- Rating
- 4(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes LatentDDM. It is a local-to-global neural operator framework that can adapt the pretrained PDE surrogate models on small subdomains to a problem domain with new dynamics settings. It first trains a neural operator on local domain and freezes it, then uses a lightweight transformer to compose the local predictions. The method is evaluated on Darcy flow with increasing domain sizes and on fast-pitching air foil flow. LatentDDM on both experiments shows prediction accuracy gain compared against transformer-based full domain baselines.
Critical strengths
1.The local-to-global transfer design is useful. Separating reusable local predictions from a learnable composition enables efficiency gain while freezing local model. In addition in section 3.2, the error decomposition into the error from local solver and the error from global corrections were helpful as a diagnosing perspective.
2.The ablation experiments were helpful in establishing the effectiveness of the method. The comparison experiment in Appendix D.4 is carefully designed, showcasing that composer-only adaptation can efficiently achieve similar observed accuracy with much lower adaptation time.
3.The experiments cover complementary settings: Darcy flow tests generalization to larger domains while airfoil flow evaluates prediction under changing conditions. The result that in low-data regime LatentDDM outperforms baselines was intuitive and validated.
4.The method uses a single learned communication step efficiently instead of repeatedly updating boundary conditions and evaluating local solver.
Critical weaknesses
1.In the Darcy experiment, the 64 labeled target domain simulations were also used despite the m=41664 training simulations. Although these simulations didn’t play a role for gradient updates they still provide substantial target supervision, which should be considered when interpreting the few-shot adaptation results.
2.In the airfoil experiment, the model is adapted using labeled transitions from each OOD trajectory and it is evaluated on rollout windows from the same trajectory. It remains unclear whether the adapted model can generalize to unseen OOD trajectories without further adaptation.
3.The efficiency claim involves many different types of cost. The author claims LatentDDM is faster at inference compared to multi-sweep SNI and requires less adaptation time compared to local adapters. However in Appendix A.1 the method’s Darcy adaptation time on 8,8 and 16,16 domains is about 10 and 44 minutes, while the Patch full fine-tuning method only takes less than one minute and three minutes. Therefore parameter and sample efficiency does not directly translate into wall-clock efficiency. There are also other costs including additional source training stages and PDE-specific engineering costs. The paper should be more careful in making these efficiency claims.
Questions for authors
-
Can the author rerun the m=4 and 16 Darcy experiments with the total budget that includes validation and the model selection? Possible designs can include validating hyperparameters on source or cross validation within the limited target examples etc.
-
Can the authors discuss why there is a large adaptation wall-clock difference when compared to Patch Fine-tuning and Lora? It might be helpful to have a comparison of test error versus actual adaptation time including validation and checkpoint selection.
-
Can the author extend the airfoil adaptation experiment to be evaluated on held out trajectories different from those used for calibration to test cross-trajectory transfer?
5 comments
Question on additional result and scalability
Thanks to the author for the rebuttal effort and additional experiment result, they have addressed most of my concern. However on the new wall-clock experiment it seems to be contradicting appendix A.1’s result of 10 and 45 minutes for Darcy adaptation?
I also have one remaining question left on the method scalability. The composition module uses full attention with quadratic complexity in the number of subdomains. For example the paper evaluates up to 16,16 tiles, however if extend to 64,64 tiles or larger then the attention matrix would have millions of elements and incur substantial computational overhead. Is this a potential scalability limitation?
If these points are sufficiently addressed, I will consider raising my score.
Response for wall-clock clarification and composer scalability
-
Wall-clock clarification. Thank you for pointing out this ambiguity. Appendix A.1 reports the duration of the complete 24-epoch adaptation; approximately 10 min for 8x8 cover and 45 min for 16x16 cover. In contrast, Q2 reports the time budget results. For LatentDDM adaptation, roughly 10min in 8x8 case would be 24-epoch, and 45min in 16x16 case would be 24-epoch. In the new table in Q2, we further extend to a longer adaptation time to see if it helps to lower the error more; where for the PatchFT and PatchLoRA, we also use the same time budget, meaning they use far more epochs than just 24 in Appendix A1, thus all target samples, epochs, and total time's effects are all controlled and compared. The result shows that when abundant time is given, the LatentDDM still have lower error than other methods after adaptation. We will clarify this distinction.
-
Scalability concern. Yes, full attention of the composition module is a scalability limitation. Increasing the domain size to 64x64 would increase the attention cost 256-fold. While it is not within the scope of this paper, since with the local patch we already enjoy a reduced, parallelizable computation pattern in LatentDDM, a natural extension is sparse window attention (SWA) with coarse global communication for the composition module. We will list it as our future work in Conclusions "further study the composition-module design especially under limited adaptation budgets and large-scale datasets".
Thank you for your advices!
Thank you for the helpful comments on adaptation cost and generalization
We thank the reviewer for the helpful feedback on transfer, label accounting, and computational cost.
Below, W denotes weaknesses and Q denotes questions.
W1. Total target labels and trajectory-specific calibration
We agree that validation cost and transfer across trajectories need separate evaluation. Q1 reports sensitivity to the checkpoint-selection budget. Q3 evaluates a shared adapted model on disjoint OOD test trajectories.
W2. Comparator-specific efficiency
We agree that fewer trainable parameters do not imply lower adaptation cost. We will clarify the comparison in the main text with the following sentence:
Q1. Darcy adaptation under a total-label constraint
To isolate validation-set size, we selected checkpoints from the same run using nested sets of 4/16/64 cases and tested the final checkpoint. Reruns use one run per setting and the same 64 test cases per scale; the paper column gives three-seed means.
Entries are test relative L2; bold marks matched training/validation counts. Using 16 rather than 64 validation cases reduces checkpoint-selection labels by 75% with nearly unchanged errors.
Q2. Adaptation wall-clock difference
Training the composer requires gradients through the frozen local operator with respect to its corrected inputs (Appendix D.1). Fewer updated parameters therefore need not reduce adaptation time.
To show the accuracy–time tradeoff, we compared test relative L2 at fixed time budgets with 16 training cases and one run per method/scale. Timings include setup, training, validation, checkpoint writing, and archival overhead.
Zero-minute rows are Table 1 zero-shot references, not rerun initial-checkpoint measurements. Positive-time entries use the best validation-selected checkpoint available, carried forward after completion. At the earliest positive budgets, LatentDDM has approximately 47%/43% lower error than the better Patch control on 8×8/16×16. Further training improves LatentDDM, while Patch runs finish earlier.
Q3. Adaptation across held-out target trajectories
To test transfer to unseen trajectories, we adapted one shared model per method on 16 transitions from each of 16 formerly reserved OOD validation trajectories (256 pairs). Each method used the same data for one adaptation run. Final fixed-duration checkpoints were evaluated on 16 disjoint OOD test trajectories without further updates.
The table reports H20 field relative L2. Shared adaptation improves LatentDDM in all three bins without test-trajectory updates, supporting cross-trajectory transfer. Case-specific calibration remains better in bins 2–3; baseline responses are mixed, indicating that LatentDDM maintains is adaptability in shared adaptation regime as well.