acceptodds
Under review as a conference paper at ICLR 2027

Embedded Neural Boundary: Learning Near-Wall Impulse for Fast Flow Prediction

Abstract

Learned correctors for coarse-grid flow solvers are typically volumetric: a large CNN corrects every cell and is trained by unrolling through a differentiable solver. Classical simulation concentrates its effort within a few cells of the wall; we embed the learned component there, as a neural boundary. For coarse embedded-boundary flow past a sphere, an impulse confined, before the pressure projection, to a three-cell near-wall band removes  60% of the near-wall velocity error, and most of what remains is static per operating condition. We shape the corrector accordingly: a small MLP predicts eight POD band coefficients per frame from the near-wall state and inflow condition, and a static bias field absorbs the rest. Its network adds 1% to the solver's wall-clock on a 24-thread CPU, against 190% for a U-Net. In closed loop it has 1.3–2.1× lower near-wall error than the best of five training variants of a volumetric solver-in-the-loop CNN with 177× its trainable parameters (21× counting fixed tables), and, with its impulse counted as body force, drag within 2–4% of the fine reference; the static field raises the margin to 2.6× and recovers 87–96% of a replayed full-field oracle's improvement. The same construction holds on 13 two-dimensional geometries and a spheroid at incidence for held-out Reynolds numbers.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
computational fluid dynamics, scientific machine learning, learned correction, embedded-boundary methods

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_sBzXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The idea of this paper is that a lot of the interesting bits of a fluid simulation happens near boundaries, and so a learning system can be designed that has higher "resolution" near boundaries and less "resolution" away from boundaries, to more efficiently capture representations of fluids. In that sense, I believe this paper is indeed about a learning-based representation of a fluid and is on topic for ICLR. The resulting system is demonstrated to be around 20x faster on a GPU, compared to a basic U-Net.

Critical strengths

The core idea of having a network architecture that is adaptive to where the high-frequency details of a problem are is not novel, but in the context of applying that to fluid simulations, I think there is potential novelty, and it seems like a nice idea.

Critical weaknesses

My main issue is the comparison point here is a CNN/U-Net from 2020. There have been many advancements since then - for instance, EddyFormer is about 30x faster than solver-in-the-loop, as well as Universal Physics Transformers, which are also very fast compared to U-Nets. While the authors' baseline comparisons could be valid, I think there is a lack of justification for using an old algorithm as a reference compared to one of these newer techniques out there.

The paper also only considers very low Reynolds number flow (Re <= 500) which by the authors' own admission in the limitations section means we don't really know how well this method applies to more chaotic/turbulent fluid simulations. My guess is that for such simulations, there is more of that high-frequency noise throughout the fluid domain, so the neural boundary technique will not work as well, but that would be interesting to see regardless.

Questions for authors

See above

Reviewer_8T91OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper corrects a coarse embedded-boundary flow solver only in a thin three-cell band around the body, rather than over the whole domain. A small MLP, trained one-step, reads spherical-harmonic features of the near-wall state plus the inflow condition and predicts 8 POD coefficients of a near-wall velocity impulse. A non-learned static bias field, averaged from reference defects, is added on top. On a sphere, a 6:1 spheroid and 13 2D geometries, the corrector has 1.3–2.1× lower near-wall error than a 177× larger solver-in-the-loop U-Net and adds only ~1% to the solver's wall-clock time.

Critical strengths

Cheap and well-motivated: Placing the correction where cut-cell errors arise is a sensible physical prior. The learned stage costs 6 ms per frame on CPU, compared with 880 ms for the U-Net (Table 2), and trains in about 20 s.

Careful controls: Appendix E changes one property at a time (band-limited U-Net, U-Net encoder with 8 band coefficients, MLP trained on the CNN's 1,280 pairs). This convincingly shows that the gain comes from the output prior rather than the backbone.

Honest analysis: The paper reports its own failure cases, such as the decelerated spheroid, the NACA airfoils where the band mean wins, and the coarse force operator's drag bias (Section 5, Appendix F).

Critical weaknesses

Narrow generalization: Basis, corrector and static field are retrained per body. Generalization is tested only by interpolating in Re between training conditions. When the inflow condition is outside the training set, the method fails: on the decelerated spheroid, the corrector is much worse than no correction (9.56 vs 1.29). The static field also needs a fine reference at or near the deployment condition.

Weak baselines: The U-Net is trained on only 1,280 pairs, compared with the corrector's 18,584 frames. The "unrolled" variants use gradient-free DAgger relabelling rather than true differentiable solver-in-the-loop training as in Um et al. (2020). The one-step CNN diverges at Re = 250 (23.08 vs 10.55 for no correction), and every CNN variant is worse than no correction under deceleration. This suggests the baselines are not well tuned.

Short evaluation window: All main errors are scored over t ∈ [135, 150], about two shedding periods. The longer rollouts cover only three sphere trajectories and report force statistics, not field errors.

Hard to read: The paper is very difficult to read, with many mode names (sotl1s, sotlsur, mlpvol, volstatic, bandstatic, cnnpod...). Key results are spread across 36 pages, with much of the evidence in the appendix. The abstract and introduction are difficult to follow without reading Appendices B–F.

** Minor Issues**:

  • There is a typo: "Eq. equation 4".
Questions for authors
  1. Could the authors add a learned-interpolation baseline (Kochkov et al., 2021) for a comparison?

  2. How does the U-Net baseline perform when trained on the same 18,584 frames, and with true differentiable unrolled training?

  3. Can the method generalize to inflows or geometries not seen in training, for example by adding dynamic inflow to the spheroid training set?

  4. Can you report near-wall field errors over longer horizons, or explain why force statistics are sufficient there?

Reviewer_dKseOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary
  • Interesting paper.
  • Idea is to train a (small memory) network that corrects a coarse flow solver. Works with embedded boundary approches + octree meshes, which is the pathway to really scaling up
  • Paper illustrates claims with numerical examples on a limited set of 3D shapes
Critical strengths

None

Critical weaknesses
  • No new AI/Learning or representation understanding: It is well known (as the authors themselves state) that error accumulates and is corrected at the boundary.
  • Neural boundary representation has also been explored recently (see Karki et al, for instance)
  • The use of spherical harmonics limits applicability. I recommend using an expansion of the eikonal representation as an alternative. This could really extend the applicability of the work
  • No complex shapes (or generalization to complex shapes) are shown. While the examples are canonical, existing solvers do an excellent job with them (so the value prop of AI methods is limited)
  • More damagingly, the simulations are limited to laminar flow. This method would have been very powerfully shown (and I would be far more likely to accept it) if it showed turbulent flows (or transitional flows); even for the sphericla case, could they show, say, Re = 3700; or even better, capture the drag crises?
Questions for authors

I am hopeful that the authors can extend this to transitional or turbulent flows

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.