- 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 , 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 ; 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
-
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.
-
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 , 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 alone, and CG-QLEAR with and without pairing.
Lemma 1 implies that 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 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 and , while the deployed pipeline uses followed by refinement. Proposition 3 and Theorem 4 therefore do NOT describe the deployed pipeline, and the bound 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 (AdaField) and (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 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
-
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?
-
At a fixed budget, how does Gain Gate selection compare with selection by , by locator score, at random, and with applying QLEAR everywhere?
-
What is the effect of the symmetry component alone ( applied to the backbone output), and of removing pairing from CG-QLEAR? What fraction are in hard-exact pairs, and how large is on those pairs in the training data?
-
Were the Gain Gate targets computed on vehicles used to train QLEAR? If so, have you tried cross-fitted gains, and how well does predict on test vehicles (either correlation and sign accuracy)?
-
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?
-
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 , so that values such as an MSE around 1,300 and a MaxAE around 12,000 can be compared with prior work.