acceptodds
Under review as a conference paper at ICLR 2027

AneumoBench: A Source-Linked Benchmark for Synthetic-Geometry Transfer in Aneurysm CFD

Abstract

Scientific machine learning uses simulation data to train surrogate models for fast physical-field prediction across geometries. Local shape editing can expand limited geometry collections, but whether its variants improve prediction on unseen geometries, and how to allocate them across sources, require controlled evaluation. We introduce **AneumoBench**, a dataset and benchmark linking 401 source aneurysm geometries to 9,693 locally edited descendant records, with computational fluid dynamics (CFD) fields computed on both. It contains 80,752 steady velocity–pressure cases across eight inlet conditions and 9,715 transient sequences of velocity, pressure, and wall shear stress (WSS). Each sequence contains 100 frames sampled at 0.01-s intervals from a 1-s cardiac cycle. Mesh, point, and voxel interfaces support steady field prediction and WSS forecasting from four observed frames. With family-disjoint splits, we compare source-only training, descendant training, and descendant pretraining followed by source fine-tuning across nine architectures on 79 held-out sources. Under the reported schedules, two-stage training lowers steady-field and reset-window WSS errors relative to source-only training. With the number of sampled fields and training updates fixed within each comparison, GraphSAGE benefits from descendant training and from distributing a fixed number of descendants across more sources. For WSS, reset-window gains do not consistently persist through 96-step rollout, and lower trajectory error need not improve cycle-level shear metrics or hotspot localization. These data and protocols enable researchers to compare descendant selection and training strategies on the same unseen source geometries.

Primary area:
datasets and benchmarks
Keywords:
Scientific machine learning, Datasets and benchmarks, Computational fluid dynamics, Intracranial aneurysms, Synthetic geometry, Transfer learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_hqeROpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces AneumoBench, a dataset that links 401 real aneurysm geometries to 9,693 locally edited variants, with steady velocity–pressure simulations and transient wall shear stress sequences for each. Using splits that hold out whole source families, the authors compare training on sources, on variants, and on variants followed by source fine-tuning across nine architectures, and study how many variants to use and how to spread them across sources.

Critical strengths
  1. Large new transient dataset covering full cardiac cycles, backed by substantial simulation effort.

  2. Careful protocol: whole source families are held out, and every test geometry counts equally.

  3. Honest reporting of negative results and data issues.

Critical weaknesses
  1. The reference models barely outperform a trivial baseline on the steady task. The paper notes that predicting zero velocity gives a RelL2 of about 1, yet source-only velocity errors range from 0.83 to 1.02, and some recipes exceed 1 (e.g., F-FNO R2 velocity, FNO R1 pressure). In addition, all reported steady results include about 6% zero-valued placeholder points inside the fluid domain, and only six GraphSAGE checkpoints were re-evaluated without them. This makes the R1–R3 comparisons difficult to interpret.

  2. Several of the main conclusions are confounded or do not hold up. R3 receives extra fine-tuning beyond R2, and the paper notes that sample and update counts differ across recipes, so only the GraphSAGE controls compare at matched compute. The allocation preference reverses for Transolver on airfoils. In WSS rollout, both Transolver and U-Net perform worse than simple persistence, and a geometry-free PointwiseMLP outperforms all geometry-aware models.

  3. The descendants differ structurally from the source aneurysms. Each one is made by removing the original aneurysm and displacing a surface patch outward by only 0.5–1.0 mm, so the resulting geometries look more like vessels with small bulges than like edited aneurysms. This weakens the paper's framing of aneurysm CFD transfer and raises the question of whether the observed gains come from learning general vessel flow rather than aneurysm hemodynamics. The editing is also operator-guided in Geomagic, which makes the construction hard to reproduce or extend, and 60 descendant records are exact duplicates.

I suggest the authors to report results with the corrected point domain and physical coordinates for all models, compare recipes at matched compute, and strengthen the baselines so that they clearly outperform trivial predictors before drawing conclusions about the value of the variants.

Questions for authors

Please see the Weaknesses section.

Reviewer_JC8fOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a new benchmark for evaluating the impact of data augmentation on physical simulations, using aneurysms as the main application. The paper uses different metrics and notices that, while the synthetic data improve the prediction quality in the short term, their effectiveness on long-term predictions severely decreases as errors accumulate.

Critical strengths
  • Novel and little-studied application
  • Generally, a clear and linear explanation, albeit with minor missed points (see weaknesses)
Critical weaknesses
  • Some design choices are not well clarified. For instance, it is unclear to me why the aneurysm is removed, and no study has been done on the impact of its removal on the overall geometry

  • The fact that more sources help when the total number of descendants is the same seems to reflect the well-known idea that synthetic data generators benefit from diverse sources.

  • The overall goal of the work is a bit unclear, and a comparison with similar studies in other fields would help.

Questions for authors
  1. How does the removal of the aneurysm affect the geometries?

  2. Can you add an analysis of how the findings compare with similar studies in other domains?

  3. What is the failure mode of long-term predictions?

Reviewer_uEMCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces AneumoBench, linking 401 source aneurysm geometries to 9,693 synthetic descendants with CFD fields. It evaluates synthetic-data transfer for steady-field prediction and WSS forecasting, including controlled experiments on data quantity and allocation.

Critical strengths

The source-linked dataset and family-disjoint splits provide a useful resource for studying synthetic-data transfer. The matched-budget GraphSAGE experiments offer evidence that synthetic data can improve prediction beyond simply increasing training time.

Critical weaknesses

The main steady-state models were trained on data containing invalid zero-valued points. Reevaluating fixed checkpoints does not establish what happens when training uses cleaned data. The main comparisons use different training budgets, while matched-budget controls are largely limited to GraphSAGE. This limits conclusions across architectures.

Questions for authors

Do the main improvements remain after retraining on the cleaned, cell-connected data? Do matched-budget experiments on another architecture support the same transfer gains?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.