Topology-Inspired Fine-Tuning for Wind-Field Super-Resolution

What it claims

Fine-tuning the PhIRE wind-field SR CNN (Stengel et al.) with gradient-magnitude supervision plus fixed-pair supervision lowers persistence-diagram distances without computing any persistence diagram during training. Relative to the pretrained CNN, mean bottleneck, W2,∞ and W2,2 distances fall 31%, 25% and 27%. Relative to a matched control fine-tuned on reconstruction loss alone, they fall 32–35%. The control by itself makes all three PD distances worse. So the topology gains come from the auxiliary losses, not from extra training.

How it works

The network reconstructs the full wind vector [u, v]. All topological losses act on the derived scalar wind speed.

• Candidate losses studied:
• scalar speed
• gradient magnitude
• soft high-speed level sets
• high-speed local maxima
• fixed-pair supervision
• Fixed-pair supervision:
• The Topology ToolKit (TTK) runs offline on ground truth to locate persistence-pair endpoints.
• A critical-value term L_CV penalizes predicted values at those fixed locations. It follows the first-phase constraint of Kissi et al.
• A persistence term L_pers constrains the scalar contrast between the two endpoints of each pair.
• Unlike Kissi et al.'s second phase, the prediction's diagram is never recomputed per iteration. PD and MT distances are used only for evaluation.
• Final objective: reconstruction + gradient + fixed-pair.

flowchart LR
  GT[Ground-truth speed] -->|TTK, offline| P[Pair endpoints]
  P --> FP[Fixed-pair loss]
  Pred[CNN prediction speed] --> FP
  Pred --> G[Gradient loss]
  Pred --> R[Reconstruction loss]
  FP --> L[Fine-tuning objective]
  G --> L
  R --> L

• Filtration convention:
• All descriptors use TTK's sublevel convention.
• Maxima therefore appear as death endpoints of saddle–maximum (D1) pairs.
• The paper notes that not every maximum targeted by the local-maximum loss ends a long-lived pair.

Evidence

Benchmark setup

• 168 fields, held out from all fine-tuning.
• Topology is evaluated on a fixed 160×160 subregion of each field.

Baseline comparison (pretrained CNN vs GAN)

• The CNN wins PSNR on all 168 fields and speed SSIM on 165.
• MT distance favors the CNN on 148 fields; bottleneck favors it on 129.
• The order-2 Wasserstein distances favor the GAN on 155 and 158 fields.
• The GAN is also stronger on spectral behavior, gradient distributions and extreme-exceedance measures.

Loss study

• 16 configurations, each against its own matched control (same initialization, data, optimizer, schedule).
• Includes all 8 combinations of the speed, gradient and level-set losses, with factorial analyses.
• Gradient loss gives the largest PD gain among the scalar-field losses but does not improve mean MT distance.
• Fixed-pair supervision gives the most consistent joint PD+MT gain.

Final model

• All three PD distances drop on 150 of 168 fields.
• PSNR rises 1.30 dB, about half of standard fine-tuning's gain.
• Mean MT distance improves, but MT gets worse on 59 of those 150 PD-improved fields.

How to use it

The page names no code release, weights or commands. The practical requirements it does state:

• the PhIRE pretrained CNN as a starting point
• TTK for offline extraction of ground-truth pairs
• for evaluation, explicit implementations of bottleneck, W2,∞ and W2,2

The paper says "W2" is used ambiguously in the literature and always writes both indices. Check which ground norm any comparison number uses.

Limits

• The authors frame it as a controlled proof of concept.
• It covers one dataset family (PhIRE wind), one architecture (CNN) and a single 168-field benchmark.
• Topology is measured only on a 160×160 subregion.
• PD and MT improvements are often decoupled per field. Optimizing PD does not reliably buy merge-hierarchy fidelity.
• You trade away roughly half the PSNR gain of plain fine-tuning.
• This version corrects the fixed-pair implementation and the PD-distance computation relative to the TopoInVis Connect 2026 poster. Numbers from the workshop version should not be relied on.
• The fetched text cuts off mid-related-work, after the first of three stated differences from Kissi et al. Detailed results tables and the authors' own limitations section were not visible.