acceptodds
Under review as a conference paper at ICLR 2027

Train Once for Budget-Adaptive Deployment: Learning Nested Sensor Layouts for Physical Field Reconstruction

Abstract

Accurately reconstructing physical fields from sparse observations depends on model capability and sensor placement. In practice, sensor placement must also accommodate budgets that may change with resource availability or sensing requirements. However, existing placement methods typically optimize layouts for a single sensor budget, so budget changes require reoptimizing all sensor locations. We therefore propose Budget-Adaptive Sensor Placement (BASP), a differentiable framework that learns nested layouts across budgets in a single training run. BASP first encodes the physical field prior into spatial features, supplementing reconstruction feedback with explicit structural guidance. Then, ordered sensor queries sequentially generate continuous coordinates by combining cross attention to these features with spatial suppression, forming nested sensor layouts. During training, a pretrained, frozen reconstructor jointly supervises reconstruction at a randomly sampled budget and the maximum budget, encouraging early sensors to remain informative across layout sizes. Experiments on four datasets with four reconstructors show BASP achieves lower reconstruction errors than independent coordinate optimization, random and uniform sampling, D-, A-, and E-optimal placement in most settings. Beyond accuracy gains, BASP supports multiple budgets in one training run and enables rapid budget adjustment without relocating retained sensors.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
physical field reconstruction, sensor placement, machine learning

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_kZ7zOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes BASP, which learns a fixed, ordered sensor layout for physical field reconstruction. Different prefixes support different sensor budgets without relocating retained sensors.

Critical strengths

The evaluation covers four datasets and four reconstructors. The OrderedCoords baseline provides a useful control for assessing PlaceNet beyond the shared training objective and nested structure.

Critical weaknesses

The novelty relative to sequential sensor placement is not sufficiently established. The idea of adding observations in sequence is well established, and a comparison with existing methods that naturally produce nested layouts would strengthen the paper.

The paper also provides limited insight into why joint training sometimes outperforms independent optimization. The current results do not clearly distinguish benefits from network parameterization, optimization, or training across budgets.

Questions for authors

How does BASP differ from existing sequential sensor placement methods, and could a representative nested placement baseline be included? Are the improvements over OrderedCoords consistent across optimization seeds?

1 comment

We thank Reviewer kZ7z for highlighting the evaluation across four datasets and four reconstructors and the value of OrderedCoords in assessing PlaceNet’s contribution. We address the reviewer’s questions below.

Weakness1: The novelty relative to sequential sensor placement is not sufficiently established. The idea of adding observations in sequence is well established, and a comparison with existing methods that naturally produce nested layouts would strengthen the paper.

We agree that sequential sensor addition is an established way to construct nested layouts. BASP’s contribution is to jointly learn a single sequence of continuous sensor coordinates from reconstruction feedback across budgets. PlaceNet generates this sequence using a physical field prior, ordered queries, and spatial suppression. During training, the losses at the sampled and maximum budgets allow early sensor positions to adapt to their contributions to both shorter prefixes and the full layout. At deployment, each budget uses a prefix of the learned sequence.

Table A compares BASP with two nested baselines using the same frozen reconstructor. Nested pivoted QR derives a sensor ordering from a fixed POD basis [1]. FixedCoords (Greedy) uses the same per-budget reconstruction MSE as BASP and optimizes only newly added coordinates at each budget increase, keeping earlier locations fixed. BASP achieves lower relative L2 error than nested pivoted QR and FixedCoords (Greedy) at all budget expansions.

Table A. Relative L2 errors of BASP, FixedCoords (Greedy), and nested pivoted QR.

DatasetSensorsBASPFixedCoords (Greedy)Nested pivoted QRSensorsBASPFixedCoords (Greedy)Nested pivoted QR
Turbulent flow#250.27650.30810.7839#500.17410.19280.6156
Turbulent flow#1000.13360.14150.3644#2000.11250.11830.1421
Cylinder wake#100.01830.02270.2644#150.01680.02070.2253
Cylinder wake#200.01590.01920.1922#300.01510.01740.1112
Heat conduction#100.00200.00200.1384#200.00180.00180.0059
Heat conduction#300.00170.00180.0047#400.00160.00180.0033
Aerofoil flow#80.01520.01910.2523#120.01440.01540.2596
Aerofoil flow#160.01430.01510.1491#200.01400.01450.1406

Weakness2: The paper also provides limited insight into why joint training sometimes outperforms independent optimization. The current results do not clearly distinguish benefits from network parameterization, optimization, or training across budgets.

BASP and OrderedCoords use the same cross-budget objective and nesting constraint. BASP achieves lower errors in all eight comparisons in Table 3, supporting the value of the PlaceNet parameterization under the same objective. Table B shows that independently trained BASP (single budget) also outperforms FixedCoords in all eight reported comparisons, so the benefit of PlaceNet is also observed under fixed-budget training.

Cross-budget supervision allows early sensor positions to be optimized for their contributions to both sampled prefixes and the full layout. The deployment order is learned through this shared reconstruction supervision without explicit importance labels. Removing the maximum-budget loss increases reconstruction error in all eight settings in Table 3, supporting the role of full-layout supervision in learning effective prefixes.

Table B. Single-budget comparison.

Cylinder#10#15#20#30Heat conduction#10#20#30#40
BASP1.5341.4841.4291.297BASP0.34230.27270.25130.2430
FixedCoords1.5041.3881.3021.287FixedCoords0.33900.24970.24800.2391
BASP(single budget)1.4311.3511.2921.241BASP(single budget)0.33830.24700.23450.2296

Question1: How does BASP differ from existing sequential sensor placement methods, and could a representative nested placement baseline be included? Are the improvements over OrderedCoords consistent across optimization seeds?

See our response to Weakness 1 and Table A for the nested baselines. BASP’s reported error is below the five-seed OrderedCoords mean in all eight settings (Table C).

Table C. Relative L2 errors (×10⁻²) of BASP and OrderedCoords.

Cylinder #10#15#20#30Aerofoil #8#12#16#20
BASP1.8301.6791.5931.5071.5241.4411.4291.400
OrderedCoords2.4762.2762.0751.7601.7891.5431.4601.419

[1] Manohar, Krithika, et al."Data-driven sparse sensor placement for reconstruction: Demonstrating the benefits of exploiting known patterns."IEEE Control Systems Magazine38.3 (2018): 63-86.

Reviewer_9dMbOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Budget-Adaptive Sensor Placement (BASP), a differentiable framework that learns nested sensor placements in a single training run for multiple budgets. It uses a placement network that combines physical field prior, spatial suppression and cross attention to generate ordered sequence of sensor locations. The network is jointly optimized across budgets using the reconstruction feedback from a pretrained reconstructor. The nested formulation enables flexible sensor deployment without requiring retraining or relocating. BASP is evaluated on four physical field datasets and four reconstruction architectures. It achieves improved reconstruction performance in most evaluation settings compared with conventional placement strategies and independently optimized layouts for individual budgets. Ablation studies further evaluate the contribution of each component.

Critical strengths
  • This paper proposes a practical reconstruction-aware sensor placement framework for multiple budgets. It learns nested layouts in a single training run, allowing budget changes without re-optimizing sensor placements or relocating existing sensors.
  • The evaluation on four datasets and four reconstructor architectures demonstrates advantage over baselines in most settings.
Critical weaknesses
  • The baselines are not sufficient to demonstrate the advantage of BASP over simpler nested placement strategies. It would be good to compare with a reconstruction-aware greedy baseline that sequentially optimizes newly added sensors while keeping previously selected sensors fixed. It also lacks fixed-budget PlaceNet models trained with the same architecture and similar optimization settings to compare with independently optimized placements.
  • The paper does not sufficiently justify the tradeoff between reconstruction accuracy, training cost, and the benefit of avoiding sensor relocation. The reported computational savings are estimated assuming independent optimization for every integer budget while only four budgets per dataset are evaluated. Given that accuracy improvements over FixedCoords are sometimes marginal or inconsistent, the practical benefits and accuracy tradeoff remain unclear.
Questions for authors
  1. How does BASP compare against an independent reconstruction-aware greedy baseline that sequentially optimizes newly added sensors using the same frozen reconstructor and reconstruction loss?
  2. Since independently optimal sensor layouts at different budgets need not to be nested, could the authors quantify the accuracy tradeoff introduced by this constraint? How does BASP compare with independently trained PlaceNet models for each budget, using the same architecture and comparable training settings?
  3. Equation (7) assigns substantial weight to the max budget loss, but the max trained budgets are not included in the main evaluation. What's the motivation of this weighting strategy and how sensitive is the performance to it, and to the max budget setting?
  4. Could the authors report results over multiple seeds, including standard deviations or confidence intervals, particularly for settings where BASP only marginally outperforms FixedCoords?
  5. In section 3.2 it states the gradients are stopped through both the coverage and occupancy map updates. What is the motivation of this design choice and how does it affect joint optimization of the ordered sensor sequence?
4 comments

We thank Reviewer 9dMb for recognizing that BASP learns nested layouts in one run and improves reconstruction in most settings. We address the reviewer’s questions below.

Weakness1: The baselines are not sufficient to demonstrate the advantage of BASP over simpler nested placement strategies.

Question1: How does BASP compare against an independent reconstruction-aware greedy baseline that sequentially optimizes newly added sensors using the same frozen reconstructor and reconstruction loss?

Question2: Since independently optimal sensor layouts at different budgets need not to be nested, could the authors quantify the accuracy tradeoff introduced by this constraint? How does BASP compare with independently trained PlaceNet models for each budget, using the same architecture and comparable training settings?

Nested pivoted QR uses prefixes of a single pivot ordering computed from a fixed POD basis [1]. To our knowledge, no published reconstruction-aware greedy baseline matches our setting of learning nested nested layouts across budgets through a frozen neural reconstructor. We therefore implemented FixedCoords (Greedy), which starts from the smallest-budget FixedCoords layout, retains all previously selected coordinates, and optimizes only the newly added sensors at each budget increase. With the same frozen VoronoiCNN, BASP yields lower errors than both baselines in all 16 settings (Table A).

We selected Cylinder and Heat with RecFNO because FixedCoords outperforms multi-budget BASP in both settings. To examine the cost of sharing a nested sequence across budgets, we trained the same PlaceNet architecture separately for each budget using same seetings. BASP (single budget) achieves lower error than both FixedCoords and multi-budget BASP in all eight reported comparisons (Table B). Its gains over FixedCoords support the value of the PlaceNet architecture, and its error reduction relative to multi-budget BASP quantifies the observed accuracy trade-off of jointly learning a shared nested sequence in these settings.

Table A. Relative L2 errors of BASP, FixedCoords (Greedy), and nested pivoted QR.

DatasetSensorsBASPFixedCoords (Greedy)Nested pivoted QRSensorsBASPFixedCoords (Greedy)Nested pivoted QR
Turbulent#250.27650.30810.7839#500.17410.19280.6156
Turbulent#1000.13360.14150.3644#2000.11250.11830.1421
Cylinder#100.01830.02270.2644#150.01680.02070.2253
Cylinder#200.01590.01920.1922#300.01510.01740.1112
Heat#100.00200.00200.1384#200.00180.00180.0059
Heat#300.00170.00180.0047#400.00160.00180.0033
Aerofoil#80.01520.01910.2523#120.01440.01540.2596
Aerofoil#160.01430.01510.1491#200.01400.01450.1406

Table B. Single-budget comparison.

Cylinder#10#15#20#30Heat#10#20#30#40
BASP1.5341.4841.4291.297BASP0.34230.27270.25130.2430
FixedCoords1.5041.3881.3021.287FixedCoords0.33900.24970.24800.2391
BASP (single budget)1.4311.3511.2921.241BASP (single budget)0.33830.24700.23450.2296

Weakness2: The reported computational savings are estimated assuming independent optimization for every integer budget while only four budgets per dataset are evaluated. Given that accuracy improvements over FixedCoords are sometimes marginal or inconsistent, the practical benefits and accuracy tradeoff remain unclear.

One BASP run takes less time than the four-budget FixedCoords cost in 14 of 16 settings (Table C). The exceptions are turbulent flow with VoronoiCNN and RecFNO. The full-range FixedCoords estimate covers all supported integer budgets (Table 6). BASP learns layouts for the full supported range in one run, and budget changes require no further placement optimization or relocation of retained sensors.

Table C. Placement optimization time (min).

DatasetBASPFixedCoords one budgetFixedCoords four budgetsFixedCoords full-range budgets
VoronoiCNNCylinder46.9916.6466.56416.00
Turbulent388.8521.5286.085939.52
Aerofoil45.6520.0480.16340.68
Heat117.1440.93163.721678.13
RecFNOCylinder67.0427.24108.96681.00
Turbulent436.9254.37217.4815006.12
Aerofoil62.5930.46121.84517.82
Heat146.8965.04260.162666.64
PhySenseCylinder85.5738.83155.32970.83
Turbulent672.77248.90995.6068696.40
Aerofoil165.3379.97319.881359.43
Heat455.73212.33849.328705.67
POD-FMCylinder87.6039.93159.72998.33
Turbulent814.77329.431317.7290923.60
Aerofoil170.5082.00328.001394.00
Heat479.60222.70890.809130.70

Question3: Equation (7) assigns substantial weight to the max budget loss, but the max trained budgets are not included in the main evaluation. What's the motivation of this weighting strategy and how sensitive is the performance to it, and to the max budget setting?

In the main experiments, we set K to the largest sensor count used to pretrain the corresponding reconstructor. Table D reports BASP’s reconstruction errors at these maximum budgets using the existing checkpoints. Equation (7) takes the arithmetic mean of the losses at the sampled and maximum budgets. The maximum-budget loss supervises all sensor queries at every update and encourages early sensors to remain useful in the full layout. Removing this term increases reconstruction error in all eight settings in Table 3.

Table E reports the relative L2 errors of models trained with the original and reduced values of K, evaluated at the same sensor budgets. In these settings, the larger K supports a wider range of budgets while maintaining comparable reconstruction accuracy at the shared budgets.

Table D. Relative L2 errors (×10⁻²) of BASP at the maximum sensor budgets.

DatasetMax. sensorsVoronoiCNNRecFNOPhySensePOD-FM
Cylinder wake321.5091.5763.2962.390
Turbulent flow30010.12712.64110.1168.279
Aerofoil flow241.3700.1840.6020.489
Heat conduction500.1600.2420.2380.246

Table E. Relative L2 errors (×10⁻²) at shared sensor budgets for the original and reduced K.

Turbulent flow#25#50#100Heat conduction#10#20#30
BASP (K=300)27.6517.4113.36BASP (K=50)0.20030.17550.1658
BASP (K=100)27.5817.5813.17BASP (K=30)0.19770.17520.1652

Question4: Could the authors report results over multiple seeds, including standard deviations or confidence intervals, particularly for settings where BASP only marginally outperforms FixedCoords?

We repeated placement optimization for both BASP and FixedCoords using four initialization seeds. For each dataset and method comparison, we used the same frozen reconstructor and evaluation protocol. In each run, we averaged the relative L2 error over all evaluation samples. Table F reports the mean ± std of these averages across the multiple seeds.

Table F. Reconstruction errors (mean ± std) across multiple initialization seeds.

ReconstructorSensorsBASPFixedCoords
VoronoiCNN#250.276516±0.0001740.307329±0.006807
#500.174147±0.0003200.184364±0.001185
#1000.133261±0.0002480.141337±0.000233
#2000.112822±0.0002160.118122±0.000414
PhySense#250.26598±0.004180.34895±0.04695
#500.15609±0.001340.16273±0.00407
#1000.12433±0.000600.12602±0.00054
#2000.11004±0.000660.11009±0.00065

Question5: In section 3.2 it states the gradients are stopped through both the coverage and occupancy map updates. What is the motivation of this design choice and how does it affect joint optimization of the ordered sensor sequence?

We stop gradients through the coverage and occupancy updates to avoid backpropagating through a chain of recursive map updates. During the forward pass, later queries still use the maps constructed from earlier sensor locations. Detaching the maps removes the indirect gradient path through which earlier queries influence the loss by changing later sensor locations. Each query remains trainable through its own coordinates and the reconstruction loss. The sampled-budget loss supervises the corresponding prefix, while the maximum budget loss supervises all queries at every update.

[1] Manohar, Krithika, et al."Data-driven sparse sensor placement for reconstruction: Demonstrating the benefits of exploiting known patterns."IEEE Control Systems Magazine38.3 (2018): 63-86.

Thank you for the additional experiments. The experiments and explanation mostly address my concerns. Remaining limitations include the narrow coverage of baseline and seed comparisons, untested loss-weight sensitivity, and an inconsistency in the FixedCoords seed statistics. Overall, the rebuttal strengthens the paper. I am raising the score to 3.

Thank you for recognizing the value of the additional experiments and for raising your score. We acknowledge the limited coverage of the baseline and seed comparisons. Due to an oversight, the FixedCoords statistics in Table F included only the four additional initialization seeds. The original seed-42 result is reported separately in Table 1. Thank you for pointing out this inconsistency. We have corrected the text and table labels.

We also ran a loss-weight sensitivity experiment on the heat conduction dataset. As shown in Table G, increasing the weight on the maximum-budget loss slightly decreased the reconstruction error at the maximum budget, while decreasing the weight slightly increased it.

Table G. Sensitivity to the maximum-budget loss weight on heat conduction. Relative L2 errors are reported in units of ×10⁻². The lowest error at each budget is bolded.

Maximum-budget loss weight10 sensors20 sensors30 sensors40 sensors50 sensors (KK)
1/30.19880.17240.16650.16280.1610
1/2 (default)0.20030.17550.16580.16190.1597
2/30.21300.17590.16970.16270.1581
Reviewer_zdu4OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper presents the budget adaptive sensor placement (BASP) method, which learns ordered sequence of sensor positions for physical field reconstruction. The budget is for the deployment of K sensors to use the first K locations, so increasing the budget will not require retraining or moving the exsiting sensors. BASP learns sequences from training set field statisitcs and feedback from reconstruction model under sampling and maximum budget. The proposed method is evaluated on four datasets using foour reconstruction models.

Critical strengths

1.When the sensing budget increases, the cost of moving existing sensors may be high. Learning an ordered sequence can expand the deployment without relocating sensors or retraining the deployment model. BASP solves this problem.

2.Using the same frozen voronoicnn, BASP has lower reconstruction error than greedy fixedcoords and nested pivot QR on all 16 data sets and budget test combinations. This comparison is particularly useful because all three methods must retain previously placed sensors.

3.Table 8 clarifies the trade-off between flexibility and accuracy. In the recfno of cylindrical wake and heat conduction, multi budget BASP tracks fixed coordinates in all eight test budgets, but training BASP separately for each budget will make the result opposite. This shows the precision cost of sharing a layout sequence in these settings.

Critical weaknesses

1.The division of turbulence data is not clear. The paper describes 10000 snapshots for training and testing, while the rebuttal describes 10000 test snapshots. Since the number of data partitions is not provided, nor does it explain how snapshots are divided by time, it is difficult to judge the generalization ability of the learned layout in an unseen flow state.

2.Table 7 compares BASP with nested baseline only under voronoiCNN. Some differences are very small: in the heat conduction experiments of 10 sensors, BASP and greedy fixedcoords reported errors of 0.2003 and 0.2004 respectively. Additional seed experiments did not determine whether this particular difference was repeatable or whether this advantage extended to other reconstructions.

3.The reported training time savings are based on the assumption that fixedcoords will be optimized separately under each supported integer budget. In the settings of each report, a single BASP run takes longer and consumes more memory than a single fixedcoordinates run. Therefore, the actual savings depend on the number of different budgets actually required for deployment.

Questions for authors
  1. Dataset partitioning and generalization: What are the exact training, validation, and testing partitions for all four datasets, particularly turbulent flow? Are the test samples temporally or physically independent of the training samples? For aerofoil flow, are the test operating conditions independent of the training conditions?

  2. Statistical reliability: Have the experiments been repeated with multiple random initializations of PlaceNet and multiple reconstruction sampling seeds? Could the authors provide mean and standard deviation of the reconstruction errors, especially for the comparisons where BASP and FixedCoords achieve very similar results?

  3. Performance across different budgets: Could the authors provide reconstruction error curves over the entire supported budget range rather than evaluating only four selected budgets? Does BASP maintain its performance advantage at intermediate budgets, particularly for the turbulent-flow setting with a maximum budget of 300 sensors?

  4. Effectiveness of the flow-matching objective: How strongly does minimizing the flow-matching velocity prediction loss correlate with the final generated-field reconstruction error? Could the authors provide an analysis demonstrating this relationship and clarify the inference solver and sampling configurations?

  5. Comparison with coordinate optimization: Given that FixedCoords initialized from BASP produces better results than BASP on every evaluated aerofoil budget, how much of the performance difference is attributable to the nested-layout constraint rather than the proposed query and suppression architecture? Could the authors provide comparisons using multiple FixedCoords initializations and more extensive optimization?

  6. Reproducibility and computational efficiency: Could the authors provide the missing implementation details, including the temperature parameter, Gaussian suppression bandwidths, and other suppression hyperparameters? Could they also report actual placement optimization time and GPU memory consumption for PhySense and POD-FM, rather than only VoronoiCNN?

3 comments

We thank Reviewer zdu4 for recognizing the practical value of adapting sensor budgets without relocating retained sensors and the value of learning nested layouts in a single training run. We address the reviewer’s questions and concerns point by point below.

Weakness1: BASP does not consistently outperform this baseline.

Question5: how much of the performance difference is attributable to the nested-layout constraint rather than the proposed query and suppression architecture? Could the authors provide comparisons using multiple FixedCoords initializations and more extensive optimization?"

BASP balances reconstruction accuracy at individual budgets with maintaining nested layouts across budgets, while FixedCoords optimizes each budget independently without this constraint. To examine this trade-off, we trained BASP separately for each budget using RecFNO on cylinder wake and heat conduction, where FixedCoords outperformed multi-budget BASP. Single-budget BASP achieves lower errors than FixedCoords in all reported comparisons (Table A), demonstrating PlaceNet’s effectiveness and illustrating the trade-off between accuracy and nested layouts. For POD-FM, D/A/E’s advantage on turbulent flow at larger budgets may reflect the alignment between POD-based sensor selection and POD-FM’s conditioning on POD coefficients, as discussed in Section 4.1.

FixedCoords (BASP init.) can improve single-budget accuracy by allowing all sensor coordinates to move, providing greater flexibility than nested deployment permits. When preceding-budget BASP sensors are retained, the frozen-prefix control yields higher errors than BASP in all evaluated cases (Section 4.2, Table 3). All our FixedCoords have reached full optimization and convergence by the 200th epoch. Further results using additional random initialization seeds can be found in Table B below.

Table A. Single-budget comparison on RecFNO

Cylinder wake10 sensors#15#20#30Heat conduction10 sensors#20#30#40
BASP1.5341.4841.4291.297BASP0.34230.27270.25130.2430
FixedCoords1.5041.3881.3021.287FixedCoords0.33900.24970.24800.2391
BASP (single budget)1.4311.3511.2921.241BASP (single budget)0.33830.24700.23450.2296

Weakness2: Lack of statistical evaluation and robustness analysis.

Question2: Have the experiments been repeated with multiple random initializations of PlaceNet and multiple reconstruction sampling seeds? Could the authors provide mean and standard deviation of the reconstruction errors, especially for the comparisons where BASP and FixedCoords achieve very similar results?

All reported errors are mean relative L2 errors over the entire test set. The original experiments used initialization seed 42. We additionally repeated BASP and FixedCoords placement optimization with four new initialization seeds. Table B reports the mean and std across all four independent runs, showing limited variation in reconstruction performance in the tested settings. As shown as Table B, BASP exhibits consistently lower mean error and std than FixedCoords across multiple seeds.

For PhySense and POD-FM, all placement methods use the same initial-noise random seeds and identical sampling settings during evaluation. For both Physense and POD-FM, all experimental tests employ an ODE sampler based on the explicit Euler method and 10 sampling steps.

Table B. Reconstruction errors (mean ± std) across multiple initialization seeds.

VoronoiCNNBASPFixedCoords
#250.2765±0.0001740.3073±0.00680
#500.1741±0.0003200.1843±0.00118
#1000.1332±0.0002480.1413±0.00023
#2000.1128±0.0002160.1181±0.00041
PhysenseBASPFixedCoords
#250.26598±0.004180.34895±0.04695
#500.15609±0.001340.16273±0.00407
#1000.12433±0.000600.12602±0.00054
#2000.11004±0.000660.11009±0.00065

Weakness3: Insufficient evaluation of the central budget-adaptive capability.

Question3: Could the authors provide reconstruction error curves over the entire supported budget range rather than evaluating only four selected budgets? ......

We additionally evaluated BASP at a denser set of integer budgets (Table C). All layouts are prefixes of the same learned sequence and require no further placement optimization. FixedCoords comparisons remain at the original budgets, since each additional budget requires a separate optimization run.

In our experiments, K matches the maximum sensor count used during reconstructor pretraining. Changing the deployment budget within this range simply selects a different prefix and requires no retraining. Extending beyond K would require extending PlaceNet and validating the reconstructor over the expanded budget range, potentially followed by fine-tuning or retraining.

Table C. Test results of denser set of integer budgets.

kErrorkErrorkErrorkErrorkErrorkErrorkError
250.2765650.15751050.13211450.12231850.11492250.10952650.1050
300.2438700.15361100.13071500.12121900.11412300.10892700.1045
350.2201750.15031150.12941550.12031950.11332350.10832750.1039
400.2019800.14671200.12801600.11942000.11252400.10772800.1034
450.1865850.14301250.12681650.11862050.11192450.10722850.1029
500.1741900.13961300.12571700.11752100.11122500.10662900.1024
550.1681950.13651350.12441750.11642150.11062550.10612950.1018
600.16301000.13361400.12331800.11572200.11002600.10563000.1013

Weakness4: Incomplete dataset partition and generalization details.

Question1: What are the exact training...... are the test operating conditions independent of the training conditions?

For cylinder wake, aerofoil flow, and heat conduction, we randomly split the samples into 70% for training and 30% for testing.10,000 snapshots of turbulent flow are used for testing, consistent with the approach in existing literature[1-3]. Each aerofoil snapshot corresponds to a distinct combination of angle of attack and inlet velocity, so the held-out samples contain operating-condition combinations absent from training. The sensor layout learned from the training cases is applied unchanged to these test cases, providing evidence of generalization across operating conditions within the benchmark distribution.

Weakness5: The optimization objective for generative reconstructors needs further justification.

Question4: How strongly does minimizing the flow-matching velocity prediction loss correlate with the final generated-field reconstruction error? Could the authors provide an analysis demonstrating this relationship and clarify the inference solver and sampling configurations?"

We use the flow-matching velocity loss because directly optimizing the final reconstruction MSE requires backpropagation through the full multi-step sampling process, incurring substantial memory and computational costs. This choice follows PhySense, which approve the relationship between flow loss and A-optimal conditional variance under its stated assumptions (Section 3.2, Theorem 3.5; Appendix A.2).

Both PhySense and POD-FM use straight interpolation during training. At inference, we integrate the learned velocity field from Gaussian initial noise using 10 explicit Euler steps.

Weakness6: Limited comparison with other reconstruction-aware sensor placement methods.""There is also a potential issue with how the D/A/E results are summarized."

To our knowledge, BASP is the first framework to jointly learn continuous nested sensor layouts across budgets in one training run using feedback from the neural reconstructor. DSPO’s original protocol alternates reconstructor training and coordinate optimization, whereas our comparisons use the same frozen checkpoint. FixedCoords denotes PhySense’s placement strategy. Table 3 includes OrderedCoords to assess PlaceNet’s contribution beyond nested ordering and shared supervision, and a frozen-prefix control to evaluate sensor addition while retaining existing BASP locations. In order to further compare our nested sensor deployment method, we have added a nested FixedCoords based on a greedy algorithm and nested pivoted QR based on fixed POD modes [4], with the results shown in Figure D.

The individual D-, A-, and E-optimality results are provided in Appendix E, Figures 3–6. The three criteria generally yield similar reconstruction errors, so we report only the lowest test error among them for each configuration to save space in the main tables and make comparisons easier.

Table D. Relative L2 errors of BASP, FixedCoords (Greedy), and nested pivoted QR.

DatasetSensorsBASPFixedCoords (Greedy)Nested pivoted QRSensorsBASPFixedCoords (Greedy)Nested pivoted QR
Turbulent flow#250.27650.30810.7839#500.17410.19280.6156
Turbulent flow#1000.13360.14150.3644#2000.11250.11830.1421
Cylinder wake#100.01830.02270.2644#150.01680.02070.2253
Cylinder wake#200.01590.01920.1922#300.01510.01740.1112
Heat conduction#100.00200.00200.1384#200.00180.00180.0059
Heat conduction#300.00170.00180.0047#400.00160.00180.0033
Aerofoil flow#80.01520.01910.2523#120.01440.01540.2596
Aerofoil flow#160.01430.01510.1491#200.01400.01450.1406

Weakness7: The computational efficiency claims need more comprehensive validation.

Table E reports placement optimization time and GPU memory for all four reconstructors. We estimate the full-range FixedCoords cost by scaling its per-budget time to all supported integer budgets. Relative to this estimate, BASP reduces total placement time by 86.60–99.10%. In all 16 settings, one BASP run takes longer than the reported per-budget FixedCoords time and uses more GPU memory.

Table E . Training cost and GPU memory consumption.

ReconstructorDatasetBASP one run(min)FixedCoords perbudget(min)FixedCoords full-range estimate(min)Time reduction(%)GPU memory(GiB),BASP / FixedCoords
VoronoiCNNCylinder wake46.9916.64416.0088.704.64/4.03
Turbulent flow388.8521.525939.5293.454.09/3.82
Aerofoil flow45.6520.04340.6886.609.82/6.62
Heat117.1440.931678.1393.0212.96/10.60
RecFNOCylinder wake67.0427.24681.0090.163.54/2.14
Turbulent flow436.9254.3715006.1297.093.89/2.71
Aerofoil flow62.5930.46517.8287.919.75/5.98
Heat146.8965.042666.6494.497.31/4.41
PhySenseCylinder wake85.5738.83970.8391.196.64/4.35
Turbulent flow672.77248.9068696.4099.0228.68/14.86
Aerofoil flow165.3379.971359.4387.8425.22/13.23
Heat455.73212.338705.6794.7723.43/12.24
POD-FMCylinder wake87.6039.93998.3391.236.64/4.35
Turbulent flow814.77329.4390923.6099.1014.70/10.14
Aerofoil flow170.5082.001394.0087.7725.22/13.23
Heat479.60222.709130.7094.7523.44/12.21

Weankess8: Limited investigation of practical sensing constraints.

Thank you for the insightful question. Although our current work focuses on nested sensor deployment, additionally considering aspects such as 'involving obstacles, inaccessible regions, minimum sensor separation......' has practical application value, and we view this as a promising direction for future work. For aerofoil flow, the field prior is computed from the training set and held fixed during evaluation. Table 2 shows how the learned layouts perform on held-out combinations of angle of attack and inlet velocity, providing evidence of generalization within the benchmark.

Weakness9: Limited theoretical justification beyond the nested-layout construction.

Under uniform budget sampling, Eq.(7) minimizes the average placement loss across all supported budgets, with additional weight on the maximum budget. Every supported prefix is therefore directly supervised in expectation, and each sensor is jointly optimized across the budgets in which it appears. The maximum-budget term additionally supervises all queries at every iteration. Table 3 supports this design empirically: removing maximum-budget supervision increases errors in all eight evaluated settings. Table C above further evaluates prefix reconstruction errors over a denser set of intermediate budgets. These analyses explain and empirically assess how the objective encourages informative prefixes.

Question6:"Could the authors provide the missing implementation details, including the temperature parameter......""Could they also report actual placement optimization time and GPU memory consumption for PhySense and POD-FM"

Thank you for this comment. Based on some simple hyperparameter experiments, we set the attention temperature to 0.65 for cylinders,0.70 for airfoils, and 0.75 for turbulence and heat. The coverage gaussian suppression bandwidths was 0.08 in normalized coordinates. According to the original paper [5], the soft Voronoi temperature is 1.0, and the bandwidth of the Gaussian sensor mask is 1.5 grid pixels. We show the sensor optimization time and GPU memory usage on Physense and POD-FM in above Table E.

[1] Santos, Javier E., et al."Development of the senseiver for efficient field reconstruction from sparse observations."Nature Machine Intelligence5.11 (2023): 1317-1325.

[2] Liu, Xu, et al."Enhancing deep learning-based field reconstruction with a differentiable learning framework."Nature Machine Intelligence7.7 (2025): 1129-1140.

[3] Ma, Yuezhou, et al."Physense: Sensor placement optimization for accurate physics sensing."Advances in Neural Information Processing Systems38 (2025): 35697-35725.

[4] Manohar, Krithika, et al."Data-driven sparse sensor placement for reconstruction: Demonstrating the benefits of exploiting known patterns."IEEE Control Systems Magazine38.3 (2018): 63-86.

[5] Xiao, Renjie, et al."Flow Field Reconstruction via Voronoi-Enhanced Physics-Informed Neural Networks with End-to-End Sensor Placement Optimization."arXiv preprint arXiv:2603.09371(2026).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.