betaFeedback

WorksComputer Science

Semi-Supervised Neural Super-Resolution for Mesh-Based Simulations

Abstract

Mesh-based simulations provide high-fidelity solutions to partial differential equations (PDEs), but achieving such accuracy typically requires fine meshes, leading to substantial computational overhead. Super-resolution techniques aim to mitigate this cost by reconstructing high-resolution (HR), high-fidelity solutions from low-cost, low-resolution (LR) counterparts. However, training neural networks for super-resolution often demands large amounts of expensive HR supervision data. To address this challenge, we propose SuperMeshNet, an HR data-efficient super-resolution framework for mesh-based simulations aided by message passing neural networks (MPNNs). At its core, SuperMeshNet introduces complementary learning, a semi-supervised approach that effectively leverages both 1) a small amount of paired LR-HR data and 2) abundant unpaired LR data via two jointly trained, complementary MPNN-based models. Additionally, our model is enriched by inductive biases, which are empirically shown to further improve super-resolution performance. Extensive experiments demonstrate that SuperMeshNet requires 90% less HR data to achieve even lower root mean square error (RMSE) than that of the fully supervised benchmark without the inductive biases. The source code and datasets are available at https://github.com/jykim-git/SuperMeshNet.git.

Introduction

Mesh-based simulations—such as the finite element method (FEM), or finite volume method (FVM)—are widely used to obtain high-fidelity solutions to partial differential equations (PDEs) across a range of scientific and engineering domains. In mesh-based simulations, the mesh size is carefully chosen to balance computational costs against solution fidelity: finer meshes offer higher fidelity but incur significantly greater computational expenses ([22]). Super-resolution techniques are developed to alleviate this trade-off by predicting high-resolution (HR) simulation results from low-resolution (LR) counterparts, thereby aiming to deliver high-fidelity solutions at a reduced cost ([5, 22]). However, training super-resolution models via conventional fully supervised learning demands substantial quantities of computationally expensive HR data, making data collection a significant bottleneck ([22]). In this context, improving the HR data efficiency of super-resolution model training is of paramount importance in reality.

As summarized in Table 1, several unsupervised learning approaches have tackled this challenge but have their own limitations. For example, PhySRNet ([2]) performs super-resolution without any HR data by incorporating PDEs and constraints into its loss function; however, PhySRNet uses a finite-difference scheme for derivative calculations, which limits its applicability to irregular meshes. MAgNet ([6]) offers an alternative with zero-shot super-resolution through an interpolator trained on LR data; yet, the prediction error of MAgNet is much larger than that of supervised methods (see Appendix § 16.§ 1). To the best of our knowledge, semi-supervised learning has never been applied to the super-resolution task for mesh-based simulations. This may be partly due to the limited exploration of semi-supervised regression methods, especially those compatible with message passing neural networks (MPNNs), compared to semi-supervised classification.

ReferenceLearning methodModel
(Li & McComb, 2022), (Yonekura et al., 2023), [22]Fully supervised (r=1r = 1)CNN
[11], [5]Fully supervised (r=1r = 1)MPNN
[2]Unsupervised (r=0r = 0)CNN
[6]Unsupervised (r=0r = 0)MPNN
SuperMeshNet (ours)Semi-supervised (0<r≪10 < r \ll1) (complementary)MPNN (inductive biases)

Table 1. Comparison between prior studies and our work. Here, r=number of HR data samplesnumber of LR data samplesr = \frac{\text{number of HR data samples}}{\text{number of LR data samples}}.

On one hand, many related studies ([2, 22] (Yonekura et al. 2023; Li & McComb 2022)) on super-resolution for mesh-based simulations rely heavily on convolutional neural networks (CNNs), which cannot directly handle irregular mesh structures. CNNs require interpolating irregular mesh-based data onto a regular grid, which often necessitates a significantly larger number of nodes to achieve the same fidelity as an irregular mesh, leading to relatively lower computational efficiency. It is worth noting that while regular-grid settings in fluid dynamics can

Proceedings of the 43rd International Conference on Machine Learning, Seoul, South Korea. PMLR 306, 2026. Copyright 2026 by the author(s). lead to significant LR–HR discrepancies, our primary focus is on discrepancies induced by complex geometries, which are poorly handled by conventional CNNs. On the other hand, some studies have adopted MPNNs, such as graph convolutional networks (GCNs) [11] or SRGNN [5], which can directly handle irregular mesh data. However, the design and impact of inductive biases on enhancing mesh-based super-resolution performance remain largely underexplored.

To address these limitations, we propose SuperMeshNet, an HR data-efficient super-resolution framework tailored for mesh-based simulations under very scarce HR supervision, which differs from the supervised and unsupervised settings. This is the first general framework that can be applied across diverse MPNN architectures for the super-resolution task. Specifically, SuperMeshNet introduces two key components: complementary learning and inductive biases for MPNNs. First, the complementary learning is a semi-supervised learning method that exploits a small amount of paired LR–HR data for supervised learning, while leveraging a large pool of unpaired LR data in an unsupervised manner. Our complementary learning is built upon two models; an MPNN-based primary model predicts HR solutions from LR counterparts, while an MPNN-based auxiliary model predicts the difference between two HR solutions corresponding to two LR counterparts. The predictions from each model are utilized to calculate pseudo-ground truths, which serve as the ground truth for the other, enabling mutual supervision. Since conventional semi-supervised methods typically employ two identical models [32, 10], they often produce highly similar pseudo-ground truths, making them less informative. On the other hand, owing to distinct but interrelated input–output configurations of our complementary learning, the auxiliary model is capable of capturing intra-resolution relationships while the primary model focuses on inter-resolution relationships. This division of roles fosters synergies in mutual supervision, enhancing super-resolution performance while reducing training time compared to prior semi-supervised strategies.

Second, to further improve the performance of mesh-based super-resolution, we introduce inductive biases for MPNNs, guided by our empirical observation. Specifically, we employ two MPNN-architecture-agnostic inductive biases: node-level centering and message-level centering. The node-level centering centers each node embedding by subtracting the global mean of all node embeddings from each node embedding, while the message-level centering performs a similar centering operation over aggregated messages.

We carry out extensive experiments to validate the effectiveness of our two components in SuperMeshNet. Our results demonstrate that, even with only a small portion (e.g., 10%) of paired LR–HR data, SuperMeshNet surpasses a fully supervised (e.g., 100% paired) benchmark method lacking inductive biases in terms of the root mean square error (RMSE). We also prove that the injected inductive biases consistently reduce the RMSE across six different MPNN architectures, underscoring their general applicability. Finally, our main contributions are summarized as follows:

  • MPNN-agnostic applicability. SuperMeshNet provides a general super-resolution framework for mesh-based simulations, applicable to various MPNNs, under very scarce HR supervision scenarios.

  • Complementary learning. To the best of our knowledge, this is the first attempt to incorporate semi-supervised learning compatible with MPNNs into super-resolution for mesh-based simulations.

  • Inductive biases. We introduce node-level centering and message-level centering, which can substantially enhance super-resolution performance across different MPNN types.

Methodology

Problem Definition

Briefly, we aim to predict an HR solution u^h\hat{u}_h from an LR solution ulu_l of the same PDE, while relying on as few HR solutions uhu_h as possible for training. Formally, let Ω⊂RD\Omega\subset\mathbb{R}^{D} be the computational domain on which the PDE is solved. Here, DD denotes the spatial dimension. A parameter μ\mu represents all variations of PDE instances, such as material coefficients, domain geometry, or boundary conditions. For example, μ\mu could be an angle of an applied force or an aspect ratio of an elliptical hole (see Figure 13–Figure 15 in Appendix § 14). Each choice of μ\mu defines a different PDE instance. As depicted in Figure 1, we discretize Ω\Omega with an LR mesh Ml=(Pl,El)M_l=(P_l,E_l) and an HR mesh Mh=(Ph,Eh)M_h=(P_h,E_h), where Pl∈Rnl×DP_l\in\mathbb{R}^{n_l\times D} and Ph∈Rnh×DP_h\in\mathbb{R}^{n_h\times D} are the positions of the nodes on MlM_l and MhM_h, respectively, and ElE_l and EhE_h are edges on MlM_l and MhM_h, respectively. Running a PDE solver on these meshes yields LR and HR solutions, which we regard as an LR data sample ulu_l and an HR data sample uhu_h, defined on the nodes on MlM_l and MhM_h, respectively. Our objective is to predict u^h∈Rnh×d\hat{u}_h\in\mathbb{R}^{n_h\times d} from ulu_l as closely as possible to uhu_h, while minimizing the amount of HR data uhu_h required for training, where dd denotes the dimension of the solution field.

Figure 1. Problem setting. We aim to predict u^h\hat{u}_h on HR mesh MhM_h, containing nodes at positions PhP_h and edges EhE_h, from LR data sample ulu_l defined on LR mesh MlM_l, comprising nodes at positions PlP_l and edges ElE_l.

Examples of LR and HR data samples with various angles of applied force relative to the x-axis from Dataset 1

Figure 13. Examples of LR and HR data samples with various angles of applied force relative to the x-axis from Dataset 1.

Examples of LR and HR data samples with various ratios between the lengths of the major and minor axes from Dataset 3

Figure 15. Examples of LR and HR data samples with various ratios between the lengths of the major and minor axes from Dataset 3.

Complementary Learning

Dataset Setting

As depicted in Figure 2, complementary learning in SuperMeshNet leverages both a paired LR–HR training dataset Da={(ulq,uhq)∣q=1,2,…,Nh}\mathcal{D}_a=\{(u_l^q,u_h^q)\mid q=1,2,\ldots,N_h\} containing NhN_h LR–HR data pairs (ulq,uhq)(u_l^q,u_h^q) and an unpaired LR training dataset Db={ulq∣q=Nh+1,Nh+2,…,N}\mathcal{D}_b=\{u_l^q\mid q=N_h+1,N_h+2,\ldots,N\} having N−NhN-N_h LR data samples. The total number of LR data samples is NN, among which only NhN_h have HR counterparts, with Nh≪NN_h\ll N in practice. In other words, N−NhN-N_h fewer HR data samples are required compared to fully supervised learning. Here, the superscript qq is simply an index to distinguish different samples corresponding to different parameters μ\mu. For instance, if μ\mu is the angle of an applied force, then (ul1,uh1)(u_l^1,u_h^1) corresponds to one angle μ1\mu^1, and (ul2,uh2)(u_l^2,u_h^2) corresponds to another μ2\mu^2.

Dataset setting showing paired LR–HR and unpaired LR training data

Figure 2. Dataset setting. Complementary learning utilizes a paired LR–HR training dataset, including NhN_h paired data samples (green hexagons), and an unpaired LR training dataset, containing N−NhN-N_h unpaired LR data samples (white hexagons). In total, complementary learning can reduce N−NhN-N_h HR data samples compared to the case of fully supervised learning.

The Two Models: FθF_\theta and GϕG_\phi

To fully exploit unpaired LR data, our complementary learning framework leverages mutual supervision between two models, a primary model FθF_\theta and an auxiliary model GϕG_\phi, which are jointly trained with distinct roles. Unlike conventional semi-supervised methods, where multiple models generate pseudo-labels for the same prediction target—often resulting in correlated errors and confirmation bias—our framework decomposes learning into two structurally distinct tasks: inter-resolution mapping (LR →\to HR) and intra-resolution difference modeling (LR–LR →\to HR–HR). Because the two models predict different targets, their errors are inherently less correlated than in conventional co-training (see Appendices § 9 and 1.7.2). Consequently, the auxiliary model GϕG_\phi provides complementary supervisory signals that improve the overall effectiveness of the framework. From a physical perspective, the two HR solutions are governed by the same PDE and differ only in a parameter μ\mu. Their difference therefore captures the system’s physical response to parameter variation. Accordingly, GϕG_\phi learns the solution’s sensitivity to parameter perturbations and guides FθF_\theta to produce physically consistent predictions under varying conditions.

The primary model FθF_\theta, which is used for inference only, predicts an HR solution u^hq\hat{u}_h^q from its LR counterpart ulqu_l^q:

Fθ(ulq)=u^hq(ground truth : uhq).(1)F_\theta(u_l^q)=\hat{u}_h^q \quad\text{(ground truth : }u_h^q\text{)}. \tag*{(1)}

On the other hand, the auxiliary model GϕG_\phi, which is used only during training, predicts the difference between two HR solutions u^hrs\hat{u}_h^{rs} corresponding to two LR input samples ulr,ulsu_l^r,u_l^s to further utilize intra-resolution relations. Here, rr and ss indicate two LR samples corresponding to different values of parameter μ\mu. Since computational geometry may vary across samples with different μ\mu, HR samples uhru_h^r and uhsu_h^s may be defined on different positions PhrP_h^r and PhsP_h^s. Thus, direct subtraction is not possible. To resolve this, we apply kk-nearest neighbor (kNN) interpolation (Qi et al., 2017) to project solutions defined on PhsP_h^s onto PhrP_h^r:

Gϕ(ulr,uls)=u^hrs(ground truth : uhr−kNN(uhs;Phs→Phr)).(2)\begin{aligned} G_\phi(u_l^r,u_l^s)&=\hat{u}_h^{rs}\\ \text{(ground truth : }u_h^r-kNN(u_h^s;P_h^s\to P_h^r)\text{)}. \tag*{(2)} \end{aligned}

In kkNN interpolation, each target point is assigned a distance-weighted average of its kk nearest source points (see Appendix § 11 for the details).

Learning Procedure

As depicted in Figure 3, each training step combines supervised and unsupervised learning of the two models FθF_\theta and GϕG_\phi. In other words, the loss functions for training FθF_\theta and GϕG_\phi, denoted by LF\mathcal{L}_F and LG\mathcal{L}_G, respectively, are expressed as:

Figure 3. A schematic overview of complementary learning in SuperMeshNet. It first samples paired LR–HR data (ulα,uhα)(u_l^\alpha, u_h^\alpha), (ulβ,uhβ)(u_l^\beta, u_h^\beta) and unpaired LR data ulγu_l^\gamma. Complementary learning leverages both supervised and unsupervised learning to jointly train two neural network models, FθF_\theta and GϕG_\phi. FθF_\theta predicts an HR solution from its LR counterpart, while GϕG_\phi predicts the difference between two HR solutions from two LR counterparts to enable synergistic mutual supervision. More specifically, for supervised learning, FθF_\theta and GϕG_\phi are trained with pairs of LR–HR data (green hexagons). In unsupervised learning, the prediction of one model (yellow and purple hexagons predicted by FθF_\theta and GϕG_\phi, respectively) is used to calculate a pseudo-ground truth that serves as the target for training another model (as depicted by solid and dotted red arrows). The model parameters of FθF_\theta and GϕG_\phi are updated using the total loss, including both supervised and unsupervised losses based on predictions from both FθF_\theta and GϕG_\phi.

LF=LF,sup+LF,unsup(3)\mathcal{L}_F = \mathcal{L}_{F,\mathrm{sup}} + \mathcal{L}_{F,\mathrm{unsup}} \tag*{(3)}
LG=LG,sup+LG,unsup,(4)\mathcal{L}_G = \mathcal{L}_{G,\mathrm{sup}} + \mathcal{L}_{G,\mathrm{unsup}}, \tag*{(4)}

where the subscripts supsup and unsupunsup represent supervised and unsupervised learning, respectively. To this end, three samples are randomly sampled: 1) two paired LR samples ulαu_l^\alpha and ulβu_l^\beta from the paired LR–HR dataset DaD_a for supervised learning and 2) one additional unpaired LR sample ulγu_l^\gamma from the unpaired LR dataset DbD_b for unsupervised learning. Here, α\alpha, β\beta, and γ\gamma are the indices referring to distinct parameters μ\mu.

In supervised learning, an HR data sample uhαu_h^\alpha is available. As depicted in Figure 3, FθF_\theta is trained to reduce the MSE between its prediction u^hα\hat{u}_h^\alpha and its target, which is the ground truth uhαu_h^\alpha. An analogous procedure is applied to β\beta, thus resulting in:

LF,sup=ℓ(u^hα,uhα)+ℓ(u^hβ,uhβ),(5)\mathcal{L}_{F,\mathrm{sup}} = \ell(\hat{u}_h^\alpha, u_h^\alpha) + \ell(\hat{u}_h^\beta, u_h^\beta), \tag*{(5)}

where ℓ(⋅,⋅)\ell(\cdot,\cdot) denotes the MSE. Similarly, as expressed in Eq. (2), GϕG_\phi is trained to reduce the MSE between its predictions u^hαβ\hat{u}_h^{\alpha\beta} and its target, which is the ground truth difference between uhαu_h^\alpha and uhβu_h^\beta using the following loss:

LG,sup=ℓ(u^hαβ,uhα−kNN(uhβ;Phβ→Phα)).(6)\mathcal{L}_{G,\mathrm{sup}} = \ell(\hat{u}_h^{\alpha\beta}, u_h^\alpha- kNN(u_h^\beta; P_h^\beta\to P_h^\alpha)). \tag*{(6)}

In unsupervised learning, the ground truth HR data sample uhγu_h^\gamma is unavailable. Under this circumstance, we leverage mutual supervision between FθF_\theta and GϕG_\phi. For example, as depicted in Figure 3, if Gϕ(ulα,ulγ)G_\phi(u_l^\alpha,u_l^\gamma) predicts u^hγα\hat{u}_h^{\gamma\alpha} that approximates the difference between two HR samples uhγ−uhαu_h^\gamma-u_h^\alpha, then adding this to the known uhαu_h^\alpha yields an estimate of uhγu_h^\gamma. This pseudo-ground truth can serve as a target for Fθ(ulγ)F_\theta(u_l^\gamma). Similarly, if Fθ(ulγ)F_\theta(u_l^\gamma) produces u^hγ\hat{u}_h^\gamma close to uhγu_h^\gamma, then subtracting known uhαu_h^\alpha from u^hγ\hat{u}_h^\gamma provides an approximation of uhγ−uhαu_h^\gamma-u_h^\alpha, which can serve as a target for Gϕ(ulγ,ulα)G_\phi(u_l^\gamma,u_l^\alpha), accordingly. An analogous procedure is applied to the pair (β,γ)(\beta,\gamma). It should be noted that a thorough treatment of kNNkNN interpolation is also required to effectively handle mesh mismatches. For example, in Eq. (7), u^hβ−u^hγβ\hat{u}_h^\beta-\hat{u}_h^{\gamma\beta} can be used to approximate uhγu_h^\gamma. However, the values uhβu_h^\beta are defined on position PhβP_h^\beta, whereas the prediction u^hγ\hat{u}_h^\gamma is defined on position PhγP_h^\gamma. To reconcile this discrepancy, kNNkNN interpolation is employed to project the values from PhγP_h^\gamma onto PhβP_h^\beta. The resultant loss functions are:

LF,unsup=ℓ(u^hγ,u^hγα+kNN(uhα;Phα→Phγ))+ℓ(u^hγ,kNN(uhβ−u^hβγ;Phβ→Phγ))(7)\begin{aligned} \mathcal{L}_{F,\mathrm{unsup}} ={}& \ell(\hat{u}_h^\gamma, \hat{u}_h^{\gamma\alpha} + kNN(u_h^\alpha; P_h^\alpha\to P_h^\gamma)) \\ &+ \ell(\hat{u}_h^\gamma, kNN(u_h^\beta- \hat{u}_h^{\beta\gamma}; P_h^\beta\to P_h^\gamma)) \tag*{(7)} \end{aligned}
LG,unsup=ℓ(u^hγα,u^hγ−kNN(uhα;Phα→Phγ))+ℓ(u^hβγ,uhβ−kNN(u^hγ;Phγ→Phβ)).(8)\begin{aligned} \mathcal{L}_{G,\mathrm{unsup}} ={}& \ell(\hat{u}_h^{\gamma\alpha}, \hat{u}_h^\gamma- kNN(u_h^\alpha; P_h^\alpha\to P_h^\gamma)) \\ &+ \ell(\hat{u}_h^{\beta\gamma}, u_h^\beta- kNN(\hat{u}_h^\gamma; P_h^\gamma\to P_h^\beta)). \tag*{(8)} \end{aligned}

Model Architecture

The architecture of the primary model FθF_\theta, illustrated in Figure 4, is built upon SRGNN ([5]). The role of FθF_\theta is to transform LR data into HR data, which is conducted by the lowermost upsampler in Figure 4. To surpass the performance of kkNN interpolation by the lowermost upsampler, we introduce additional upsampling in latent space. Specifically, an encoder maps the physical quantities into high-dimensional latent space. The LR processor applies message passing to refine LR representations, which are then upsampled to HR latent embeddings. Subsequently, the HR processor applies additional message passing to further enhance the HR representations. Finally, a decoder maps the latent embeddings back to the physical space. The final HR output is obtained by adding the two upsampled HR fields: one from the kkNN-based upsampler and the other from the latent-space upsampling pathway.

Figure 4. Model architecture of FθF_\theta.

The auxiliary model GϕG_\phi, visualized in Figure 5, is responsible for predicting the difference between two HR samples corresponding to two LR input samples. It extends FθF_\theta to accommodate two input samples, maintaining a comparable structure. A notable architectural feature is the use of a shared feature extractor between FθF_\theta and GϕG_\phi, which helps reduce computational costs during training. First, latent embeddings are extracted from the two LR inputs using the shared feature extractor. Then, one embedding is subtracted from another and the result is decoded to predict the HR difference. The final HR output is obtained by adding the two upsampled HR fields: one from the kkNN-based upsampler and the other from the latent-space upsampling pathway. The interpolators in Figure 5 serve only to address mesh mismatches when the two LR samples are defined on different meshes.

Figure 5. Model architecture of GϕG_\phi.

Inductive Biases for MPNNs

The key difference between the two models, FθF_\theta and GϕG_\phi, and SRGNN ([5]) is that the MPNNs in the LR and HR processors are enriched with inductive biases: node-level centering and message-level centering. We illustrate these inductive biases using an MPNN based on MeshGraphNet (MGN) ([23]), which serves as the default MPNN architecture throughout our experiments. Implementation details for various types of MPNNs (including MGN) are provided in Appendix § 13.

In an MPNN layer, a target node ii receives messages msgijmsg_{ij} from its neighboring nodes j∈N(i)j \in\mathcal{N}(i). In MGN, the message is defined as the edge embedding eije_{ij}, where eije_{ij} is computed from the node embeddings xix_i and xjx_j of nodes ii and jj using a learnable multilayer perceptron (MLP) MLPeMLP_e:

eij←MLPe(xi,xj,eij),e_{ij} \leftarrow MLP_e(x_i,x_j,e_{ij}),
msgij=eij,(9)msg_{ij} = e_{ij}, \tag*{(9)}

where the exact definition of msgijmsg_{ij} depends on the MPNN types.

Each node ii then aggregates messages msgijmsg_{ij} from its neighbors jj, and the aggregated message aggiagg_i is used to update its node embedding by employing a learnable MLP MLPxMLP_x:

aggi=∑j∈N(i)msgij,agg_i = \sum_{j\in\mathcal{N}(i)} msg_{ij},
xi←MLPx(xi,aggi).(10)x_i \leftarrow MLP_x(x_i,agg_i). \tag*{(10)}

We empirically find that injecting the following two inductive biases into the message passing mechanism substantially improves super-resolution performance. The first inductive bias, node-level centering, subtracts the mean of all node embeddings xix_i from each individual xix_i after the node embedding update in Eq. (10):

xi←xi−1n∑i=1nxi.(11)x_i \leftarrow x_i - \frac{1}{n}\sum_{i=1}^{n}x_i. \tag*{(11)}

where nn denotes the number of nodes in the LR mesh MlM_l or the HR mesh MhM_h. For MPNNs that explicitly compute aggregated messages aggiagg_i, such as MGN, message-level centering subtracts the mean of aggiagg_i from each individual aggregated message between the message aggregation and the node embedding update in Eq. (10):

aggi←aggi−1n∑i=1naggi.(12)agg_i \leftarrow agg_i - \frac{1}{n}\sum_{i=1}^{n}agg_i. \tag*{(12)}

The centering operations tend to smooth the loss landscape, similar to the effect reported for batch normalization ([28]), which may facilitate optimization. However, because centering removes global mean information, it is beneficial only for tasks that do not rely heavily on such information (e.g., super-resolution). Experimental validations of these arguments are provided in Appendix § 16.3.

Experimental Results and Analyses2

Datasets3

FEM Datasets. Table 2 summarizes three FEM datasets used in our experiments, detailing their governing PDEs, the quantities derived from solving these equations, the parameters that vary across the samples, and the number of nodes in LR and HR meshes. Dirichlet boundary conditions are utilized for all datasets. Datasets 2 and 3 include elliptical geometries with varying aspect ratios, leading to various magnitude differences between LR and HR fields despite similar spatial patterns, making these datasets particularly challenging for super-resolution.

DatasetEquationSolutionParameterLR nodesHR nodes
1Linear elasticityvon Mises stressForce angle3334,053
2Linear elasticityvon Mises stressHole shape329–3873,959–4,157
3Poisson equationElectric fieldHole shape324–3883,959–4,154

Table 2. Summary of FEM datasets.

CFD Datasets. To validate the applicability of SuperMeshNet to a complex real-world geometry and a time-dependent PDE, we adopt three computational fluid dynamics (CFD) datasets. As summarized in Table 3, the real-world geometry dataset is constructed by solving the incompressible Navier-Stokes equation around a motorbike with a rider. Since our primary focus is on handling LR–HR field discrepancies arising from complex geometries, this real-world benchmark dataset is particularly challenging. The time-dependent PDE dataset 1 is obtained by solving the incompressible Navier–Stokes equations for flow around a cylinder. The time-dependent PDE dataset 2 is generated following Kochkov et al. [18] to construct a dataset in which the HR vorticity differs substantially from the LR vorticity. Meanwhile, for the time-dependent PDE dataset 1, LR data samples are generated by downsampling HR data onto LR meshes because we found that time synchronization between independently simulated LR and HR trajectories is not guaranteed. For all other datasets, LR data samples are obtained by independently solving the governing PDEs on LR meshes.

DatasetEquationSolutionParameterLR nodesHR nodes
Real-world geometry datasetIncompressible Navier-StokesVelocity, pressureAngle of attack15,61986,087
Time-dependent PDE dataset 1Incompressible Navier-StokesVelocity, pressureTime5,20020,400
Time-dependent PDE dataset 2Incompressible Navier-StokesVorticityTime1,0241,048,576

Table 3. Summary of CFD datasets.

Experimental Setup

We evaluate our methodology using six representative MPNNs, including GCN (Kipf & Welling, [17]), Graph-SAGE (SAGE) (Hamilton et al., [13]), GAT (Veličković et al., [33]), Graph Transformer (GTR) (Shi et al., [31]), GIN (Xu et al., 2019), and MGN (Pfaff et al., [23]). Each MPNN consists of three layers for LR processing and additional three layers for HR processing. The default hidden dimension is set to 30; however, for larger datasets—specifically, the time-dependent PDE 2 dataset and the real-world geometry dataset—it is increased to 150 and 120, respectively. For datasets that predict both velocity and pressure—namely, the time-dependent PDE dataset 1 and the real-world geometry dataset—we replace the standard MSE loss with a weighted MSE loss to account for the scale difference between velocity and pressure. The weights assigned to velocity and pressure are 99 : 1 and 10−810^{-8} : 1 for the time-dependent PDE dataset 1 and the real-world geometry dataset, respectively. Throughout the experiments, we adopt the Adam optimizer with a learning rate of 1×10−31 \times10^{-3} and PyTorch’s automatic mixed precision training to improve computational efficiency. All experiments are carried out on a machine with Intel (R) Core (TM) i9-10920X [email protected] GHz and an NVIDIA RTX A6000 GPU. The RMSE is used as a metric where lower values indicate better performance.

Comparison with Full Supervision

Table 4 compares the RMSE of each MPNN integrated with our framework SuperMeshNet, and its variant without inductive biases, SuperMeshNet-O, against two fully supervised baselines—the same type of MPNNs but trained with full supervision without inductive biases. SuperMeshNet-O trained with 20 HR data samples (Nh=20N_h = 20) and 200 LR data samples (N=200N = 200) achieves a significantly lower RMSE compared to the MPNNs trained exclusively on 20 paired LR–HR samples (Nh=N=20N_h = N = 20). The improvement is attributed to complementary learning, which inherently designed to effectively leverage the 180 unpaired LR samples that fully supervised learning cannot utilize. Remarkably, despite being trained only with 20 HR data samples, SuperMeshNet-O achieves RMSE values that are on par with the second fully supervised baseline trained with the entire 200 HR data samples (i.e., Nh=N=200N_h = N = 200). SuperMeshNet, enriched by inductive biases, surpasses the second fully supervised baseline (Nh=N=200N_h = N = 200) in most cases, highlighting the efficacy of the proposed inductive biases tailored for super-resolution in improving performance. This implies the potential to reduce up to 90% of the effort required to generate HR data. It is worth noting that the main advantage of our approach lies in reducing HR data requirements by 90% while maintaining the accuracy of full supervision, rather than in improving absolute RMSE itself. Furthermore, our findings consistently demonstrate the improvement of SuperMeshNet in terms of HR data efficiency across all six MPNNs. This underscores its versatility and effectiveness in enhancing super-resolution performance, regardless of the type of underlying MPNN architecture.

MethodNhN_h, NNGCNSAGEGATGTRGINMGN
Dataset 1
Fully supervised20, 200.08740.08760.08260.07580.08190.0655
Fully supervised200, 2000.05750.05440.05120.04500.03810.0228
SuperMeshNet-O20, 2000.06130.05890.05440.04510.04040.0269
SuperMeshNet20, 2000.04310.04500.04570.03850.02770.0226
Dataset 2
Fully supervised20, 200.09720.10250.09830.09830.07750.0730
Fully supervised200, 2000.06240.06330.06370.05720.05340.0461
SuperMeshNet-O20, 2000.06360.06640.06800.06310.05690.0514
SuperMeshNet20, 2000.05740.06240.06340.06000.05370.0507
Dataset 3
Fully supervised20, 200.05870.06110.06160.05130.05690.0523
Fully supervised200, 2000.03700.03400.03740.03290.03170.0243
SuperMeshNet-O20, 2000.03800.03660.03750.03630.03160.0281
SuperMeshNet20, 2000.02970.02970.03100.02940.02580.0245

Table 4. The RMSE of SuperMeshNet (with inductive biases) and SuperMeshNet-O (without inductive biases) trained with Nh=20N_h = 20 HR data samples and N=200N = 200 LR data samples across six MPNNs and three datasets, in comparison with two fully supervised MPNNs including 1) Nh=N=20N_h = N = 20 and 2) Nh=N=200N_h = N = 200. The best performer is highlighted in bold. The main advantage of our approach lies in reducing HR data requirements by 90% (200 → 20) while maintaining the accuracy of full supervision, rather than in improving absolute RMSE itself.

Overall Training Cost

SuperMeshNet alleviates the reliance on expensive HR data samples, reducing training data generation time. However, it incurs longer training time compared to the fully supervised baseline. Figure 6 displays how the training time increase and data generation time decrease—resulting from the use of SuperMeshNet (Nh=20N_h = 20, N=200N = 200)—scale with mesh size, relative to fully supervised learning (Nh=N=200N_h = N = 200) on Dataset 1 (the full set of results on all FEM datasets as well as device settings are available in Appendix 1.5.1). Comparison of the two slopes, which characterize how training time increases and data generation time decreases with mesh size, respectively, reveals that data generation time grows more rapidly as mesh size decreases. This trend is particularly important in our target regime, where fine meshes lead to significant computational cost. We expect that, for sufficiently fine meshes, the savings in data generation time will outweigh the additional training time, resulting in an overall reduction in computational costs.

Training time increase and data generation time decrease

Figure 6. Training time increase (left) and data generation time decrease (right), resulting from the use of SuperMeshNet (Nh=20N_h = 20, N=200N = 200), relative to fully supervised learning (Nh=N=200N_h = N = 200) on Dataset 1 and its mesh-size variants. All experiments use MGN as the underlying MPNN architecture.

Comparison with Benchmark Semi-Supervised Regression Methods

Table 5 compares complementary learning in SuperMeshNet against benchmark semi-supervised regression methods. As presented in Table 5, SuperMeshNet achieves the lowest RMSE while also exhibiting the shortest training time among all benchmark semi-supervised regression methods. The performance improvements achieved by SuperMeshNet likely stem from its inherent characteristics of using two complementary models. Mean-Teacher (Tarvainen & Valpola, [32]) and UCVME (Dai et al., [10]) employ two models to predict the same target, i.e., an HR data sample. Similarly, TNNR (Wetzel et al., [34]) uses one twin neural network to predict the difference between two HR data samples. On the other hand, SuperMeshNet employs the complementary learning mechanism that leverages two distinct yet cooperative models: primary model FθF_{\theta}, which learns to predict an HR data sample, and auxiliary GϕG_{\phi}, which learns to predict the difference between two HR data samples, as formulated in Eq. (2). While FθF_{\theta} operates on a single LR input, GϕG_{\phi} utilizes two LR data samples along with one HR data sample, enabling the two models to make predictions from distinct informational viewpoints. This design reduces the correlation of errors between the two models’ predictions, thereby enabling complementary mutual supervision.

MethodsRMSETraining time (s)
Mean-Teacher (Tarvainen & Valpola, [32])0.0325693.84
TNNR (Wetzel et al., [34])0.0624477.48
UCVME (Dai et al., [10])0.02931122.62
SuperMeshNet-O0.0269503.2
SuperMeshNet0.0226421

Table 5. Comparison with benchmark semi-supervised regression methods in terms of the RMSE and training time (in seconds). Here, MGN is employed as an MPNN for each method. Training is conducted when Nh=20N_h = 20 and N=200N = 200 for Dataset 1. The best value in each metric is highlighted in bold.

Ablation Studies on Inductive Biases

Table 6 presents ablation results on the two inductive biases in SuperMeshNet, demonstrating their effect on super-resolution performance in terms of the RMSE across six different MPNNs4. For all MPNNs, the incorporation of node-level centering (N) and message-level centering (M) into the MPNN architecture leads to substantial improvements in super-resolution performance compared to MPNNs without inductive biases (O). We provide further analysis on the underlying factors contributing to the performance improvement in Appendix § 16.3.

MPNNRMSE
ONMN + M
GCN0.06130.0431--
SAGE0.05890.04930.05280.0450
GAT0.05440.0457--
GTR0.04510.04050.04380.0385
GIN0.04040.02900.02810.0277
MGN0.02690.02370.02470.0226

Table 6. Ablation studies on inductive biases. The RMSE of SuperMeshNet across six MPNNs under four inductive bias conditions (O: without inductive biases, N: node-level centering, M: message-level centering, and N+M: both node-level and message-level centering operations) trained with Nh=20N_h = 20 and N=200N = 200 for Dataset 1 is compared. For each MPNN, the lowest RMSE value among the four inductive bias conditions is highlighted in bold.

Application to Real-World Geometry and Time-Dependent PDE

Figure 7 demonstrates the applicability of SuperMeshNet to real-world geometry when MGN is used as the MPNN backbone. It visualizes the error distribution on the real-world geometry dataset, where darker blue indicates higher error. Notably, SuperMeshNet, trained with only 40 HR samples, achieves even lower errors than the case of a fully supervised model trained with 200 HR samples, particularly around the back of the rider. These results indicate that SuperMeshNet remains effective in handling complex geometric structures.

Comparison of the squared error of pressure between SuperMeshNet and fully supervised baselines on a real-world geometry dataset

Figure 7. Comparison of the squared error of pressure between SuperMeshNet and fully supervised baselines on a real-world geometry dataset. Here, NhN_h and NN denote the numbers of HR and LR data samples, respectively. For all cases, MGN is used as the underlying MPNN.

To further validate the superiority of SuperMeshNet from both physical and practical perspectives, we evaluate the drag and lift coefficients, where the drag coefficient quantifies the resistance force acting on the motorbike opposing the flow direction, and the lift coefficient measures the force perpendicular to it. Table 7 shows that SuperMeshNet achieves a lower relative error than the case of the fully supervised baseline in predicting the drag coefficient, despite using significantly fewer HR samples. In contrast, SuperMeshNet exhibits a slightly larger relative error for the lift coefficient, which can be attributed to its small magnitude—small absolute deviations can lead to large relative errors. Since the absolute error is less than 0.01 for both coefficients, these results indicate that SuperMeshNet remains effective for deriving meaningful physical quantities.

MethodsDrag coefficient (relative error)Lift coefficient (relative error)
Ground truth HR data0.37240.0368
SuperMeshNet (Nh=40,N=200N_h = 40, N = 200)0.3778 (0.014)0.0433 (0.177)
Fully supervised (Nh=N=200N_h = N = 200)0.3653 (0.019)0.0380 (0.033)

Table 7. Comparison with the fully supervised baseline in predicting physical quantities (drag and lift coefficients) on the real-world geometry dataset. Here, MGN is used as the MPNN backbone for all methods.

Next, we evaluate the applicability of SuperMeshNet to time-dependent PDEs. Figure 8 presents qualitative comparisons on the time-dependent PDE dataset 1. The predictions of SuperMeshNet, trained with only 40 HR samples, closely match the ground truth and are comparable to those of the fully supervised model trained with 200 HR samples. Furthermore, we evaluate the amplitudes of the drag and lift coefficients of the cylinder, which characterize the wake flow behind it. Table 8 shows that SuperMeshNet predicts these amplitudes with an absolute error below 0.01 and a relative error under 3%, while outperforming the fully supervised baseline despite using fewer HR samples. In addition, Figure 9 presents qualitative comparisons on the time-dependent PDE dataset 2. SuperMeshNet successfully predicts the HR counterpart even when the ground truth differs substantially from the LR input, whereas conventional fully supervised learning fails despite having access to significantly more HR data samples. This advantage stems from our auxiliary model GϕG_{\phi}, which leverages HR–HR relationships during training, while fully supervised baselines

MethodsAmplitude of drag coefficient (relative error)Amplitude of lift coefficient (relative error)
Ground truth HR data0.04510.6984
SuperMeshNet (Nh=40N_h = 40, N=200N = 200)0.0441 (0.0232)0.6907 (0.0111)
Fully supervised (Nh=N=200N_h = N = 200)0.0435 (0.0360)0.6568 (0.0595)

Table 8. Comparison with the fully supervised baseline in predicting physical quantities (amplitudes of drag and lift coefficients) on the time-dependent PDE dataset 1. Here, MGN is used as the MPNN backbone for all methods.

Figure 8. Comparison of the ground truth and predicted fluid flow speeds on the time-dependent PDE dataset 1. Here, NhN_h and NN represent the number of HR and LR data samples, respectively. For all cases, MGN is utilized as the underlying MPNN.

Figure 9. Comparison of the LR input, ground truth HR data, and two vorticity predictions produced by SuperMeshNet and full supervision on the time-dependent PDE dataset 2. Here, NhN_h and NN represent the number of HR and LR data samples, respectively. For all cases, MGN is utilized as the underlying MPNN.

rely solely on LR–HR mappings that are more challenging to learn due to the large discrepancies between LR and HR fields.

Conclusions and Limitations

In this paper, we address the open problem of super-resolution for mesh-based simulations by proposing SuperMeshNet, which leverages complementary learning and inductive biases to achieve high HR data efficiency. Complementary learning enables the effective use of unpaired LR data, while the inductive biases further improve performance across a range of MPNN architectures. While SuperMeshNet substantially reduces the reliance on HR data, several challenges remain for making complementary learning more efficient, effective, and better understood. Complementary learning is slower than fully supervised training. We expect that, for sufficiently fine meshes, the reduction in HR data generation cost will outweigh this additional training overhead; however, further improving the computational efficiency of complementary learning remains an important direction for future work. In addition, while Appendix § 16.13 provides a preliminary empirical investigation of training stability, a rigorous theoretical characterization is needed to understand when complementary learning remains stable without mutual error amplification. Moreover, Appendix § 16.12 shows that HR data selection strategies can significantly influence performance, suggesting the need for a more principled understanding of how HR samples should be chosen. Finally, because the auxiliary model learns solution variations under parameter perturbations, the framework may be less effective in regimes with strong nonlinearities or bifurcations. Understanding these failure modes is an important step toward extending SuperMeshNet to more complex physical systems.

Acknowledgments

The work of W.-Y. Shin was supported by the National Research Foundation of Korea (NRF), South Korea grant funded by the Korea government (MSIT) (RS-2021-NR059723), and by SMEs Technology Innovation Development Program through the Technology Innovation and Promotion Agency (TIPA), funded by Ministry of SMEs and Startups (RS-2024-00511332). The work of Y. Hong was supported by Basic Science Research Program through the NRF funded by the Korea government (MSIT) (RS-2023-00219980), and by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) [NO.RS-2021-II211343, Artificial Intelligence Graduate School Program (Seoul National University)].

Impact Statement

This paper introduces SuperMeshNet, a super-resolution framework designed to advance the field of machine learning-aided physical simulations. By providing a cost-effective alternative to traditional simulation methods, SuperMeshNet substantially reduces computational and resource expenses, making high-fidelity simulations more accessible. This efficiency can facilitate the broader adoption of simulations across various engineering disciplines including, but not limited to, solid mechanics, enabling more extensive exploration and optimization while minimizing the need for costly trial-and-error experiments. Ultimately, SuperMeshNet has the potential to accelerate innovation by lowering the barriers to conducting complex simulations.

Notations

Table 9 summarizes the notations used throughout the paper.

NotationDescription
μ\muPDE parameter
Fθ,GϕF_{\theta}, G_{\phi}neural network models
ul,ulq,ulr,uls,ulα,ulβ,ulγu_l, u_l^{q}, u_l^{r}, u_l^{s}, u_l^{\alpha}, u_l^{\beta}, u_l^{\gamma}LR data samples
uh,uhq,uhr,uhs,uhα,uhβ,uhγu_h, u_h^{q}, u_h^{r}, u_h^{s}, u_h^{\alpha}, u_h^{\beta}, u_h^{\gamma}HR data samples
u^h,u^hα,u^hβ,u^hγ\hat{u}_h, \hat{u}_h^{\alpha}, \hat{u}_h^{\beta}, \hat{u}_h^{\gamma}prediction by FθF_{\theta}
u^hαβ,u^hβγ,u^hγα\hat{u}_h^{\alpha\beta}, \hat{u}_h^{\beta\gamma}, \hat{u}_h^{\gamma\alpha}prediction by GϕG_{\phi}
MlM_lLR mesh
MhM_hHR mesh
PlP_lnodal positions of the LR mesh
Ph,Phα,Phβ,PhγP_h, P_h^{\alpha}, P_h^{\beta}, P_h^{\gamma}nodal positions of the HR mesh
ElE_ledges of the LR mesh
EhE_hedges of the HR mesh
nln_lnumber of nodes in the LR mesh
nhn_hnumber of nodes in the HR mesh
nnnumber of nodes in the underlying graph
Da\mathcal{D}_apaired LR–HR training dataset
Db\mathcal{D}_bunpaired LR training dataset
NNtotal number of data samples
NlN_l= number of LR data samples
NpN_pnumber of paired LR–HR data samples
NhN_h= number of HR data samples
LF\mathcal{L}_Floss function for training FθF_{\theta}
LG\mathcal{L}_Gloss function for training GϕG_{\phi}
ℓ\ellmean squared error
kNNk\mathrm{NN}kk-nearest neighbor interpolation
xix_inode embedding of node ii
xjx_jnode embedding of node jj
eije_{ij}edge embedding between nodes ii and jj
msg⁡ij\operatorname{msg}_{ij}message between nodes ii and jj
agg⁡i\operatorname{agg}_iaggregated message of node ii
fmf_mmessage function
fxf_xnode embedding update function
N(i)\mathcal{N}(i)set of neighboring nodes of ii

Table 9. Summary of notations.

Related Work

Super-resolution for simulations. Similar to surrogate models for simulations, early super-resolution models for simulations predominantly employed CNN-based image super-resolution architectures, such as SRGAN (Li & McComb, 2022) and UNet (Yonekura et al., 2023). As a pioneering work, CFD-GCN [11] introduced GCNs for the super-resolution of computational fluid dynamics (CFD) simulations. This method demonstrated both improved generalization to unseen data and enhanced cost efficiency. More recently, advanced MPNN architectures such as SRGNN were applied to the super-resolution of fluid flows [5]. Despite these advancements, inductive biases tailored for MPNNs in the context of super-resolution for mesh-based simulations are largely underexplored.

Semi-supervised regression. Semi-supervised regression involves predicting real-valued output using both labeled and unlabeled datasets. Compared to semi-supervised classification, semi-supervised regression remains largely underexplored [20]. A co-training approach typically splits input features into groups, with each group used to train a separate model [7]. However, in scenarios with limited features, such as FEM-relevant data including only nodal positions and nodal values, splitting features can lead to insufficient information for accurate predictions. As an alternative, CoREG [37] was presented to eliminate the need for feature splitting by using two kk-nearest neighbor (kkNN) regressors with different distance metrics. Unfortunately, this approach is restricted to kkNN regressors and is unsuitable for predicting values at the node level. A recent method, Rankup [14], reformulated regression tasks into classification tasks to leverage a rich set of methodologies developed for semi-supervised classification. However, this technique lacks generalizability for regression tasks involving mesh-based graph data. The Mean-Teacher framework [32], though not originally designed for mesh-based data, has potential for dealing with mesh-based graph data. It involves teacher and student models with the teacher’s weights updated as the exponential moving average of the student’s weights. In contrast, the recently proposed UCVME framework [10] has demonstrated superior performance over Mean-Teacher [32] by incorporating uncertainty consistency and utilizing a variational model ensemble. However, because Mean-Teacher [32] and UCVME [10] both employ identically structured models predicting the same target, they exhibit reduced pseudo-label diversity (correlated errors) during training. This uniformity (correlation) diminishes synergy between the models and consequently hinders learning efficiency. Additionally, Twin Neural Network Regression (TNNR) [34] is applicable to mesh-based predictions when an appropriate model architecture is involved. However, it involves only a single twin neural network, thus lacking the synergistic benefits of mutual supervision.

Pseudo-code for Complementary Learning

We present pseudo-code of our complementary learning mechanism in SuperMeshNet.

Algorithm 1 Complementary learning

Input: paired LR–HR dataset: Da={(ulq,uhq)∣q=1,2,⋯ ,Nh}\mathcal{D}_a = \{(u_l^q,u_h^q) \mid q = 1,2,\cdots,N_h\}, unpaired LR dataset: Db={ulq∣q=Nh+1,Nh+2,⋯ ,N}\mathcal{D}_b = \{u_l^q \mid q = N_{h+1},N_{h+2},\cdots,N\}, neural network models: feature extractor EcE_c, FθF_\theta’s decoder DFD_F, and GϕG_\phi’s decoder DGD_G, maximum number of epochs: epep, learning rate: η\eta, early stopping criterion

Output: Trained neural network models: EcE_c, DFD_F, and DGD_G

for epoch ←1\leftarrow1 to epep do

for step ←1\leftarrow1 to NN do

Sample paired LR–HR data (ulα,uhα),(ulβ,uhβ)∈Da(u_l^\alpha,u_h^\alpha),(u_l^\beta,u_h^\beta) \in\mathcal{D}_a

Sample unpaired LR data ulγ∈Dbu_l^\gamma\in\mathcal{D}_b

Compute node embeddings by EcE_c:

xα←Ec(ulα),xβ←Ec(ulβ),xγ←Ec(ulγ)x^\alpha\leftarrow E_c(u_l^\alpha),\quad x^\beta\leftarrow E_c(u_l^\beta),\quad x^\gamma\leftarrow E_c(u_l^\gamma)

Compute prediction by DFD_F:

u^hα←DF(xα),u^hβ←DF(xβ),u^hγ←DF(xγ)\hat{u}_h^\alpha\leftarrow D_F(x^\alpha),\quad\hat{u}_h^\beta\leftarrow D_F(x^\beta),\quad\hat{u}_h^\gamma\leftarrow D_F(x^\gamma)

Compute prediction by DGD_G:

u^hαβ←DG(xα,xβ),u^hβγ←DG(xβ,xγ),u^hγα←DG(xγ,xα)\hat{u}_h^{\alpha\beta} \leftarrow D_G(x^\alpha,x^\beta),\quad\hat{u}_h^{\beta\gamma} \leftarrow D_G(x^\beta,x^\gamma),\quad\hat{u}_h^{\gamma\alpha} \leftarrow D_G(x^\gamma,x^\alpha)

Compute loss for FθF_\theta:

LF=ℓ(u^hα,uhα)+ℓ(u^hβ,uhβ)+ℓ(u^hγ,uhγ+kNN(uhα;Phα→Phγ))+ℓ(u^hγ,kNN(uhβ−u^hβγ;Phβ→Phγ))\mathcal{L}_F = \ell(\hat{u}_h^\alpha,u_h^\alpha) + \ell(\hat{u}_h^\beta,u_h^\beta) + \ell(\hat{u}_h^\gamma,u_h^\gamma+ kNN(u_h^\alpha;P_h^\alpha\to P_h^\gamma)) + \ell(\hat{u}_h^\gamma,kNN(u_h^\beta-\hat{u}_h^{\beta\gamma};P_h^\beta\to P_h^\gamma))

Compute loss for GϕG_\phi:

LG=ℓ(u^hαβ,uhα−kNN(uhβ;Phβ→Phα))+ℓ(u^hγα,uhγ−kNN(uhα;Phα→Phγ))+ℓ(u^hβγ,uhβ−kNN(u^hγ;Phγ→Phβ))\mathcal{L}_G = \ell(\hat{u}_h^{\alpha\beta},u_h^\alpha-kNN(u_h^\beta;P_h^\beta\to P_h^\alpha)) + \ell(\hat{u}_h^{\gamma\alpha},u_h^\gamma-kNN(u_h^\alpha;P_h^\alpha\to P_h^\gamma)) + \ell(\hat{u}_h^{\beta\gamma},u_h^\beta-kNN(\hat{u}_h^\gamma;P_h^\gamma\to P_h^\beta))

Compute total loss:

L=LF+LG\mathcal{L} = \mathcal{L}_F + \mathcal{L}_G

Compute gradients: ∇ψL\nabla_\psi\mathcal{L} where ψ\psi is the parameters of EcE_c, DFD_F, and DGD_G

Update weights:

ψ←ψ−η∇ψL\psi\leftarrow\psi-\eta\nabla_\psi\mathcal{L}

end for

if early stopping criterion is met then

Break

end if

end for

return EcE_c, DFD_F, and $D_G

kkNN Interpolation in SuperMeshNet

We provide a brief explanation of the kkNN interpolation procedure in SuperMeshNet, which projects values defined on nodes of a source mesh onto nodes of a target mesh.

  1. Find kk nearest neighbors (kkNN). For each node in the target mesh, identify the kk closest nodes in the source mesh. For example, as illustrated in Figure 10, the darker blue nodes represent the three nearest neighbors of the darker yellow node.

Figure 10. Schematic illustration of kkNN interpolation with k=3k = 3. Yellow nodes belong to the target mesh, blue nodes to the source mesh, and the darker blue nodes indicate the kk nearest neighbors of the darker yellow node. Given the positions of the kk nearest source nodes pip_i (1≤i≤k1 \le i \le k), their corresponding values yiy_i, and the target node position p0p_0, the value at the target node y0y_0 can be estimated via weighted averaging.

  1. Known information. The nodal positions of the kk nearest source nodes pip_i (1≤i≤k1 \le i \le k), their values yiy_i, and the target node position p0p_0.

  1. Unknown quantity. The value at the target node, denoted by y0y_0.

  1. Compute the target node value via weighted averaging. The interpolation weight for each neighbor is defined as the inverse squared distance from the target node:

wi=1d(p0,pi)2,w_i = \frac{1}{d(p_0,p_i)^2},

where d(⋅,⋅)d(\cdot,\cdot) denotes the Euclidean distance.

The interpolated value at the target node y0y_0 is then obtained as

y0=∑i=1kwiyi∑i=1kwi.y_0 = \frac{\sum_{i=1}^{k} w_i y_i}{\sum_{i=1}^{k} w_i}.

Model Architectures

The architecture of FθF_\theta, illustrated in Figure 11, is basically built upon SRGNN ([5]), with the key difference that the MPNNs in the LR and HR processors are enriched with our proposed inductive biases. The GϕG_\phi, visualized in Figure 12, extends FθF_\theta to accommodate two input samples, maintaining a comparable structure. A notable architectural feature is the use of a shared feature extractor between FθF_\theta and GϕG_\phi, which helps reduce computational costs during training. A detailed description of the model architectures follows.

Figure 11. The schematic overview of the primary model FθF_\theta. The model FθF_\theta aims to predict u^hq\hat{u}_h^q targeting HR data sample uhqu_h^q from LR data sample ulqu_l^q. The LR data sample ulqu_l^q is input to FθF_\theta as a part of node feature 0Xlq^0X_l^q of an input graph glqg_l^q.

Figure 12. The schematic overview of the auxiliary model GϕG_{\phi}. The GϕG_{\phi} aims to predict u^hrs\hat{u}^{rs}_{h} targeting difference between two input LR data samples ulru^{r}_{l} and ulsu^{s}_{l}. The two input LR data samples ulru^{r}_{l} and ulsu^{s}_{l} are fed into the GϕG_{\phi} as parts of node features of two input graphs glrg^{r}_{l} and glsg^{s}_{l}, respectively. In order to reduce computational cost, FθF_{\theta} and GϕG_{\phi} share a feature extractor in Figure 11 consisting of an encoder, an LR processor, an upsampler, and an HR processor.

Model architecture of FθF_\theta

The role of FθF_\theta is to transform LR data into HR data, which is conducted by the lowermost upsampler in Figure 11. To surpass the performance of kNN interpolation by the lowermost upsampler, we introduce additional upsampling in latent space. Specifically, an encoder maps the physical quantities into high-dimensional latent space. The LR processor applies message passing to refine LR representations, which are then upsampled to HR latent embeddings. Subsequently, the HR processor applies additional message passing to further enhance the HR representations. Finally, a decoder maps the latent embeddings back to the physical space. The final HR output is obtained by adding the two upsampled HR fields: one from the kNN-based upsampler and the other from the latent-space upsampling pathway.

More precisely, FθF_\theta is designed to make prediction u^hq\hat{u}_h^q from an LR data sample ulqu_l^q. The LR data sample ulqu_l^q is input to the FθF_\theta as a form of an input graph glqg_l^q. More specifically, input graph glqg_l^q’s node feature 0Xlq^0X_l^q is the concatenation of LR data sample ulqu_l^q and node position PlqP_l^q. Depending on MPNN types used in the LR and HR processors (refer to Appendix § 13), the glqg_l^q may further include edge feature, which is the concatenation of positions of source and target nodes of the edge ElqE_l^q. The FθF_\theta comprises the encoder, the LR processor, upsamplers, the HR processor, and the decoder. The encoder can be a multi-layer perceptron (MLP) that can convert low-dimensional 0Xlq^0X_l^q to high-dimensional node embeddings 0xlq^0x_l^q. The LR processor updates the node embedding 0xlq^0x_l^q to Hlxlq^{H_l}x_l^q through stacked HlH_l MPNN layers enriched by our inductive biases. Here, the prescript 0 and HlH_l indicate the index of the MPNN layers. The node embedding Hlxlq^{H_l}x_l^q defined on nodes located at PlqP_l^q is upsampled onto nodes of ghqg_h^q positioned at PhqP_h^q by using kNN interpolation. The HR processor similarly updates 0xhq^0x_h^q to Hhxhq^{H_h}x_h^q through stacked HhH_h MPNN layers equipped with our inductive biases. Then, the decoder, which is an MLP, predicts low-dimensional output from Hhxhq^{H_h}x_h^q. Finally, the LR data sample ulqu_l^q upsampled onto PhqP_h^q by kNN interpolation is added to the output of the decoder to obtain the final prediction u^hq\hat{u}_h^q. The upsampled LR data sample serves as a rough estimation of the prediction, enabling the super-resolution model to focus on learning the finer details, thereby simplifying the learning task.

Model architecture of GϕG_\phi

The model GϕG_\phi is responsible for predicting the difference between two HR samples corresponding to two LR inputs. Again, to go beyond simple kNN-based upsampling by upsampler in Figure 12, we further perform latent-space processing. We extract latent embeddings from the two LR inputs using a shared encoder. The shared encoder is the one used for the model FθF_\theta depicted in Figure 11. Then, we subtract the embeddings, and decode the result to predict the HR difference. Here, we incorporate subtraction because the goal is to predict the difference between two HR samples. The final HR output is obtained by adding the two upsampled HR fields: one from the kNN-based upsampler and the other from the latent-space upsampling pathway. The interpolators in Figure 12 serve only to address mesh mismatches when the two LR samples are defined on different meshes. Since the underlying computational domain geometry may vary across samples, direct point-wise operations, such as subtraction or addition, are generally infeasible. To overcome this, we apply kNNk\mathrm{NN} interpolation to project one mesh onto another, enabling consistent alignment between mesh structures.

More precisely, the auxiliary model GϕG_{\phi} is designed to make prediction u^hrs\hat{u}^{rs}_{h} from a pair of LR data samples ulru^{r}_{l} and ulsu^{s}_{l}. More specifically, the two input LR data samples ulru^{r}_{l} and ulsu^{s}_{l} are fed into the GϕG_{\phi} as parts of node features of two input graphs glrg^{r}_{l} and glsg^{s}_{l}, respectively. In order to reduce computational costs, FθF_{\theta} and GϕG_{\phi} share a feature extractor comprising the encoder, the LR processor, the interpolator, and the HR processor. The shared feature extractor returns node embeddings xhrx^{r}_{h} and xhsx^{s}_{h} from input graphs glrg^{r}_{l} and glsg^{s}_{l}, respectively. Then, xhsx^{s}_{h} is subtracted from xhrx^{r}_{h} to yield xhrsx^{rs}_{h}. Here, kNNk\mathrm{NN} interpolator is used to enable subtraction operation between two node embeddings xhrx^{r}_{h} and xhsx^{s}_{h} defined at different nodal positions. The xhrsx^{rs}_{h} is fed into the decoder, and the upsampled difference between ulru^{r}_{l} and ulsu^{s}_{l} through kNNk\mathrm{NN} interpolation is added to the decoder’s output. Here, the interpolated difference between ulru^{r}_{l} and ulsu^{s}_{l} also serves as a rough estimation of the prediction u^hrs\hat{u}^{rs}_{h}. Again, a kNNk\mathrm{NN} interpolator is used to enable subtraction operation between ulru^{r}_{l} and ulsu^{s}_{l} defined on different nodal positions.

Incorporation of Inductive Biases into MPNN Architectures

This section describes how inductive biases are incorporated into each of MPNN models.

Inductive biases-enriched GCN [17]

msgij=ΘTejixjxi=aggi=∑j∈N(i)msgijxi←xi−1n∑i=1nxi(13)\begin{aligned} msg_{ij} &= \Theta^{T}e_{ji}x_j \\ x_i = agg_i &= \sum_{j \in\mathcal{N}(i)} msg_{ij} \\ x_i &\leftarrow x_i - \frac{1}{n}\sum_{i=1}^{n} x_i \tag*{(13)} \end{aligned}

where Θ\Theta is a learnable parameter and ejie_{ji} is an edge weight.

Inductive biases-enriched GraphSAGE (SAGE) [13]

msgij=xjaggi=1∣N(i)∣∑j∈N(i)msgijaggi←aggi−1n∑i=1naggixi←W1xi+W2aggixi←xi−1n∑i=1nxi(14)\begin{aligned} msg_{ij} &= x_j \\ agg_i &= \frac{1}{\lvert\mathcal{N}(i) \rvert}\sum_{j \in\mathcal{N}(i)} msg_{ij} \\ agg_i &\leftarrow agg_i - \frac{1}{n}\sum_{i=1}^{n} agg_i \\ x_i &\leftarrow W_1x_i + W_2agg_i \\ x_i &\leftarrow x_i - \frac{1}{n}\sum_{i=1}^{n} x_i \tag*{(14)} \end{aligned}

where W1W_1 and W2W_2 are learnable parameters.

Inductive biases-enriched GAT [33]

αij=exp⁡(LeakyReLU⁡(asTΘsxi+atTΘtxj))∑k∈N(i)exp⁡(LeakyReLU⁡(asTΘsxi+atTΘtxk))msgij=αijΘtxjxi=aggi=∑j∈N(i)msgijxi←xi−1n∑i=1nxi(15)\begin{aligned} \alpha_{ij} &= \frac{\exp\left(\operatorname{LeakyReLU}\left(a_s^{T}\Theta_s x_i + a_t^{T}\Theta_t x_j\right)\right)}{\sum_{k \in\mathcal{N}(i)} \exp\left(\operatorname{LeakyReLU}\left(a_s^{T}\Theta_s x_i + a_t^{T}\Theta_t x_k\right)\right)} \\ msg_{ij} &= \alpha_{ij}\Theta_t x_j \\ x_i = agg_i &= \sum_{j \in\mathcal{N}(i)} msg_{ij} \\ x_i &\leftarrow x_i - \frac{1}{n}\sum_{i=1}^{n} x_i \tag*{(15)} \end{aligned}

where asa_s, ata_t, Θs\Theta_s, and Θt\Theta_t are learnable parameters.

Inductive biases-enriched Graph Transformer (GTR)

αij=softmax⁡((W3xi)T(W4xj))msgij=αijW2xjaggi=∑j∈N(i)msgijaggi←aggi−1n∑i=1naggixi←W1xi+aggixi←xi−1n∑i=1nxi(16)\begin{aligned} \alpha_{ij} &= \operatorname{softmax}\left((W_3x_i)^{T}(W_4x_j)\right) \\ msg_{ij} &= \alpha_{ij}W_2x_j \\ agg_i &= \sum_{j\in\mathcal{N}(i)} msg_{ij} \\ agg_i &\leftarrow agg_i-\frac{1}{n}\sum_{i=1}^{n}agg_i \\ x_i &\leftarrow W_1x_i+agg_i \\ x_i &\leftarrow x_i-\frac{1}{n}\sum_{i=1}^{n}x_i \tag*{(16)} \end{aligned}

where W1W_1, W2W_2, W3W_3 and W4W_4 are learnable parameters.

Inductive biases-enriched GIN

msgij=xjaggi=∑j∈N(i)msgijaggi←aggi−1n∑i=1naggixi←MLP⁡Θ((1+ϵ)xi+aggi)xi←xi−1n∑i=1nxi(17)\begin{aligned} msg_{ij} &= x_j \\ agg_i &= \sum_{j\in\mathcal{N}(i)} msg_{ij} \\ agg_i &\leftarrow agg_i-\frac{1}{n}\sum_{i=1}^{n}agg_i \\ x_i &\leftarrow\operatorname{MLP}_{\Theta}\left((1+\epsilon)x_i+agg_i\right) \\ x_i &\leftarrow x_i-\frac{1}{n}\sum_{i=1}^{n}x_i \tag*{(17)} \end{aligned}

where MLP⁡Θ\operatorname{MLP}_{\Theta} is a learnable MLP and ϵ\epsilon is a learnable parameter.

Inductive biases-enriched MeshGraphNet (MGN)

eij←MLP⁡e(xi,xj,eij)msgij=eijaggi=∑j∈N(i)msgijaggi←aggi−1n∑i=1naggixi←MLP⁡x(xi,aggi)xi←xi−1n∑i=1nxi(18)\begin{aligned} e_{ij} &\leftarrow\operatorname{MLP}_{e}(x_i,x_j,e_{ij}) \\ msg_{ij} &= e_{ij} \\ agg_i &= \sum_{j\in\mathcal{N}(i)} msg_{ij} \\ agg_i &\leftarrow agg_i-\frac{1}{n}\sum_{i=1}^{n}agg_i \\ x_i &\leftarrow\operatorname{MLP}_{x}(x_i,agg_i) \\ x_i &\leftarrow x_i-\frac{1}{n}\sum_{i=1}^{n}x_i \tag*{(18)} \end{aligned}

where MLP⁡e\operatorname{MLP}_{e} and MLP⁡x\operatorname{MLP}_{x} are learnable MLPs.

Datasets for Experimental Evaluations

Dataset 1

The first dataset is inspired by simustruct [26], the dataset for machine learning-based methods in structural analysis. Examples of HR and LR data samples from Dataset 1 are visualized in Figure 13. As depicted in the figure, the computational domain is a rectangle measuring 0.25×0.50.25 \times0.5 in the xx- and yy-directions, containing six circular holes, each with a diameter of 0.050.05. For the HR mesh, the mesh size around outer four sides is 10×10−310 \times10^{-3}, while the mesh size around the circular holes is set to be 4×10−34 \times10^{-3}. For the LR mesh, the mesh size around the outer sides is 40×10−340 \times10^{-3}, and the mesh size around the circular holes is 16×10−316 \times10^{-3}.

On the computation domain, the following linear elasticity equation is solved:

−∇⋅σ(u)=0σ(u)=λtr⁡(ϵ(u))I+2Gϵ(u),ϵ(u)=12(∇u+(∇u)T),(19)\begin{aligned} -\nabla\cdot\sigma(u) &= 0 \\ \sigma(u) &= \lambda\operatorname{tr}(\epsilon(u))I + 2G\epsilon(u), \\ \epsilon(u) &= \frac{1}{2}\left(\nabla u + (\nabla u)^{T}\right), \tag*{(19)} \end{aligned}

where σ(u)\sigma(u) is the stress tensor, λ\lambda and GG are Lamé’s elasticity parameters for the material, II is the identity tensor, tr⁡\operatorname{tr} is the trace operator on a tensor, ϵ(u)\epsilon(u) is the symmetric strain tensor (symmetric gradient), and uu is the displacement vector field.

A force of 1×1081 \times10^{8} is applied to the top side of the rectangle in angles between 40∘40^\circ and 140∘140^\circ relative to the xx-axis, while the bottom side of the rectangle is fixed to zero displacement. Lamé’s first and second parameters are 1.251.25, and 80.8×10980.8 \times10^{9}, respectively. Von Mises stress is evaluated at each node of the meshes. In order to solve the equation for each dataset, we leverage FEniCSx (Baratta et al., 2023; Scroggs et al., 2022b;a; Alnaes et al., 2014), an open-source computing platform for solving PDEs with the FEM.

Dataset 2

The geometry of the second dataset resembles that of the first dataset, with the primary difference being the shapes of the holes. Specifically, the holes in the second dataset are elliptical, with varying ratios between the lengths of the major and minor axes. The mesh sizes remain the same as those in Dataset 1. Similarly as in Dataset 1, the linear elasticity equation in Eq. (19) is solved. The applied force is directed along the yy-axis, while all other conditions and constants remain identical to those in Dataset 1. Examples of HR and LR data samples from Dataset 2 are visualized in Figure 14.

LR and HR data samples from Dataset 2 with varying ratios between the lengths of the major and minor axes

Figure 14. Examples of LR and HR data samples with various ratios between the lengths of the major and minor axes from Dataset 2.

Dataset 3

The geometry and mesh sizes of Dataset 3 are identical to those of Dataset 2. However, instead of solving the linear elasticity equation, the following Poisson equation is solved.

∇2u=0,(20)\nabla^{2}u = 0, \tag*{(20)}

where uu is an electrical potential.

The boundary conditions are defined as follows: the four outer sides are set to 0, while the elliptical holes have alternating boundary values. Specifically, the holes centered at (0.08, 0.15), (0.17, 0.25), and (0.08, 0.35) are assigned a value of -1, whereas the holes centered at (0.17, 0.15), (0.08, 0.25), and (0.17, 0.35) are assigned a value of 1. The magnitude of the electric field is calculated at each node of the mesh. Examples of HR and LR data samples from the Dataset 3 are visualized in Figure 15.

Real-World Geometry Dataset

An Example of HR and LR data samples from the real-world geometry dataset are visualized in Figure 16. As depicted in the figure, the computational domain is a rectangular box of size 20×8×820 \times8 \times8 in the xx-, yy- and zz-directions, containing a rider on a motorbike. The mesh size is set to be finer around the rider and the motorbike. The dataset is built upon a bike tutorial of OpenFOAM [16] by varying angle of attack from 0∘0^\circ to −90∘-90^\circ. The speed of fluid at the left side of the rectangular box is set to be 20.

LR and HR data samples corresponding to an angle of attack of 0 degrees

Figure 16. An example of LR and HR data samples corresponding to an angle of attack of 0∘0^\circ from the real-world geometry dataset.

To obtain the solution, we use OpenFOAM [16] in two stages. First, a potential-flow initialization is performed by solving

∇2ϕ=0,(21)\nabla^{2}\phi= 0, \tag*{(21)}

and setting

u=∇ϕ,(22)u = \nabla\phi, \tag*{(22)}

where uu denotes velocity and ϕ\phi is the velocity potential. This velocity field is used as an initial guess. Then, a steady-state incompressible RANS solver is run to obtain a more accurate velocity and pressure field.

Time-Dependent PDE Dataset 1

Examples of HR and LR data samples from the time-dependent PDE dataset 1 are visualized in Figure 17. As depicted in the figure, the computational domain is a square measuring 2×22 \times2 in the xx- and yy-directions, containing one cylinder at the center of the domain with a diameter of 0.050.05. The mesh size is set to be finer around the cylinder.

Example velocity magnitudes of LR and HR data samples at Time=20, Time=20.16, and Time=20.32

Figure 17. Example velocity magnitudes of LR and HR data samples corresponding to multiple timestamps from the time-dependent PDE dataset 1.

On the computation domain, the following incompressible Navier-Stokes equation is solved:

∇⋅u=0,(23)\nabla\cdot\mathbf{u} = 0, \tag*{(23)}
ρ(∂u∂t+u⋅∇u)=−∇p+η∇2u,(24)\rho\left(\frac{\partial\mathbf{u}}{\partial t} + \mathbf{u} \cdot\nabla\mathbf{u}\right) = -\nabla p + \eta\nabla^{2}\mathbf{u}, \tag*{(24)}

where u\mathbf{u} is velocity, pp is pressure, and ρ\rho is density, η\eta is dynamic viscosity. The velocity at the left boundary of the square increases from 0 to 1 over the time interval t∈[0,0.5]t \in[0, 0.5], and remains constant thereafter. The time step is set to 5×10−45 \times10^{-4}, and the total simulation time is 40; only data for t≥20t \ge20 are used. The density and dynamic viscosity are set to 1 and 5×10−45 \times10^{-4}, respectively. The governing equations are solved using OpenFOAM ([16]), an open-source CFD toolbox.

Time-Dependent PDE Dataset 2

The time-dependent PDE dataset 2 is generated using the JAX-CFD library [19], following the Kolmogorov-type forced 2D incompressible Navier–Stokes dynamics. Examples of HR and LR vorticity fields are visualized in Figure 18. In contrast to the time-dependent PDE dataset 1, this dataset is designed so that the HR and LR trajectories exhibit markedly different flow structures.

Examples of LR and HR data samples corresponding to multiple timestamps from the time-dependent PDE dataset 2

Figure 18. Examples of LR and HR data samples corresponding to multiple timestamps from the time-dependent PDE dataset 2.

The computation domain is a periodic square domain (0,2π)×(0,2π)(0, 2\pi) \times(0, 2\pi). On this domain, we solve the 2D incompressible Navier–Stokes equations under external forcing:

∇⋅u=0,(25)\nabla\cdot\mathbf{u} = 0, \tag*{(25)}
∂u∂t+u⋅∇u=−∇p+ν∇2u+f,(26)\frac{\partial\mathbf{u}}{\partial t} + \mathbf{u} \cdot\nabla\mathbf{u} = -\nabla p + \nu\nabla^{2} \mathbf{u} + \mathbf{f}, \tag*{(26)}

where u\mathbf{u} is velocity, pp is pressure, and ν\nu is the kinematic viscosity. We set ν=10−2\nu= 10^{-2} and maximum velocity to be 7. f\mathbf{f} is the external forcing term.

The HR simulation uses a 1024×10241024 \times1024 spectral grid, while the LR simulation uses a 32×3232 \times32 grid. The HR time step is determined using the stability condition provided in JAX-CFD, and the LR time step is scaled proportionally to the grid resolution ratio. Both simulations are advanced using the Crank–Nicolson RK4 integrator implemented in the JAX-CFD spectral module.

The LR initial condition is downsampled in the Fourier space. We truncate the HR Fourier coefficients to retain only the lowest 32×3232 \times32 modes and rescale the amplitudes to preserve energy. We evaluate the vorticity, ω=∇×u\omega= \nabla\times\mathbf{u}, at every grid point for all time steps. A total of 300 frames are collected for both HR and LR simulations.

Physical Validation of Datasets

FEM Datasets

To verify the physical validity of the three FEM datasets, we conduct mesh convergence tests on the stress and electric field in high-concentration regions. For Dataset 1, the high-concentration regions are defined as circles of radius 0.025 located to the left and right of the bottom-right hole. For Dataset 2, the region is defined as a circle of radius 0.035 located to the right of the bottom-right hole. For Dataset 3, the regions are defined as circles of radius 0.026 located to the left of the bottom-right hole and to the right of the bottom-left hole. As shown in Figure 19, these quantities exhibit clear saturation as the mesh is refined, indicating convergence toward a mesh-independent regime.

Mesh convergence tests of stress and electric field in high-concentration regions across three FEM datasets

Figure 19. Mesh convergence tests of stress and electric field in high-concentration regions across three FEM datasets.

Our HR FEM datasets contain approximately 4000 nodes, which lie near the onset of this convergence plateau. While further refinement may yield slight accuracy improvements, the gain is marginal relative to the substantial increase in data generation cost. Given that our objective is to study data-efficient super-resolution, we consider this resolution regime to be appropriate. Moreover, if the HR solutions were fully mesh-independent, the super-resolution task would become trivial.

Real-World Geometry Dataset

The mesh convergence results in Figure 20 demonstrate that both drag and lift coefficients exhibit clear convergence trends with mesh refinement, verifying physical validity of the real-world geometry dataset. The HR data of the real-world geometry dataset, comprising approximately 80,000 nodes, lies near the onset of the convergence plateau.

Mesh convergence tests of the drag and lift coefficients for the real-world geometry dataset

Figure 20. Mesh convergence tests of the drag and lift coefficients for the real-world geometry dataset.

Time-Dependent PDE Dataset 1

To verify the physical validity of the Time-dependent PDE dataset 1, we conduct mesh convergence tests on mean and amplitude of drag coefficients and amplitude of lift coefficients, which reflect vortex shedding behind a cylinder. As depicted in Figure 21, those values approach constant values with mesh refinement. Our HR data contain approximately 20,000 nodes so that the HR data fall near the onset of the plateau regime, consistent with the FEM datasets.

Mesh convergence tests of the mean and amplitude of drag coefficients, and the amplitude of lift coefficients

Figure 21. Mesh convergence tests of the mean and amplitude of drag coefficients, and the amplitude of lift coefficients, for the time-dependent PDE dataset 1.

Additional Experimental Results and Analyses

The following Table 10 summarizes the contents of this section.

Comparison of MAgNet, Full Supervision, and SuperMeshNet

We empirically show the effect of HR supervision by comparing no HR supervision (MAgNet [6]), partial supervision (SuperMeshNet), and full supervision cases. Table 11 demonstrates that MAgNet’s RMSE is far apart from that of the fully supervised baseline and SuperMeshNet across all three datasets, while being approximately up to four times higher.

Dataset 1Dataset 2Dataset 3
MAgNet0.0979±0.00090.0979 \pm0.00090.1305±0.00070.1305 \pm0.00070.0754±0.00140.0754 \pm0.0014
Fully supervised0.0228±0.00150.0228 \pm0.00150.0461±0.00040.0461 \pm0.00040.0243±0.00170.0243 \pm0.0017
SuperMeshNet0.0226±0.00070.0226 \pm0.00070.0507±0.00110.0507 \pm0.00110.0245±0.00050.0245 \pm0.0005

Table 11. RMSE comparison of MAgNet [6], the baseline using fully supervised learning, and SuperMeshNet. SuperMeshNet is MGN with inductive biases trained with Nh=20N_h = 20 and N=200N = 200, and the baseline is MGN without inductive biases trained with Nh=N=200N_h = N = 200. MAgNet is a zero-shot super-resolution method trained with Nh=0N_h = 0 (i.e., no HR data) and N=200N = 200.

Analysis on Model Architectures

Effect of Our Shared Feature Extractor

As addressed in subsection 2.3, in SuperMeshNet, the two models FθF_{\theta} and GϕG_{\phi} use a shared feature extractor. To assess its effect, we compare our default shared-extractor implementation with a variant that employs separate extractors for the two models. As shown in Table 12, using separate extractors improves accuracy, but at a substantially higher computational cost. The higher accuracy is because each extractor can specialize more strongly for the distinct roles of the two models. Therefore, the two models’ pseudo-label errors are less correlated, resulting in synergy between the two. In contrast, sharing the extractor reduces training time by more than a factor of three by avoiding redundant feature extraction on the same input under complementary learning. Given this trade-off between accuracy and computational complexity, we adopt the shared feature extractor in SuperMeshNet.

MethodsRMSETraining time (s)
Separate feature extractors0.0192962.85
Shared feature extractor0.0226263.72

Table 12. Comparison of a shared feature extractor with separate feature extractors in terms of RMSE and training time. Here, MGN is employed as an MPNN for each method. Training is conducted when Nh=20N_h=20 and N=200N=200 for Dataset 1.

Application to a CNN-Based Architecture or Image Super-Resolution

To test whether our complementary learning can be generalized beyond MPNN architectures, we additionally apply complementary learning to a CNN-based architecture. Specifically, we replace both the LR and HR MPNN processors with ResNet blocks. For regular-grid data, we also replace kkNN interpolation with bilinear interpolation for computational efficiency. Table 13 demonstrates that complementary learning not only improves HR-data efficiency but also even outperforms the fully supervised baseline on the regular-grid setting for the time-dependent PDE dataset 2. This result indicates that complementary learning is architecture-agnostic and can be flexibly extended beyond MPNNs.

MethodsNhN_hNNRMSE
Complementary learning + ResNet402000.0339
Full supervision + ResNet2002000.0417

Table 13. Application of complementary learning to a CNN-based architecture.

In principle, the proposed complementary learning is not inherently restricted to mesh-based simulations. At its core, the method leverages unpaired LR data samples through mutual supervision between two complementary models, which can be naturally extended to other super-resolution tasks such as image super-resolution. That being said, we would like to emphasize simulation domains, where data scarcity, corresponding to our target problem setting, is significantly more severe compared to the case of image super-resolution.

Analysis on Inductive Biases

We observe that effects of our inductive biases depends on the task.

  • For tasks where the global mean information is less important (e.g., super-resolution), inductive biases tend to smooth the optimization landscape and can improve RMSE performance.

  • For tasks where the global mean is important (e.g., norm prediction), inductive biases removes useful information and may hinder optimization, leading to degraded performance.

In this section, we first consider super-resolution and norm prediction as representative tasks where global mean information is, respectively, less important and more important. We further present additional exemplary tasks that may benefit from our inductive biases. Furthermore, we compare our inductive biases with standard normalization techniques.

Example 1: Super-resolution

Task description

The goal of super-resolution is to predict HR fields from LR fields.

Importance of the global mean on the super-resolution task

As summarized in Table 14, subtracting the mean from LR data (i.e., using the deviation from the mean as input) results in only a minor change in RMSE, whereas providing only the global mean (i.e., using the mean of the LR data as input) significantly degrades performance. This suggests that the global mean contributes little to prediction accuracy. Super-resolution primarily involves learning high-frequency discrepancies between LR and HR fields [12]. Moreover, our architecture (Figure 11 and Figure 12) generates HR predictions by refining a kkNN-interpolated coarse estimate, which may further reduce the need to model the global mean explicitly.

InputRMSE
LR data0.0228
Deviation of LR data0.0235
Mean of LR data0.0617

Table 14. RMSE comparison of MGN-based SuperMeshNet trained for super-resolution task on Dataset 1, using the following three cases as input: the LR data sample, its deviation from the mean, and the mean.

We further clarify that whether mean information matters for a task is determined by how strongly the ground truth output is influenced by the mean of the input. In Figure 22, the super-resolution task does not exhibits a clear dependency between the input mean and the target output, demonstrating that global mean information is relatively unimportant for this task.

Relationship between the mean of the input and the ground truth output for the super-resolution task

Figure 22. Relationship between the mean of the input and the ground truth output for the super-resolution task.

Effect of inductive biases on loss landscape

As illustrated in Figure 23, compared to the case without our inductive biases (w/o IB), the cases with our inductive biases (w IB) yield a smoother loss landscape over training iterations in the super-resolution task. As reported in prior work on batch normalization [28], a smoother loss landscape stabilizes and improves optimization, possibly improving the performance.

Loss landscapes for SuperMeshNet with and without inductive biases

Figure 23. Effect of inductive biases on the loss landscape for the super-resolution task when SuperMeshNet with and without inductive biases (corresponding to w IB and w/o IB, respectively) is used. Here, MGN is employed as an MPNN for each method and is trained when Nh=20N_h=20 and N=200N=200 for Dataset 1.

To measure the loss landscape, following the prior work in [28], we perturb the model parameters by taking a step along the gradient direction and evaluate the loss at the perturbed point. The perturbation magnitude is set to a fixed multiple of the learning rate (four times the learning rate in our experiments).

Effect of inductive biases on RMSE performance

As summarized in Table 15, our inductive bias empirically improves RMSE performance in super-resolution. Although it removes global mean information, this information is relatively less important for super-resolution tasks. Moreover, the inductive bias smooths the loss landscape, facilitating improved performance.

Inductive biasesRMSE
O0.0226
X0.0269

Table 15. Effect of inductive biases on the performance of super-resolution task. All experiments are conducted using MGN-based SuperMeshNet and Dataset 1.

Example 2: Norm prediction

Task description

We consider an exemplary task, whose goal is to predict the (normalized) norm of LR input fields, which is constant across nodes.

Importance of the global mean on the norm prediction task

For the norm prediction, the global mean appears to contain essential information. As summarized in Table 16, when only the mean is provided as input, performance gets improved, whereas providing only the deviation significantly degrades performance.

InputRMSE
LR data0.00269
Deviation of LR data0.00378
Mean of LR data0.00149

Table 16. RMSE comparison of MGN-based SuperMeshNet trained for norm prediction task on Dataset 1, using the following three cases as input: the LR data sample, its deviation from the mean, and the mean.

Furthermore, as illustrated in Figure 24, the norm prediction task exhibits a clear dependency between the input mean and the target output (the norm), indicating that the mean indeed carries essential information.

Relationship between the mean of the input and the ground truth output for the norm prediction task

Figure 24. Relationship between the mean of the input and the ground truth output for the norm prediction task.

Effect of inductive biases on loss landscape

As illustrated in Figure 25, the case with inductive biases (w IB) yields a noticeably rougher loss landscape in norm prediction.

Loss landscapes for norm prediction with and without inductive biases

Figure 25. Effect of inductive biases on the loss landscape for the norm prediction task when SuperMeshNet with and without inductive biases (corresponding to w IB and w/o IB, respectively) is used. Here, MGN is employed as an MPNN for each method and is trained when NhN_h=20 and NN=200 for Dataset 1.

Effect of inductive biases on RMSE performance

As summarized in Table 17, our inductive bias significantly degrades accuracy in norm prediction. This is because it removes essential information—the global mean—and makes the loss landscape rougher, resulting in challenging optimization.

Inductive biasesRMSE
O0.00924
X0.00647

Table 17. Effect of inductive biases on the performance of norm prediction task. All experiments are conducted using MGN-based SuperMeshNet and Dataset 1.

Further Examples

Below, we additionally provide concrete examples of tasks where the mean is typically less important:

  1. Scientific tasks

  • Vorticity prediction: the vorticity field depends on the local rotational structure, not the mean velocity.

  • Force prediction: forces depend on relative distances, not absolute positions.

  1. General graph learning tasks

  • Node classification: predictions rely mainly on relational structures, not absolute feature scales.

  • Link prediction: predictions depend on similarity between features rather than the absolute feature scale.

These tasks may benefit from our inductive biases.

Comparison with Standard Normalization

Similarly as in our inductive biases, existing normalization techniques such as layer normalization (Ba et al., 2016) and batch normalization [15] also involve mean subtraction. In the context of MPNNs, layer normalization computes the mean across features within each node embedding, whereas batch normalization computes the mean across node embeddings, which is also leveraged in our inductive biases. Thus, rather than layer normalization (Ba et al., 2016), our inductive biases are closer to batch normalization [15]. Standard normalization techniques typically include additional operations such as scaling by the standard deviation and the use of learnable shift and scale parameters. The results in Table 18 demonstrate that, without these additional components, SuperMeshNet consistently outperforms layer normalization and, in some cases, even surpasses batch normalization. This suggests that centering alone, namely mean subtraction, acts as a crucial inductive bias for the super-resolution task.

DatasetMethodsRMSE
1Layer normalization (Ba et al., 2016)0.0310 ± 0.0014
1Batch normalization [15]0.0165 ± 0.0009
1SuperMeshNet0.0226 ± 0.0007
2Layer normalization (Ba et al., 2016)0.0587 ± 0.0008
2Batch normalization [15]0.0508 ± 0.0007
2SuperMeshNet0.0507 ± 0.0011
3Layer normalization (Ba et al., 2016)0.0297 ± 0.0013
3Batch normalization [15]0.0250 ± 0.0003
3SuperMeshNet0.0245 ± 0.0005

Table 18. Comparison of our inductive biases in SuperMeshNet with layer normalization (Ba et al., 2016) and batch normalization [15]. All models are based on the MGN architecture and trained using complementary learning with Nh=20N_h = 20 and N=200N = 200. The best-performing result in each case is highlighted in bold.

Although our inductive biases are rather simple and closely related to conventional normalization methods, they offer several key insights. First, computing the mean across nodes (as in batch normalization and SuperMeshNet) is proven to be more effective than computing the mean within each node (as in layer normalization), likely because super-resolution benefits more from local deviations from the global mean. Second, the additional components of standard normalization, such as division by the standard deviation and learnable affine parameters, do not necessarily yield performance gains. This underscores that centering itself is the most essential component for super-resolution tasks.

Comparison with Full Supervision

RMSE

In Table 19, we present the full version of Table 4 from the main manuscript, now including the standard deviation of RMSE values. In some cases, SuperMeshNet (Nh=20N_h=20, N=200N=200) yields slightly worse RMSEs than the fully supervised baselines (Nh=N=200N_h=N=200). In such cases, Table 20 summarizes additional results using SuperMeshNet with Nh=40N_h=40 and N=200N=200, where NhN_h is set to a slightly larger value than our default setting (i.e., Nh=20N_h=20) but still significantly smaller than 200, in comparison with the fully supervised (Nh=N=200N_h=N=200) baselines. These results demonstrate that using only 20% of HR data samples (i.e., Nh=40N_h=40) in SuperMeshNet is sufficient to outperform the fully supervised baseline.

Method (Nh,N)(N_h, N)GCNSAGEGATGTRGINMGN
Fully supervised (20, 20)0.0874 ± 0.00390.0876 ± 0.00150.0826 ± 0.00420.0758 ± 0.00680.0819 ± 0.00470.0655 ± 0.0030
Fully supervised (200, 200)0.0575 ± 0.00350.0544 ± 0.00250.0512 ± 0.00160.0450 ± 0.00230.0381 ± 0.00270.0228 ± 0.0015
SuperMeshNet-O (20, 200)0.0613 ± 0.00200.0589 ± 0.00210.0544 ± 0.00080.0451 ± 0.00200.0404 ± 0.00280.0269 ± 0.0019
SuperMeshNet (20, 200)0.0431 ± 0.00090.0450 ± 0.00100.0457 ± 0.00160.0385 ± 0.00290.0277 ± 0.00060.0226 ± 0.0007
Fully supervised (20, 20)0.0972 ± 0.00820.1025 ± 0.00520.0983 ± 0.00260.0983 ± 0.00160.0775 ± 0.00730.0730 ± 0.0075
Fully supervised (200, 200)0.0624 ± 0.00220.0633 ± 0.00320.0637 ± 0.00130.0572 ± 0.00160.0534 ± 0.00090.0461 ± 0.0004
SuperMeshNet-O (20, 200)0.0636 ± 0.00130.0664 ± 0.00320.0680 ± 0.00230.0631 ± 0.00180.0569 ± 0.00230.0514 ± 0.0003
SuperMeshNet (20, 200)0.0574 ± 0.00030.0624 ± 0.00060.0634 ± 0.00180.0600 ± 0.00120.0537 ± 0.00150.0507 ± 0.0011
Fully supervised (20, 20)0.0587 ± 0.00380.0611 ± 0.00430.0616 ± 0.00500.0513 ± 0.00520.0569 ± 0.00160.0523 ± 0.0055
Fully supervised (200, 200)0.0370 ± 0.00290.0340 ± 0.00150.0374 ± 0.00120.0329 ± 0.00210.0317 ± 0.00220.0243 ± 0.0017
SuperMeshNet-O (20, 200)0.0380 ± 0.00180.0366 ± 0.00210.0375 ± 0.00090.0363 ± 0.00230.0316 ± 0.00100.0281 ± 0.0006
SuperMeshNet (20, 200)0.0297 ± 0.00080.0297 ± 0.00140.0310 ± 0.00120.0294 ± 0.00110.0258 ± 0.00080.0245 ± 0.0005

Table 19. The RMSE of SuperMeshNet (with inductive biases) and SuperMeshNet-O (without inductive biases) trained with Nh=20N_h=20 HR data samples and N=200N=200 LR data samples across six MPNNs and three datasets, in comparison with two fully supervised MPNNs including 1) Nh=N=20N_h=N=20 and 2) Nh=N=200N_h=N=200. The best performer is highlighted in bold.

DatasetMethodNhN_hNNMPNNRMSE
2SuperMeshNet40200GTR0.0568 ± 0.0015
2Fully supervised200200GTR0.0572 ± 0.0016
2SuperMeshNet40200GIN0.0501 ± 0.0004
2Fully supervised200200GIN0.0534 ± 0.0009
2SuperMeshNet40200MGN0.0461 ± 0.0011
2Fully supervised200200MGN0.0461 ± 0.0004
3SuperMeshNet40200MGN0.0225 ± 0.0003
3Fully supervised200200MGN0.0243 ± 0.0017

Table 20. The RMSE and its standard deviation of SuperMeshNet trained with Nh=40N_h=40 HR data samples and N=200N=200 LR data samples, in comparison with a fully supervised MPNN trained with Nh=N=200N_h=N=200. Experiments are conducted only for cases where using Nh=20N_h=20 is insufficient to outperform the fully supervised baseline.

Physics-aware evaluation

We compare SuperMeshNet with a fully supervised baseline by evaluating prediction within regions of interest, rather than averaging over the entire domain. These regions correspond to areas where stress and electric field are highly concentrated. For Dataset 1, the high-concentration regions are defined as circles of radius 0.025 located to the left and right of the bottom-right hole. For Dataset 2, the region is defined as a circle of radius 0.035 located to the right of the bottom-right hole. For Dataset 3, the regions are defined as circles of radius 0.026 located to the left of the bottom-right hole and to the right of the bottom-left hole. As shown in Table 21, SuperMeshNet predicts stress and electric field in these concentrated regions with less than 5% relative error, while outperforming the fully supervised baseline despite using fewer HR samples.

MethodsDataset 1Dataset 2Dataset 3
Stress (relative error)Stress (relative error)Electric field (relative error)
Ground truth HR data1.65×1081.65 \times10^{8}2.00×1082.00 \times10^{8}47.4
SuperMeshNet (Nh=40N_h = 40, N=200N = 200)1.70×1081.70 \times10^{8} (0.0255)1.98×1081.98 \times10^{8} (0.0100)45.4 (0.0429)
Fully supervised (N=Nh=200N = N_h = 200)1.72×1081.72 \times10^{8} (0.0382)1.95×1081.95 \times10^{8} (0.0223)43.8 (0.0770)

Table 21. Comparison between SuperMeshNet, trained with Nh=40N_h = 40 high-resolution (HR) samples and N=200N = 200 low-resolution (LR) samples, and a fully supervised MPNN trained with Nh=N=200N_h = N = 200, in terms of stress prediction (Datasets 1 and 2) and electric field prediction in highly concentrated regions.

Computational Cost

Training Cost

SuperMeshNet alleviates the reliance on expensive HR data samples but incurs longer training time compared to the fully supervised baseline. Figure 26–Figure 28 display how the training time increases and data generation time decreases—resulting from the use of SuperMeshNet (Nh=20N_h = 20, N=200N = 200)—scale with mesh size, relative to fully supervised learning (Nh=N=200N_h = N = 200). To estimate the data generation time decrease, we measure the time required to generate 180 HR data samples by solving the PDE using a direct solver on an Intel(R) Core(TM) i7-9700K CPU @ 3.60GHz. For the training time increase, we compute the difference between the training times of MGN under fully supervised learning and SuperMeshNet, using an Intel (R) Core (TM) i9-10920X [email protected] GHz and an NVIDIA RTX A6000 GPU. The slopes in each figure are computed using the least square method. Comparison of the two slopes, which characterize how training time increases and data generation time decreases with mesh size, respectively, consistently reveals that, on all three datasets, data generation time grows more rapidly as mesh size decreases.

Training time increase and data generation time decrease for Dataset 1

Figure 26. Training time increase (left) and data generation time decrease (right), resulting from the use of SuperMeshNet (Nh=20N_h = 20, N=200N = 200), relative to fully supervised learning (Nh=N=200N_h = N = 200) on Dataset 1 and its mesh-size variants. All experiments use MGN as the underlying MPNN architecture.

Two-panel plot showing log(training time increase) versus -log(mesh size) with slope=0.7646 on the left, and log(-data generation time) versus -log(mesh size) with slope=1.0105 on the right

Figure 28. Training time increase (left) and data generation time decrease (right), resulting from the use of SuperMeshNet (Nh=20N_h = 20, N=200N = 200), relative to fully supervised learning (Nh=N=200N_h = N = 200) on Dataset 3 and its mesh-size variants. All experiments use MGN as the underlying MPNN architecture.

Training time increase and data generation time decrease for Dataset 2

Figure 27. Training time increase (left) and data generation time decrease (right), resulting from the use of SuperMeshNet (Nh=20N_h = 20, N=200N = 200), relative to fully supervised learning (Nh=N=200N_h = N = 200) on Dataset 2 and its mesh-size variants. All experiments use MGN as the underlying MPNN architecture.

Inference Cost

Figure 29 demonstrates that SuperMeshNet substantially reduces the computational cost of HR simulation. Specifically, the combined computational cost of LR simulation and subsequent SuperMeshNet inference is significantly lower than the cost of HR simulation. The time saving becomes even more pronounced for small mesh sizes, where the computational cost required for HR simulation grows rapidly.

Comparison of HR simulation time with the combined computational cost of LR simulation and subsequent SuperMeshNet inference

Figure 29. Comparison of the HR simulation time with the combined computational cost of LR simulation and subsequent SuperMeshNet’s inference. Each value is measured over 1,000 data samples.

Scalability

Mesh Size and Magnification Ratio

In Table 22, we evaluate the scalability of our framework SuperMeshNet by measuring the training time and RMSE as a function of mesh size (while fixing the magnification ratio). These experiments are conducted using MGN, trained with 40 HR and 200 LR data samples from Dataset 1. According to the results, the training time increases moderately as the mesh becomes finer. The RMSE tends to increase and then decrease again as the mesh becomes extremely fine. This is because we fix the magnification ratio (the ratio between the LR mesh size and the HR mesh size) while varying both the LR mesh size and the HR mesh size. When the mesh becomes finer, the HR data samples contain more detailed features, making the LR-to-HR transformation more challenging. However, once the mesh resolution exceeds the characteristic length scale of the domain geometry, no additional details can be represented in the HR mesh. At the same time, the LR data samples already capture most of the meaningful features, and the LR-to-HR transformation becomes easier again. For comparison, we provide Table 23, presenting experimental results where the LR mesh size is fixed while the magnification ratio (the HR mesh size) varies. In this case, the RMSE tends to monotonically increase as the HR mesh becomes finer since the LR mesh size is fixed.

Smallest mesh sizeTraining time (s)RMSE
0.016383.280.0112
0.008443.340.0158
0.004587.500.0194
0.002710.960.0121

Table 22. Training time (in seconds) and RMSE of MGN-based SuperMeshNet trained with 40 HR and 200 LR data samples from Dataset 1 as a function of mesh size while fixing the magnification ratio.

Magnification ratioRMSE
2×22 \times20.0099
4×44 \times40.0112
8×88 \times80.0689
16×1616 \times160.0796

Table 23. RMSE of MGN-based SuperMeshNet trained with 40 HR and 200 LR data samples from Dataset 1 as a function of magnification ratio (HR mesh size) while fixing the LR mesh size.

Large Dataset

We further validate scalability on BlastNet 2.0 [9], which contains approximately 2 million nodes. We conduct experiments on two subsets (Forced Hit and Parametric Variation) of BlastNet 2.0, which include regular grid-based multi-physics (reaction–flow) and turbulent flow simulations. For the experiments, we set the hidden dimension to 10 and use two MPNN layers for both the LR processor and the HR processor. The results in Table 24 verify the scalability of SuperMeshNet on large-scale datasets with over one million nodes.

Sub-datasetFully supervised (Nh=N=50N_h=N=50)SuperMeshNet (Nh=10 N=50N_h=10\ N=50)
Forced hit0.04510.0230
Parametric variation0.07850.0643

Table 24. RMSE comparison between a fully supervised baseline and SuperMeshNet on two subsets of BlastNet 2.0 [9].

Comparison with Benchmark Semi-Supervised Regression Methods

ACCURACY AND TRAINING TIME

Table 25 presents the full version of Table 5 from the main manuscript, now including the mean and standard deviation of RMSE and training time for all datasets under the setting of Nh=20N_h = 20 HR data samples and N=200N = 200 LR data samples. SuperMeshNet consistently results in the shortest training time across all datasets, in comparison with benchmark semi-supervised regression methods. It also achieves the lowest RMSE on Datasets 1 and 3. For Dataset 2, while SuperMeshNet yields a slightly higher RMSE, it significantly reduces training time compared to all the benchmarks. Overall, SuperMeshNet is shown to reveal strong potential in terms of both predictive accuracy and training efficiency. In Table 26, we additionally report that, for Dataset 2 with Nh=40N_h = 40 and N=200N = 200, SuperMeshNet still exhibits both the lowest RMSE and the shortest training time among all evaluated methods.

DatasetMethodsRMSETraining time (s)
2Mean-Teacher (Tarvainen & Valpola, [32])0.0474 ± 0.0007427 ± 37
2TNNR (Wetzel et al., [34])0.0807 ± 0.0074409 ± 143
2UCVME (Dai et al., [10])0.0474 ± 0.0013780 ± 89
2SuperMeshNet-O0.0479 ± 0.0005348 ± 15
2SuperMeshNet0.0461 ± 0.0011292 ± 25

Table 26. Comparison with benchmark semi-supervised regression methods in terms of the RMSE and training time (in seconds). Here, MGN is employed as an MPNN architecture for each method. Training is conducted when Nh=40N_h = 40 and N=200N = 200 for each dataset. The best value in each metric is highlighted in bold.

DatasetMethodsRMSETraining time (s)
1Mean-Teacher (Tarvainen & Valpola, [32])0.0325 ± 0.0016694 ± 50
1TNNR (Wetzel et al., [34])0.0624 ± 0.0202477 ± 333
1UCVME (Dai et al., [10])0.0293 ± 0.00121123 ± 102
1SuperMeshNet-O0.0269 ± 0.0019503 ± 64
1SuperMeshNet0.0226 ± 0.0007421 ± 93
2Mean-Teacher (Tarvainen & Valpola, [32])0.0499 ± 0.0007402 ± 28
2TNNR (Wetzel et al., [34])0.0823 ± 0.0055350 ± 90
2UCVME (Dai et al., [10])0.0484 ± 0.0006739 ± 59
2SuperMeshNet-O0.0514 ± 0.0003306 ± 16
2SuperMeshNet0.0507 ± 0.0011250 ± 9
3Mean-Teacher (Tarvainen & Valpola, [32])0.0270 ± 0.0003446 ± 33
3TNNR (Wetzel et al., [34])0.0393 ± 0.0027437 ± 83
3UCVME (Dai et al., [10])0.0281 ± 0.0013700 ± 89
3SuperMeshNet-O0.0281 ± 0.0006345 ± 10
3SuperMeshNet0.0245 ± 0.0005286 ± 18

Table 25. Comparison with benchmark semi-supervised regression methods in terms of the RMSE and training time (in seconds). Here, MGN is employed as an MPNN architecture for each method. Training is conducted when Nh=20N_h = 20 and N=200N = 200 for each dataset. The best value in each metric is highlighted in bold.

Prediction Diversity

The results in Table 27 demonstrate that SuperMeshNet exhibits consistently greater prediction diversity across all datasets, compared to the case of UCVME [10]. The prediction diversity is quantified as the root mean square of the difference between two predictions made by MGN model pairs, each trained with Nh=20N_h = 20 and N=200N = 200 on each dataset. This implies that our architectural design for complementary learning apparently promotes greater diversity in the learning process than the case of UCVME.

DatasetMethodsPrediction diversity
1UCVME [10]0.0017 ± 0.0001
1SuperMeshNet0.0200 ± 0.0014
2UCVME [10]0.0024 ± 0.0002
2SuperMeshNet0.0514 ± 0.0011
3UCVME [10]0.0014 ± 0.0001
3SuperMeshNet0.0294 ± 0.0004

Table 27. Comparison of prediction diversity between UCVME and SuperMeshNet. The prediction diversity is quantified as the root mean square of the difference between two predictions made by MGN model pairs trained by each method with Nh=20N_h=20 and N=200N=200 on each dataset. The dropout probability in UCVME is set to 0.1.

Comparison with a Super-Resolution Competitor

Although the primary objective of SuperMeshNet is to improve super-resolution performance across a wide range of MPNNs rather than to outperform a specific state-of-the-art method, we compare a special case of SuperMeshNet using MGN with the most recent and relevant benchmarks, SRGNN [5], to further validate its effectiveness. The results in Table 28 signify that SuperMeshNet, even when trained with only 20 HR data samples, outperforms SRGNN [5] trained with 200 HR data samples, underscoring its superior data efficiency.

DatasetMethodsNhN_hRMSE
1SuperMeshNet50.0447±0.00100.0447 \pm0.0010
1SuperMeshNet100.0280±0.00130.0280 \pm0.0013
1SuperMeshNet200.0226±0.00070.0226 \pm0.0007
1SuperMeshNet400.0191±0.00210.0191 \pm0.0021
1SRGNN [5]2000.0247±0.00130.0247 \pm0.0013
2SuperMeshNet50.0723±0.00180.0723 \pm0.0018
2SuperMeshNet100.0645±0.00220.0645 \pm0.0022
2SuperMeshNet200.0507±0.00110.0507 \pm0.0011
2SuperMeshNet400.0461±0.00110.0461 \pm0.0011
2SRGNN [5]2000.0487±0.00110.0487 \pm0.0011
3SuperMeshNet50.0353±0.00150.0353 \pm0.0015
3SuperMeshNet100.0294±0.00100.0294 \pm0.0010
3SuperMeshNet200.0245±0.00050.0245 \pm0.0005
3SuperMeshNet400.0225±0.00030.0225 \pm0.0003
3SRGNN [5]2000.0254±0.00110.0254 \pm0.0011

Table 28. The RMSE and its standard deviation of MGN-based SuperMeshNet trained with varying numbers of HR data samples NhN_h and a fixed N=200N=200 LR data samples in comparison with SRGNN with full supervision (N=Nh=200N=N_h=200).

Ablation Studies on Inductive Biases

Table 29 presents the full version of Table 6 from the main manuscript, now including the mean and standard deviation of RMSE values for all datasets. On Dataset 2, the combination of both inductive biases (N+M) occasionally results in a higher RMSE than using a single inductive bias (N). To further investigate this, Table 30 reports additional results on Dataset 2 using SuperMeshNet with Nh=80N_h = 80 and N=200N = 200, where NhN_h is set to a slightly larger value than our default setting (i.e., Nh=20N_h = 20) yet still significantly smaller than 200. With the larger NhN_h, the incorporation of both inductive biases consistently yields the lowest RMSE among the four inductive bias settings. These findings reveal the existence of a dataset-dependent threshold for NhN_h above which leveraging both inductive biases becomes beneficial.

DatasetMPNNONMN + M
1GCN0.0613 ± 0.00200.0431 ± 0.0009--
1SAGE0.0589 ± 0.00210.0493 ± 0.00240.0528 ± 0.00180.0450 ± 0.0010
1GAT0.0544 ± 0.00080.0457 ± 0.0016--
1GTR0.0451 ± 0.00200.0405 ± 0.00250.0438 ± 0.00100.0385 ± 0.0029
1GIN0.0404 ± 0.00280.0290 ± 0.00260.0281 ± 0.00150.0277 ± 0.0006
1MGN0.0269 ± 0.00190.0237 ± 0.00100.0247 ± 0.00140.0226 ± 0.0007
2GCN0.0636 ± 0.00130.0574 ± 0.0003--
2SAGE0.0664 ± 0.00320.0623 ± 0.00050.0652 ± 0.00090.0624 ± 0.0006
2GAT0.0680 ± 0.00230.0634 ± 0.0018--
2GTR0.0631 ± 0.00180.0607 ± 0.00090.0617 ± 0.00250.0600 ± 0.0012
2GIN0.0569 ± 0.00230.0523 ± 0.00090.0549 ± 0.00080.0537 ± 0.0015
2MGN0.0514 ± 0.00030.0488 ± 0.00130.0509 ± 0.00040.0507 ± 0.0011
3GCN0.0380 ± 0.00180.0297 ± 0.0008--
3SAGE0.0366 ± 0.00210.0309 ± 0.00180.0346 ± 0.00180.0297 ± 0.0014
3GAT0.0375 ± 0.00090.0310 ± 0.0012--
3GTR0.0363 ± 0.00230.0312 ± 0.00080.0327 ± 0.00060.0294 ± 0.0011
3GIN0.0316 ± 0.00100.0261 ± 0.00020.0268 ± 0.00070.0258 ± 0.0008
3MGN0.0281 ± 0.00060.0245 ± 0.00030.0246 ± 0.00060.0245 ± 0.0005

Table 29. Ablation studies on inductive biases. The RMSE of SuperMeshNet across six MPNNs under four inductive bias conditions (O: without inductive biases, N: node-level centering, M: message-level centering, and N+M: both node-level and message-level centering operations) trained with Nh=20N_h = 20 and N=200N = 200 for each dataset is compared. For each MPNN, the lowest RMSE value among the four inductive bias conditions is highlighted in bold.

DatasetMPNNONMN + M
2GCN0.0630 ± 0.00170.0556 ± 0.0006--
2SAGE0.0626 ± 0.00230.0606 ± 0.00160.0628 ± 0.00260.0606 ± 0.0011
2GAT0.0637 ± 0.00050.0606 ± 0.0016--
2GTR0.0606 ± 0.00110.0560 ± 0.00100.0574 ± 0.00120.0554 ± 0.0013
2GIN0.0528 ± 0.00110.0481 ± 0.00070.0470 ± 0.00060.0468 ± 0.0006
2MGN0.0463 ± 0.00050.0447 ± 0.00090.0448 ± 0.00110.0432 ± 0.0003

Table 30. Additional ablation studies on inductive biases. The RMSE of SuperMeshNet across six MPNNs under four inductive bias conditions (O: without inductive biases, N: node-level centering, M: message-level centering, and N+M: both node-level and message-level centering operations) trained with Nh=80N_h = 80 and N=200N = 200 for Dataset 2 is compared. For each MPNN, the lowest RMSE value among the four inductive bias conditions is highlighted in bold.

Ablation Study on Core Components

To disentangle the individual contributions of complementary learning (CL) and inductive biases (IB), we evaluate the performance of all four combinations (no CL + no IB, CL only, IB only, and CL + IB) under the same data configuration (i.e., Nh=20N_h = 20, N=200N = 200), while also reporting the standard fully supervised baseline trained only on 20 paired samples (Nh=N=20N_h = N = 20) for reference. For comparison, the fully supervised baseline is assumed to use all 200 paired HR–LR samples (i.e., Nh=N=200N_h = N = 200).

The results in Table 31 demonstrate that:

MethodsIBCLNhN_hNNRMSE
No CL + no IB (20,20)XX20200.0655
No CL + no IB (20, 200)XX202000.0454
CL onlyXO202000.0269
IB onlyOX202000.0371
CL + IB (SuperMeshNet)OO202000.0226
Baseline: fully supervisedXX2002000.0228

Table 31. Ablation study on complementary learning (CL) and inductive biases (IB). All experiments use the MGN-based SuperMeshNet and Dataset 1.

  • both CL and IB indeed contribute to improved performance relative to the case of no CL + no IB (20, 200).

  • Both components are necessary for SuperMeshNet to outperform the fully supervised baseline trained with 180 additional HR samples.

In particular, in regard to the settings where CL is not used ((i) no CL + no IB (20, 200) and (ii) IB only), we pretrain models to reconstruct LR data from noised LR data (via denoising reconstruction), and fine-tune the models on the 20 paired LR–HR samples. This approach provides a stronger control than the case of no CL + no IB (20, 20) involving training only on the 20 paired samples, as reflected in the lower RMSE.

Application to Real-World Geometry and Time-Dependent PDE.

Qualitative Analysis of Results on Time-Dependent PDE Dataset 2

The visualization in Figure 30 shows that SuperMeshNet successfully predicts the HR counterpart even when the HR data differ substantially from the LR input, whereas conventional fully supervised learning fails despite having access to more HR data. Due to the relatively large grid size of the time-dependent PDE dataset 2, we increased the hidden dimension to 150, whereas the default setting is 30. In addition, because the dataset is defined on a regular grid, we used standard bilinear interpolation instead of kkNN interpolation.

LR input, ground truth HR data, and two vorticity predictions produced by SuperMeshNet and full supervision over multiple timestamps

Figure 30. LR input, ground truth HR data, and two vorticity predictions produced by SuperMeshNet and full supervision for the time-dependent PDE dataset 2 over multiple timestamps. Here, NhN_h and NN represent the number of HR and LR data samples, respectively. For all cases, MGN is utilized as the underlying MPNN.

Quantitative Analysis

We further quantitatively evaluate the applicability of SuperMeshNet to CFD datasets, including one real-world geometry dataset and two time-dependent PDE datasets. The results in Table 32 show that SuperMeshNet, trained with only 40 HR data samples, achieves even a lower error than the case of the fully supervised model trained with 200 HR data samples, demonstrating its effectiveness on real-world geometry and time-dependent PDE datasets.

DatasetMethodsNhN_hNNRMSE
Real-world geometrySuperMeshNet402000.0559
Real-world geometryfully supervised2002000.0584
Time-dependent PDE 1SuperMeshNet402000.0700
Time-dependent PDE 1fully supervised2002000.0793
Time-dependent PDE 2SuperMeshNet402000.0543
Time-dependent PDE 2fully supervised2002000.1613

Table 32. The RMSE of MGN-based SuperMeshNet trained with Nh=40N_h = 40 HR data samples and N=200N = 200 LR data samples for the real-world geometry and time-dependent PDE datasets, in comparison with fully supervised MGNs trained with Nh=N=200N_h = N = 200. The best-performing result in each case is highlighted in bold.

Analysis on HR Data Sampling Strategies

In this section, we empirically analyze how HR data sampling strategies affect performance in terms of RMSE.

Impact of Different 20 HR Subsets

Table 33 shows that using different subsets of 20 HR samples (i.e., Nh=20N_h = 20) introduces a moderate variance: the RMSE standard deviation (STD) is about 6.7% of the mean RMSE and is roughly twice as large as in the case of using the same 20 HR samples. In the same-samples setting, all five training runs use the same 20 HR samples drawn once from a uniform distribution, whereas, in the different-samples setting, each run uses a distinct set of 20 HR samples independently drawn from the same distribution. For both settings, we measure the RMSE from each training run and report the mean and STD across the five runs. These results indicate that, under a fixed sampling strategy, the variability induced by using different subsets of 20 HR samples remains moderate.

StrategyRMSE meanRMSE STD
Same random samples0.02260.0007
Different random samples0.02080.0014

Table 33. Effect of different HR subsets on the mean and STD of RMSE for MGN-based SuperMeshNet trained with Nh=20N_h = 20 and N=200N = 200 on Dataset 1.

Impact of Distribution Mismatch

We further examine how the distributional mismatch between two cases using all 200 LR samples and the 20 LR samples corresponding to selected HR samples affects performance. The mismatch is quantified using maximum mean discrepancy (MMD). As summarized in Table 34, MMD strongly correlates with RMSE: subsets that better match the overall LR distribution yield lower RMSE.

StrategyRMSEMMD
Same random samples0.02260.147
MMD-minimizing greedy sampling0.01850.058
MMD-maximizing greedy sampling0.02590.330

Table 34. Impact of distribution mismatch-based HR data selection strategies on RMSE for MGN-based SuperMeshNet trained with Nh=20N_h = 20 and N=200N = 200 on Dataset 1.

To study this, we compare three sampling strategies. MMD-minimizing greedy sampling (kernel herding [8]) begins by randomly selecting one HR sample from the pool of 200 candidates and then iteratively chooses the next HR sample whose corresponding LR counterpart most reduces the MMD between the selected subset and the full LR distribution. This procedure ensures that the selected HR samples closely represent the overall dataset. In contrast, MMD-maximizing greedy sampling intentionally selects the next point that most increases the MMD, pushing the subset away from the full distribution. As demonstrated below, the selected sampling strategy has a significant impact on the RMSE of SuperMeshNet.

Impact of Active Learning and Hybrid Strategy

We investigate an inconsistency-based active learning strategy that selects HR samples based on the discrepancy between pseudo-labels generated by the main model FθF_{\theta} and the auxiliary model GϕG_{\phi}. Training is initialized with 10 HR samples selected using an MMD-minimizing strategy at the first epoch. Subsequently, one HR sample is added at each epoch by selecting the sample with the highest inconsistency loss, until the total number of HR samples reaches 20. We further evaluate a hybrid strategy that combines the inconsistency-based active learning with MMD-based selection. Specifically, when selecting a next HR sample during active learning, the top 10 candidates are first selected to minimize MMD, and the most inconsistent sample is then chosen from this subset.

As shown in Table 35, inconsistency-based active learning improves performance over random sampling. Moreover, the hybrid strategy achieves the lowest RMSE, indicating that combining inconsistency-based selection with distribution-aware sampling leads to more effective HR data acquisition.

StrategyRMSE
Same random samples0.0226
MMD-minimizing greedy sampling0.0185
Inconsistency-based active learning0.0206
Hybrid(inconsistency-based active learning + MMD-minimizing sampling)0.0174

Table 35. Impact of inconsistency-based active learning and MMD-based HR sample selection on RMSE for MGN-based SuperMeshNet, trained with Nh=20N_h = 20 and N=200N = 200 on Dataset 1.

Training Stability

To assess training stability, we plot the loss curves over epochs for five different random seeds. As shown in Figure 31, the loss consistently decreases in all runs without exhibiting divergence.

Loss curves for five different random seeds

Figure 31. Loss curves for five different random seeds, each shown in a different color. For all cases, the MGN-based SuperMeshNet is trained using Nh=20N_h = 20 and N=200N = 200 from Dataset 1.

Although a pseudo-label generated by one model may contain potential errors, such errors do not lead to catastrophic mutual reinforcement. This is because every loss computation judiciously incorporates both pseudo-label-based mutual supervision and ground truth–based supervision. The presence of true labels at every iteration constrains error propagation, preventing any amplification caused by imperfect pseudo-labels. Consequently, complementary learning in our SuperMeshNet framework maintains stable optimization dynamics.

To further investigate training stability, we conduct controlled experiments by injecting errors into pseudo-labels. As illustrated in Figure 32(a), error amplification emerges once the pseudo-label error exceeds a threshold of 0.02. This phenomenon becomes substantially more pronounced when training relies solely on unsupervised loss without access to true labels, as shown in Figure 32(b). These observations highlight the critical role of supervised loss in stabilizing training, as it anchors predictions to ground truth at each iteration and mitigates uncontrolled error propagation.

Stability analysis via controlled pseudo-label perturbations

Figure 32. Stability analysis via controlled pseudo-label perturbations. For all cases, the MGN-based SuperMeshNet is trained using Nh=20N_h = 20 and N=200N = 200 from Dataset 1.

Effect of pseudo-label error magnitude on training stability

Figure 32a. Effect of pseudo-label error magnitude on training stability.

Effect of supervised loss on training stability

Figure 32b. Effect of supervised loss on training stability.

Analysis on kkNN Interpolation

Relative Computational Cost of kkNN Interpolation

We acknowledge that kkNN interpolation introduces additional overhead, particularly for very small mesh sizes. As presented in Table 36, the relative interpolation cost indeed grows as the task difficulty increases. However, even after including kkNN interpolation, the total inference cost remains far lower than the cost of running the HR simulation for all mesh sizes, as indicated in Figure 29. Thus, while kkNN interpolation may act as a localized bottleneck (particularly for very small mesh sizes), it does not diminish the substantial overall speed-up achieved by SuperMeshNet. We note that our framework is not tied to kkNN and can leverage more efficient interpolation (e.g., bilinear interpolation) on structured grid datasets. Developing a lighter interpolation method for irregular mesh datasets remains an important direction for future work.

Mesh size(kkNN interpolation time) / (total inference time)
817.0%
418.6%
229.4%
171.2%

Table 36. Relative overhead of kkNN interpolation across different mesh sizes. All experiments are conducted using the MGN-based SuperMeshNet trained with Nh=20N_h = 20 and N=200N = 200 on Dataset 1.

Singular Point Analysis

To evaluate the effect of geometric singularity on super-resolution performance, we additionally generate a dataset that explicitly includes a singular point. Representative HR and LR samples from this dataset are shown in Figure 33.

Figure 33. Examples of LR and HR data samples from the dataset including a singular point.

The computational domain is defined by six vertices: (0,−0.25)(0, -0.25), (−0.25,−0.25)(-0.25, -0.25), (−0.25,0.25)(-0.25, 0.25), (0.25,0.25)(0.25, 0.25), (0.25,0)(0.25,0), and (x0,y0)(x_0, y_0), where (x0,y0)(x_0, y_0) denotes the location of the singular point. Both x0x_0 and y0y_0 are independently sampled from a uniform distribution on [−0.125,0.125][-0.125, 0.125]. The default mesh sizes for LR and HR meshes are set to 0.04 and 0.01, respectively. However, within the vicinity of the singular point, the mesh is refined by a factor of four for each resolution.

On this domain, the Poisson equation below is solved:

∇2u=((x−x0)2+(y−y0)2)−1,(27)\nabla^2 u = \left((x-x_0)^2 + (y-y_0)^2\right)^{-1}, \tag*{(27)}

where uu denotes the electrical potential. The boundary condition is fixed at 0. After solving the equation, the magnitude of the electric field is calculated at each node of the mesh. Examples of LR and HR fields are visualized in Figure 33.

We compare the HR fields predicted by SuperMeshNet with those obtained via pure kkNN interpolation without neural network-based prediction in Figure 34. Both methods exhibit relatively high errors at the singular point. Nevertheless, the MPNN in SuperMeshNet noticeably mitigates the severity of these errors: both the magnitude and spatial extent of the high error region are reduced compared to the case of pure interpolation. Specifically, the maximum error near the singularity is 4.18 for pure kkNN interpolation, whereas SuperMeshNet reduces this peak error to 1.57.

Figure 34. Comparison of predictions by SuperMeshNet and pure kkNN interpolation.

Comparison with a PINN

As summarized in Table 37, we compare a physics-informed neural network (PINN) [25] and SuperMeshNet in terms of computational time and RMSE performance.

MethodsData preparation time + training time (s)Inference time (s)RMSE
PINN5.80.00038.0×10−48.0 \times10^{-4}
SuperMeshNet16467.30.00611.0×10−41.0 \times10^{-4}

Table 37. Comparison between PINN and MGN-based SuperMeshNet.

First, we describe the experimental setting. We conduct evaluations on the Poisson equation:

∇2u=x2+y2,\nabla^{2}u = \sqrt{x^{2} + y^{2}},

where uu is the solution and (x,y)(x,y) denotes the spatial coordinate. The computational domain is an L-shaped region with six vertices:

(0,−0.25), (−0.25,−0.25), (−0.25,0.25), (0.25,0.25), (0.25,0), (0,0),(0,-0.25),\ (-0.25,-0.25),\ (-0.25,0.25),\ (0.25,0.25),\ (0.25,0),\ (0,0),

and homogeneous Dirichlet boundary conditions (u=0u = 0) are applied on all boundaries.

For training SuperMeshNet, we construct a dataset consisting of 20 HR and 200 LR samples. The domain is the same L-shape, except that the final vertex is replaced by a variable point (x0,y0)(x_0,y_0), where both x0x_0 and y0y_0 are independently sampled from a uniform distribution on [−0.125,0.125][-0.125, 0.125]. The default sizes for LR and HR meshes are set to 0.04 and 0.01, respectively; however, within the vicinity of the singular point, the mesh is refined by a factor of four for each resolution. On this domain, to generate LR and HR data, we solve the following Poisson equation:

∇2u=(x−x0)2+(y−y0)2.\nabla^{2}u = \sqrt{(x-x_0)^{2} + (y-y_0)^{2}}.

Next, we show experimental results, which exhibit a clear trade-off: SuperMeshNet achieves substantially lower RMSE, while PINN attains much lower total computation time per instance.

However, these results are not conclusive. Once SuperMeshNet is trained, it can be applied to any new instance without retraining, whereas a PINN must be retrained for every new instance. Thus, when a large number of parameterized PDE instances must be solved, the initial training cost of SuperMeshNet becomes amortized. Likewise, the RMSE of a PINN can potentially be improved by recent advanced optimization techniques. A more extensive comparison is left for future work.

References

  1. [2]Arora, R. PhySRNet: Physics informed super-resolution network for application in computational solid mechanics. In 2022 IEEE/ACM International Workshop on Artificial Intelligence and Machine Learning for Scientific Applications (AI4S), pp. 13–18, 2022. doi: 10.1109/AI4S56813.2022.00008.DOI
  2. [5]Barwey, S., Pal, P., Patel, S., Balin, R., Lusch, B., Vishwanath, V., Maulik, R., and Balakrishnan, R. Mesh-based super-resolution of fluid flows with multiscale graph neural networks, 2024. URL https://arxiv.org/abs/2409.07769.
  3. [6]Boussif, O., Bengio, Y., Benabbou, L., and Assouline, D. MagNet: Mesh agnostic neural PDE solver. In Oh, A. H., Agarwal, A., Belgrave, D., and Cho, K. (eds.), Advances in Neural Information Processing Systems, 2022. URL https://openreview.net/forum?id=bx2roi8hca8.arxiv.org/abs/2210.05495
  4. [7]Brefeld, U., Gärtner, T., Scheffer, T., and Wrobel, S. Efficient co-regularised least squares regression. Proceedings of the 23rd International Conference on Machine Learning, 2006. URL https://api.semanticscholar.org/CorpusID:2025415.DOI
  5. [8]Chen, Y., Welling, M., and Smola, A. Super-samples from kernel handling. In Proceedings of the Twenty-Sixth Conference on Uncertainty in Artificial Intelligence, UAI’10, pp. 109–116, Arlington, Virginia, USA, 2010. AUAI Press. ISBN 9780974903965.
  6. [9]Chung, W. T., Akoush, B., Sharma, P., Tamkin, A., Jung, K. S., Chen, J., Guo, J., Brouzet, D., Talei, M., Savard, B., Poludnenko, A. Y., and Ihme, M. Turbulence in focus: Benchmarking scaling behavior of 3D volumetric super-resolution with BLASTNet 2.0 data. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. URL https://openreview.net/forum?id=ugRnNHKMK95.arxiv.org/abs/2309.13457
  7. [10]Dai, W., Li, X., and Cheng, K.-T. Semi-supervised deep regression with uncertainty consistency and variational model ensembling via Bayesian neural networks. Proceedings of the AAAI Conference on Artificial Intelligence, 37(6):7304–7313, 2023.
  8. [11]de Avila Belbute-Peres, F., Economon, T. D., and Kolter, J. Z. Combining differentiable PDE solvers and graph neural networks for fluid flow prediction. In International Conference on Machine Learning, 2020. URL https://api.semanticscholar.org/CorpusID:220424832.
  9. [12]Guo, Y., Song, J., Cao, X., Zhao, C., and Leng, H. Physics field super-resolution reconstruction via enhanced diffusion model and Fourier neural operator. Theoretical and Applied Mechanics Letters, 15(5):100604, 2025. ISSN 2095-0349. doi: https://doi.org/10.1016/j.taml.2025.100604. URL https://www.sciencedirect.com/science/article/pii/S2095034925000364.
  10. [13]Hamilton, W., Ying, Z., and Leskovec, J. Inductive representation learning on large graphs. Advances in Neural Information Processing Systems, 30, 2017.arxiv.org/abs/1706.02216
  11. [14]Huang, P.-Y., Fu, S.-W., and Tsao, Y. RankUp: Boosting semi-supervised regression with an auxiliary ranking classifier. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=d2lPM1ACzs.arxiv.org/abs/2410.22124
  12. [15]Ioffe, S. and Szegedy, C. Batch normalization: accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32nd International Conference on Machine Learning - Volume , ICML’15, pp. 448–456. JMLR.org, 2015.
  13. [16]Jasak, H. OpenFOAM: Open source CFD in research and industry. International Journal of Naval Architecture and Ocean Engineering, 1(2):89–94, 2009. ISSN 2092-6782. doi: https://doi.org/10.2478/IJNAOE-2013-0011. URL https://www.sciencedirect.com/science/article/pii/S2092678216303879.
  14. [17]Kipf, T. N. and Welling, M. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations (ICLR), 2017.arxiv.org/abs/1609.02907
  15. [18]Kochkov, D., Smith, J. A., Aliev, A., Wang, Q., Brenner, M. P., and Hoyer, S. Machine learning–accelerated computational fluid dynamics. Proceedings of the National Academy of Sciences, 118(21):e2101784118, 2021a. doi: 10.1073/pnas.2101784118. URL https://www.pnas.org/doi/abs/10.1073/pnas.2101784118.DOI
  16. [19]Kochkov, D., Smith, J. A., Alieva, A., Wang, Q., Brenner, M. P., and Hoyer, S. Machine learning–accelerated computational fluid dynamics. Proceedings of the National Academy of Sciences, 118(21), 2021b. ISSN 0027-8424. doi: 10.1073/pnas.2101784118. URL https://www.pnas.org/content/118/21/e2101784118.DOI
  17. [20]Kostopoulos, G., Karlos, S., Kotsiantis, S., Ragos, O., Tiwari, S., Trivedi, M., and Kohle, M. L. Semi-supervised regression: A recent review. J. Intell. Fuzzy Syst., 35(2):1483–1500, January 2018. ISSN 1064-1246. doi: 10.3233/JIFS-169689. URL https://doi.org/10.3233/JIFS-169689.
  18. [22]Obiols-Sales, O., Vishnu, A., Malaya, N. P., and Chandramowlishwaran, A. SURFNet: Super-resolution of turbulent flows with transfer learning using small datasets. In Proceedings of the 30th International Conference on Parallel Architectures and Compilation Techniques, PACT ’21, pp. 331–344. IEEE Press, 2024. ISBN 9781665442787. doi: 10.1109/PACT52795.2021.00031. URL https://doi.org/10.1109/PACT52795.2021.00031.
  19. [23]Pfaff, T., Fortunato, M., Sanchez-Gonzalez, A., and Battaglia, P. Learning mesh-based simulation with graph networks. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=roNqYlO_XP.arxiv.org/abs/2010.03409
  20. [25]Raissi, M., Perdikaris, P., and Karniadakis, G. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686–707, 2019. ISSN 0021-9991. doi: https://doi.org/10.1016/j.jcp.2018.10.045. URL https://www.sciencedirect.com/science/article/pii/S0021999118307125.
  21. [26]Ribeiro, B. A., Ribeiro, J. A., Ahmed, F., Pedonenes, H., Belinha, J., Sarmento, L., Bessa, M. A., and Tavares, S. SimuStruct: Simulated structural plate with holes dataset with machine learning applications. In Workshop on ”Machine Learning for Materials” ICLR 2023, 2023. URL https://openreview.net/forum?id=s3tOuyR1vM7.
  22. [28]Santurkar, S., Tsipras, D., Ilyas, A., and Mądry, A. How does batch normalization help optimization? In Proceedings of the 32nd International Conference on Neural Information Processing Systems, NIPS ’18, pp. 2488–2498, Red Hook, NY, USA, 2018. Curran Associates Inc.arxiv.org/abs/1805.11604
  23. [31]Shi, Y., Huang, Z., Feng, S., Zhong, H., Wang, W., and Sun, Y. Masked label prediction: Unified message passing model for semi-supervised classification, 2021. URL https://openreview.net/forum?id=B9t708KMr9d.arxiv.org/abs/2009.03509
  24. [32]Tarvainen, A. and Valpola, H. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS ’17, pp. 1195–1204, Red Hook, NY, USA, 2017. Curran Associates Inc. ISBN 9781510860964.arxiv.org/abs/1703.01780
  25. [33]Veličković, P., Cucurull, G., Casanova, A., Romero, A., Liò, P., and Bengio, Y. Graph attention networks. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id=rJXMpikCZ.
  26. [34]Wetzel, S. J., Melko, R. G., and Tamblyn, I. Twin neural network regression is a semi-supervised regression algorithm. Machine Learning: Science and Technology, 3(4):045007, oct 2022. doi: 10.1088/2632-2153/ac9885. URL https://dx.doi.org/10.1088/2632-2153/ac9885.
  27. [37]Zhou, Z.-H. and Li, M. Semi-supervised regression with co-training. In Proceedings of the 19th International Joint Conference on Artificial Intelligence, IJCAI’05, pp. 908–913, San Francisco, CA, USA, 2005. Morgan Kaufmann Publishers Inc.

Paper details

vlnc:W_689hu43z
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 8

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