acceptodds
Under review as a conference paper at ICLR 2027

CG-QLEAR: Counterfactual Gain Guided Residual Correction for Vehicle Surface Pressure Modeling

Abstract

Accurate vehicle surface pressure prediction requires resolving local errors while respecting applicable geometric symmetries. Yet a large prediction error or candidate residual does not necessarily indicate a beneficial correction. We introduce CG-QLEAR, a framework that selects residual updates by their predicted counterfactual gains. A candidate generator proposes local pressure updates, and a gain estimator learns the reduction in absolute error obtained by applying each update. Correction decisions and residual values are coupled across eligible mirror pairs, while asymmetric regions remain independent. A fixed point budget controls the selected residual-correction stages before final refinement. We compare structured correction with continued backbone optimization from the same 55 epoch checkpoint. The protocol uses 5,817 training, 1,148 validation, and 1,154 test vehicles from DrivAerNet++, with 32,768 surface points per vehicle. Against 15 additional backbone epochs, CG-QLEAR reduces pooled MSE by 18.48%, pooled MAE by 3.77%, and mean per vehicle maximum absolute error by 12.52%, while using 4.45% less logged optimization and validation time after the shared checkpoint, excluding one-time setup. The same correction interface also improves all five reported metrics on a Transolver backbone, supporting counterfactual gain as a useful criterion for allocating local pressure updates.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Scientific machine learning, surface pressure prediction, reflection symmetry, sparse residual correction, counterfactual gain estimation

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_LjyxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CG-QLEAR, a post-hoc residual correction pipeline on top of a frozen surface-pressure surrogate, evaluated on DrivAerNet++ (5,817/1,148/1,154 vehicles, 32,768 surface points each).

It has four stages: (i) QLEAR, a small pointwise network on 13 hand-crafted features that outputs a signed residual and a locator logit for large errors; (ii) a Gain Gate that regresses the counterfactual gain gi=∣pi−yi∣−∣pi−(yi+δi)∣g_i = |p_i - y_i| - |p_i - (y_i + \delta_i)|, that is, how much applying the candidate reduces the absolute error; (iii) budgeted selection, consisting of a fixed 1% locator-ranked set plus greedy addition of units with positive predicted gain, where "hard-exact" left/right mirror pairs share values and selection decisions through a projection Π\Pi; and (iv) a bounded MLP ("Median Refinement") applied to the remaining points.

The theory is elementary: pair averaging cannot increase MAE/MSE when paired pressures are equal (Lemma 1), selection preserves pair symmetry (Proposition 2), the utility decomposes additively (Proposition 3), and the regret is bounded linearly in the gain estimation error (Theorem 4).

The experiments compare against continued backbone training from the same checkpoint with similar logged time. On AdaField, CG-QLEAR reduces MSE by 18.5%, MAE by 3.8%, global MaxAE by 0.75% and mean per-vehicle MaxAE by 12.5%. On Transolver, MSE drops by 28.7% and MAE by 6.4%.

Critical strengths
  1. Choosing corrections by predicted error reduction, rather than by error or residual magnitude, is a sensible idea and addresses a real-world failure, since large candidates can overshoot or have the wrong sign. The formulation is cleanly justified: Proposition 3 links the gain target to the pre-refinement MAE, and Theorem 4 separates gain estimation error from optimization error. The idea is general and potentially would be useful in other post-hoc refinement settings.

  2. The gains are consistent across two backbones with a small add-on (about 57K parameters). All reported metrics improve on the approx. 1K held-out vehicles for both AdaField and Transolver, and the Transolver experiment includes paired per-vehicle comparisons and bootstrap CIs (excluding zero). The large reduction in per-vehicle MaxAE (12.5% and 20.8%) suggests that the method does fix localized worst-case errors.

Critical weaknesses

1. The baseline and ablations do NOT isolate the claimed contributions, so it is unclear what "really" drives the improvement (causally).

(a) The only control is continued backbone training from epoch 55 (15 more epochs for AdaField, 99 for Transolver). For AdaField, continuation reduces MSE by only about 3%, so the comparison depends on the learning rate schedule and on how converged the backbone is. There is no fully trained backbone, no published AdaField/Transolver result on DrivAerNet++, and no comparison with other recent surface-pressure models. It is therefore impossible to tell whether CG-QLEAR improves on the SOTA or simply compensates for an under-trained checkpoint.

(b) Table 3 adds the stages cumulatively, and most of the gain comes from the first stage (QLEAR: MSE 1689.6 to 1413.8). The Gain Gate adds about 3% and Median Refinement another 3%. In the budget sweep (Table 5), going from the QLEAR-only 1% budget to the best Gate budget changes MSE by only about 1% (1348.6 to 1333.3). The measured effect of the main idea, counterfactual-gain selection, is therefore small. The following ablations are needed, such as but not limited to:

(i) comparing Gain Gate selection vs. selection by ∣δi∣|\delta_i|, by locator score, and at random under the same budget;

(ii) applying the QLEAR residual everywhere without selection;

(iii) reporting Median Refinement alone on the backbone output, and;

(iv) reporting the projection Π\Pi alone, and CG-QLEAR with and without pairing.

Lemma 1 implies that Π\Pi by itself can reduce error, yet Sec. 4.5 contains no quantitative results on the symmetry component.

2. The gain targets and several protocol details weaken the efficiency and fairness claims.

(a) The training gains gig_i are computed on the vehicles that QLEAR was trained on, so they are optimistic compared to test-time gains. Cross-fitting or held-out vehicles would give unbiased targets. This is not discussed, although it directly affects what the Gate learns.

(b) The gains are defined with respect to the raw yy and Πd\Pi d, while the deployed pipeline uses b=Πyb = \Pi y followed by refinement. Proposition 3 and Theorem 4 therefore do NOT describe the deployed pipeline, and the bound 2Kεmax⁡/cmin⁡2K\varepsilon_{\max}/c_{\min} is not empirically checked.

(c) The compute-matching argument is weak. The training time of the Transolver correction route is a "target", not a measurement. Inference latency is compared across different protocols (cached outputs vs. the full pipeline), and the observed slowdown is about 2.3×2.3 \times (AdaField) and 4.9×4.9 \times (Transolver). Matching post-checkpoint training time is also not the most relevant efficiency comparison for a surrogate meant to replace CFD.

(d) The AdaField results come from a single run with no variance or CI. Since every vehicle has the same number of points, pooled and macro MSE/MAE are identical by construction, so Table 2 adds little beyond MaxAE.

(e) Assumption 1 (equal pressure on mirror pairs) is never checked on the data. CFD fields are rarely exactly symmetric (unsteady wake, meshing). The paper should report the fraction of points in hard-exact pairs and the observed ∣pi−pj∣\vert p_i - p_j \vert on those pairs.

Overall, my assessment is between weak and clear rejection. Neither claim is supported by empriical evidence: there are no quantitative results for the partial symmetry component, and the measured contribution of counterfactual-gain selection is small. Together with the limited theory expansion, the paper reads as a well-curated pipeline, not a methodological advance. I would need the ablations in Q2-Q3 and a comparison against converged backbones (Q1).

Questions for authors
  1. What are the test results of AdaField and Transolver trained to convergence with their original schedules, and of other recent surface-pressure baselines on DrivAerNet++? Does your proposed method still improve over a converged backbone?

  2. At a fixed budget, how does Gain Gate selection compare with selection by ∣δi∣\vert \delta_i \vert, by locator score, at random, and with applying QLEAR everywhere?

  3. What is the effect of the symmetry component alone (Π\Pi applied to the backbone output), and of removing pairing from CG-QLEAR? What fraction are in hard-exact pairs, and how large is ∣pi−pj∣\vert p_i - p_j \vert on those pairs in the training data?

  4. Were the Gain Gate targets computed on vehicles used to train QLEAR? If so, have you tried cross-fitted gains, and how well does g^\hat{g} predict gg on test vehicles (either correlation and sign accuracy)?

  5. Can you report the measured training time of the Transolver correction route, and the inference latency of both routes under the same end-to-end protocol?

  6. Please explain why the last stage is called "Median" Refinement, since (as far as I understand) the architecture described is a bounded MLP.

Suggestions: please report seeds & variance for AdaField, and consider normalized or physically meaningful units like CpC_p, so that values such as an MSE around 1,300 and a MaxAE around 12,000 can be compared with prior work.

Reviewer_LWUsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CG-QLEAR to refine vehicle surface-pressure predictions from a frozen backbone. A small residual model proposes corrections, while a Gain Gate predicts their error reduction and selects updates under a point budget. The method also applies partial symmetry constraints and a final MLP refinement. Experiments on DrivAerNet++Surf report improvements over continued training of AdaField and Transolver.

Critical strengths
  1. Reducing local errors in surface-pressure prediction is a practically relevant problem for aerodynamic surrogate modeling. Evaluating the benefit of a proposed correction is also a reasonable idea, since a large original error does not necessarily imply that the correction will help.

  2. The complete pipeline improves results on two backbones. The per-vehicle error analysis and paired bootstrap results on Transolver provide useful evidence for these improvements.

Critical weaknesses
  1. The methodological novelty appears limited, and the value of the main new component is insufficiently established. The framework primarily combines residual learning, learned gating, and symmetry-based post-processing. Most of the sequential MSE reduction in Table 3 occurs before the Gain Gate is added. Without comparisons against a standard residual MLP and simpler selection policies under the same budget and post-processing, it is unclear whether predicting counterfactual gain offers a meaningful advantage.

  2. The gain target does not directly match the deployed correction process. Eq. (8) measures improvement relative to the raw prediction y, whereas deployment uses the symmetry-projected base Pi y. A correction can therefore have zero labeled gain but improve the projected prediction. Proposition 3 is valid for its stated pre-projection setting, but does not resolve this mismatch. The authors should justify this choice or evaluate gains defined on the actual deployed base.

Questions for authors
  1. With the same candidates, point budget, symmetry handling, and final refinement, how does Gain Gate compare with locator-based, residual-magnitude, and random selection? How much improvement remains over a standard residual-MLP correction baseline?

  2. Why is gain supervision computed from y rather than Pi y? Does training the gate using the projected-base gain change the results?

  3. Does the point budget reduce measured inference cost, or only the number of selected updates? The final MLP can modify unselected points, and the Transolver pipeline takes 0.3291 seconds per vehicle versus 0.0667 for the baseline.

Reviewer_P7MHOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper tackled the challenge of missing symmetry in the case of surrogate modeling for vehicle surface pressure prediction on a benchmark car dataset. The authors proposed a method that learns target corrections to a trained model predictions and present it as a better way to use limited additional compute.

Critical strengths
  • The question regarding where a limited amount of additional go is relevant and interesting for surrogate modeling performance improvement.
Critical weaknesses
  • I had a very hard time reading the paper, especially the method section. Even after several readings, the research gap and the reason why the method is needed did not become clearer. The abstract also needs to explain this better. The same applies for the research-gap statements at the end of each related work paragraph.
  • The paper rely mainly on the mirror pairs to justify their method. It is not clear to me why this requires the proposed framework or why a simpler approach would not be sufficient.
  • The method introduces too much notation without necessarily justifying the need for each element and at least justifying this design choice over others. This makes the paper very hard to follow. Equation 2 for example could be said in one line without the complex notation.
  • The QLEAR section introduces the e_i, for training. But it’s still not clear for me how the method is used at inference time. A diagram of the different block would be helpful to understand where are each component taking as input and how they are related.
  • The main baseline comparison is not sufficient. Given the emphasis on mirror pairs, why not learn the backbone model only on half the cars and use symmetry of the predictions and report the different errors.
  • The paper should report the different architecture and optimization choices. From the main paper, it is currently difficult to reproduce the results in the paper.
  • Why is QLEAR trained for 20 epochs, the gain gate for 10 epochs and median refinement for 15 epochs ?
  • Figure 3 should be replaced with an error based plot comparing plot delta in pressure between CFD and the proposed method and between CFD and the backbone model.
  • Figures 1 and 2 do not appear to be referenced in the main text.
  • The Gino reference at lines 39-40 should be put before introducing Transolver.
Questions for authors

See above

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.