acceptodds
Under review as a conference paper at ICLR 2027

Learning Transferable Aerodynamic Laws across Physical Domains for Heterogeneous Flapping-Wing Systems

Abstract

Flapping-wing flight involves strongly coupled interactions among rapid bilateral wing motion, body dynamics, and unsteady aerodynamic responses. Physical data that reveal how these processes interact are essential for data-driven aerodynamic modeling, system identification, state estimation, reduced-order dynamics, control-oriented analysis, and scientific machine learning. We introduce FlapKAD-State, a state-resolved synthetic dataset comprising 2,000 curated rigid-wing flight episodes and 720,152 valid time steps generated using synchronized rigid-body integration and quasi-steady aerodynamic modeling. Our dataset provides comprehensive bilateral flap and wing-rotation kinematics, together with their corresponding angular velocities, local body linear velocity, body angular velocity, and independently resolved aerodynamic forces for the left and right wings. These variables provide complementary descriptions of bilateral wing kinematics, body motion, and aerodynamic response within a unified episode structure. To facilitate the use and evaluation of the dataset, we provide standardized episode-level partitions, variable-length masks, channel definitions, physical units, coordinate conventions, and data-quality documentation, together with reference benchmarks that illustrate representative learning scenarios.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
transfer learning, flapping-wing aerodynamics, physics-informed machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 5 reviews, above 27% of the others.
Reviewer_YnfUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors have tried to tackle a transfer learning problem in the domain of flapping wings. Such systems are complex as they involve interaction between a moving body whose aerodynamic force generation surfaces are continuously changing configuration unlike fixed wing systems like airplanes. The idea is to use simulation data (which is abundant and cheaper compared to experimental data) to train a model to predict the instantaneous aerodynamics forces and then transfer it to unseen configurations of the flying robots like change in the medium, flapping frequency etc. This paper proposes CLAW, a physics informed transfer learning framework that learns aerodynamic force relationships from simulation and adapts them to experimental flapping wing systems using limited data. Its main strength is improved data efficiency through shared aerodynamic representations.

Critical strengths
  1. The authors aim to tackle a real gap in the domain, where limited experimental data is available.
  2. The authors have used a physics informed model design. The decomposition into shared and platform specific components is motivated by aerodynamic force scaling rather than an arbitrary network architecture.
  3. The authors have demonstrated meaningful gains over training from scratch
  4. The CLAW model used here is compact and fast to train.
  5. One important result identified here is that many domain adaptation methods try to make the source and target feature distributions similar. CLAW argues that this can be physically inappropriate.
  6. Detailed comparisons are provided against other domain adaptation and fine tuning strategies.
Critical weaknesses
  1. The source of the simulation data is not full field CFD simulations. Instead it is a semi-analytical, physics based model, which is comparatively cheap to evaluate. This can limit how well the forces are resolved in the training data.
  2. The authors do not show how well the training data compares against experiment or full field CFD.
  3. The authors transfer the learned model on experimental data, but only uses 183 trajectories due to lack of complete cycles. So the transfer learning part is evaluated only on the highly periodic portion of the experimental data, which may not be fully representative of real world flight performance.
  4. The transfer experiment is very limited. It does not cover multiple geometry configurations. It could partly be due to lack of experimental data. But if available, it will be interesting to see the performance on another set of independent experimental data.
  5. In zero shot test target, CLAW is not the best model, but the authors show it is cheaper in terms of time. But in real world application, it is not clear if that timesaving is relevant or such systems will benefit more from a more accurate model.
  6. The few shot transfer results does not establish the superiority of CLAW over other strategies.
Questions for authors
  1. If the model used in this study for generating simulating data does not require full field CFD calculation, which means it is cheap to evaluate, what is the need for a trained model? Can't we use the cheaper analytical model to predict for unseen scenarios?
  2. Can the authors demonstrate that the filtered target dataset is representative of the complete target dataset?
  3. Why does table 2 and 3 appear before table 1? Is that a formatting error?
  4. Can the authors provide more transfer experiments with additional wing geometries, Re and reduced frequencies?
Reviewer_t2wjOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper solves aerodynamic force prediction task with limited data for heterogeneous flapping-wing systems. The authors propose CLAW, which is a training strategy that firstly pretrains an MLP on synthetic data and then finetunes the model with limited experimental data.

Critical strengths

The authors address a practical problem of aerodynamic force prediction with limited experimental data.

Critical weaknesses

1. Limited novelty and generalizability.

The proposed method is essentially an MLP pretraining and fine-tuning framework. The network architecture and training strategy have limited technical novelty. The theory also doesn't establish a substantially new principle.

More importantly, the experiments only demonstrate transfer between one selected pair of physical systems. The source and target aerodynamic responses already show high correlation after scalar calibration. So the results do not really support the claimed 'generalizable physical-law transfer'. The results are more likely to reflect the benefits of pretraining with more compatible data.

2. Insufficient justification and missing physics-based baselines.

The synthetic training data are generated using an explicit, computationally cheap analytical model of aerodynamics. However, the authors haven't compared their MLP directly against this analytical model to the target operating conditions either with or without physical-parameter calibration. This is an essential baseline for proving the necessity of the pretraining and finetuning a data-driven model. With current results only, whether training an additional neural model makes any sense over the existing simulator is unclear.

In addition, Appendix 13.4 indicates that the simulator configurations were selected based on their correlation with the target aerodynamic data. It's unclear that if this design also involved validation and test set data.

3. Limited performance advantage over existing methods.

CLAW does not show a consistent or substantial improvement over existing fine-tuning strategies. For example, at 128 target trajectories, L2-SP performs better at several intermediate data budgets. In the zero-shot setting, RSD also achieves comparable performance. These results don't show a meaningful advantage of CLAW over existing methods.

4. Inconsistent zero-shot evaluation claim.

This paper claims an R2R^2 of 0.761 without target force labels. However, this result requires fitting the target-domain output calibration parameters using labeled target samples. Without this calibration, the reported R2R^2 is −2.79. So the result is not strict zero-shot transfer. Meanwhile, the number and use of target labels in this evaluation are not sufficiently clarified.

5. Poor presentation quality.

Several figures have poor layout, small text and inconsistent visual styles. The overview figures are cluttered and difficult to read. The overall quality of the figures and manuscript presentation needs significant improvement.

Questions for authors

1 . Why not directly apply the analytical aerodynamic simulator to the target system? How does its performance compare with CLAW, either with and without parameter calibration?

2 . Why generate synthetic data under operating conditions so different from the data of the target system rather than matching the target conditions more closely?

3 . Which target data were used to select the source simulator configurations in Appendix 13.4? Were any validation or test labels involved?

4 . How is the reported result considered zero-shot when output calibration requires target force labels? How many labeled samples were used?

5 . What specific advantage does CLAW show over existing fine-tuning methods such as L2-SP and WiSE-FT?

1 comment

The authors have provided additional justifications, but I don't think they have addressed my concerns. Instead, they add new concerns:

  1. The authors insist on presenting their MLP finetuning strategy as a physically motivated factorization, and claim that this special MLP architecture is key to superior performance. However, results in Table 5 don't support a substantial advantage, where improvements over the unfactorized MLP are only 0.008/0.001 at b=64/128.
  2. The authors provide a new claim of why alignment-based methods fail based on an inequality derinved from Lipschitz continuity. But this doesn't automatically mean CLAW is a better approach.
  3. The authors provide shuffle-label pretraining experiments to justify the importance of CLAW's pretraining. While the results support the benefit of correct source labels, from scratch/shuffled labels = 0.844/0.841 at b=64 and from scratch/shuffled labels = 0.864/0.860 at b=128. It actually means that the full finetuning can almost get rid of the negative impact of shuffled source data. It weakens the claim that CLAW has the physical law transfer capability.
  4. The authors also provide shuffle-condition/phase experiments. At b=128, correct labels/shuffled configuration/shuffled phase=0.917/0.914/0.915. It doubts the importance of the factorized condition and phase information.
  5. The newly provided corrected-label result is 0.917 at b=128, which is different from the reported 0.958 in Table 4. Same gaps exist at other b settings.
  6. For authors reply to W2/Q1, my question was why not construct an analytical aerodynamic model directly for the target system, instead of using an adjusted model with only affine parameters. This does not answer my question nor support the necessity of using a data-driven model for prediction.
  7. For the reply to W2/Q3, the authors confirmed that test labels were used in simulator selection. It's still unclear if the selection process has been revised accordingly.
  8. I cannot see any labeled changes in the current PDF manuscript. I hope the authors simply forgot to upload it.

9. As an invited emergency reviewer, I expect the authors to show at least basic courtesy and respect for the time I spent reviewing this submission on the extremely short notice.

Reviewer_d1Y8OpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This work investigates transferring simulated aerodynamic force predictors to a physical flapping-wing rig to cut target-domain data requirements. CLAW adopts a phase-centric representation and factorizes the model into private input adapters, a shared nonlinear operator and private output calibrators. It outperforms scratch training across target-data regimes and validates simulation pretraining’s data efficiency.

Critical strengths
  • The problem has clear practical value for lowering experimental modeling overhead in flapping-wing control.
  • The design is clean: phase representation plus input/output adaptations target separate cross-domain mismatches.
  • Experiments span varied data budgets, fine-tuning schemes and ablations, with steady gains over target-only training.
Critical weaknesses

Top performance requires full-network fine-tuning and is comparable to generic methods like L2-SP. It is hard to separate gains from shared aerodynamic structure versus general pretraining effects.

Questions for authors
  1. Which parameters are trainable in the adapter-only setting of Table 4? If the shared operator is frozen, how does this result support transferable shared aerodynamic laws?
  2. Are tests conducted on unseen frequencies, amplitudes or phase offsets? What extra advantages does CLAW bring over a standard MLP with nondimensionalized physical inputs?
Reviewer_M4qMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors proposed to do transfer learning on flapping wing systems. The source domain is simulation, and the target domain is experiments. The model is an input encoder and an MLP. The author showed that a simple scaling can effectively make the target mapping transferable, which is the mapping between angle of attack and normalized force.

Critical strengths
  • The phase domain representation and the scaling property are simple and seem to be effective for the transfer learning task.
Critical weaknesses
  • The authors tried to impose latent matching loss which is presented as a significant part of the paper, but it does not seem to work in practice, although some theoretical arguments are provided.
  • It is not clear whether the factorized design (freeze MLP, only fine-tune encoder) is indeed useful.
Questions for authors
  • Is the scaling behavior just because source and target domains used different units or it is something that holds in general?
  • Proposition 1 and 2 introduced some mathematical proofs. What are they trying to show intuitively?
  • Did you try the full fine-tuning? E.g., fine-tune both the adapter and the MLP together, both from the pre-trained weights.
  • Since simulation is easy to get, why is the angle of attach from simulation covering a smaller range than experimental data?
  • The training seems to be very light-weight (seems to take minutes to train). Is using a neural network model necessary?
  • L2-SP seems to work well in Table 4. What is this method?
Reviewer_EuBxOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1(−1)
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes CLAW for aerodynamic force prediction across flapping wing systems. CLAW represents kinematics in stroke phase and factorizes the predictor into a domain-private input map, a shared law operator, and a domain-private output scale. The model pretrains on simulation data and finetunes on expensive to collect experimental data, which show consistent large improvements than training from scratch on the limited experimental data. Authors compare different domain-adaption and finetune methods, show improvements of performance and time cost with using a 12k-param small model.

Critical strengths
  1. The paper is well motivated, aim to address a practical problem where experimental force measurements are expensive. The phase-based representation and factorization provide a physically motivated way to transfer knowledge from simulation to experimental settings.
Critical weaknesses
  1. The evaluation is limited to one simulation-to-experiment transfer setting, and the experiment dataset size further reduced because of the cycle detection. Though there are rich transfer methods experiments, architecture ablation, it's still not well supported whether CLAW is a generally working method.
  2. CLAW is claimed to outperform selected baselines, but comparisons are limited to non-strict zero-shot transfer. There's no clear accuracy improvements over RSD, and the following few-shot experiments are limited to CLAW with different transfer methods. This makes it unclear whether CLAW still outperforms the 5 baselines after fine-tuning.
Questions for authors
  1. If I'm not wrong, Figure 4 is supposed to present "From Scratch" and "CLAW, full FT" numbers in Table 4, but from Figures, "Ours(transfer)" R$^2$ numbers looks lower than those in Table 4, please clarify, there are other formatting issues of the paper can be improved (like Table 2 and 3), please consider fix.
  2. Could authors report the preprocessing cost required to apply CLAW?
  3. Could authors provide training and validation curves to show whether scratch training and fine-tuning converged under the 40-epoch target budget? Given CLAW’s additional source pretraining, would giving scratch training a longer budget gives better results?
  4. See weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.