acceptodds
Under review as a conference paper at ICLR 2027

LGO: A LOCAL GEOMETRY BRANCH FOR AUGMENTING NEURAL OPERATORS IN 3D INDOOR AIRFLOW PREDICTION

Abstract

Predicting indoor airflow across changing diffuser layouts requires models to capture local supply-jet behaviour, governed by nearby flow-driving boundaries, and room-scale recirculation set by the enclosure. Neural operators efficiently model global flow fields but typically lack explicit local boundary information at each query point. We introduce the Local–Global Operator (LGO), a lightweight local-geometry module using query-relative boundary coordinates stratified by boundary type to preserve small but important supply and return patches. A zeroinitialised projection enables controlled comparison with the original operator. We evaluate LGO on quality-controlled CFD simulations of a single room with diffuser layout as the only varying design parameter. Generalisation is tested using separation-controlled train–test splits based on a dimensionless layout distance, with the threshold set where a nearest-layout retrieval baseline loses predictive skill. Across three neural-operator architectures, LGO adds about 1% more parameters, improves full-field velocity R2 on both splits, and remains more stable as test layouts become more separated from training. Improvements concentrate near supply jets, while no consistent gain is observed in the ASHRAE Standard 55 occupied zone, where model differences are comparable to CFD repeatability and a simple lookup baseline. These results show that query-relative boundary information improves geometric transfer where local boundary forcing is important, at the cost of additional inference time.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural operators, Indoor airflow prediction, Local geometric encoding, Layout generalization, Surrogate modeling.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_dJhBOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
1
Presentation
2(+1)
Soundness
2(+1)
Read review
Paper summary

This paper proposes LGO, an additional module designed to be integrated into existing neural operators to provide geometric and boundary condition information for conditioning the models. For each query point, the module extracts boundary points by type and computes relative geometric information to learn a representation.

Critical strengths
  1. In the work, the authors identify the importance of conditioning neural operators using the boundary conditions of the simulations as an important source of information.
Critical weaknesses
  1. Authors claim that "Neural operators efficiently model global flow fields but typically lack explicit local boundary information at each query point.". However, neural operators usually receive boundary information as input in the form of mask, representing, for instance, fluid node, wall node, source node, etc. That information passing through the inductive bias mechanism of the model is usually enough to represent the boundary information of the problem. This has been represented in several previous works, for example, MeshGraphNets [1]. Authors do not clarify if they included this information as input to their "Host" models, and if so, in which particular way they did it.

  2. I consider the baselines and experiments insufficient. There are several stronger baselines that, given the correct input information (e.g. boundary condition information, mesh info, etc) could potentially be stronger than the ones used in the paper. To name a few: Transolver++ [2], AB-UPT [3], GeoTransolver [4], in addition to the Transolver model the authors mention in the paper, which you could find an easy to use implementation in PhysicsNeMo library. I suspect that the attention-based inductive bias of these models, in addition to correct input information, could potentially model small scales with good performance.

[1] Pfaff, Tobias, et al. "Learning mesh-based simulation with graph networks." arXiv preprint arXiv:2010.03409 (2020).

[2] Luo, Huakun, et al. "Transolver++: An accurate neural solver for pdes on million-scale geometries." arXiv preprint arXiv:2502.02414 (2025).

[3] Alkin, Benedikt, et al. "AB-UPT: Scaling neural CFD surrogates for high-fidelity automotive aerodynamics simulations via anchored-branched universal physics transformers." arXiv preprint arXiv:2502.09692 (2025).

[4] Adams, Corey, et al. "Geotransolver: Learning physics on irregular domains using multi-scale geometry aware physics attention transformer." arXiv preprint arXiv:2512.20399 (2025).

Questions for authors

In addition to the weaknesses mentioned above, please clarify:

  1. What is the reason you decided to use R^2 as the primary metric for baseline comparison, rather than other more conventional metrics in the surrogate modeling field such as L2 error? I think you could benefit from including additional metrics beyond R^2.
5 comments
AuthorsOpenReview ↗

Thanks for the detailed review (1/4)

We thank the reviewer for the detailed review. Several points come from information that the paper should have stated, and the revision now states it. Line numbers refer to the revised PDF. Host A is Geom-DeepONet, Host B is GINOT (cross-attention) and Host C is GINO (a graph kernel with an FNO).

R3-W1: "Authors claim that "Neural operators efficiently model global flow fields but typically lack explicit local boundary information at each query point.". However, neural operators usually receive boundary information as input in the form of mask, representing, for instance, fluid node, wall node, source node, etc. [...] Authors do not clarify if they included this information as input to their "Host" models, and if so, in which particular way they did it."

Yes, every host receives this information, and the paper should have said so. Section 4 now has a paragraph on what every model receives (line 309), and we revised the sentence you quote from the abstract (lines 15 to 16), which now says that neural operators typically do not give each query point the geometry of nearby boundaries explicitly. Each model, with or without the branch, gets the same boundary point cloud of 15,000 points with 13 channels:

ChannelsContentSet to the field mean (zero), because the solver computes it
3position
5one-hot boundary type: leak, wall, supply, return, and a fifth slot left over from the code we built on, always zero
3velocity u,v,wu, v, wat return vents and the leak
1pressureeverywhere
1temperatureeverywhere except supply vents

The hidden values are outputs of the CFD solve, so passing them in would give the model part of the answer. How each host reads the cloud:

  • Host A (Geom-DeepONet): a Perceiver-style encoder reads all 13 channels of every point and is pooled into the coefficients of the DeepONet expansion.
  • Host B (GINOT): the same encoder produces latent tokens, and each query attends to them.
  • Host C (GINO): a graph kernel integral maps the 13-channel cloud onto a latent grid, and an FNO runs on that grid.

So the hosts have the mask information the reviewer describes, as in MeshGraphNets. What the mask does not give a host is each query point's own view of the nearby boundary, in coordinates relative to that point. That is what the branch adds, and it is where the gains appear. Near-jet R2R^2 on the δ≈0.97\delta\approx0.97 test split (Table 3, line 379):

Host AHost BHost C
Host alone (with the mask)0.2410.3380.627
Host with the branch0.9240.9690.973

The hosts are not weak baselines chosen to make the branch look good. Each was tuned over the same four learning rates and trained for the same budget as every baseline, and Host C alone (0.627) is already above every attention baseline on this split (0.513 to 0.584, next comment).

AuthorsOpenReview ↗

Response to Reviewer dJhB (2/4)

R3-W2: "I consider the baselines and experiments insufficient. There are several stronger baselines that, given the correct input information (e.g. boundary condition information, mesh info, etc) could potentially be stronger than the ones used in the paper. To name a few: Transolver++ [2], AB-UPT [3], GeoTransolver [4], in addition to the Transolver model the authors mention in the paper [...]. I suspect that the attention-based inductive bias of these models, in addition to correct input information, could potentially model small scales with good performance."

We added GeoTransolver and Transolver++ as baselines, using the PhysicsNeMo code unchanged. Each gets the same 13-channel boundary input as every other model, the same search over four learning rates (1e-3 was best for both), 450 epochs, and one scoring pass on test. GeoTransolver's local features are off by default in PhysicsNeMo; we switched them on, with ball radii of 0.05 and 0.15 of the normalised room and 8 and 32 neighbours, as the Table 3 caption now states. We did not find a setting that brings it level with Transolver here; at δ≈0.50\delta\approx0.50 the two are within 0.03. Held-out test, single seed:

ModelParametersFull-field R2R^2, δ≈0.50\delta\approx0.50 (near jet)Full-field R2R^2, δ≈0.97\delta\approx0.97 (near jet)Relative L2L^2, δ≈0.50\delta\approx0.50 / δ≈0.97\delta\approx0.97Time per full field
Transolver (already in the paper)5.99 M0.967 (0.992)0.582 (0.584)0.179 / 0.6152.6 s
GeoTransolver5.96 M0.940 (0.976)0.524 (0.513)0.243 / 0.6651.3 s
Transolver++5.47 M0.961 (0.987)0.542 (0.539)0.195 / 0.6523.1 s
Hosts A, B, C with LGO5.4 to 5.9 M0.939 to 0.973 (0.992 to 0.997)0.835 to 0.917 (0.924 to 0.973)0.157 to 0.196 / 0.284 to 0.37311.3 to 13.3 s
GeoTransolver with LGO (new)6.02 M0.952 (0.994)0.870 (0.945)0.213 / 0.34510.7 s

Layouts on which each of our hosts with the branch beats Transolver on full-field R2R^2:

Host AHost BHost C
δ≈0.50\delta\approx0.50, 21 layouts182011
δ≈0.97\delta\approx0.97, 25 layouts222324

The reviewer's expectation holds for test layouts close to the training layouts: on the δ≈0.50\delta\approx0.50 split the attention models match our hosts with the branch, and Host C with the branch beats Transolver on only 11 of 21 layouts. It does not hold for layouts far from them. On the δ≈0.97\delta\approx0.97 split the attention models' near-jet R2R^2 falls to between 0.51 and 0.58, while our three hosts with the branch stay between 0.92 and 0.97. Their near-jet relative L2L^2 rises from 0.085 to 0.596 (Transolver), from 0.112 to 0.641 (Transolver++) and from 0.149 to 0.657 (GeoTransolver), against a rise from 0.053–0.074 to 0.159–0.234 for our hosts with the branch.

Times are for one full field of 566,700 points on one GPU (our hosts from Appendix E; the other rows from one later timing job, which reproduces Transolver's 2.6 s). The branch is not free: our hosts with it take 11.3 to 13.3 s, about 4 to 5 times Transolver, and it makes GeoTransolver about 8 times slower (1.3 s to 10.7 s). Every model here is still more than 120 times faster than the 28-minute mean CFD solve.

AuthorsOpenReview ↗

Response to Reviewer dJhB (3/4)

R3-W2, continued. Because the branch is an add-on, these models can also be hosts. We added it to GeoTransolver in the same way as to Hosts A and C, and it is bit-identical to GeoTransolver at initialisation. On the δ≈0.97\delta\approx0.97 test split, GeoTransolver's R2R^2 rises from 0.524 to 0.870, better on 24 of 25 layouts, and its near-jet R2R^2 from 0.513 to 0.945, better on all 25. The branch adds 1% to its parameters.

How these models differ from the branch. GeoTransolver comes closest, since it gathers boundary points in balls around each query. Two differences remain: in the PhysicsNeMo code the gathered points enter in absolute coordinates, which we checked by shifting the whole input, and they are not separated by boundary type. Transolver and Transolver++ have no neighbourhood around each query.

Would an explicit distance to the nearest supply vent be enough for an attention model? We have three pieces of evidence, each limited. On ShapeNet-Car, an explicit distance-to-boundary input cuts the host's error by 11.9% but does nothing once the branch is present, and the host with that input (0.0792 relative L2L^2) stays well behind the host with the branch (0.0515). On our data, giving Host A an unsigned distance to the nearest supply vent gave no gain (validation δ≈0.50\delta\approx0.50, single seed: full-field R2R^2 0.684, against 0.699 to 0.753 for Host A over three seeds). And the per-type control in our reply to Reviewer qmPS shows that seeing the right points matters as well as how many. We also gave Transolver this input directly: a per-point distance to the nearest supply vent as an extra channel (+512 parameters, same training and learning rate, one test pass). It did not help. Test full-field R2R^2 went from 0.967 to 0.963 at δ≈0.50\delta\approx0.50 and from 0.582 to 0.512 at δ≈0.97\delta\approx0.97, where it was better on only 8 of 25 layouts; our hosts with the branch still beat it on 24 of 25.

We did not include AB-UPT, because its input interface does not fit this task. It was built for external aerodynamics with a single surface, and its public implementation takes point coordinates only, with no per-point input for boundary type or boundary values. Our task is defined by those values: which boundary points are supply vents, and the velocity and temperature of the supply air. Running it here would require redesigning its input path, so the result would no longer be AB-UPT as published. We discuss it in Related Work, and we can add an adapted version if the reviewer considers it useful.

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

The paper proposes to add extra modeling to improve neural simulation accuracy with CFD simulation of airflow as groundtrutth. This is done by adding an extra branch to allow query aware injection of embedding. Numerical results and subjective visualization conform the efficacy of the proposed method.

Critical strengths

The paper proposes an enhancement to neural simulation fidelity with CFD as groundtruth. Introducing local branch allows easier incorporation of local geometry information and help improve modelling, and the experiments results is convincing.

Critical weaknesses

The proposed method and problem need some general introduction first to help readers not familiar with the settings. Also, design choices alternative to the one given in Figure 2, like using In-Context Learning, or more injections (for example, at every layer rather than only at the input) like in dual-stream, may be at least discussed (if not compared in experiments). In particular, the design choices should be discussed in terms of model-complexity vs. modelling fidelity.

A minor point is that for R2 in Table 2, the plus sign seems to not indicate incremental value, which is confusing.

Questions for authors

No particular questions.

4 comments
AuthorsOpenReview ↗

Thanks for the positive assessment and suggestions

We thank the reviewer for the positive assessment and the useful suggestions. We address each of them below. Line numbers refer to the revised PDF.

R2-W1: "The proposed method and problem need some general introduction first to help readers not familiar with the settings."

We agree. Section 3 now opens with a short description of the problem (line 174). It explains what a vent layout is, what the model takes as input (the room's boundary as a point cloud, with each point's type and the velocity and temperature of the supply air) and what it predicts (velocity and temperature at any point in the room). It also explains why the jets and the room-scale circulation need different information, which is why we add a local branch to a global operator.

R2-W2: "Also, design choices alternative to the one given in Figure 2, like using In-Context Learning, or more injections (for example, at every layer rather than only at the input) like in dual-stream, may be at least discussed (if not compared in experiments). In particular, the design choices should be discussed in terms of model-complexity vs. modelling fidelity."

We agree, and we added this discussion with a small experiment (Section 4, line 316). First, a clarification: in the paper the branch enters once, at the end rather than at the input. Its output is added to the host's per-query state after the host's last layer, just before the host's output head; for Host B it also seeds the query. The figure caption (Figure 1 in the revised PDF) now says this.

We tried two other entry points (δ≈0.97\delta\approx0.97 validation, 450 epochs, single seed):

VariantParametersTime per full field, paper design to variantFull-field R2R^2, paper design to variantLayouts better / worse
Host A (Geom-DeepONet): branch output added to the host's input, so it passes through every host layer (close to a dual-stream design)5,888,438, unchanged10.1 s to 9.6 s0.759 to 0.869 (+0.110, 2.8 SE)23 / 2
Host B (GINOT): the host's global features steer the branch's attention5,472,694, unchanged11.0 s to 11.6 s0.876 to 0.876 (−0.0001-0.0001, 0.02 SE)14 / 11

Times are for one full field (566,700 points) on one GPU, measured in one job. Both variants run the branch once per query, so complexity barely changes: same parameters, time within 5%. For Host A, feeding the branch in at the input is clearly better at no extra cost. For Host B the coupling changes nothing on average and costs slightly more time. Moving the room and the query points together by 0.16 m changes the predicted velocity by 6.6% of the mean speed for the paper design and 7.7% for the coupled one, against 57% for Host B alone. So neither design is exactly translation-equivariant as a whole, since Host B's encoder reads absolute boundary positions; the branch itself is (Table 2 now says this). We keep one entry rule for every host, adding the branch output at the host's last per-query layer, so that the hosts are compared under the same rule and the Host A and Host C pairs stay exact ablations. Tuning the entry point per host is a separate choice (Section 4 says the late point is not the best for every host); for Host A it would only raise the reported gain, so the paper's Host A results are, if anything, conservative.

Injection at every layer. Running the branch at every layer multiplies its cost by the number of layers (about 1.6 s per 100,000 queries per run on Host A). Caching the neighbour search would save only the search; the attention over the neighbours would still run at every layer. The input-side variant above already lets every layer use the branch for the cost of one run.

In-context learning. Training layouts already have CFD solutions, so they could serve as context at prediction time without new solves. Whether that helps depends on how close the context layouts are to the new one. A nearest-layout lookup, the simplest way to use such context, scores −0.081-0.081 full-field R2R^2 on the δ≈0.97\delta\approx0.97 test split, because no training layout is close enough to copy. A learned in-context model could do better, but on this split its context would be layouts at least 0.58 m away.

R2-W3: "A minor point is that for R2 in Table 2, the plus sign seems to not indicate incremental value, which is confusing."

Thank you. The sign was meant to mark values that can be negative, since R2R^2 is negative when a prediction is worse than the mean of the case (the lookup baseline scores −0.081-0.081). It reads as an increment, so we removed it, and the caption of the main results table (now Table 3, line 379) explains that R2R^2 can be negative.

AuthorsOpenReview ↗

Following-Up comments

Dear Reviewer xDus,

Thank you again for the positive assessment. Before the discussion closes on 18 November, we wanted to check whether our response answers your points:

  • Section 3 now opens with a short description of the problem for readers outside indoor airflow (line 174).
  • Section 4 discusses where the branch output can enter the host, with two tested variants and their cost, and discusses in-context learning (line 316).
  • The plus sign is removed from the R2R^2 values in the main results table (Table 3, line 379).

We would be glad to answer any further question.

Reviewer_xDusOpenReview ↗

Readability improves. And question on zero-initialization

Thanks for taking the efforts to improve readability, at least to me! I will raise the relevant scores accordingly.

The +0.110 R^2 gain result seems to suggest that location to introduce the local branch matters, and may worth more study efforts. One thing to be sure, having extra branch at input with Control-Net like technique allows zero-init exact ablation. It may be left as future work given the stringent timeline.

For discussion about ICL, please make the distinction between direct borrowing by looking up and neural ICL as in few-shot learning more blatant (just review the wordings, no need for more experiments).

AuthorsOpenReview ↗

Thanks for helpful suggestions

Dear Reviewer xDus,

Thank you for the follow-up and the helpful suggestions, and for reconsidering the scores.

We agree that a zero-initialised input-side entry, as in ControlNet, also keeps the host bit-identical at initialisation (our unit test checks this for Host A), so exact ablation does not require the late entry point and our earlier reply overstated that reason. We will update Section 4 to say so, and to state that we kept the late point so that all hosts follow one rule. Given the +0.110 on Host A, choosing the entry point per host deserves its own study.

We will also revise the in-context learning discussion in Section 4 to separate copying the nearest training layout (the retrieval baseline, which has no skill at δ≈0.97\delta\approx0.97) from few-shot neural in-context learning, where a model learns to condition on a set of solved layouts. The retrieval result says nothing about the latter, which we will state as future work.

Thank you again; these changes will make the paper clearer.

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

This paper proposes the Local–Global Operator (LGO), which augments existing neural operators with explicit local boundary geometry information for 3D indoor airflow prediction. The proposed module uses query-relative boundary coordinates and stratified neighborhood sampling to capture local flow structures, particularly around supply diffusers. The authors further introduce a separation-controlled evaluation protocol to assess generalization across different diffuser layouts. Experiments on three neural operator architectures show consistent improvements in full-field velocity prediction with a small increase in model parameters. The paper also evaluates the method on ShapeNet-Car to demonstrate applicability beyond indoor airflow prediction.

Critical strengths
  1. The results show substantial improvements in full-field velocity prediction across all three backbone architectures. The gains are particularly large for geometrically separated layouts. The additional ShapeNet-Car experiment also provides useful evidence that the proposed module may generalize beyond the original application.

  2. I appreciate the authors' effort to design separation-controlled train–test splits and evaluate different spatial regions separately. The ablations across three architectures, the analysis of CFD repeatability, and the discussion of inference overhead help provide a more complete understanding of the method.

Critical weaknesses
  1. The main idea of combining local geometric representations with global neural operators has been explored in previous work, including GINO, GNOT, and more recent graph-based neural operators such as RIGNO. PINTO also uses cross-attention to incorporate boundary-condition information into query representations. The proposed stratified neighborhood sampling is reasonable, but the overall architecture largely combines existing techniques such as nearest-neighbor retrieval, relative positional encoding, cross-attention, and residual injection. I would like to see a clearer explanation of the methodological novelty relative to these closely related approaches.

  2. The main experiments are conducted on only one room, with fixed geometry, boundary conditions, and diffuser counts. Only diffuser locations vary. Therefore, the evaluation mainly demonstrates interpolation or extrapolation within a restricted layout-design space, rather than generalization across different physical environments. Although the ShapeNet-Car experiment partially addresses this concern, it would be more convincing to evaluate different room geometries, diffuser counts, or operating conditions. In addition, the CFD solutions are not fully converged, which raises concerns about the reliability of the training targets, particularly in the occupied zone.

  3. Although LGO introduces only around 1% additional parameters, the inference overhead is substantial, with the paper reporting a 5–45× reduction in the host model's speedup. Therefore, parameter efficiency does not necessarily translate into computational efficiency. More importantly, the improvements mainly concentrate near supply jets, while the paper reports no consistent improvement in the occupied zone, which is particularly relevant for ventilation applications.

My major concern is that the paper seems more focused on an application-specific improvement for indoor airflow prediction, with limited methodological novelty or broader insights into neural operator learning, making its fit for ICLR somewhat questionable.

Questions for authors

See weakness

7 comments
AuthorsOpenReview ↗

Thanks for the detailed review (1/4)

We thank the reviewer for the detailed review and for the positive comments on the results, the evaluation splits and the ablations. We address each concern below. All numbers match the revised PDF, and line numbers refer to it. Host A is Geom-DeepONet, Host B is GINOT (a cross-attention model) and Host C is GINO (a graph kernel with an FNO).

R1-W1: "The main idea of combining local geometric representations with global neural operators has been explored in previous work, including GINO, GNOT, and more recent graph-based neural operators such as RIGNO. PINTO also uses cross-attention to incorporate boundary-condition information into query representations. [...] I would like to see a clearer explanation of the methodological novelty relative to these closely related approaches."

We agree that the building blocks are standard. What we add is one input that none of these operators gives a query point: the nearest boundary points of each boundary type, described only by their offsets from the query point. The branch has no room-scale mechanism of its own; it relies on its host for that. Table 1 of the paper (line 109) gives a shorter version of this table.

MethodWhat a query point reads near itselfCoordinates of what it readsBoundary types kept apartRoom-scale mechanismForm
GINOlatent grid nodes within a radius, after the boundary is mapped onto the gridabsolute positions of both points enter the kernelnoFNO on the latent gridstandalone model
GNOTall input tokens, through attentionabsolutenoattention over all tokensstandalone model
RIGNOnearby regional latent nodesrelative on edges, absolute on nodesnomulti-scale regional graphstandalone model
PINTOall boundary points, through cross-attentionabsolutenocross-attention over the whole boundarystandalone model
BNOthe whole boundary, through a low-rank boundary integral at every layerabsolute (kernel of the form ϕk(x)ϕk(y)\phi_k(x)\phi_k(y))no (one boundary type in its experiments)the boundary integralstandalone model
GeoTransolverboundary points in balls of two radiiabsolute (PhysicsNeMo code)noslice attention with a pooled geometry contextstandalone model
LGO (ours)the nearest boundary points of each typerelative onlyyesnone of its own; uses the host'sadd-on to a host

Because the branch is an add-on with a zero-initialised output, adding it to Hosts A and C leaves them bit-identical at initialisation, so those gains are controlled comparisons (Section 4, line 300). For Host B the branch also replaces the host's absolute-position query, so that pair changes two things at once and we treat it as weaker evidence. The measurements:

  • Host A: on test, +0.299+0.299 full-field R2R^2 at δ≈0.50\delta\approx0.50 (9.7 SE, better on 21 of 21 layouts) and +0.570+0.570 at δ≈0.97\delta\approx0.97 (9.8 SE, 24 of 25).
  • Host C, which is GINO: +0.040+0.040 (6.1 SE, 20 of 21) and +0.273+0.273 (8.9 SE, 25 of 25). Over three validation seeds the gains are +0.049+0.049 (t=29.3t = 29.3) and +0.224+0.224 (t=16.8t = 16.8).
  • Host B, which attends to the whole boundary from absolute positions as PINTO does: +0.301+0.301 (7.6 SE, 21 of 21) and +0.552+0.552 (16.0 SE, 25 of 25), with the caveat above.
  • Moving the room and its vents together changes the branch's output features (dimensionless network units) by at most 1.2×10−71.2\times10^{-7}, which is float32 round-off. When the branch's attention query is instead taken from the host's absolute-position features, the same test gives 3.9×10−33.9\times10^{-3}, about 30,000 times larger.
  • On ShapeNet-Car, an explicit distance-to-boundary input reduces the host's error by 11.9%. With the branch present it changes the error by −3.2%-3.2\% and +1.2%+1.2\% on two folds, so the branch already provides what the distance gives.

The per-type control follows in the next comment.

AuthorsOpenReview ↗

Response to Reviewer qmPS (2/4)

R1-W1, continued. A fixed number of points per boundary type matters (new experiment, validation, single seed). Walls outnumber a supply panel by more than ten to one, so a plain nearest-neighbour search returns only wall points for most of the room. On Host B we kept the branch and its 88 neighbours and replaced the per-type selection with one nearest-neighbour search over the whole boundary, with each point keeping its type label. Full-field R2R^2:

SplitHost alonePlain nearest neighboursFixed number per type (LGO)Layouts where LGO wins
δ≈0.50\delta\approx0.500.7430.8690.97328 of 28
δ≈0.97\delta\approx0.970.3390.7380.87622 of 25

Without the per-type selection, the branch still helps near the jets, but far from the vents it makes the host worse (far-field R2R^2 0.760 to 0.619 at δ≈0.50\delta\approx0.50). The control is in Section 6 (line 481). In the paper, removing the supply group alone, with the parameter count fixed, lowers R2R^2 by 0.101 and makes every one of the 28 validation layouts worse (Appendix E; Host A, 150 epochs, δ≈0.50\delta\approx0.50).

R1-W2: "The main experiments are conducted on only one room, with fixed geometry, boundary conditions, and diffuser counts. Only diffuser locations vary. [...] In addition, the CFD solutions are not fully converged, which raises concerns about the reliability of the training targets, particularly in the occupied zone."

The single room is the main limitation of the study, and the Limitations paragraph (line 524) and Appendix D state it. Holding the room fixed also helps the analysis: the vent layout is the only thing that changes, so any difference in accuracy comes from the layout. For changing geometry we rely on ShapeNet-Car, where every sample has a different shape; there the branch lowers relative L2L^2 error by 40.8% on a published fold. We are not aware of a public 3D CFD dataset in which vent positions vary, which is why we built ours, and it is released with the code at the anonymous repository linked in the paper. New rooms or vent counts need new CFD runs, which we could not complete in the rebuttal period.

On convergence: every model is trained and scored on one snapshot per case, the solution at iteration 3,000 of a fixed solver protocol, not on a time average. We measured how much the solution moves between snapshots of the same case:

MeasurementValue
Full field: correlation between snapshots 1,500 iterations apartρ\rho = 0.932
Bulk circulation: spread between layouts divided by spread within one run5.4 to 6.9
Occupied zone: correlation between snapshots 1,500 iterations apartρ\rho = 0.166

The full field is stable, and our claims rest on it. The occupied zone is not, which is why we make no claim there (Appendix C).

AuthorsOpenReview ↗

Response to Reviewer qmPS (3/4)

R1-W3: "Although LGO introduces only around 1% additional parameters, the inference overhead is substantial, with the paper reporting a 5–45× reduction in the host model's speedup. [...] More importantly, the improvements mainly concentrate near supply jets, while the paper reports no consistent improvement in the occupied zone, which is particularly relevant for ventilation applications."

We agree that the parameter count says little about compute cost, and the branch dominates prediction time. We timed Host A in one job on one GPU, per 100,000 query points:

ConfigurationNeighbours per queryTimeShare taken by the branch
Host A alone00.05 s
With the branch, paper setting881.62 s97%
With the branch, 8 supply neighbours641.25 s96%

In the paper's timing table (Appendix E, a separate run) the branch takes 98%, 86% and 78% of decode time for Hosts A, B and C, and costs 1.65 to 1.73 s per 100,000 queries. Accuracy does not change between 8 and 128 supply neighbours (−0.0015-0.0015 R2R^2 at 8, Appendix E), so using 8 cuts the branch's cost by 24% with no measurable loss; we have not tested smaller counts. A full field of 566,700 points takes 11.3 to 13.3 s including case setup. Dividing the 28-minute mean CFD solve (1,680 s) by these times gives a speed-up of 126 to 149 times. The Limitations paragraph states this trade-off (line 529).

Why the near field matters for design: the supply jets carry the conditioned air into the room, so their direction and reach decide which parts of the room receive it. A surrogate that misplaces them misjudges the layout it is meant to compare.

In the occupied zone, the CFD target itself is not stable (ρ\rho = 0.166 between snapshots of one case, above). A nearest-layout lookup with no trained parameters scores within the spread of the trained models there: occupied-zone ρ\rho on the δ≈0.50\delta\approx0.50 test split is 0.729 for the lookup, 0.727 for Transolver and 0.739 for Host C with the branch. With the current data this zone cannot rank architectures, so we claim no gain there, and we see no consistent change either way. Resolving it needs converged or time-averaged CFD. The solution's autocorrelation time is about 850 iterations, so cutting this noise five-fold would take roughly 48,000 iterations per case, against 3,000 now.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Thanks to the reviewers and summary of our responses

Dear AC and Reviewers,

Thank you for the careful reviews. We were glad that the reviewers found several parts of the work useful:

  • Reviewer qmPS noted substantial gains in full-field velocity prediction across all three host architectures, particularly large on geometrically separated layouts, and found the ShapeNet-Car experiment useful evidence that the module may generalise beyond indoor airflow.
  • Reviewer qmPS also appreciated the separation-controlled splits, the separate evaluation of different regions of the room, the ablations across three architectures, the analysis of CFD repeatability and the discussion of inference overhead.
  • Reviewer xDus found that the local branch makes it easier to bring in local geometry, and found the experimental results convincing.
  • Reviewer dJhB agreed that the boundary conditions of a simulation are an important source of information for conditioning a neural operator.

Throughout the replies, the three hosts are Host A (Geom-DeepONet), Host B (GINOT, a cross-attention model) and Host C (GINO, a graph kernel with an FNO). The new experiments in this rebuttal use a single seed. The control and design experiments are scored on validation; the new baselines are scored once on the held-out test sets. All changes are marked in blue in the revised PDF, and line numbers below refer to it.

  • What the hosts receive (Reviewer dJhB). Every host, with or without the branch, already receives the boundary as a point cloud with a one-hot mask of the boundary type and the boundary values. Section 4 now says so (line 309).
  • Stronger baselines (Reviewer dJhB). We trained GeoTransolver and Transolver++ from PhysicsNeMo on the same input, with the same learning-rate search and training budget as every other model (Table 3, line 379). Both do well on test layouts close to the training layouts and lose most of their accuracy on layouts far from them. We also added our branch to GeoTransolver, which raised its test R2R^2 on the far layouts from 0.524 to 0.870.
  • Novelty (Reviewer qmPS). Related Work now has a table comparing what each related operator reads around a query point (Table 1, line 109). A new control shows that taking a fixed number of neighbours from each boundary type matters: with one plain nearest-neighbour search over the whole boundary instead, accuracy drops on every validation layout of the δ≈0.50\delta\approx0.50 split (line 481).
  • Design alternatives (Reviewer xDus). We added a discussion of where the branch output can enter the host, with a small experiment, and of in-context learning (line 316).
  • Metrics (Reviewer dJhB). Appendix G repeats the two main results tables in relative L2L^2 error, and adds the occupied-zone correlation, amplitude ratio and R2R^2. Every conclusion holds.
  • Inference cost (Reviewer qmPS). We timed the branch with different numbers of neighbours. Using 8 instead of 32 supply-vent neighbours cuts the branch's cost by 24% with no measurable loss in accuracy (line 531).
  • Readability (Reviewer xDus). Section 3 now opens with a short description of the problem for readers outside indoor airflow (line 174), and we removed the confusing plus sign from the R2R^2 values in the main results table.

What has not changed. The study still covers one room with one vent count, and we still make no claim in the occupied zone, where the CFD target itself is not stable between solver iterations. We state both limits in the Limitations paragraph (line 524) and discuss them in our reply to Reviewer qmPS.

We answer each review in detail in its own thread.

AuthorsOpenReview ↗

Briefing for the Area Chairs: Summary of Discussion phase

Dear Area Chairs,

Thank you for handling our submission. Below is a short summary of the discussion. All three reviewers responded to our rebuttal.

  • Reviewer xDus (weak accept) asked for a short introduction to the problem, a cost-versus-accuracy discussion of other designs (injection points, in-context learning), and a fix to a confusing plus sign. We made all three changes, with a small experiment on where the branch enters each host. The reviewer said readability improved and that they will raise the relevant scores. They also noted that a zero-initialised input-side entry keeps the ablation exact, and asked us to separate nearest-layout lookup from few-shot neural in-context learning; we agreed and will update Section 4.

  • Reviewer qmPS (weak reject) raised novelty relative to GINO, GNOT, RIGNO and PINTO, the single room, CFD convergence, inference cost and fit for ICLR. We added a comparison table and a control showing that a fixed number of neighbours per boundary type matters (one plain nearest-neighbour search is worse on all 28 validation layouts). Using 8 instead of 32 supply neighbours cuts the branch's cost by 24% with no measurable loss. The reviewer found the results promising and the concerns partially addressed, and kept their score. We answered their two remaining points: the 48,000-iteration figure applies only to the occupied zone, where we make no claim, while the full field is stable between snapshots (ρ\rho = 0.932); and the branch is an add-on that improved four hosts and ShapeNet-Car.

  • Reviewer dJhB (clear reject, now weak accept) asked whether the hosts receive boundary-type information, requested stronger attention baselines, and asked why we use R2R^2. Every host already receives the boundary mask and values, and the paper now states this. We added GeoTransolver and Transolver++ from PhysicsNeMo with the same input, tuning and budget. They match our models on test layouts close to training, but on far layouts reach full-field R2R^2 0.524 and 0.542, against 0.835 to 0.917 for our hosts with the branch. The branch raises GeoTransolver to 0.870; an explicit distance to the supply vents did not help Transolver. Appendix G repeats the main tables in relative L2L^2 error, with the same conclusions. AB-UPT takes coordinates only as input, so it does not fit this task without redesign. The reviewer considered their main concerns addressed and raised their score to weak accept.

All changes are marked in blue in the revised PDF. We kindly ask you to consider our paper, rebuttal and follow-up responses together, including the additional results, bearing in mind that we have responded to every point the reviewers raised. Please feel free to contact us; we would be glad to answer any question about the submission or the discussion.

Finally, we thank all three reviewers for their time and constructive feedback, which has improved the paper, and we thank you for overseeing the discussion. We are grateful for the time and care that reviewers and Area Chairs give to ICLR and to the research community.

Warm regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.