org/10.1126/sciadv.abi8605.
  41. [41]Z. XU, Y. GUO, AND J. H. SALEH, Accurate remaining useful life prediction with uncertainty quantification: a deep learning and nonstationary gaussian process approach, IEEE Transactions on Reliability, 71 (2021), pp. 443–456.arxiv.org/abs/2109.12111
  42. [42]Y. YANG AND P. PERDIKARIS, Adversarial uncertainty quantification in physics-informed neural networks, Journal of Computational Physics, 394 (2019), pp. 136–152.DOI
  43. [43]J. YU, L. LU, X. MENG, AND G. E. KARNIADAKIS, Gradient-enhanced physics-informed neural networks for forward and inverse PDE problems, Computer Methods in Applied Mechanics and Engineering, 393 (2022), p. 114823.DOI

Paper details

vlnc:W_9sv6c5fp
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents