Graph Fourier Neural Kernels: Learning Full PDE Trajectories Across Geometries and Anisotropic Parameter Fields
Authors: Shane E. Loeffler, Zan Ahmad, Shiyi Chen, Syed Yusuf Ali, Carolyna Yamamoto, Kensuke Sakata, Dan M. Popescu, Alana Yee, Yash Lal, Natalia Trayanova, Mauro Maggioni
Organizations: Johns Hopkins University, Department of Biomedical Engineering · Johns Hopkins University, Department of Applied Mathematics and Statistics
High-fidelity simulations of time-dependent nonlinear partial differential equations (PDEs) are expensive to repeat across families of domains and parameter fields, a cost that recurs in design optimization, uncertainty quantification, and computational medicine. Most neural-operator surrogates require each domain to be mapped onto a common grid or reference shape, and few predict the full time course. We introduce Graph Fourier Neural Kernels (G-FuNK), a neural operator for second-order semilinear PDEs with a diffusive principal part that transfers across geometries and anisotropic diffusion tensors without coordinate maps or diffeomorphisms. A weighted graph and its Laplacian supply a spectral basis that encodes geometry and anisotropy. Shared layers act in this basis, at a cost linear in mesh size after one partial eigendecomposition per geometry, and a neural ODE integrator propagates the learned generator to full spatiotemporal trajectories on unseen domains. Across anisotropic heat flow, stiff reaction-diffusion, FitzHugh-Nagumo, and patient-specific left-atrial electrophysiology on CT-derived surfaces, G-FuNK reproduces held-out trajectories with relative L2 error between 0.046 and 0.160. Its accuracy is comparable to that of a geometry-aware Fourier neural operator on regular domains, and its predictions are unchanged when a rectangular domain is rotated. On left atria, whose venous and valvular openings prevent mapping onto a reference shape, it predicts an 85-ms trajectory in under half a second, against about 5 minutes for finite elements, with a mean activation-time error of 1.73 ms. Accuracy depends on graph construction and is highest when the learned generator is integrated at its training time step.
Figures & tables
Figure 1: Overview of the G-FuNK framework for data-driven cardiac electrophysiology. Patient-specific anatomies are obtained from medical imaging and processed into computational domains with associated myocardial fiber fields. High-fidelity numerical simulations are used to generate training data. These geometries are then downsampled and converted into graphs, from which graph Laplacians and their spectral decompositions are computed to define domain- and parameter-adapted bases. G-FuNK learns in this spectral domain, combining graph Fourier projections with learned kernel operators to model the temporal evolution of the transmembrane potential. Time integration is performed via a Neural ODE formulation, enabling the prediction of full spatiotemporal trajectories. Once trained, the model provides rapid predictions on previously unseen geometries and parameters, from which clinically relevant functional quantities can be derived. The framework is general and is evaluated across increasingly complex PDE systems, culminating in patient-specific cardiac electrophysiology applications.
Figure 2: Heat Equation: G-FuNK vs Target (Two Test Examples). Two representative test examples corresponding to different diffusion fields β . In each example, the top-left panel shows a stream plot of the primary direction of diffusion, F^0,β(1) , as defined in SI Appendix, Eq. S1. The bottom-left panel shows the initial condition at t0 with a scale bar of 0.5 to show the size of the domain. The remaining panels show snapshots of the solution at three later time points, with the G-FuNK predictions shown above the corresponding ground truth.
Figure 3: Reaction–Diffusion on Random Rectangle: G-FuNK vs Target (Two Test Examples). Two representative test examples corresponding to different domains with anisotropic diffusion. In each example, the top-left panel shows a stream plot of the primary direction of diffusion (5:1 ratio), F^α,β(1) , which is constant for all trajectories in this example. The bottom-left panel shows the initial condition at t0 with a scale bar of 10mm to show the various sizes of the test domains. The remaining panels show snapshots of the solution at three later time points, with the G-FuNK predictions shown above the corresponding ground truth targets.
Figure 4: FitzHugh–Nagumo with Various Domain and Diffusive Parameters: G-FuNK vs Target (Two Test Examples). Two representative test examples corresponding to different domains with different Gaussian-defined anisotropic diffusive fields. In each example, the top-left panel shows the pointwise multiplying function ϕα,β(x) that combines with Σ and defines the local strength of the anisotropic (ratio 2:1) diffusive field Kα,β(x) , as defined in ( 5 ). The bottom-left panel shows the initial condition at t0=2.5 . The remaining panels show snapshots of the solution at three later time points, with the G-FuNK predictions shown above the corresponding ground truth targets.
Figure 5: Cardiac Electrophysiology: G-FuNK vs Target (Out-of-Training Test Geometry). Comparison of G-FuNK predictions and ground truth targets on a representative out-of-training patient-specific atrial geometry. The top-left panel shows the myocardial fiber field, where the black lines indicate the primary direction of electrical conduction throughout the left atrium and define the orientation of the anisotropic diffusion tensor Kα,0(x) . The bottom-left panel shows the initial condition of transmembrane potential for u(x,t=0\text{,}\mathrm{ms})=u0 . The middle columns show transmembrane potential snapshots at t=5\text{,}\mathrm{ms} and $t=$15\text{\,}\mathrm{ms} , with G-FuNK predictions displayed above the corresponding ground truth targets. The rightmost column shows the local activation time (LAT) functional computed from the predicted solution and from the corresponding ground truth. A full temporal roll-out example is provided in SI Appendix, Fig. S5.
Problem
G-FuNK
Geo-FNO
GNN
Anisotropic heat
Parameter adaptation
0.056±0.018
0.064±0.025
0.298±0.075
Reaction–diffusion
Domain adaptation
0.100±0.025
0.112±0.015
0.335±0.010
FitzHugh–Nagumo
Domain + parameter adaptation
0.046±0.017
0.344±0.164
0.448±0.139
Table 1: Prediction performance across benchmark problems. Performance of G-FuNK, Geo-FNO, and a GNN equipped with a neural ODE on the problems in Section 2 . Reported values are relative L2 errors on held-out test sets with unseen initial conditions, domains, and, where applicable, diffusion fields, unless otherwise indicated; where ± appears for the first three benchmarks, the spread is over test trajectories. Cardiac electrophysiology used five single-geometry hold-out splits; entries are mean ± standard deviation across splits, with Pearson correlation, LAT relative L2 , and mean absolute error (MAE, ms) reported in addition to relative L2 error.
Figure 6: Top: Network structure of the Neural Operator. The input a is a (possibly vector-valued) function on Gα,β . (1) It is lifted to a higher dimensional space via a lifting operator P . (2) Multiple G-FuNK layers are then applied. (3) The network projects to the output dimensional space via Q . Bottom: Single G-FuNK layer structure. Starting from input k , Spectral branch (1) F denotes the GFT of the input which is reduced to the kmax -lowest modes. (2) The function is transformed with a learned spectral operator R which incorporates the eigenvalues. (3) F−1 represents the inverse GFT, mapping back to the original domain. (4) The input function k undergoes a linear mapping in W . (5) The outputs of the top and bottom branches are combined and then passed through an activation function σ . The output of the G-FuNK layer is z and is operated on by the next G-FuNK layer.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure S1 : Three different and unique examples of the primary diffusion vector fields used for the heat equation example in section 2.2.1 of the main text. Each example represents a diffusive field that was randomly determined for each trajectory.
Section
Parameter formula
Ex. 1, 3
Ex. 2, EP
Computational cost
(1) Input projection
2dp
400
300
O(nαdudp)
(2) Spectral branch
Fourier transform (and inverse)
O(nαkmaxdp)
Self-attention blocks ( R0--R3 )
7dSA+1
1796
1796
O(kmax2dSA+kmaxdSA2)
(2) Spatial branch
Linear projections ( W0--W3 )
dp(dp+1)
160800
90600
O(nαdp2)
Appendix
Table S1: Parameter counts and computational costs at prediction time for G-FuNK in examples 1, 2, 3 and cardiac electrophysiology (EP). In the spectral and spatial branches, formulas are per block/layer, while the numerical parameter counts include all four G-FuNK layers.
Figure S2 : FitzHugh–Nagumo with Various Domain and Diffusive Parameters: Geo-FNO vs Target (Three Test Examples). Three representative test trajectories corresponding to different initial conditions, different Gaussian-defined anisotropic diffusive fields and domains. In each example, the top-left panel shows the pointwise multiplying function ϕα,β(x) that defines the local strength of the anisotropic (ratio 2:1) diffusive field Kα,β(x) , as defined in Eq. 5 of the main text. The bottom-left panel shows the initial condition at t0=2.5 . The remaining panels show snapshots of the solution at three later time points, with the Geo-FNO predictions shown above the corresponding ground truth targets. Panel (a) and panel (b) are the same domain and parameter set but different initial conditions. Panel (c) shows a domain and scaling factor that are more skewed compared with panels (a) and (b).
Figure S3 : Temporal evolution of relative L2 error. Time-dependent relative L2 errors for G-FuNK across the four benchmark problems. Solid lines indicate the mean across test trajectories and shaded regions indicate ± one standard deviation. For cardiac electrophysiology, statistics are computed across the five single-geometry hold-out evaluations.
Figure S4 : Geo-FNO Performance: Target vs Prediction for the random rectangle case. Top Geo-FNO performance on the original test set. Bottom Geo-FNO performance on the same test set but rotated 90∘ . Times are measured from the initial condition at t=0 . When the domain is rotated 90∘ the performance of Geo-FNO decreased as the rel. L2 error increased from 0.112±0.015 to 0.450±0.039 .
Anisotropic Heat - Parameter Adaptation
Weighting
Normalization
Rel. L2 Error
Additional Notes
Gaussian
Symmetric
0.056
Radius = 0.25
Gaussian
Symmetric
0.070
Radius = 0.2
Edge-Conductivity
Symmetric
0.032
Radius = 0.25
FitzHugh–Nagumo
Weighting
Normalization
Rel. L2 Error
Additional Notes
Appendix
Table S2 : Comparison of graph construction and Laplacian normalization strategies across benchmark problems. The first configuration listed for each example corresponds to the model reported in the main text. dt refers to the time step used for training the model.
Number of nodes
Relative L2 error
∼ 2,500
0.255
∼ 4,000
0.187
∼ 5,000
0.173
∼ 6,000
0.163
∼ 10,000
0.165
Appendix
Table S3 : Resolution transferability of G-FuNK. Relative L2 error of a cardiac G-FuNK model trained at approximately 5,000 nodes and evaluated without retraining at different graph resolutions on the same held-out geometries.
Training trajectories
Rel. L2 error
Std.
50
0.120
0.037
100
0.099
0.030
250
0.073
0.025
500
0.056
0.018
Appendix
Table S4 : Convergence with training-set size for the anisotropic heat equation. Rel. L2 error of G-FuNK on 100 held-out test trajectories (21 time points each) as a function of the number of training trajectories. Std. is the standard deviation across test trajectories. The 500-trajectory value is the model reported in the main text.
Benchmark
Inference dt′/dt
Forward Euler
RK4
FHN
4
7.995
0.595
Trained with dt=0.2
2
0.1292
0.171
1
0.046
0.153
21
0.103
0.161
41
0.133
0.161
FHN
4
1.968
0.301
Appendix
Table S5 : Sensitivity to inference temporal resolution and integration scheme. Relative L2 errors for the learned FitzHugh–Nagumo and cardiac electrophysiology generators across inference step sizes and integration schemes. The FitzHugh–Nagumo G-FuNK model was trained with both dt=0.2 and dt=0.05 time units; the sensitivity analysis for the dt=0.05 model was evaluated on 50 test trajectories. The cardiac electrophysiology G-FuNK model was trained with dt=0.2ms . All models were trained using Forward Euler, and inference was performed using Forward Euler and fourth-order Runge–Kutta (RK4).
Figure S5 : Full-trajectory G-FuNK predictions versus ground truth from two different views.
Mathematics for Materials Modelling (MatMat), Institute of Mathematics & Institute of Materials, École Polytechnique Fédérale de Lausanne, 1015 Lausanne, Switzerland