acceptodds
Under review as a conference paper at ICLR 2027

Operator learning with point-set local transformer

Abstract

Scaling neural operators to high-resolution physical fields on irregular geometries requires efficient processing of large point sets. We introduce the Point-Set Local Transformer (PLT), which combines quadrature-weighted local attention with sampled global communication to capture both local structure and long-range dependencies – without dense point-to-point attention. Its separation of source processing and query decoding enables memory-efficient prediction at independently specified output locations. We evaluate PLT on several benchmarks, including HiLiftAeroML, where full-field aircraft predictions cover 126–151 million native surface points. PLT provides accurate full-field predictions for high-lift aircraft configurations, achieving a 6.47% mean relative error in the pre-stall benchmark. On the Plasticity benchmark, PLT achieves the lowest prediction error among all considered methods, improving the state-of-the-art by 41.46%. On the SuperWing benchmark, PLT achieves a 2.91% mean relative error for the surface pressure coefficient, while chunked full-field inference reduces peak memory by 45.6%. In scaling tests, PLT supports training with 4.19 million input points using 41.4 GiB on a single GPU and achieves the lowest peak memory among the evaluated implementations. These results demonstrate accurate physical-field prediction and memory-efficient evaluation of PLT at industrial resolution.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural operators, point-set local transformer, memory efficiency, large-scale PDE surrogates

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ieNPOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Point-Set Local Transformer (PLT), a neural operator architecture for irregular point sets and high-resolution physical fields, aiming to reduce the computational and GPU memory overhead of large-scale physical field prediction. PLT adopts an encoder–decoder structure and mainly includes the following designs:

  1. Local Weighted Attention: Center points are selected via Farthest Point Sampling (FPS); spherical neighborhoods and center-to-neighbor cross-attention are used to aggregate local information; and geometric positional biases and quadrature weights are introduced into the attention logits.
  2. Global Information Interaction: Low-dimensional global token aggregation based on Efficient Attention is used to enable cross-region information propagation while avoiding a full point-to-point attention matrix.
  3. Independent Query Decoding: The encoder produces fixed center-point representations, and the decoder independently predicts outputs for arbitrary query locations. Since there is no direct interaction among query points, the model can use chunked inference to reduce peak GPU memory without changing the prediction results (within numerical precision).
Critical strengths
  1. A practically valuable architecture design, especially low-memory prediction for large-scale output fields. PLT combines point-set local geometric aggregation, global communication, and independent query decoding, enabling the encoded representation to be reused across different output queries. Compared with architectures that must jointly process all output points, this design offers clear practical advantages in high-resolution physical field prediction. In particular, on SuperWing, after partitioning the complete output into multiple query chunks, peak GPU memory decreases from 0.591 GiB to 0.322 GiB, while the prediction results remain consistent within numerical tolerance. The authors also report a corresponding increase in inference latency, making the trade-off between memory and speed relatively transparent. This design is of practical significance for scientific computing tasks whose output grids are much larger than the model's internal representation size.

  2. The experiments cover a relatively broad range and include large-scale aerodynamics tasks of practical relevance. The paper evaluates not only standard operator learning tasks such as Elasticity and Plasticity but also validates physical field prediction on real 3D complex geometries in SuperWing and HiLiftAeroML. The error reduction on Plasticity is relatively significant, and the pressure prediction results on SuperWing are also competitive. The native-surface full-field prediction on HiLiftAeroML further demonstrates the architecture's ability to handle ultra-high-resolution outputs. Compared with works that only demonstrate synthetic data or smaller-scale PDE benchmarks, this experimental design combining standard tasks with industrial-scale problems has high practical value.

Critical weaknesses
  • HiLiftAeroML lacks baseline comparisons under a common evaluation protocol.

    HiLiftAeroML is a central experiment supporting the paper’s industrial relevance. However, the results from prior work shown in Figure 12 differ from those of PLT in terms of data splits, evaluation locations, and error-weighting schemes. Although the authors acknowledge these limitations, the comparison is still insufficient to establish PLT’s relative performance. At a minimum, several representative baselines should be evaluated directly using the same data splits, input sampling, and evaluation metrics.

  • The core novelty needs to be clarified.

    The main components of PLT are closely related to existing methods such as PointNet++, Efficient Attention, OFormer, UPT, and SMART, and its contributions appear to lie primarily in technical integration and efficient implementation. The authors should further clarify which design choices are critical to the performance gains and what distinctive advantages PLT offers over existing local–global hybrid architectures. Comparisons and discussion with other local–global multiscale architectures are also missing from the baseline evaluation.

  • The independent contribution of quadrature weighting remains unclear.

    Table 5 removes both the quadrature weights and the relative positional bias simultaneously, making it impossible to distinguish their individual contributions. Separate ablations are recommended. Moreover, since the primary motivation for this mechanism is to reduce bias caused by nonuniform sampling, its benefits would be more appropriately evaluated under varying sampling densities, local oversampling, and cross-discretization settings. Appendix C states that the downsampled center points directly inherit their original weights. However, FPS changes the spatial distribution of the points, so the authors should also explain how these inherited weights continue to provide a reasonable quadrature approximation at the center level.

  • The conditional modulation mechanism lacks experimental validation.

    The paper uses adaLN-Zero to incorporate temporal or physical conditions but does not compare it with simpler alternatives, such as concatenating the conditioning variables with the input features. It therefore remains unclear how much this design contributes to performance.

  • The computational cost comparison remains incomplete.

    The model uses FPS and neighborhood construction, but the scalability experiments adopt uniform random sampling and exclude offline sampling and neighborhood search from the timing measurements. Consequently, the reported runtimes do not fully reflect the cost of using the model on real geometries. The authors should report geometry preprocessing times for actual tasks and distinguish between processing a new geometry for the first time and reusing an existing geometry. In addition, baseline GPU memory measurements are missing from Table 6, preventing a comprehensive assessment of PLT’s overall resource advantages. The reported TFLOPs for PLT differ from those of the baselines by an order of magnitude. The authors should standardize the FLOP-counting methodology or explicitly clarify where the figures are not directly comparable. They should also distinguish whether reductions in arithmetic cost arise from architectural design, compression budgets, or differences in the scope of the computation being counted.

  • Sampling budgets and accuracy–cost trade-offs have not been validated on realistic aerodynamic tasks.

    Figure 7 measures only the computational cost on synthetic point clouds, while the ablation on Elasticity is insufficient to establish how accuracy changes on complex aerodynamic tasks. The authors should vary the number of input points, the number of centers, and the neighborhood budget on SuperWing or HiLiftAeroML, while jointly reporting error, GPU memory usage, and runtime. This would help validate the accuracy–cost trade-off and robustness to changes in sampling density.

Questions for authors

If the authors can address my core concerns, I will consider raising my score.

4 comments
AuthorsOpenReview ↗

Response to Reviewer ieNP - Part 1 of 2

We thank the reviewer for their careful review and for recognizing our method's practical value.

ieNP-W1 — HiLiftAeroML comparisons under a common evaluation protocol

We agree that the originally cited results do not provide a direct benchmark comparison. We have therefore retrained AeroJEPA on the official 4∘4^\circ split with a comparable model size and aligned training settings, and evaluated both models on the same complete native surfaces using identical area-weighted metrics. PLT achieves lower mean errors across all four reported accuracy metrics. The results and protocol are provided in our response to hHD9-W1.

ieNP-W2 — Novelty and comparisons with local–global and multiscale architectures

PLT’s contribution is a geometry-anchored local–global operator that combines quadrature-weighted local attention with sampled global communication while retaining spatially resolved center states for independent query decoding. As such, PLT really provides a novel approach for operator learning. In more detail:

Although PLT shares geometric grouping with PointNet++, neighborhood aggregation is performed through quadrature-weighted, content-dependent attention rather than pooling. Unlike the standard Perceiver-pooled representation in UPT, PLT retains physical center states for local processing and decoding. Unlike SMART’s dense geometry-to-latent cross-attention, PLT separates bounded local interactions from compact global communication. Erwin captures long-range dependencies through hierarchical coarsening, refinement, and cross-ball interactions. PLT instead combines overlapping radius-based 1×K1\times K cross-attention with an explicit sampled global mixing path. Independent query evaluation is shared with OFormer, UPT, and SMART and is not claimed as a standalone novelty.

Our comparisons also include multiscale architectures. These comprise FigConvUNet on SuperWing, U-FNO and U-NO on Elasticity and Plasticity, and HT-Net on Plasticity.

ieNP-W3 — Quadrature ablations, sampling robustness, and inherited center weights

The separate ablations reported in our response to Dhxf-W1 demonstrate an additional accuracy benefit from quadrature weighting beyond the learned geometric bias on both Elasticity and SuperWing. These experiments isolate the effect of weighting at fixed sampling budgets, but do not separately evaluate local oversampling or cross-discretization.

Regarding center weights, as discussed in our response to Dhxf-Q1, we retain the original weights as relative geometric weights associated with the selected locations. A center’s expanding receptive field describes where its features gather information, rather than a physical cell assigned to that center. Since neighborhoods overlap, their areas cannot simply be assigned as distinct center-cell areas. Nevertheless, FPS changes the sampling distribution, and inherited weights are not a recomputed quadrature rule for the coarsened point set. We will clarify this distinction in Appendix C.

ieNP-W4 — Validation of adaLN-Zero conditioning

Our choice of adaLN-Zero follows Peebles and Xie (2023), who found that it outperformed cross-attention and conditioning through additional input tokens in diffusion transformers. We have now compared adaLN-Zero with direct concatenation of angle of attack and Mach number on SuperWing, keeping the backbone, sampling budgets, and training settings fixed. The table reports mean test relative L2L_2 errors.

ConditioningCpC_pCf,τC_{f,\tau}Cf,zC_{f,z}
adaLN-Zero0.02910.02730.0638
Concatenation0.03520.03220.0735

adaLN-Zero achieves lower errors for all three quantities, supporting its use in PLT. These results will be included in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer ieNP - Part 2 of 2

ieNP-W5 — Geometry preprocessing, GPU memory, and FLOP comparisons

For fixed geometry and sampling, FPS and source neighborhoods can be precomputed and reused across epochs and predictions. We profiled three HiLiftAeroML test geometries on an NVIDIA A100 80GB PCIe, reporting medians over three repetitions. Input preparation (including FPS and neighborhood searches) takes 0.350–0.367 s for 1,048,576 sampled inputs and 8,192 centers (K=48K=48). Full-field query-neighborhood construction takes 63.75–73.09 s for 126–140 million output locations in 8,192-query chunks. Peak allocated GPU memory is 0.141 GiB for preparation and 0.115 GiB for query processing. These timings will be added to the revised manuscript.

The baseline results in Table 6 are taken from Giral et al. (2026, Table 2), who did not report GPU memory usage. Accordingly, we leave these entries unreported.

Our FLOP counting follows the released AeroJEPA evaluator's forward-inference convention (NVIDIA PhysicsNeMo): two operations per multiply–add, including PLT's fused local-attention arithmetic. Model sizes and compression budgets differ: our instantiation of AeroJEPA using Giral et al.'s Table 3 dimensions has 11.81M inference parameters, versus 1.31M for PLT. PLT's compact architecture achieves lower pressure error in Table 6 with substantially fewer parameters and operations, highlighting its architectural effectiveness. We will clarify this distinction in the revision.

ieNP-W6 — Accuracy–cost trade-offs across sampling budgets

We have extended the SuperWing evaluation to four center budgets and three neighborhood caps, using the same backbone, training settings, and 32,768 input points. These experiments show that reducing the neighborhood cap provides substantial memory and runtime savings with modest changes in error. The center sweep follows the density-based radius rule and adjusts the global-source budget with the center count, whereas the neighborhood sweep changes only KK.

The table reports mean test relative L2L_2 errors across all 2,871 fields using the final epoch-200 checkpoints, together with one-pass inference costs measured on an NVIDIA A100 80GB PCIe GPU. We follow the inference profiling protocol in the SuperWing appendix, reporting peak allocated GPU memory and median complete-field forward time.

Centers NcN_cKKCpC_pCf,τC_{f,\tau}Cf,zC_{f,z}Peak memory (GiB)Forward time (ms)
1,024480.03900.03650.08340.54534.08
2,048480.03500.03420.07750.55135.76
4,096480.03380.03230.07430.56534.82
8,192120.03100.02840.06690.21822.46
8,192240.03030.02790.06640.34227.00
8,192480.02910.02730.06380.59136.70

At 8,192 centers, reducing KK from 48 to 12 lowers peak memory by 63.1% and forward time by 38.8%, while increasing pressure error by 6.7%. With K=24K=24, the corresponding memory and time reductions are 42.1% and 26.4%, with a 4.2% increase in pressure error. Both reduced-neighborhood configurations retain pressure-prediction accuracy competitive with the strongest published baselines in Table 6. Increasing the center budget improves all three predicted quantities, while reducing centers alone provides smaller resource savings. These accuracy and resource results will be included in the revised manuscript.

The SuperWing sweeps hold the input count fixed. To examine input-resolution dependence separately, we have prepared a HiLiftAeroML study comparing fixed-center and scaling-center policies across different input resolutions, using the same independently sampled test queries for every configuration. Results from this separate study will be added upon completion.

References

William Peebles and Saining Xie. Scalable Diffusion Models with Transformers. ICCV, 2023.

Francisco Giral et al. AeroJEPA: Learning Semantic Latent Representations for Scalable 3D Aerodynamic Field Modeling. arXiv:2605.05586, 2026, version 1.

NVIDIA PhysicsNeMo. AeroJEPA SuperWing FLOP evaluator. Released implementation.


We thank the reviewer once again for their detailed feedback and the time devoted to our work. We would be pleased to clarify any remaining questions during the rebuttal phase.

Thank you for the additional relevant experiments and supplementary works. My questions have been partially addressed, and I will raise my score to 3.

We thank the reviewer for recognizing the value of our contribution. The newly completed HiLiftAeroML comparisons show that PLT outperforms the retrained AeroJEPA and Transolver-3 baselines at 4° and 12°, while achieving the lowest peak allocated GPU memory in our matched profiling. These results and the completed sampling-budget studies are presented in our comment, Summary of revisions and new results, and are detailed in the revised manuscript, with changes highlighted in blue.

Reviewer_hHD9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Point-Set Local Transformer (PLT), a neural operator architecture for high-resolution physical field prediction on irregular geometries, aiming to effectively capture local geometric features and global dependencies while reducing the computational cost of large-scale point-set modeling. Specifically, PLT employs quadrature-weighted local attention to aggregate information from geometric neighborhoods and sampled global communication to capture long-range dependencies without dense point-to-point attention. In addition, the architecture separates source encoding from query decoding and uses query chunking to reduce peak GPU memory usage during large-scale field prediction. Experiments on multiple benchmarks demonstrate competitive predictive performance, while computational scalability tests highlight PLT's potential for handling large-scale physical field prediction tasks.

Critical strengths

S1. By combining local cross-attention with global information exchange, PLT captures local geometric features and long-range physical dependencies while controlling computational costs.

S2. Furthermore, by separating source encoding from query decoding, PLT enables predictions at independently specified output locations, while query chunking improves its ability to handle large-scale physical field prediction. Given the computational challenges posed by high-resolution datasets such as HiLiftAeroML, this work provides a promising approach to large-scale physical field modeling.

Critical weaknesses

W1. The results on HiLiftAeroML demonstrate PLT's potential for industrial-scale applications. However, the paper lacks direct baseline comparisons under consistent experimental settings. Specifically, PLT is evaluated using separate models for three fixed angles of attack, whereas the reported baseline results are obtained under different data splits and evaluation locations. These differences prevent a direct assessment of PLT's relative predictive performance. Furthermore, the absence of comparisons with recent scalable neural operators such as Transolver-3 limits the completeness of the experimental evaluation, particularly regarding computational scalability.

W2. The ablation studies are primarily conducted on the Elasticity benchmark, without corresponding component-level ablations on large-scale tasks involving complex geometries, such as SuperWing and HiLiftAeroML. Given the relatively small point-set size of the Elasticity benchmark, the current ablation results are insufficient to fully establish the effectiveness of the proposed components in large-scale, complex-geometry scenarios. Further ablation studies on these challenging benchmarks would strengthen the experimental validation.

Questions for authors

Q1. In the computational scalability experiments, the compared models use different memory optimization and execution strategies. Specifically, PLT employs techniques such as active-source compaction and activation checkpointing, while some baselines do not use comparable optimizations. Could the authors clarify the rationale for these experimental settings?

Q2. Furthermore, how do the current experiments distinguish the scalability benefits of PLT's architectural design from those introduced by additional implementation-level optimizations?

2 comments
AuthorsOpenReview ↗

Response to Reviewer hHD9 - Part 1 of 2

We thank the reviewer for their constructive feedback and for recognizing PLT's potential for large-scale physical-field prediction.

hHD9-W1 — Matched HiLiftAeroML baseline

To the best of our knowledge, at the time of submission, our work was the first to report results on HiLiftAeroML using its official single-angle train–test splits.

We acknowledge that differences in data splits, evaluation locations, and metric weighting prevent a direct comparison with the AeroJEPA results cited in our submission. To address this, we have trained AeroJEPA on the official 4∘4^\circ split with the same surface CpC_p target as PLT and a comparable model size. Both use 1,000 training epochs, the same AdamW optimizer settings and warmup–cosine learning-rate schedule, and 8,192 reconstruction queries per training case. AeroJEPA retains its additional auxiliary training objectives.

Both models are evaluated on the complete native surfaces of the same 36 test geometries using identical area-weighted metrics. The table reports mean errors across these geometries.

ModelRelative L2L_2Relative L1L_1RMSEMAE
PLT0.06470.03440.03000.0108
AeroJEPA (retrained)0.07230.05120.03290.0159

PLT achieves lower mean errors across all four metrics in this comparison. We have also included Transolver-3 in the controlled computational scaling study, with results reported in our response to Dhxf-W2.

hHD9-W2 — Component ablations on complex aerodynamic geometries

The completed SuperWing ablations support the complementary contributions of local aggregation and global communication on complex aerodynamic geometries. We use the training and evaluation settings described in the SuperWing appendix, with only the indicated components removed. The table reports mean test relative L2L_2 errors at the final training epoch.

VariantCpC_pCf,τC_{f,\tau}Cf,zC_{f,z}
Full PLT0.02910.02730.0638
Local only, without global mixing0.10250.07550.1703
Global only, without local attention0.06060.05770.1250

The local-only and global-only variants have approximately 3.53 and 2.09 times the pressure error of full PLT, respectively. Separate and joint ablations of quadrature weighting and geometric bias are reported in our response to Dhxf-W1. Together, these results extend the component-level evidence to SuperWing and will be included in the revised manuscript's ablation study.

AuthorsOpenReview ↗

Response to Reviewer hHD9 - Part 2 of 2

hHD9-Q1 — Rationale for model-specific memory optimizations

The original experiments measured complete training-step costs under the execution policies documented in Section 4, retaining the memory-efficient attention implementations provided in the compared models' respective released codebases. We agree that the different checkpointing policies limit attribution of the measured memory differences to architectural design alone.

PLT's active-source compaction removes only source states unused by the selected local neighborhoods, retained centers, and sampled global paths. It preserves the model's computational dependencies and does not introduce additional subsampling. This optimization exploits PLT's sparse connectivity, whereas other architectures retain their own execution strategies.

The controlled checkpointing experiments reported in our response to Dhxf-W2 address the difference in checkpointing policies.

hHD9-Q2 — Architectural benefits versus implementation optimizations

PLT's architectural design enables its efficient implementation. As described in Section 2, bounded local interactions, sampled global communication, and center-based processing restrict computational dependencies. This structure permits active-source compaction without discarding information used by the selected operator. Independent query decoding further enables chunked evaluation while preserving predictions. These opportunities are architecture-dependent. For example, the same source compaction is unavailable when every source point contributes to every update, and independently evaluating query partitions cannot generally preserve predictions when queries interact with one another.

The controlled experiments in our response to Dhxf-W2 demonstrate that PLT's large-scale efficiency persists without activation checkpointing. With checkpointing disabled for every model, both PLT variants use substantially less peak GPU memory than the Transolver variants and AeroJEPA at shared feasible input sizes. At 262,144 input points, where all models are feasible, fixed-center PLT also achieves the lowest peak memory and training-step time among all compared models. These results support the effectiveness of PLT's architectural choices in enabling efficient execution without relying on activation checkpointing.


We thank the reviewer for their careful assessment and constructive recommendations. We look forward to continuing the discussion.

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

This paper investigates how to predict high-resolution physical fields on irregular geometries and proposes the Point-Set Local Transformer (PLT). The model first aggregates information from input points to a sampled center point via local attention, then establishes long-range connections using global information extracted from the sampled point.

Experiments on Plasticity, Elasticity, SuperWing, and HiLiftAeroML demonstrate that PLT achieves good prediction performance across a variety of tasks. Additionally, the model can process output positions in batches, reducing GPU memory usage while maintaining prediction results largely unchanged, offering practical value in balancing prediction accuracy and memory requirements.

Critical strengths
  1. Good performance, and the method is indeed applied to data of practical scale. On the Plasticity task, the error is 6.44×10−46.44×10^{−4}, roughly 41%41\% lower than the strongest baseline in Table 2, which is a clear improvement and one the paper highlights. The authors also predicte pressure across the entire aircraft surface on the HiLiftAeroML dataset, where each sample requires 126–151 million outputs, yet peak GPU memory stayed below 1.30 GiB. Being able to run predictions at this scale is a solid achievement.

  2. The study clearly demonstrates how much GPU memory batch prediction saves and what the trade-offs are. Across all 2,871 test cases in SuperWing, the maximum difference between batch and single-pass predictions is only 4.77×10−74.77×10^{−7}. The measurements in Table 9 show that peak VRAM usage dropped from 0.591 GiB to 0.322 GiB, while the median execution time increases from 36.70 ms to 51.36 ms. When GPU memory is insufficient, spending a little more time yields nearly identical prediction results. This is particularly useful for users.

Critical weaknesses
  1. The respective roles of the quadrature weights and geometric bias terms are not yet clear. Figure 9 shows the results when both of these design elements are removed. While the error increases, it is not possible to determine which term was primarily responsible for the improvement. It is recommended to conduct tests by removing the quadrature weights and geometric bias separately to more clearly illustrate their respective contributions.

  2. The comparison of training GPU memory usage employed different memory-saving settings. In Figure 7, PLT and Transolver++ used activation checkpointing, whereas the other comparison models did not. Therefore, the difference in GPU memory usage cannot be entirely attributed to model design. It is recommended to supplement the analysis with comparisons of memory usage and runtime after uniformly enabling or disabling activation checkpointing, to determine whether PLT's advantages still hold under identical settings.

Questions for authors
  1. After a center point aggregates information from surrounding points, why do subsequent calculations still use the area of the original small block it corresponds to, rather than re-estimating the area it now represents?
  2. Could the experiments in Table 5 be broken down to remove only the quadrature weights or only the geometric bias, respectively, to see how much each factor contributes?
  3. Are there comparisons on SuperWing under the same training settings? This would help determine whether the lower error in Table 6 stems primarily from model design or differences in training settings.
2 comments
AuthorsOpenReview ↗

Response to Reviewer Dhxf - Part 1 of 2

We thank the reviewer for their thoughtful review and for recognizing PLT's accuracy and practical scalability.

Dhxf-W1 — Separate contributions of quadrature weighting and geometric bias

We have conducted separate ablations of quadrature weighting and geometric bias on both Elasticity and SuperWing. Each component plays an important role in improving accuracy, with geometric bias providing the larger contribution in these experiments. Within each benchmark, all variants retain the same data split, sampling budgets, optimizer, training duration, and random seed.

The table reports mean test relative L2L_2 errors at the final training epoch.

VariantElasticitySuperWing CpC_pSuperWing Cf,τC_{f,\tau}SuperWing Cf,zC_{f,z}
Full PLT0.00510.02910.02730.0638
Without quadrature weighting0.00560.03030.02870.0668
Without geometric bias0.00730.04300.04700.0911
Without both0.00830.04010.04160.0870

Quadrature weighting accounts for the area or volume associated with each source point. Quadrature ablation removes weighting from both local and global attention. The learned geometric bias adjusts local attention scores using radius-normalized relative displacement and distance. The geometric-bias ablation removes these score terms while retaining spatial neighborhoods and coordinate inputs.

On Elasticity, removing geometric bias increases the error by approximately 44%, compared with 10% when removing quadrature weighting. On SuperWing, the corresponding error increases reach approximately 72% and 5%, respectively, across pressure and both skin-friction components. Thus, geometric bias contributes to the larger accuracy improvement in these runs, while quadrature weighting provides an additional benefit in the full model.

We have also completed a full SuperWing component ablation study, including separate removal of the local and global paths, as reported in our response to hHD9-W2. These completed results will be added to the ablation study in the revised manuscript.

Dhxf-W2 — Controlled scaling with checkpointing enabled and disabled

We have repeated the scaling experiments with activation checkpointing uniformly disabled and uniformly enabled across all models, following the profiling protocol in Section 4 and details provided in Appendix H. Other execution settings remain fixed within each model's off/on pair, with architecture-specific checkpoint boundaries. We also included Transolver-3 in response to the request in hHD9-W1.

For the models included in the comparison, we retain the memory-efficient attention implementations available in their released code. In particular, UPT, Transolver++, and Transolver-3 use PyTorch’s scaled dot-product attention, with the memory-efficient backend selected in our profiling runs. The attention implementations and backend settings remain unchanged between the checkpointing-enabled and disabled experiments.

The table below reports median peak allocated GPU memory and training-step time over three repeats at 262,144 input points, the largest tested size where all models fit without checkpointing. Step time includes the forward pass, loss computation, backward pass, and optimizer update.

ModelPeak memory (GiB)Checkpointing off\substack{\text{Peak memory (GiB)}\\\\\text{Checkpointing off}}Peak memory (GiB)Checkpointing on\substack{\text{Peak memory (GiB)}\\\\\text{Checkpointing on}}Step time (s)Checkpointing off\substack{\text{Step time (s)}\\\\\text{Checkpointing off}}Step time (s)Checkpointing on\substack{\text{Step time (s)}\\\\\text{Checkpointing on}}
PLT, scaling centers8.203.060.1900.295
PLT, fixed centers4.692.930.1510.228
UPT5.985.070.2640.344
Transolver75.7913.760.7420.988
Transolver++75.9825.140.7901.095
AeroJEPA25.9310.110.2840.416
Transolver-375.0311.940.8261.007

With checkpointing enabled, both PLT variants use less memory than every baseline at shared feasible input sizes and support 4.19 million inputs. Without checkpointing, fixed-center PLT still supports 4.19 million inputs using 41.33 GiB, whereas the largest feasible baseline reaches 2.10 million. UPT uses less memory than scaling-center PLT in the disabled setting.

These results demonstrate that PLT's large-scale efficiency does not rely solely on activation checkpointing. We will add the memory and runtime curves for both policies to the revised manuscript and qualify the memory comparisons accordingly.

AuthorsOpenReview ↗

Response to Reviewer Dhxf - Part 2 of 2

Dhxf-Q1 — Why retain the original quadrature weights at centers?

We distinguish the neighborhood radius from the point-associated quadrature weight. After downsampling, we increase the radius to account for the larger spacing between centers. The quadrature weights are inherited from the selected source points as a design choice. Each center retains its spatial location while its learned features incorporate information from increasingly broad regions through successive blocks. The expanding receptive field describes the region from which a center gathers information, rather than a larger physical cell assigned to that center. Since different centers can gather information from overlapping regions, their receptive fields do not partition the domain into cells with distinct areas. We therefore retain the original weights as relative geometric weights rather than assigning the area of the entire neighborhood to each center.

Dhxf-Q2 — Separate quadrature-weight and geometric-bias ablations

We have conducted separate ablations of quadrature weighting and geometric bias on both Elasticity and SuperWing, as reported in our response to W1. Removing either component increases the error relative to full PLT, with geometric-bias removal producing the larger increase. We will include these separate ablations alongside the joint-removal results in the revised manuscript.

Dhxf-Q3 — Comparability of SuperWing training settings

We selected PLT's training settings to align with the reported SuperWing setup in terms of epoch budget and learning-rate schedule. PLT follows the official SuperWing train–test split and uses the same 200-epoch training budget reported by Giral et al. for AeroJEPA and its baselines. PLT and AeroJEPA also use a learning rate of 0.001 with warmup and cosine decay, and predict the same pressure and skin-friction quantities conditioned on geometry, angle of attack, and Mach number. PLT contains 1.31M parameters, compared with 13.21M during training and 11.81M at inference in our instantiation of the released AeroJEPA implementation using the architectural dimensions in their Table 3.


We appreciate the reviewer's time and helpful suggestions. We welcome any further questions or comments during the rebuttal discussion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions and new results

We thank all reviewers for their careful assessment and constructive feedback, which have helped improve the quality and clarity of our work. We have revised the manuscript to incorporate the rebuttal findings and additional experiments completed since our earlier responses.

  1. We have substantially expanded the HiLiftAeroML evaluation with direct full-field comparisons against retrained AeroJEPA and Transolver-3 baselines at 4°, 12°, and 22°, using the official splits, comparable model sizes, and aligned training settings. PLT outperforms both baselines at 4° and 12°. At 12°, PLT reduces the mean relative L2 error by approximately 37% compared with AeroJEPA and 34% compared with Transolver-3. PLT also achieves lower error than AeroJEPA in the challenging post-stall regime at 22°. These results, reported in Table 1, provide strong additional evidence of PLT’s accuracy at industrial scale.

  2. We now report peak GPU memory, FLOPs, and complete-field forward time for all three models under a common profiling protocol on the same test geometry and GPU. PLT uses the lowest peak allocated GPU memory, 0.191 GiB, for a geometry with approximately 147 million output locations. The inference settings are documented in Appendix F, alongside the separate geometry-preprocessing measurements reported during the rebuttal.

  3. The previously pending HiLiftAeroML sampling-budget study is now included in Appendix G. It examines center and neighborhood budgets and isolates the effect of input resolution. With the centers, global sources, radii, and neighborhood cap fixed, reducing the input count fourfold increases the mean relative L2 error by only 1.9%, showing that PLT retains comparable accuracy with fewer input points.

  4. We have incorporated the Elasticity and SuperWing component ablations, SuperWing accuracy–cost and conditioning studies, and controlled scaling experiments with activation checkpointing enabled and disabled, including Transolver-3. A short ablation summary has been added to the main text, with detailed studies in the appendices.

We have updated the manuscript PDF, with revisions highlighted in blue. We thank the reviewers once again for their time and valuable contributions to improving the manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.