acceptodds
Under review as a conference paper at ICLR 2027

BambooSR: Super-Resolving Dark Matter Detection with Physics-Aware Implicit Neural Field

Abstract

Dark matter direct-detection experiments reconstruct rare particle interactions from faint optical signals. In the PandaX-4T detector, the transverse position of an interaction is encoded in a secondary-scintillation (S2) light pattern that is sparsely and irregularly sampled by 169 top-array photomultiplier tubes, with inactive channels and incomplete coverage further degrading reconstruction near the detector wall. We introduce BambooSR, which formulates this readout as a continuous-field super-resolution problem. Rather than mapping sensor counts directly to a position, BambooSR reconstructs an event-conditioned light field over the readout plane, guided by physics priors on direct propagation and boundary reflection, and integrates it over finite sensor footprints to produce readouts on a denser virtual layout, including responses at non-operational sensors. To enable supervised training and evaluation, we construct the first signal super-resolution benchmark for a large liquid-xenon time-projection chamber, comprising 112,890 paired low- and high-resolution readouts from photon-level simulation, together with 5,583 real PandaX-4T Run 2 wall events for downstream evaluation. BambooSR outperforms implicit neural representation and geometric baselines, with the largest gains near the detector wall and around the inactive sensor. On real PandaX-4T wall events, BambooSR improves calibrated position-reconstruction resolution by 39.6%. We will release the code and dataset.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI for Science, AI for Physics, Implicit Neural Representations, Super Resolving Physical Signal, Particle Physics, Dark Matter, Signal Reconstruction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_jGd6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents BambooSR, a physics-aware implicit neural representation framework for super-resolving sparse and heterogeneous detector readouts in the PandaX-4T experiment. The method reconstructs a continuous signal field from 169 physical PMT measurements and integrates it over a denser virtual layout of 1261 PMTs. The paper also introduces a new paired simulation benchmark for this task and evaluates the method on both simulated data and real PandaX-4T wall events.

Critical strengths

1.The paper studies a relatively novel application and provides a dedicated benchmark. It formulates super-resolution for sparse and heterogeneous PandaX-4T detector readouts and introduces a paired LR-HR simulation dataset with real detector data for downstream evaluation.

2.The method uses physical priors as soft guidance and validates their contribution with controlled experiments. Physics-derived features are injected through a learnable adapter, while the neural control suggests that the gains are not simply due to increased model capacity.

Critical weaknesses

1.Although the application setting is new, the underlying task remains fundamentally a super-resolution problem. The main difference is the sparse and heterogeneous sensor layout, while existing INR methods such as FFN, SIREN, WIRE, and SASNet can still be adapted to this setting. Thus, the novelty lies mainly in the application, sensing geometry, and physics-aware design rather than a fundamentally new learning paradigm.

2.The quantitative improvements are clear, but the qualitative differences are relatively subtle. Even in the enlarged regions, BambooSR looks similar to the stronger INR baselines, making the practical significance of some numerical gains less visually evident. More challenging examples could better demonstrate the advantage of the proposed method.

3.Paper does not report inference efficiency. Although BambooSR performs feed-forward inference, each HR PMT requires multiple quadrature-point evaluations for finite-area integration. Reporting inference time, memory usage, or computational cost relative to the main INR baselines would better clarify its practicality for detector reconstruction.

Questions for authors

1.Since existing INR methods can already be adapted to this signal super-resolution setting, could the authors clarify more explicitly what aspects of the task fundamentally require the proposed formulation beyond adapting standard continuous SR methods to irregular sensor geometry?

2.The numerical improvements over the strongest baselines are measurable, but the qualitative differences appear relatively small in the provided visualizations. Could the authors provide additional examples, particularly difficult cases near the detector boundary or inactive sensors, where the advantage of BambooSR is visually more evident?

6 comments

We sincerely thank the reviewer for recognizing the novelty of the detector-specific super-resolution task, the contribution of our paired LR–HR benchmark, and the role of physics-informed features supported by controlled experiments. We also appreciate the thoughtful questions regarding the positioning of our methodological contributions, the qualitative presentation of challenging reconstructions, and the computational efficiency of our approach. These constructive comments have helped us improve the presentation and provide additional evidence for the proposed framework. All modifications to the manuscript are highlighted in Bamboo Green for ease of your review.

Response to W1 / Q1

We thank the reviewer for this thoughtful assessment of the novelty of our work. We agree that BambooSR does not introduce a fundamentally new learning paradigm, and that existing INR architectures such as FFN, SIREN, WIRE, and SASNet can be adapted to the detector super-resolution task. We therefore position our contribution not as a new general-purpose INR architecture, but as a physics-aware continuous super-resolution framework tailored to the sparse and heterogeneous sensing conditions of particle detectors.

Importantly, the physics-aware design is not merely an architectural addition; its contribution is supported by controlled ablation experiments. We evaluate two complementary variants. First, Direct-only removes the explicit boundary-response feature while retaining the other geometric and propagation-related features. Restoring the boundary-response feature reduces RMSE from 2.576 to 2.185 PE/channel, edge RMSE from 2.707 to 2.311 PE/channel, and PMT3-region RMSE from 2.965 to 2.745 PE/channel. Second, Neural control preserves the same head and adapter dimensions but replaces the six physics-derived features with generic event- and coordinate-dependent features. The full BambooSR model improves all eight reported metrics over this control, reducing RMSE from 2.421 to 2.185 PE/channel. These comparisons demonstrate that the explicit boundary-response feature and the physics-informed feature content contribute beyond the use of an auxiliary adaptation branch alone.

We further examine physical reliability beyond aggregate channel-level errors. Appendix E.3 (Figure 10) presents event-level HR total-PE calibration, comparing predicted and ground-truth event totals across the test set. BambooSR maintains a compact, nearly diagonal calibration relation across the signal range, with a global signed total bias of approximately +0.05%. Importantly, the two-dimensional distributions also expose additional low-PE failure populations in several competing methods that can be obscured by aggregate signed bias through cancellation between over- and under-predicted events. This provides a complementary assessment of event-level signal preservation, particularly in the weak-signal regime.

Spatial plausibility is examined through representative challenging cases in Appendix E.4. These examples cover weak and strong signals near the masked physical PMT and detector boundary. In response to the reviewer's suggestion in W3.2, we have further added a common-scale two-dimensional continuous-field comparison and a corresponding one-dimensional profile, making differences in spatial localization, peak morphology, and local variations easier to inspect. These visualizations complement the quantitative results by providing qualitative checks of the reconstructed spatial structures, without treating visual similarity as pointwise ground-truth evidence.

Finally, the relevance of this detector-specific formulation is supported by our real-data downstream evaluation. Using the same 5,583 PandaX-4T Run 2 wall events and a common COG estimator, BambooSR reduces the calibrated wall-event spread from 11.01 to 6.65 mm, corresponding to a 39.6% improvement in position resolution. This demonstrates that the proposed signal reconstruction framework provides benefits beyond the simulated super-resolution benchmark and can improve a practically relevant downstream detector-reconstruction task.

The corresponding real-data evaluation is presented in Section 5.2 (Figure 5 and Table 3), with the common downstream reconstruction and cross-calibration protocol detailed in Appendix F.2.

Taken together, we appreciate the reviewer's characterization of the contribution. We do not claim a fundamentally new INR learning paradigm. Rather, our contribution lies in formulating detector-specific continuous super-resolution under sparse and heterogeneous sensing, introducing physics-informed features that provide measurable improvements over generic learned adaptation, and establishing their relevance through both event-level signal diagnostics and real-detector evaluation.

Response to W2 / Q2

We thank the reviewer for this helpful suggestion. We agree that the qualitative visualizations in our original submission did not make the differences among reconstruction methods sufficiently clear, particularly in challenging boundary cases. We have therefore improved the visualization by adding a side-by-side comparison of the reconstructed two-dimensional fields and a corresponding one-dimensional profile, providing a more detailed view of the differences among methods.

For our scientific task, the primary objective is to improve quantitative reconstruction accuracy rather than visual appearance alone. In detector reconstruction, spatial fields that appear visually similar may still produce different channel-integrated signals and downstream position estimates. Therefore, quantitative reconstruction errors and position-resolution metrics remain our primary evaluation criteria, while qualitative visualizations serve as complementary checks of spatial morphology and physical plausibility.

The practical relevance of these numerical improvements can be illustrated using the PandaX-4T detector. In our main seed-42 results, BambooSR combined with COG reduces the calibrated wall-event spread (σL\sigma_{\mathrm L}) from 11.01 mm (original LR COG) to 6.65 mm, corresponding to a 39.6% reduction. The resulting w68w_{68} and MAE are 4.19 mm and 3.57 mm, respectively. As an illustrative estimate, if the 4.36 mm reduction in radial spread could translate into an equivalent reduction in the radial fiducial-boundary margin, the corresponding cylindrical side-wall shell would contain approximately 18.9 L, or 54 kg, of liquid xenon. At an indicative semiconductor-grade xenon benchmark of approximately USD 2,650/kg, this represents a material-equivalent value of approximately USD 143,000. This is a conditional geometric illustration rather than a demonstrated fiducial-volume gain or actual procurement-cost saving, as the achievable improvement depends on the background distribution and physics-selection requirements.

Following the reviewer's suggestion, we have expanded Appendix E.4 (Representative Reconstructions Under Challenging Conditions) with two additional qualitative visualizations:

  1. Local continuous-field comparison: We compare the reconstructed fields of BambooSR, FFN, WIRE, SIREN, and SASNet for the same representative boundary event, together with the corresponding LR observation. All reconstructed fields share a common spatial window, color scale, and contour levels, making differences in peak morphology and local spatial variations easier to inspect.
  2. One-dimensional field profile: We further compare the reconstructed fields along a common section passing through the detector origin and the observed LR charge centroid. The profiles are obtained by interpolating the cached field grids, allowing peak shapes and local variations to be examined along a shared spatial coordinate.

These visualizations provide qualitative complements to the quantitative evaluation. Since pointwise ground truth for the continuous fields is unavailable, we do not interpret visual similarity as direct evidence of reconstruction accuracy or claim that any particular method is closest to the true underlying field.

The two additional figures and their accompanying descriptions have been incorporated into Appendix E.4 (Figures 15 and 16).

Response to W3.3

We thank the reviewer for encouraging us to characterize the computational efficiency of the proposed method. We first clarify that the position-reconstruction task considered in this work is intended for offline detector-data analysis rather than online triggering. Therefore, strict real-time latency constraints are not a requirement of the current application. Nevertheless, inference efficiency is important for assessing the practical computational cost of the proposed finite-area integration.

To provide a controlled comparison, we benchmark BambooSR and four INR baselines on the same NVIDIA V100 32GB GPU. All methods use the same input pool, 1,261 HR output channels, and 128 integration points per channel, corresponding to 161,408 spatial evaluation points per event. The reported inference time includes conditioning, continuous-field evaluation, and weighted integration and reduction, thereby accounting for the computational cost of evaluating the field at multiple integration points.

MethodSingle-event median (ms) ↓Single-event P95 (ms) ↓Batch-8 latency (ms) ↓Batch-8 throughput (events/s) ↑Peak tensor memory (MiB) ↓
BambooSR36.2436.8656.19143.5115.8
FFN12.5713.3135.21228.9122.3
WIRE16.3816.7780.2297.3271.6
SIREN11.9812.6036.85218.6115.1
SASNet116.74118.22117.7568.3121.0

BambooSR achieves a single-event median latency of 36.24 ms and a batch-8 throughput of 143.5 events/s, with a peak PyTorch tensor allocation of 115.8 MiB. Its batch throughput is lower than those of FFN and SIREN, but higher than those of WIRE and SASNet. Its memory requirement remains comparable to the other methods and substantially lower than that of WIRE.

At this measured throughput, inference for the complete set of 5,583 real wall events would take approximately 39 seconds. This estimate covers model inference only and excludes data transfer, file I/O, and subsequent position reconstruction and calibration. Overall, these measurements indicate that the finite-area integration used by BambooSR introduces a measurable computational cost, while remaining practical for offline reconstruction at the scale considered in this study.

We sincerely appreciate your time and thoughtful feedback, which have helped us further strengthen this work.

Reviewer_hLW2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a PandaX-4T-based super-resolution dataset with a simple yet effective implicit function method. Experiments show results are state of the art.

Critical strengths

The writing is clear and the dataset has certain contributions.

Critical weaknesses

It would be better for the authors to provide super-resolution evaluations across varying channel densities (e.g., 300 or 2,000 channels) to substantiate the claimed continuous querying capability.

The downstream real-data evaluation couples the reconstructed readout with a naive COG estimator. Does this two-stage pipeline actually outperform a direct neural estimator mapping the physical 169 PMTs directly to 2D coordinates? It would be better for the authors to benchmark BambooSR against established end-to-end position reconstruction baselines on the same real-data sample to justify its practical utility.

Questions for authors

see weakness

3 comments

We sincerely thank the reviewer for the encouraging assessment of our work, particularly for recognizing the clarity of the presentation and the contributions of the dataset. We also appreciate the constructive suggestions, which motivated us to extend our evaluation across different HR sensing layouts and to introduce an additional direct position-prediction baseline. These additions have helped us further clarify the applicability and advantages of the proposed reconstruction framework. All modifications in the revised manuscript are highlighted in Bamboo Green for ease of your review.

Response to W1

We extended the evaluation to two additional HR layouts (397 and 1,951 channels), alongside the original 1,261-channel configuration. The original 25-mm photocathode diameter, guided by the Hamamatsu R7057, was selected as the primary reference because its relevant physical characteristics and performance most closely match those of the PMTs installed in PandaX-4T. The additional 46-mm and 20-mm diameters were based on the effective photocathode specifications of the Hamamatsu R13089 and R4998, respectively. Thus, all three layouts are grounded in existing PMT specifications rather than arbitrarily selected upsampling scales.

Importantly, all five INR methods (BambooSR, FFN, WIRE, SIREN, and SASNet) are evaluated using their original frozen checkpoints trained on the 1,261-channel layout, without retraining, fine-tuning, or checkpoint reselection. The input remains the same masked 169-channel physical LR readout. HR targets are generated by aggregating the same underlying photon realizations over the corresponding photocathode regions. We also regenerated the original 1,261-channel targets and verified exact event-wise agreement with the original labels. All evaluations use the same 11,280 simulated test events.

The RMSE results across the three HR layouts are summarized below (lower is better):

Method397 channels1,261 channels1,951 channels
BambooSR5.7272.1851.630
FFN6.8292.5461.895
WIRE5.8622.4111.757
SIREN7.6833.1802.250
SASNet10.3313.5283.492

BambooSR achieves the lowest RMSE across all three layouts, including both additional configurations not used during training. Compared with WIRE, the second-best method in RMSE, BambooSR reduces RMSE by 2.3%, 9.4%, and 7.2% on the 397-, 1,261-, and 1,951-channel layouts, respectively.

These results demonstrate that the frozen BambooSR representation can be queried and integrated over different HR photocathode regions to generate readouts at varying channel densities without retraining.

In response to this suggestion, we also revised the Scope and Limitations section (Appendix H) to acknowledge the demonstrated cross-layout evaluation across three HR configurations, while retaining the remaining limitations concerning different physical LR sensor configurations and detector geometries.

The complete layout specifications and evaluation protocol are provided in Appendix J.1–J.2, including Figure 21 (layout illustrations) and Table 13 (layout specifications). The reconstruction results are presented in Appendix J.3, Figure 22 and Table 14, covering all eight metrics. A brief summary of the additional evaluation has also been added to Section 5.1 (p. 7, lines 376–377).

Response to W2

To directly examine the benefit of the proposed two-stage reconstruction pipeline, we added a direct position-prediction MLP that maps the masked and normalized charges from the 169 physical PMTs directly to two-dimensional event positions.

The network contains 85,602 trainable parameters, comparable to the 85,447 parameters of BambooSR. It is trained using true two-dimensional positions from simulated events as supervision, with the same source-location-disjoint data split. Real wall events are not used for training or checkpoint selection.

Both methods are evaluated on the same 5,583 real wall events using an identical two-fold, 24-sector radial cross-calibration protocol. The event list and calibration folds are kept identical, and each evaluation event is calibrated using the opposite fold.

The results are summarized below.

MethodσL\sigma_L (mm) ↓w68w_{68} (mm) ↓MAE (mm) ↓
Direct Position MLP7.526.415.23
BambooSR + COG6.654.193.57

BambooSR + COG achieves lower values across all three metrics, including both the fitted residual width and fit-free evaluation metrics. This provides direct evidence that, under the evaluated real-data wall-event protocol, reconstructing the spatial signal field before position estimation is more effective than directly predicting positions from the original sparse readout using the tested MLP baseline.

We emphasize that these metrics characterize calibrated radial wall residuals rather than absolute two-dimensional position errors. The revised manuscript provides the complete network architecture, training protocol, calibration procedure, and detailed results in Appendix I.2, Table 12 (p. 43). The direct-position comparison is also referenced in Section 5.2 (p. 8, lines 418–420).

We sincerely appreciate your time and thoughtful feedback, which have helped us further strengthen this work.

AuthorsOpenReview ↗

Summary of Responses and Manuscript Revisions

Dear Reviewer hLW2,

We sincerely thank you for your positive assessment of our work, particularly your recognition of the clarity of the writing and the contribution of the dataset, as well as your constructive suggestions.

In response to your suggestions, we have conducted two additional experiments:

  • Cross-layout evaluation: Extended the evaluation to HR layouts with 397 and 1,951 channels, demonstrating reconstruction across different HR sensor densities using the original frozen checkpoints without retraining.
  • Direct position-prediction comparison: Introduced a direct position-prediction MLP and compared it against BambooSR followed by the common COG estimator on the same real wall-event dataset.

These experiments further demonstrate the flexibility of the reconstruction framework across sensor layouts and its downstream value for detector reconstruction.

For ease of review, changes in the revised manuscript are highlighted in green. We hope these additions adequately address your concerns and appreciate your further consideration.

Thank you again for your valuable time and helpful suggestions.

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

BambooSR utilizes masked 169-channel readout data from the PandaX-4T experiment to reconstruct a continuous, event-conditioned S2 light field. The method incorporates source-relative geometry, direct response characteristics, and boundary response characteristics into an implicit neural representation (INR) based on Fourier features, employing an additive adapter. By integrating this light field over the region covered by the sensors, the method generates 1261-channel virtual readout data. The model was trained on 112,890 simulated datasets, categorized by source location; additionally, 5,583 real-world "wall events" were used for subsequent unpaired evaluation (pp. 4–6). On the simulated data, the method reduced the root-mean-square error (RMSE) from 2.411 PE/channel—achieved by the WIRE algorithm—to 2.185 PE/channel. Furthermore, for real-world "wall events," the method narrowed the calibrated residual width from 11.01 mm to 6.65 mm (relative to the LR COG method) (pp. 7–8, Tables 2–3). It should be noted that the residual width mentioned here serves merely as a proxy metric for boundary scenarios and does not represent the actual measured 2D positional error.

Critical strengths
  • The task and the benchmark setup are well-justified. Given that the sensor coverage areas are non-nested and contain gaps, explicit modeling of finite-area readouts is necessary. Both the simulated low-resolution (LR) and high-resolution (HR) observations are aggregated from the same underlying photon realizations, and all ten realizations corresponding to a single source location are assigned to the same data split. Ground-truth location information is neither used as an input for the reconstruction process nor as a target for positional supervision (p. 4; pp. 15–16, Appendix A.3–A.4).
  • The controlled feature-comparison experiments presented in the paper are highly informative. The "Direct-only" approach retains the source geometry while removing explicit boundary responses, whereas the "Neural control" approach replaces structured features while preserving information about the detector head/adapter dimensions. The full model achieves a root-mean-square error (RMSE) of 2.185, compared to 2.576 and 2.421 for the other two approaches, respectively (p. 9, Table 4; pp. 22–23, Appendix C.3). These results validate the rationale for incorporating physical features into the tested architecture.
  • In the experiments using real-world data, the authors employ robust control settings and complementary evaluation metrics. All readout data are derived from the same events and center-of-gravity (COG) estimators, and are processed through a standardized "two-fold, 24-sector" calibration pipeline. The evaluation results—based on both "fit-free containment width" and mean absolute error (MAE)—are consistent with the rankings derived from fit-based width metrics (p. 8, Table 3; pp. 31–34, Appendix F). Furthermore, the paper explicitly discusses its limitations regarding the supervision strategy, practical applications, and physical interpretation.
Critical weaknesses
  1. The reported gains need stronger evidence of robustness. Tables 2–3 show improvements across several metrics, but the neural models are evaluated from single training runs and the differences between leading methods have no uncertainty estimates. The authors also disclose that test diagnostics were consulted during development, despite using validation data for checkpoint selection (p.26, D.3; p.36). Repeated runs would help establish the stability of the gains, while a fresh holdout evaluated after fixing the final method would provide more independent confirmation.
  2. A direct learned count-prediction baseline is missing. The comparisons with other INRs and the physics-feature ablations are useful, but all experiments use one fixed sensor-layout pair (pp.19–23, Appendix C; p.36). A direct network predicting the 1,261 HR counts from the 169-channel input would help assess whether the continuous representation offers an accuracy or efficiency advantage on this task. The comparison should use the same input information and count supervision, with comparable training and model-selection budgets.
  3. The real-data validation remains limited to calibrated COG. The common estimator and cross-calibration provide meaningful evidence that the reconstructed readout improves wall-event reconstruction (pp.32–34, Appendix F). As the authors acknowledge, this does not measure absolute two-dimensional position accuracy. A comparison with the detector's operational reconstruction, if available on the same events, would clarify the practical benefit beyond the evaluated COG protocol.
  4. Real-data selection and input preparation need fuller reporting. The paper fixes the 5,583-event sample but does not specify the selection cuts or whether an existing position estimate enters selection (p.31, F.1). The model's input normalization, physical units and masking are already defined (pp.16–19, Appendices A.4 and B); what needs fuller reporting is how the real Run 2 measurements are prepared and mapped to those inputs. This correspondence matters for interpreting transfer from the simulated virtual HR targets, which omit a separate small-PMT detection and electronics response (p.15, A.3). The existing signal-strength analysis covers simulation; reporting the real S2-size distribution and results across that range would strengthen the real-data assessment (pp.26–28).
Questions for authors
  1. Can you quantify the BambooSR–WIRE and BambooSR–FFN differences in Tables 2–3, separating variability across training seeds from uncertainty due to the evaluation sample? For paired resampling, please group simulated events by source position; on real data, preserve disjoint calibration/evaluation folds, recompute calibration and refit σL. Is a new simulation holdout available for evaluating the frozen final configuration independently of development diagnostics?
  2. Can you compare with a direct 169-to-1,261 count predictor using the same input information, Poisson loss, training split and evaluation metrics, and document the capacity and model-selection budgets? This would clarify the benefit of the proposed formulation on the evaluated fixed layout.
  3. If operational reconstruction outputs are available for the same 5,583 events, can you report comparable wall-residual metrics and explain any differences in calibration? Such a comparison would place the COG-based improvement in the context of the detector's current reconstruction pipeline.
  4. What cuts define the real sample, and does any reconstructed radius enter selection? Please specify how the real measurements are calibrated and mapped to the model's defined input units, normalization and channel mask, including their correspondence to simulation. Can you report the real S2-size distribution and the principal methods' wall-residual metrics across adequately populated S2-size ranges?
7 comments

We sincerely thank the reviewer for the thoughtful and encouraging assessment of our work. We are particularly grateful for the recognition of the benchmark design, the controlled experiments supporting our physics-aware features, the rigor of our real-data evaluation, and our discussion of the study's limitations. We also appreciate the constructive questions and suggestions, which have helped us further strengthen the statistical robustness, evaluation independence, and baseline comparisons of the manuscript. We have carefully addressed these points in our revision. For ease of your review, all modifications made to the manuscript are highlighted in Bamboo Green.

Response to W1 / Q1

We have extended our evaluation with three complementary analyses: repeated training across random seeds, paired bootstrap uncertainty estimation, and a newly generated independent simulated test set.

1. Training-seed robustness.

We independently retrained BambooSR, FFN, and WIRE using three random seeds, keeping the original data splits, training configurations, and validation-based checkpoint selection unchanged. The following table reports the mean ± sample standard deviation of the key upstream and downstream metrics.

MethodUpstream RMSE (PE/ch)Wall σL (mm)Wall w68 (mm)Wall MAE (mm)
BambooSR2.1971 ± 0.01186.5201 ± 0.15254.2519 ± 0.05763.6100 ± 0.0339
FFN2.4132 ± 0.11756.7428 ± 0.20004.5467 ± 0.28343.9395 ± 0.2238
WIRE2.3572 ± 0.07356.7797 ± 0.28295.0470 ± 0.70844.3190 ± 0.5639

BambooSR achieves the lowest mean across all eight upstream and three downstream metrics, with the smallest run-to-run standard deviation in six upstream metrics and all three downstream metrics. Complete results are provided in Appendix G.2, Table 7.

2. Paired bootstrap uncertainty.

We performed 1,500 paired bootstrap resamples using the original fixed checkpoints. Simulated events were resampled by source-position groups, preserving all ten realizations per position. For real wall events, resampling was performed within the original two folds, with sector-wise cross-calibration recomputed in each replicate.

We define Δ = BambooSR − baseline; negative values indicate lower errors for BambooSR. The following table reports the original-sample differences and pointwise 95% confidence intervals.

MetricBambooSR − WIREBambooSR − FFN
Simulated RMSE (PE/ch)−0.226 [−0.288, −0.166]−0.361 [−0.390, −0.330]
Wall σL (mm)−0.384 [−0.851, −0.019]−0.304 [−0.689, −0.043]
Wall w68 (mm)−1.663 [−1.912, −1.467]−0.642 [−0.808, −0.520]
Wall MAE (mm)−1.377 [−1.515, −1.272]−0.610 [−0.693, −0.542]

All eight simulated confidence intervals favor BambooSR over FFN. Against WIRE, six of eight intervals exclude zero; the Shape TV and PMT3-region RMSE intervals include zero. All six real-wall confidence intervals lie below zero against both baselines. Full results are reported in Appendix G.2, Tables 8–9.

3. Independent evaluation and test-set usage.

We clarify that the original test set was not used for checkpoint selection, which was based on validation Poisson NLL. Aggregate test diagnostics were nevertheless examined during earlier method development; we therefore do not claim that the original test partition remained entirely unseen.

To provide additional independent evidence, we generated a new 10,000-event simulated test set after freezing the final models and evaluation protocol. It contains 1,000 new source positions, each with ten photon realizations, following the original simulation configuration. The original checkpoint hashes were recorded before data generation, and no retraining, fine-tuning, or checkpoint reselection was performed.

Metric (PE/channel)BambooSRFFNWIRE
Overall RMSE2.15732.51312.3618

BambooSR achieves the lowest overall RMSE, with reductions of 14.16% relative to FFN and 8.66% relative to WIRE, and leads in seven of eight metrics. WIRE performs slightly better in the PMT3 region. Complete results are provided in Appendix G.2, Table 10.

The clarification of test-set usage and the uncertainty-evaluation protocol is also provided in Appendix D.3, with a brief reference to the additional robustness evaluation in Section 5.1.

Response to W2 / Q2

We added a direct count-prediction MLP to compare the proposed continuous reconstruction formulation against fixed-dimensional prediction. The MLP takes the masked 169-channel LR charge fractions and a total-charge condition as input, directly predicting the 1,261 HR channel counts.

The baseline contains 84,213 trainable parameters, comparable to BambooSR (85,447). It uses the same source-location-exclusive data splits, inactive-channel masks, and HR targets. Training employs Poisson negative log-likelihood, Adam with a learning rate of 10⁻⁴, a batch size of 8, and 32 sampled HR channels per training step. Both methods are evaluated using the complete 1,261-channel output.

The quantitative results are summarized below.

MetricDirect Count MLPBambooSR
RMSE (PE/ch) ↓7.7442.185
MAE (PE/ch) ↓1.0010.638
Poisson deviance ↓1.1330.756
Total relative MAE (%) ↓8.7622.580
Absolute total bias (%) ↓6.9250.052
Shape TV ↓0.2490.169
PMT3-region RMSE (PE/ch) ↓7.4542.745
Edge RMSE (PE/ch) ↓7.1572.311

BambooSR achieves lower errors across all eight metrics, demonstrating consistent improvements in channel-level reconstruction, event-level total-signal estimation, spatial distribution recovery, and challenging detector regions. These results support the advantage of the proposed continuous reconstruction formulation over direct fixed-dimensional count prediction under the evaluated setting.

The complete baseline architecture, training configuration, and eight-metric comparison are provided in Appendix I.1, Table 11 (pp. 42–43). The additional baseline is also introduced in Section 5.1 (p. 7, lines 374–375).

Response to W3 / Q3

We would like to clarify the scientific objective of our real-data evaluation and its relevance to the detector's physics analysis. BambooSR is designed to reconstruct a more faithful spatial S2 signal representation, rather than to introduce another end-to-end position estimator or replace the existing detector-reconstruction pipeline. Our central question is whether recovering spatial information lost through sparse and incomplete sensor readout can provide a more informative representation for subsequent detector reconstruction. Accordingly, we adopt a controlled downstream evaluation on the same 5,583 PandaX-4T Run 2 wall events. The original LR readout and all reconstructed HR readouts are passed through the same charge-centroid (COG) estimator, with an identical detector-boundary model and two-fold, 24-sector cross-calibration protocol. This design isolates the contribution of improved S2 signal reconstruction while keeping the downstream estimator and evaluation procedure fixed.

An important consideration is that, for this wall-event study, position resolution rather than event-wise absolute position accuracy is the physically relevant quantity. Physics cuts identify a high-purity population of events originating from the detector boundary, but their exact interaction coordinates are experimentally unavailable. We therefore characterize how tightly this known boundary population is reconstructed relative to the detector wall, using the calibrated radial residual distribution. This provides a physically motivated measure of wall-event position resolution and is consistent with previous PandaX-4T studies of position reconstruction and surface-background modeling

The improvement in wall-event position resolution also has important implications for the detector's fiducial volume (FV), defined as the inner liquid-xenon target region retained for the physics search after accounting for surface backgrounds. Radioactive contaminants on the PTFE reflector produce surface events, whose reconstructed radial distributions are used to characterize their inward leakage and inform the selection of the FV boundary. A narrower reconstructed wall-event distribution can provide better spatial separation between boundary-associated backgrounds and the inner target region, potentially supporting a more refined FV optimization while maintaining background control.

The potential practical value can be illustrated using the PandaX-4T geometry. Our main result reduces the calibrated inward width σL\sigma_{\mathrm L} from 11.01 to 6.65 mm, an improvement of 39.6%. If the corresponding 4.36 mm reduction could translate into an equivalent reduction in the radial FV margin, the associated side-wall volume would be approximately 19 L, corresponding to about 56 kg of liquid xenon. At an indicative 2026 semiconductor-grade xenon price of approximately USD 2,650/kg, this represents a material-equivalent value of roughly USD 150,000. This is a conditional geometric estimate rather than a demonstrated FV gain or actual cost saving. Nevertheless, it illustrates the potential value of improved position resolution: retaining more effective liquid-xenon target mass can increase the exposure available for rare-event searches.

We agree that incorporating the reconstructed readouts into the complete detector position-reconstruction and physics-analysis chain would be a valuable next step. In particular, evaluating the resulting impact on surface-background modeling and the optimized FV would further establish the physics-level benefits of signal super-resolution. The present study focuses on establishing the upstream contribution through a common downstream estimator, providing a controlled basis for such subsequent analysis. We have clarified the physical motivation, common downstream evaluation protocol, and interpretation of the wall-event resolution in Appendices F.1, F.2, and F.5.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We sincerely thank the Area Chair for coordinating the review process and all reviewers for their time, careful assessment, and constructive feedback. We are grateful for the positive recognition of our work across the reviews. Reviewer RPJB acknowledged the well-justified task formulation and paired LR–HR benchmark, the informative controlled feature comparisons, and the rigorous real-data downstream evaluation. Reviewer hLW2 recognized the clarity of the writing and the contribution of the dedicated dataset. Reviewer jGd6 highlighted the novelty of the detector-specific application and benchmark, as well as the physics-informed feature design and its validation through controlled experiments. We greatly appreciate these encouraging assessments of the scientific motivation and contributions of BambooSR.

The main concerns raised and our corresponding responses are summarized below:

  • W.RPJB.1 / Q.RPJB.1 – Statistical robustness and evaluation independence: We added three-seed experiments, 1,500 paired bootstrap resamples, and a newly generated independent 10,000-event test set. Results are provided in Appendix G.2, with the evaluation protocol clarified in Appendix D.3.
  • W.RPJB.2 / Q.RPJB.2 and W.hLW2.2 / Q.hLW2.2 – Direct prediction baselines: We introduced two complementary MLP baselines for direct HR count prediction and direct 2D position prediction. Their architectures, protocols, and results are provided in Appendices I.1 and I.2, respectively.
  • W.RPJB.3 / Q.RPJB.3 and W.RPJB.4 / Q.RPJB.4 – Real-data evaluation and physical interpretation: We clarified the scientific scope of wall-event position-resolution evaluation and its relevance to fiducial-volume optimization. We also documented the event-level physics cuts, real-data input preparation, and S2-stratified reconstruction results in Appendices F.1 and F.4.
  • W.hLW2.1 / Q.hLW2.1 – Cross-layout evaluation: We extended the evaluation to three HR layouts (397, 1,261, and 1,951 channels), using five INR methods with frozen checkpoints. Results are reported in Appendix J, and the corresponding scope and limitations have been updated in Appendix H.
  • W.jGd6.1 / Q.jGd6.1 and W.jGd6.2 / Q.jGd6.2 – Methodological contribution and qualitative evidence: We clarified the detector-specific nature of our contribution and provided further evidence through controlled ablations, event-level total-PE calibration, and additional challenging-case visualizations. These results are presented in Section 6 and Appendices E.3–E.4.
  • W.jGd6.3 – Computational efficiency: We benchmarked five INR methods on the same V100 GPU, reporting inference latency, batch throughput, and peak memory usage. The supplementary results are provided in our response to W.jGd6.3.

Overall, this revision strengthens the statistical evidence, broadens the experimental comparisons, and provides a more comprehensive physical and computational assessment of BambooSR. All manuscript modifications are highlighted in Bamboo Green for ease of inspection.

We sincerely appreciate the time and thoughtful feedback of the Area Chair and all reviewers, which have helped us further strengthen this work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.