Koopman-Enhanced Graph Conv. Network

• K-GCN is a graph-based encoder–decoder that embeds node states into a latent space evolved by a learned linear Koopman operator to approximate nonlinear dynamics.
• It underpins two variants—P-K-GCN for physics-augmented super-resolution and TK-GCN for long-horizon forecasting via Transformer integration.
• The approach enhances spatial detail and temporal consistency on irregular meshes while controlling capacity and reducing error accumulation.

Searching arXiv for the cited K-GCN papers to ground the article in the latest preprints. Koopman-enhanced Graph Convolutional Network (K-GCN) denotes a graph-based encoder–decoder for spatiotemporal dynamics on irregular geometric domains in which node-wise states are embedded into a latent space whose temporal evolution is constrained by a learned linear operator $K$, interpreted as a finite-dimensional approximation of the Koopman operator. In the cited literature, K-GCN appears in two closely related roles: as the core of Physics-augmented Koopman-enhanced Graph Convolutional Network (P-K-GCN) for spatiotemporal super-resolution on irregular geometries, and as the first-stage latent representation model in Transformer with Koopman-Enhanced Graph Convolutional Network (TK-GCN) for long-horizon forecasting. Across these formulations, the central idea is to couple geometry-aware graph convolutions with approximate linearization of nonlinear temporal dynamics in latent space, thereby improving temporal consistency, stability, and fidelity on non-Euclidean meshes (Xizhuo et al., 17 Jun 2026, Wang et al., 5 Jul 2025).

1. Definition and model variants

In the formulation used for super-resolution, K-GCN is defined as a graph-based encoder–decoder architecture for irregular spatial domains, where the latent representations produced by the encoder evolve in time via a learned linear operator $\mathbf{K}$, instead of an arbitrary nonlinear recurrent block; both the instantaneous reconstructions and the Koopman-advanced predictions are decoded back to high-resolution fields and supervised in the low-resolution space. In the forecasting formulation, K-GCN is the Stage-1 model in a two-stage framework: it learns a graph-based encoder–decoder and a linear latent operator $K$, after which a Transformer models long-range temporal dependencies directly in the Koopman-encoded latent space (Xizhuo et al., 17 Jun 2026, Wang et al., 5 Jul 2025).

Variant	Defining components	Role
K-GCN	Graph-based encoder–decoder + linear Koopman latent dynamics	Baseline in P-K-GCN; Stage-1 in TK-GCN
P-K-GCN	K-GCN + physics-augmented loss	Spatiotemporal super-resolution on irregular geometries
TK-GCN	K-GCN + Transformer stage	Spatiotemporal dynamics forecasting

A central distinction in the cited works is that physics augmentation is not intrinsic to K-GCN itself. K-GCN consists of graph-based spatial encoding and decoding together with linear latent evolution. P-K-GCN extends this core by adding PDE-residual losses, while TK-GCN uses K-GCN to shape a latent space in which a Transformer operates. This separation is important because it makes clear that Koopman enhancement and physics augmentation are distinct inductive biases rather than a single inseparable mechanism.

2. Spatial representation on irregular geometries

K-GCN is designed for domains represented by irregular meshes rather than Cartesian grids. In P-K-GCN, the physical domain is a 3D heart surface represented as an undirected graph

$$\mathcal{G} = (\mathcal{V},\mathcal{E},\mathbf{W},\boldsymbol{q}),$$

where nodes correspond to mesh vertices or tissue points, edges encode mesh connectivity, and each non-zero edge attribute $\boldsymbol{w}(i,j)\in\mathbb{R}^3$ is the normalized spatial displacement between vertices $i$ and $j$. Node features are $$q_i(t)=[u(\boldsymbol{x}_i,t),v(\boldsymbol{x}_i,t)]$$. In TK-GCN, an analogous undirected attributed graph is used, with node-wise state $\mathbf{x}(t)$ defined on a 3D heart surface mesh and edge attributes given by normalized coordinate differences (Xizhuo et al., 17 Jun 2026, Wang et al., 5 Jul 2025).

The graph convolution used in both formulations follows a continuous spline-kernel construction. For edge pseudo-coordinates $\boldsymbol{w}$, the kernel is parameterized by B-spline bases: $\mathbf{K}$0 with tensor-product basis

$\mathbf{K}$1

and the corresponding continuous graph convolution at node $\mathbf{K}$2 is

$\mathbf{K}$3

In the forecasting notation, the same mechanism is written as

$\mathbf{K}$4

These operators are message-passing layers whose kernels are explicit smooth functions of relative geometry. The resulting spatial inductive bias is not merely adjacency-aware; it is geometry-aware in the stronger sense that the convolution depends continuously on normalized spatial displacements in $\mathbf{K}$5. The cited works further employ residual connections, ELU nonlinearities, and hierarchical graph pooling and unpooling via Graclus clustering and assignment matrices $\mathbf{K}$6, yielding multiscale graph representations analogous to U-Net-style encoder–decoder hierarchies.

3. Koopman latent dynamics

The defining temporal mechanism of K-GCN is the insertion of a learned linear operator into latent space. For a nonlinear dynamical system

$\mathbf{K}$7

the Koopman operator $\mathbf{K}$8 acts on observables $\mathbf{K}$9 as

$K$0

Although $K$1 is generally infinite-dimensional, K-GCN learns a finite-dimensional approximation by combining an encoder with a trainable matrix $K$2 (Xizhuo et al., 17 Jun 2026, Wang et al., 5 Jul 2025).

In P-K-GCN, the encoder produces a latent state

$K$3

which is vectorized as

$K$4

A trainable matrix $K$5 advances the latent state linearly: $K$6 After reshaping,

$K$7

The induced temporal hypothesis class is

$K$8

In TK-GCN, the same idea is expressed through an encoder–decoder pair $K$9: $$\mathcal{G} = (\mathcal{V},\mathcal{E},\mathbf{W},\boldsymbol{q}),$$0 with latent evolution

$$\mathcal{G} = (\mathcal{V},\mathcal{E},\mathbf{W},\boldsymbol{q}),$$1

Decoding after linear latent advancement gives

$$\mathcal{G} = (\mathcal{V},\mathcal{E},\mathbf{W},\boldsymbol{q}),$$2

The primary significance assigned to this construction in the cited works is twofold. First, it linearizes nonlinear temporal evolution in a compact latent space. Second, it replaces arbitrary nonlinear recurrent temporal modules with a single linear operator, which the papers argue stabilizes multi-step prediction and reduces temporal hypothesis complexity. In TK-GCN, however, the pure Koopman rollout is not the final forecasting mechanism; the learned latent space is instead passed to a Transformer, which models longer-range dependencies beyond what a single finite-dimensional linear operator captures.

4. Training objectives and physics augmentation

K-GCN is trained with reconstruction and temporal-consistency losses, and in P-K-GCN these are augmented by physics-based penalties. In the super-resolution setting, the inputs are noisy low-resolution observations

$$\mathcal{G} = (\mathcal{V},\mathcal{E},\mathbf{W},\boldsymbol{q}),$$3

on a sparse set of spatial locations $$\mathcal{G} = (\mathcal{V},\mathcal{E},\mathbf{W},\boldsymbol{q}),$$4, and the objective is to reconstruct high-resolution dynamics

$$\mathcal{G} = (\mathcal{V},\mathcal{E},\mathbf{W},\boldsymbol{q}),$$5

on a dense set $$\mathcal{G} = (\mathcal{V},\mathcal{E},\mathbf{W},\boldsymbol{q}),$$6. Training is supervised in the low-resolution space through a projection $$\mathcal{G} = (\mathcal{V},\mathcal{E},\mathbf{W},\boldsymbol{q}),$$7, while physics residuals are evaluated on high-resolution reconstructions and predictions (Xizhuo et al., 17 Jun 2026).

For a temporal block $$\mathcal{G} = (\mathcal{V},\mathcal{E},\mathbf{W},\boldsymbol{q}),$$8, the data loss is

$$\mathcal{G} = (\mathcal{V},\mathcal{E},\mathbf{W},\boldsymbol{q}),$$9

The first term supervises instantaneous autoencoder reconstructions, and the second supervises Koopman-advanced predictions.

The physics augmentation is built around a reaction–diffusion PDE on a 3D heart surface manifold $\boldsymbol{w}(i,j)\in\mathbb{R}^3$0: $\boldsymbol{w}(i,j)\in\mathbb{R}^3$1 implemented on the manifold using the Laplace–Beltrami operator $\boldsymbol{w}(i,j)\in\mathbb{R}^3$2. With predicted fields $\boldsymbol{w}(i,j)\in\mathbb{R}^3$3, the residuals are

$\boldsymbol{w}(i,j)\in\mathbb{R}^3$4

The total loss is

$\boldsymbol{w}(i,j)\in\mathbb{R}^3$5

with

$\boldsymbol{w}(i,j)\in\mathbb{R}^3$6

A notable technical point is that the physics term does not constrain $\boldsymbol{w}(i,j)\in\mathbb{R}^3$7 or $\boldsymbol{w}(i,j)\in\mathbb{R}^3$8 directly. It acts on decoded high-resolution fields, and thereby implicitly shapes the encoder, decoder, and Koopman operator to produce physically consistent trajectories. In the K-GCN ablation inside P-K-GCN, this augmentation is removed by setting $\boldsymbol{w}(i,j)\in\mathbb{R}^3$9.

In TK-GCN, the Stage-1 training objective has the same general structure but without PDE residuals: $i$0 where

$i$1

$i$2

and

$i$3

This formulation enforces multi-step latent linear consistency while regularizing the magnitude of the Koopman matrix.

5. Theoretical analysis

The most explicit theoretical treatment appears in P-K-GCN, where the authors analyze the effects of both physics augmentation and Koopman regularization on generalization and error accumulation (Xizhuo et al., 17 Jun 2026).

For super-resolution, let $i$4 denote the true high-resolution mapping and $i$5 a model in the encoder–Koopman–decoder hypothesis class $i$6. The physics-constrained hypothesis space is defined by

$i$7

where $i$8 is the expected PDE-residual loss and $i$9 satisfies $j$0. The induced low-resolution classes are

$j$1

$j$2

Under mild assumptions and sufficiently small $j$3, the empirical Rademacher complexity satisfies

$j$4

This is the formal statement that physics constraints strictly shrink the effective hypothesis class.

The paper then combines uniform convergence in the low-resolution observation space with a conditional stability assumption,

$j$5

to derive a high-resolution super-resolution error bound. The resulting statement is that even if a purely data-driven model and P-K-GCN achieve similar empirical low-resolution training errors, the physics-constrained model has a smaller Rademacher complexity term and therefore a tighter high-resolution generalization bound.

The temporal analysis isolates the effect of Koopman regularization. With Frobenius norm control

$j$6

one has

$j$7

For latent error $j$8, the cited proposition gives

$j$9

This is contrasted with a general nonlinear temporal model whose error scales like $$q_i(t)=[u(\boldsymbol{x}_i,t),v(\boldsymbol{x}_i,t)]$$0. The supplement further states that the Rademacher complexity of the Koopman temporal hypothesis space is strictly smaller than that of unconstrained nonlinear recurrent temporal models, and for bounded latent inputs with $$q_i(t)=[u(\boldsymbol{x}_i,t),v(\boldsymbol{x}_i,t)]$$1, the complexity scales as

$$q_i(t)=[u(\boldsymbol{x}_i,t),v(\boldsymbol{x}_i,t)]$$2

over a block of length $$q_i(t)=[u(\boldsymbol{x}_i,t),v(\boldsymbol{x}_i,t)]$$3.

These results formalize two claims that recur throughout the K-GCN literature in the prompt: latent linearization is not only an architectural convenience, but also a capacity-control mechanism; and physics augmentation is not only a heuristic regularizer, but a restriction of the admissible hypothesis class.

6. Applications and empirical behavior

The principal application in P-K-GCN is spatiotemporal super-resolution of cardiac electrodynamics on a 3D ventricular geometry from sparse low-resolution measurements. The state has two channels, the transmembrane potential $$q_i(t)=[u(\boldsymbol{x}_i,t),v(\boldsymbol{x}_i,t)]$$4 and recovery variable $$q_i(t)=[u(\boldsymbol{x}_i,t),v(\boldsymbol{x}_i,t)]$$5, the low-resolution mesh has 195 nodes, the high-resolution mesh has 4370 nodes, and the temporal horizon contains 1000 time steps. The baseline methods are NN, K-GCN, and PINN, and the relative error metric is

$$q_i(t)=[u(\boldsymbol{x}_i,t),v(\boldsymbol{x}_i,t)]$$6

At noise level $$q_i(t)=[u(\boldsymbol{x}_i,t),v(\boldsymbol{x}_i,t)]$$7, the reported $$q_i(t)=[u(\boldsymbol{x}_i,t),v(\boldsymbol{x}_i,t)]$$8 values are as follows (Xizhuo et al., 17 Jun 2026, Wang et al., 5 Jul 2025).

Model	$$q_i(t)=[u(\boldsymbol{x}_i,t),v(\boldsymbol{x}_i,t)]$$9
P-K-GCN	$\mathbf{x}(t)$0
K-GCN	$\mathbf{x}(t)$1
PINN	$\mathbf{x}(t)$2
NN	$\mathbf{x}(t)$3

The reported reductions for P-K-GCN are approximately $\mathbf{x}(t)$4 relative to NN, $\mathbf{x}(t)$5 relative to K-GCN, and $\mathbf{x}(t)$6 relative to PINN. The comparison NN $\mathbf{x}(t)$7 K-GCN is used to isolate the contribution of graph geometry and Koopman latent dynamics, whereas the comparison K-GCN $\mathbf{x}(t)$8 P-K-GCN isolates the contribution of physics augmentation. Qualitatively, at time step 35, NN is described as having poor spatial detail and temporal artifacts, PINN as capturing global wave patterns but producing overly smooth fields, K-GCN as preserving geometry-aware detail yet retaining artifacts and temporal inconsistency, and P-K-GCN as best matching the ground truth across noise levels while preserving fine spatial patterns and temporal coherence.

In TK-GCN, K-GCN is evaluated in a forecasting setting on a 3D ventricular geometry with 1094 nodes and 2184 mesh elements. The task is to use the first 1200 steps to predict the next 300 steps under three protocols: healthy control, localized perturbation, and remote perturbation. The comparisons include TK-GCN, Transformer + GCN, Transformer Only, Transformer + CNN-AE, and LSTM + K-GCN. In smooth dynamics, Transformer Only has the lowest short-term MSE on $\mathbf{x}(t)$9, but TK-GCN is better on $\boldsymbol{w}$0 and $\boldsymbol{w}$1. In the chaotic fragmented regime, TK-GCN is reported to dominate all horizons, with MSE $\boldsymbol{w}$2 on $\boldsymbol{w}$3 and $\boldsymbol{w}$4 on $\boldsymbol{w}$5, compared with $\boldsymbol{w}$6 and $\boldsymbol{w}$7 for Transformer Only, $\boldsymbol{w}$8 and $\boldsymbol{w}$9 for Transformer + GCN, $\mathbf{K}$00 and $\mathbf{K}$01 for Transformer + CNN-AE, and $\mathbf{K}$02 and $\mathbf{K}$03 for LSTM + K-GCN.

A further empirical distinction made in TK-GCN is between K-GCN as a latent-space regularizer and pure Koopman forecasting. The paper states that a pure Koopman baseline deteriorates quickly, with spurious oscillations and fragmented predictions, indicating that a finite-dimensional linear latent operator is useful but insufficient alone for highly complex spatiotemporal dynamics. This suggests a division of labor: K-GCN shapes a geometry-aware and temporally structured latent representation, while a more expressive temporal model handles long-range nonlinear dependencies.

7. Limitations and directions of development

The cited works identify several limitations of current K-GCN formulations (Xizhuo et al., 17 Jun 2026, Wang et al., 5 Jul 2025). One limitation is parameterization cost. In P-K-GCN, the Koopman matrix is explicitly dense, with $\mathbf{K}$04, and the paper notes that scaling such a matrix to substantially higher-dimensional latent spaces may be challenging. The combination of spline-based GCNs on large meshes with surface Laplace–Beltrami residual computation is also described as computationally intensive.

A second limitation concerns model adequacy. Both papers treat $\mathbf{K}$05 as a finite-dimensional linear operator. The forecasting paper explicitly notes that a single matrix $\mathbf{K}$06 may be insufficient to exactly capture highly nonlinear and nonstationary dynamics, which is consistent with the observed deterioration of the pure Koopman baseline. A plausible implication is that the benefit of Koopman enhancement lies as much in latent regularization as in literal linear prediction.

A third limitation is dependence on problem structure. P-K-GCN requires known governing physics, including the PDE form, diffusion coefficients, and reaction terms. The work therefore describes physics augmentation as less straightforward for systems with unknown or partially known physics. TK-GCN likewise assumes a fixed graph or mesh structure; adapting to changing topology or dynamically varying multi-resolution graphs is identified as challenging.

The theoretical guarantees also depend on assumptions that the paper characterizes as standard but potentially imperfect in practice: conditional stability, Lipschitz encoder–decoder behavior, and well-posedness of the physics residual. These assumptions support the formal error bounds, but they do not by themselves eliminate model mismatch or approximation error.

The future directions stated in the prompt are correspondingly targeted. For Koopman learning, the super-resolution paper suggests structured $\mathbf{K}$07 such as block-diagonal or low-rank forms, explicit learning of Koopman eigenfunctions and eigenvalues, and combinations of Koopman mechanisms with attention or transformer-based graph encoders. For uncertainty quantification, it suggests integrating Bayesian or Gaussian process layers in latent space. For broader applicability, it proposes extending the framework to other PDEs and geometries and applying it to real measurement data with complex noise and missingness patterns. The forecasting paper points to time-varying or state-dependent $\mathbf{K}$08, nonlinear parameterizations of the Koopman operator, joint end-to-end training of encoder, Koopman structure, and Transformer, and extensions to multi-variable or multi-graph settings.

Taken together, these directions indicate that K-GCN is best understood not as a fixed architecture, but as a design principle: graph-based spatial representation on irregular domains, paired with a latent evolution model constrained by approximate Koopman linearity, optionally combined with physics-based regularization or more expressive sequence models when the dynamics exceed what a single latent linear operator can represent.