SurfNet: Overview of Multi-Domain Systems

• SurfNet is a multifaceted label that denotes distinct systems in CFD, 3D shape generation, novel view synthesis, and quantum networking benchmarks.
• Its implementations range from transfer learning-based super-resolution models achieving up to 2.1x speedup in turbulent flow simulations to deep-residual architectures for geometry image-based 3D surface reconstruction.
• In quantum networking, SURFnet serves both as a practical real-world benchmark topology and a dual-channel, surface-code-based routing architecture addressing entanglement, error correction, and repeater placement.

SurfNet is a label used for several technically unrelated research systems, and the closely related spelling “SURFnet” denotes the Dutch research backbone that is repeatedly used as a quantum-network benchmark. In the arXiv literature considered here, the name covers a transfer learning-based super-resolution flow network for turbulent CFD, a deep residual architecture for generating 3D shape surfaces from geometry images, a sparse-radiance-field model spelled “SuRFNet,” and a dual-channel quantum-network architecture based on surface-code logical qubits. Separately, SURFnet appears as a fixed real-world topology in studies of entanglement distribution, routing, and repeater placement (Sinha et al., 2017, Obiols-Sales et al., 2021, Hamdi et al., 2022, Vardoyan et al., 2023, Hu et al., 11 Jun 2026).

1. Terminological scope and disambiguation

The term spans multiple domains, and the orthography is meaningful. One paper explicitly refers to its model as “SuRFNet (Sparse Radiance Fields Network), not SurfNet,” while several quantum-networking papers use “SURFnet” to denote the Netherlands research backbone rather than a neural model (Hamdi et al., 2022, Vardoyan et al., 2023).

Name in paper	Domain	Role
SurfNet	3D shape learning	Geometry-image surface generation
SURFNet	CFD / turbulent flows	Super-resolution flow network
SuRFNet	Novel view synthesis	Sparse radiance field completion
SURFnet	Quantum networking	Real-world benchmark topology
SurfNet	Quantum networking	Surface-code-based dual-channel architecture

This suggests that “SurfNet” functions as a reused project name rather than a single unified framework. The shared label masks substantial differences in representation, optimization, and evaluation protocol, ranging from 2D convolution on geometry images to sparse 3D voxel convolutions, mixed-integer quantum-network optimization, and surface-code routing.

2. SURFNet for super-resolution of turbulent flows

In computational fluid dynamics, SURFNet is a transfer learning-based super-resolution flow network for accelerating steady, incompressible RANS simulations with the Spalart–Allmaras turbulence model. The learned operator maps any intermediate iterate of the solver to its steady state,

$$x^N = G(x^I; \theta), \qquad x^I, x^N \in \mathbb{R}^{m \times n \times z}, \ z = 4,$$

where the four channels are the $x$-velocity $U$, $y$-velocity $V$, kinematic pressure $P$, and modified eddy viscosity $\tilde{\nu}$. The training loss is mean squared error,

$L(\theta) = \|G(x^I; \theta) - x^N\|_2^2,$

optimized with Adam at learning rate $10^{-4}$. Physics is not enforced in the loss; instead, the prediction is fed back into OpenFOAM, boundary conditions are reimposed, and the solver is refined to the same convergence criteria as baseline CFD, yielding 0% relative mean error of the final solution (Obiols-Sales et al., 2021).

The coarse model is a symmetric 8-layer convolutional–deconvolutional CNN with 4 convolution layers followed by 4 deconvolution layers; filter counts are 16, 32, 128, 256, then 256, 128, 32, 16, with kernel size $5 \times 5$, stride 1, LeakyReLU activations, and double-precision tensors. The coarse training resolution is $x$0, and target resolutions extend to $x$1, $x$2, $x$3, and $x$4. Two transfer schemes are defined. One-shot transfer learning fine-tunes directly from the coarse model to the target resolution for 1–2 epochs at learning rate $x$5. Incremental transfer learning fine-tunes stepwise across intermediate resolutions with step size 1. The coarse low-resolution dataset contains 10 ellipse geometries and 90 configurations; the transfer dataset uses one geometry with 6 variants per high-resolution target, which is reported as 15x fewer configurations than the 90-case low-resolution training set.

The reported outcome is a consistent 2–2.1x speedup over the OpenFOAM physics solver for incremental SURFNet, independent of geometry and resolution up to $x$6. One-shot transfer also accelerates the solver, but its gain declines with resolution: approximately $x$7 at $x$8, approximately $x$9 at $U$0, approximately $U$1–$U$2 at $U$3, and approximately $U$4–$U$5 at $U$6. Coarse-only SURFNet generalizes well at low resolution but degrades to approximately $U$7 at $U$8. Validation loss exhibits the same pattern: the coarse model rises from approximately $U$9 at $y$0 to approximately $y$1 at $y$2, while incremental transfer maintains approximately $y$3 across resolutions. The scope is explicitly limited to 2D steady RANS; no explicit PDE residuals or divergence-free constraints appear in the loss, and pure inference without refinement would not satisfy conservation laws.

3. SurfNet for 3D shape-surface generation

In geometric deep learning, SurfNet is a category-specific method for generating 3D shape surfaces with deep residual networks by learning in 2D on geometry images rather than in 3D on volumetric grids. A geometry image is a regular 2D grid whose pixels encode 3D surface coordinates, formalized by

$y$4

SurfNet uses $y$5 geometry images for all categories. The parameterization is authalic spherical parameterization, followed by projection of the sphere onto an octahedron and cutting along 4 of its 8 edges to unwrap the surface into a square domain (Sinha et al., 2017).

A central contribution is consistent geometry-image creation across a category. For rigid ShapeNet categories such as cars and airplanes, meshes are voxelized at $y$6, converted to $y$7-shapes with $y$8, pruned to genus-0 surfaces, and Laplacian smoothed. Pairwise D2 shape distances define a similarity matrix, spectral clustering with $y$9 selects exemplars, and Blended Intrinsic Maps establish dense correspondence to a base shape. The base shape is parameterized once; its cuts, seams, and sampling layout are then transferred to all other meshes. The resulting dataset retains 691 car models and 1490 airplane models after filtering. For the non-rigid hand category, correspondences are trivial because all meshes share identical connectivity with 1065 vertices and 2126 faces; 200,000 mesh files are generated from an 18-DOF kinematic hand model.

The network architecture uses specialized residual blocks with downsampling and upsampling. Image-to-surface networks are 102 layers deep, while code-to-surface networks are 65 layers deep. All convolutions are $V$0 unless noted; the down-residual block halves resolution with stride-2 convolution, and the up-residual block doubles resolution with $V$1 transposed convolution. A notable design choice is to train three separate networks per category, one per coordinate channel, to avoid mean-shape collapse when predicting all three coordinates jointly. Training minimizes a curvature-weighted coordinate-regression loss,

$V$2

where $V$3 is the geometry image of point-wise mean curvature. No explicit normal consistency, Laplacian smoothness, or seam continuity penalties are used.

The paper emphasizes qualitative rather than standardized quantitative evaluation. It reports single-image surface reconstruction for cars and airplanes, full hand-surface reconstruction from a single depth image, and smooth interpolation between poses and shapes. It also states that the network can invent new shape surfaces and reconstruct surfaces from previously unseen images. At the same time, the method is constrained by topology and parameterization: non-zero-genus meshes are dropped, octahedral cuts introduce seams, separate per-channel prediction can degrade smoothness on flat regions, and quality declines for views outside the training azimuth/elevation range or for images with poor texture or low contrast. The paper does not report standardized metrics such as Chamfer distance in the main text.

4. SuRFNet for sparse radiance fields

The SPARF work introduces “SuRFNet,” not SurfNet, as a sparse-radiance-field generator trained on a large-scale ShapeNet-based dataset for few-view novel view synthesis. An SRF is represented as a sparse voxel grid in COO form,

$V$4

where coordinates $V$5 and features $V$6 store density and radiance parameters. Whole SRFs use $V$7 radiance dimensions corresponding to 4 spherical-harmonic coefficients per RGB channel, while partial SRFs optimized from only 1 or 3 views use $V$8 RGB features. SuRFNet takes these partial SRFs as input and predicts a completed sparse voxel radiance field using sparse 3D convolutions with a MinkowskiNet backbone (Hamdi et al., 2022).

The architecture comprises three sparse-convolution modules with depth $V$9 and stride 2 at all stages, followed by separate heads for density and radiance. Two model sizes are reported: a small variant with approximately 13.4M parameters and a large variant with approximately 87.3M parameters. Training uses a three-part objective: a density loss $P$0, a radiance loss $P$1, and a rendering loss $P$2, combined as

$P$3

Loss sampling uses a Quantized Gaussian over voxel coordinates,

$P$4

to stabilize gradients on sparse, evolving topologies. Rendering is differentiable and follows standard alpha compositing with trilinear interpolation over occupied voxels.

The supporting dataset is unusually large. SPARF provides approximately 17,073,150 images at $P$5 resolution from approximately 39,705 shapes across 13 categories, together with 1,072,008 optimized SRFs at resolutions $P$6, $P$7, and $P$8. On the normal test track, SuRFNet reports mean validation PSNR of 16.8 dB from 1 view and 19.5 dB from 3 views; on OOD “hard” views, it reports 14.6 dB from 1 view and 17.3 dB from 3 views. These numbers exceed the reported PixelNeRF, VisionNeRF, and Plenoxels baselines under the same protocol. End-to-end throughput is also practical: both the small and large variants are reported at 15 FPS, although the large model requires roughly 90.0 ms for network forward pass versus 14.4 ms for the small model. The method remains limited by the need to first optimize a partial SRF from posed images, by the cost of high-resolution sparse voxels, and by reduced reliability on reflective or specular shapes.

5. SURFnet as a benchmark topology in quantum-network research

In quantum networking, SURFnet is the research backbone in the Netherlands and serves as a real-world benchmark topology for studies of entanglement capacity, asynchronous protocols, GHZ-based routing, and repeater placement. Different papers instantiate different abstractions of the topology: one uses a pruned 17-node overlay between Delft and Enschede, while others use a fixed SURFnet graph from Knight et al. (2011) or the Internet Topology Zoo, with experiments over all node pairs or over randomly selected user pairs (Vardoyan et al., 2023, Pouryousef et al., 2024, Chen et al., 3 Apr 2026, Pouryousef et al., 2023).

Study	SURFnet role	Main reported finding
(Vardoyan et al., 2023)	Exact bipartite-capacity case study	Delft–Enschede capacity $P$9
(Pouryousef et al., 2024)	Asynchronous protocol evaluation	Sequential and parallel SKRs are close
(Chen et al., 3 Apr 2026)	GHZ/BSM routing benchmark	Uniform GHZ comparable to BSM; hybrid slightly worse
(Pouryousef et al., 2023)	Repeater-placement planning case	End-node coherence threshold near 3.2–3.5 ms

The exact-capacity study models a pruned SURFnet graph with 17 nodes, source $\tilde{\nu}$0 Delft and terminal $\tilde{\nu}$1 Enschede, time-slotted entanglement generation, heterogeneous swapping probabilities, and a mixed-integer quadratically constrained program that maximizes normalized flow to $\tilde{\nu}$2. The overall capacity is the expectation over all link-success snapshots. The reported exact value is

$\tilde{\nu}$3

Bell pairs per time unit. The same study states that “this network would benefit from multiplexing,” and notes that a local-knowledge heuristic produced an average capacity of 0 over finite Monte Carlo samples, whereas the exact MIQCP correctly captures rare but nonzero-capacity states.

The asynchronous-protocol study evaluates 900 random SURFnet user pairs whose shortest-path length lies in the range $\tilde{\nu}$4 km and whose path includes at least 2 nodes. It compares a sequential iterative extension protocol with a parallel link-level generation protocol under link loss, classical communication delays, quantum-memory decoherence, and a per-memory cutoff strategy. The principal SURFnet-specific conclusion is “minimal difference in the SKRs with and without imposing a cutoff on quantum memories irrespective of the protocol type,” and more broadly that the parallel scheme is only marginally better than the sequential one. This supports the recommendation of the sequential scheme on SURFnet because of comparable SKR and simpler implementation.

The GHZ-routing study uses the fixed SURFnet physical layout and averages rates over all node pairs. It evaluates conventional BSM routing, pure GHZ routing under several success models, and a hybrid GHZ–BSM protocol. On SURFnet, uniform-success GHZ routing achieves average rates comparable to conventional BSM routing, hybrid GHZ–BSM performs slightly worse, and restricted $\tilde{\nu}$5-GHZ routing is “almost 0 rate.” Distance dependence is strongly controlled by the measurement success probability $\tilde{\nu}$6: for $\tilde{\nu}$7, GHZ-based strategies show evidence of distance independence, whereas for $\tilde{\nu}$8 and $\tilde{\nu}$9 all strategies exhibit significant rate decay with distance.

The repeater-placement study treats each node in the SURFnet topology as a potential repeater location and optimizes a negativity-based utility

$L(\theta) = \|G(x^I; \theta) - x^N\|_2^2,$0

under memory, coherence, and repeater-budget constraints. In the SURFnet experiments, 4 random user pairs with source–destination separation 200–250 km are selected per workload, with $L(\theta) = \|G(x^I; \theta) - x^N\|_2^2,$1 repeaters, repeater capacity $L(\theta) = \|G(x^I; \theta) - x^N\|_2^2,$2, and end-user capacity $L(\theta) = \|G(x^I; \theta) - x^N\|_2^2,$3. The key thresholds are explicit: when $L(\theta) = \|G(x^I; \theta) - x^N\|_2^2,$4 ms the optimization becomes infeasible, and once $L(\theta) = \|G(x^I; \theta) - x^N\|_2^2,$5 ms, increasing repeater coherence $L(\theta) = \|G(x^I; \theta) - x^N\|_2^2,$6 yields utility gains. Utility also saturates when either $L(\theta) = \|G(x^I; \theta) - x^N\|_2^2,$7 or $L(\theta) = \|G(x^I; \theta) - x^N\|_2^2,$8 approaches approximately 100 modes. A separate runtime study finds that weighted utility planning, $L(\theta) = \|G(x^I; \theta) - x^N\|_2^2,$9, essentially matches the oracle upper bound obtained by per-slot replanning.

Taken together, these results make SURFnet a benchmark for near-term quantum-network realism rather than for idealized topological abstraction alone. The recurring bottlenecks are low link success on long fibers, compounded swap or measurement failure, classical latency, and memory-coherence constraints. This suggests that the Dutch backbone is especially useful for stress-testing routing and resource-allocation schemes under heterogeneous, nonuniform, and physically grounded assumptions.

6. SurfNet as a surface-code-based quantum network architecture

A different quantum-network use of the name appears in “Quantum Network Routing based on Surface Code Error Correction,” where SurfNet is a dual-channel network architecture that transports quantum messages as surface-code logical qubits. The architecture combines two traditional networking modes: an entanglement-based channel and a plain direct-transmission channel. Each logical surface code is split into a “Core,” containing data qubits “most critical” to logical error events, and a “Support,” containing the remaining data qubits “less critical but still essential for error corrections.” The Core traverses the entanglement-based channel by teleportation, while the Support traverses the plain channel as photons. Users create requests, switches act as intermediate routers, and servers are switches with larger quantum memories that can perform error correction once a complete code arrives (Hu et al., 11 Jun 2026).

The code substrate is the standard planar surface code with measure-X and measure-Z ancilla on a square lattice, and the paper evaluates code distances 9, 11, 13, and 15. Link fidelity along a path is modeled as

$10^{-4}$0

and purified Core-pair fidelity is updated by

$10^{-4}$1

A lost qubit is replaced by a maximally mixed state, giving estimated fidelity $10^{-4}$2 for an erasure. Decoder edge weights are then

$10^{-4}$3

The paper proposes both a modified MWPM decoder and a new SurfNet Decoder, a union-find–style decoder with growth speeds inversely proportional to the log of edge error probabilities,

$10^{-4}$4

with default $10^{-4}$5, and for erasures specifically $10^{-4}$6. The stated worst-case complexity is $10^{-4}$7.

Routing is formulated as an offline multicommodity optimization with separate Core and Support flows, server-side correction variables, edge and node capacities, and cumulative noise budgets. The objective is

$10^{-4}$8

where $10^{-4}$9 is the number of executed surface codes for request $5 \times 5$0. Noise is linearized per edge as $5 \times 5$1, and the routing constraints separately cap Core-only noise and whole-code noise. The evaluation uses Barabási–Albert graphs with more than 20 nodes, three resource scenarios, link fidelities drawn i.i.d. from $5 \times 5$2 or $5 \times 5$3, 1080 trials per design, a fixed 15% erasure rate, and Pauli error rates from 5.0% to 8.5%, halved on the Core. Under this protocol, the SurfNet Decoder reaches a Pauli threshold of 7.25% versus 7.1% for a Union-Find baseline. At similar throughputs, SurfNet achieves substantially higher average fidelity than both a direct-transmission baseline (“Raw”) and teleportation-only baselines with per-hop purification (“Purification $5 \times 5$4”), while maintaining similar latency and throughput. The gains are strongest when facilities are abundant; they diminish when entanglement resources are scarce and links are poor.

The architecture is therefore not a topology benchmark like SURFnet, but a network design that integrates routing, fault tolerance, and decoding around surface-code modularity. A plausible implication is that the name “SurfNet” is used here to emphasize the centrality of surface codes rather than any connection to the Dutch backbone or to earlier machine-learning systems.