GLOBE: Accurate Surrogates for Boundary-Driven PDEs via Domain-Inspired Architectures and Equivariance
Organizations: NVIDIA Corporation
Abstract
We introduce GLOBE, a nonlinear neural surrogate for boundary-driven, homogeneous, weakly nonlinear PDEs that draws inductive bias from boundary-element methods and equivariant ML, representing solutions as superpositions of learnable Green's-function-like kernels evaluated from boundary faces to targets. The architecture is translation-, rotation-, and parity-equivariant; discretization-invariant; and units-invariant via rigorous nondimensionalization. On AirFRANS (steady incompressible RANS over NACA airfoils), a compact ~117k-parameter GLOBE model has lower mean-squared error than published values for the strongest reported ML surrogates on both the interpolation and low-data splits, in single-run comparisons under the released AirFRANS normalization. The model evaluates fields at arbitrary points, handles non-watertight meshes, and generalizes under Reynolds number and angle-of-attack distribution shifts. We further introduce a hierarchical Barnes-Hut acceleration that exploits the kernel's far-field decay to reduce evaluation cost from quadratic to near-linear in the number of boundary elements, with a tunable opening angle that recovers exact dense evaluation in the zero-angle limit. Quantitative validation focuses on 2D AirFRANS; on 3D DrivAerML cars, GLOBE predicts surface and volume fields on the full meshes, and we compare it with a matched Transolver baseline. Together, these results show that boundary-element-inspired inductive biases can substantially improve the accuracy of ML surrogates for boundary-driven PDEs while scaling to large 3D meshes.
Figures & tables
| Mean Squared Error, “Full” task | ||||
| Model | ||||
| GLOBE (our work) | 0.0042 | 0.0021 | 0.0023 | 0.0070 |
| enf2enf ( Catalani et al., 2025a ) | – | – | 0.1100 | 0.3200 |
| CORAL ( Serrano et al., 2023 ; Catalani et al., 2025a ) | – | – | 0.3500 | 1.2000 |
| Transolver ( Wu et al., 2024 ) | – | – | 0.3700 | 1.4200 |
| MeshGraphNet ( Pfaff et al., 2021 ; Catalani et al., 2025a ) | – | – | 2.1400 | 3.8700 |
| Mean Squared Error, “Scarce” task | ||||
| Model | ||||
| GLOBE (our work) | 0.0238 | 0.0148 | 0.0496 | 0.1889 |
| MARIO ( Catalani et al., 2025b ) | 0.1520 | 0.1130 | 0.2400 | 2.7000 |
| Transolver ( Wu et al., 2024 ; Catalani et al., 2025b ) | 2.1050 | 2.1080 | 4.4340 | 23.6700 |
| MLP ( Bonnet et al., 2023 ) | 1.6470 | 1.4510 | 3.9040 | 21.9200 |
| GraphSAGE ( Bonnet et al., 2023 ) | 1.4570 | 1.4540 | 4.6960 | 19.4500 |
Appendix figures & tables29 assets
Supplementary material from the paper’s appendix.
Appendix
| Attribute | Physical Basis | Architectural Implication |
| Physical Conservation Laws & Symmetries | ||
| Translation-equivariance | Conservation of momentum | Use only relative spatial information; no absolute coordinates |
| Rotation-equivariance | Conservation of angular momentum | Vectors via local basis reprojection, not channel concatenation |
| Parity-equivariance | Euclidean symmetry group | Proper handling of vectors vs. pseudovectors |
| Representation Invariances | ||
| Discretization-invariance | Numerical consistency | Predictions must converge as mesh is refined |
| Method | Key Strengths | Key Limitations |
|---|---|---|
| CNNs / UNets | Hierarchical features; efficient on structured grids | Structured grids only; lack geometric symmetries; costly interpolation for unstructured meshes |
| GNNs | Handle unstructured meshes; effective for parabolic PDEs | Local receptive fields require layers for elliptic PDEs; typically lack rotation equivariance |
| Neural Operators | Discretization invariance; function-space mappings; well-suited for heterogeneous PDEs that are chiefly driven by field forcing functions | Often use localized kernels to incorporate geometry information, limiting information flow; limited ability to incorporate boundary information for boundary-driven problems |
| Neural Fields | Arbitrary query points; continuous representations; discretization invariance | Global compression bottleneck; limited geometric equivariance |
| Time | Memory | |
|---|---|---|
| Dense baseline | ||
| Tree construction | ||
| Aggregate computation | ||
| Dual-tree traversal | ||
| Near-field evaluation | ||
| Far-field evaluation |
| Hyperparameter | Value |
|---|---|
| Architecture | |
| Number of communication hyperlayers ( ) | 2 |
| Hidden layer sizes | [64, 64, 64] |
| Number of latent scalars per hyperlayer | 12 |
| Number of latent vectors per hyperlayer | 6 |
| Number of spherical harmonics ( ) | 1 |
| Mean Squared Error, “Reynolds” task | ||||
| Model | ||||
| GLOBE (our work) | 0.0601 | 0.0165 | 0.0144 | 0.0473 |
| MLP [ Bonnet et al., 2023 ] | 9.5050 | 4.9240 | 4.3000 | 208.9800 |
| GraphSAGE [ Bonnet et al., 2023 ] | 7.5580 | 3.4980 | 3.8260 | 17.9700 |
| PointNet [ Bonnet et al., 2023 ] | 9.4220 | 7.1290 | 4.0110 | 20.1300 |
| GUNet [ Bonnet et al., 2023 ] | 8.3830 | 5.2500 | 4.4830 | 20.5900 |
| Mean Squared Error, “Angle of Attack” task | ||||
| Model | ||||
| GLOBE (our work) | 0.4158 | 0.2116 | 0.3536 | 1.2665 |
| MLP [ Bonnet et al., 2023 ] | 6.9650 | 10.6300 | 11.7110 | 87.6200 |
| GraphSAGE [ Bonnet et al., 2023 ] | 4.4350 | 9.4000 | 10.9080 | 76.3800 |
| PointNet [ Bonnet et al., 2023 ] | 8.6800 | 15.7960 | 16.2370 | 58.4600 |
| GUNet [ Bonnet et al., 2023 ] | 5.6890 | 10.3420 | 14.8870 | 69.6700 |
| Mean Squared Error | ||||
|---|---|---|---|---|
| Model | ||||
| GLOBE (Full) | 0.0047 | 0.0039 | 0.0031 | 0.0039 |
| GLOBE (Full, decimated) | 0.0510 | 0.0402 | 0.0311 | 0.0336 |
| GLOBE (Scarce) | 0.0208 | 0.0180 | 0.0286 | 0.0392 |
| Task | ||||
|---|---|---|---|---|
| Full | 0.0047 | 0.0039 | 0.0031 | 0.0039 |
| Scarce | 0.0208 | 0.0180 | 0.0286 | 0.0392 |
| Reynolds | 0.1897 | 0.0811 | 0.2446 | 0.2903 |
| AoA | 0.9706 | 1.1928 | 11.3786 | 21.8215 |
| Full Dataset | Scarce Dataset | |||
| Field | MSE | MAE | MSE | MAE |
| Nondimensional velocity | ||||
| Pressure coefficient | ||||
| Model | |||||
|---|---|---|---|---|---|
| GLOBE , one seed | 8.87 | 9.67 | 8.29 | 6.74 | 0.976 |
| GLOBE , 3-seed ensemble | 7.67 | 8.67 | 7.36 | 5.89 | 0.959 |
| Transolver, one seed | 4.11 | 6.15 | 5.26 | 3.81 | 0.987 |
| Transolver, 3-seed ensemble | 3.88 | 5.97 | 4.97 | 3.57 | 0.989 |
| Configuration | Total validation loss |
|---|---|
| Reference GLOBE (Padé kernel) | |
| MLP kernel (no Padé) |
| expand_far_targets | Total validation loss | |
|---|---|---|
| (dense reference) | — | |
| False | ||
| False | ||
| True | ||
| True | ||
| True |
PDE Surrogate Modeling
PDE surrogate modeling builds fast approximations of partial differential equation solutions for simulation, prediction, and design.
See all 361 papers