{"id":"706fe89d-1692-46f8-9de4-3460adca9ec2","arxiv_id":"2505.24740","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A supervised convolutional autoencoder trained on GEANT4 simulations reconstructs photon impact points in a BGO calorimeter to sub-crystal precision and estimates particle energies with a spread of roughly 75 MeV.","lead":"The paper trains convolutional autoencoder networks on simulated calorimeter data to reconstruct the position and energy of incoming photons from the pattern of energy deposits in a 29 by 29 crystal array. The method locates most hits within about a centimeter and estimates energies with a spread near 75 MeV, but it is tested only on simulation and is not compared with existing algorithms.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cluster identification efficiency and fake rate are not reported; position/energy metrics are conditional on a permissive 5x5-bin matching window, so the headline accuracy may not describe all showering photons.","rationale":"The paper's central claim is that the autoencoder reconstructs photon impact points within about one centimeter and energies with small bias. For that claim to hold, the network must not only localize matched clusters accurately but also identify the clusters in the first place. The reported distributions are conditioned on a permissive matching rule: a prediction counts as a match if any non-zero value appears in a 5×5 bin window around the true hit. Without reporting the matched fraction or the fake rate, the accuracy numbers can be dominated by easy, well-separated clusters while difficult overlapping clusters are silently dropped. This is the most load-bearing gap because it is internal to the evaluation and can be settled from the existing simulated dataset without any new experimental information. The reader's simulation-fidelity concern is legitimate but more generic: it would apply to any simulation-based study and cannot be resolved from the paper alone. The efficiency/fake-rate issue, by contrast, is directly checkable and affects the strength of the central claim even under the assumption of a perfect simulation. The paper is otherwise clear and modest, and the concern does not invalidate the method; it narrows what is actually demonstrated. Keeping the CONDITIONAL verdict is appropriate, with the requested efficiency and fake-rate metrics as a concrete acceptance condition.","tokens_in":4900,"tokens_out":3865,"duration_ms":46534,"concrete_test":"Re-run the test-set evaluation with an explicit cluster-level matching: assign each predicted peak to the nearest true hit (e.g., Hungarian algorithm), require association within one crystal, and report (1) efficiency = matched true clusters / total true clusters; (2) fake rate = unmatched predicted clusters / total predicted clusters; (3) the same Δr and ΔE distributions computed only after this stricter matching; (4) these metrics split by true inter-cluster distance (e.g., <2 cm, 2–5 cm, >5 cm). If efficiency is below about 90% or fake rate above about 10%, the headline claim should be restricted to matched clusters; if the metrics survive at high efficiency, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 defines a match as any non-zero prediction inside a 5×5 bin window (2.5×2.5 cm) around the true hit, and all reported Δr_i and ΔE_i distributions are computed only for these matched clusters. The paper never states the fraction of true clusters that are matched, the number of predicted clusters with no true match, or the number of true clusters per event that compete for one predicted peak. Consequently, \"most predicted clusters are placed less than 1 cm away\" is a statement about a selected subset. The 5×5 window itself is much larger than the claimed sub-centimeter precision, so the association criterion is lenient and can mask merging of nearby showers or hallucinated peaks. This concern is internal to the analysis: even with a perfect GEANT4 simulation, the reported accuracy is not established for all input photons. It is also directly relevant to the abstract's claim that the method is useful for separating overlapping signals, a property the paper explicitly defers to future work.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper applies convolutional autoencoders to the reconstruction of photon clusters in a simulated 29×29 BGO electromagnetic calorimeter. The network is trained on GEANT4-simulated events with upsampled sparse labels (116×116 bins) that encode photon impact positions and energies, and the output is post-processed by summing 5×5 windows around local maxima. Two model architectures are compared on an independent simulated test set, and the paper reports position residuals (sub-centimeter for most matched clusters) and energy residuals (σ ≈ 75 MeV; mean bias 3.15 MeV for Model 2). The authors claim that the method can reconstruct impact points within the same segment as the true position and suggest applicability to separating overlapping signals.","tokens_in":5055,"tokens_out":4405,"duration_ms":56012,"significance":"If the claims were fully established, the method would be a useful ML-based alternative to classical clustering for segmented calorimeters, particularly at high occupancy where overlapping showers are an issue. The paper has clear strengths: it uses a well-defined supervised training setup with sparse labels, provides architectural details sufficient for reproduction, and reports residuals on a held-out simulated sample. However, the reported metrics are conditioned on a permissive 5×5-bin matching window, and the paper does not report cluster-level efficiency, fake rate, or any baseline comparison. The significance as currently demonstrated is therefore limited to a proof-of-principle on simulated data, and the abstract overstates the demonstrated capability on overlapping signals.","major_comments":[{"comment":"The paper defines a match as any non-zero predicted value inside a 5×5-bin window around the true hit, and all reported Δr_i and ΔE_i distributions are computed only for these matched clusters. The paper never states the fraction of true clusters that are matched, the number of predicted clusters with no true match, or the number of true clusters that compete for one predicted peak. Since the 5×5 window (2.5×2.5 cm) is much larger than the claimed sub-centimeter precision, the headline accuracy is conditional on a permissive association criterion and may not hold for all input photons. Please report cluster-level efficiency, fake rate, and a multiplicity confusion matrix, and use an assignment procedure that enforces one-to-one matching.","section":"Section 3, matching criterion"},{"comment":"The abstract states that the method 'can be applied' in cases where separation of overlapping signals is key, but Section 4 explicitly defers cluster-separation studies to future work. This is an internal inconsistency: the central advertised application is not demonstrated. Either add quantitative two-photon separation metrics (e.g., resolution as a function of separation distance) or soften the abstract's claim to match the presented evidence.","section":"Abstract and Section 4"},{"comment":"The comparison between Model 1 (ΔE mean 42.39 MeV) and Model 2 (ΔE mean 3.15 MeV) is made from single test-set evaluations without error bars, bootstrap intervals, or repeated training runs. Neural network training is stochastic, and architecture comparisons based on one run can be misleading. The claim that Model 2 shows much better energy reconstruction needs run-to-run variation or confidence intervals on the reported means and widths.","section":"Section 3, energy results"},{"comment":"No comparison is made to a classical clustering algorithm, such as a local-maximum finder with a fixed energy-sum window, which would be the natural baseline for this task. Without such a baseline, the reader cannot judge whether the reported sub-centimeter position accuracy and 75 MeV energy width represent an improvement over existing methods or simply a re-implementation of a standard approach. Please add a comparison to a non-ML reference algorithm using the same matching and evaluation protocol.","section":"Section 3, no baseline comparison"},{"comment":"The entire method is developed and evaluated on GEANT4-simulated data, and the paper does not discuss how simulation-to-detector differences (shower shape, noise, thresholds, calibration) would affect the reported residuals. This is a limitation for a detector-reconstruction paper, though not a fatal one if framed correctly. Please add an explicit statement that the results are simulation-only proof-of-principle and discuss the main sources of systematic uncertainty that would need to be addressed before application to real data.","section":"Section 2, simulation-only validation"}],"minor_comments":[{"comment":"The post-processing algorithm is not fully specified when two predicted local maxima fall within the same 5×5 window or when windows overlap; please state the order in which maxima are processed and how energy assigned to one cluster is removed from consideration for subsequent clusters.","section":"Section 2, Eq. (4)"},{"comment":"The paper does not state the size of the independent test set, the number of photons per event in that set, or how the test set was generated relative to the training set; please provide these details for reproducibility.","section":"Section 3"},{"comment":"The notation Δ1x, Δ2x, and σ2(Δy) is introduced without explicit definition, and the units are missing for σ2(Δy) = 0.518; please standardize the notation and include units throughout.","section":"Section 3"},{"comment":"The GEANT4 simulation is described only as producing electromagnetic showers; please specify the physics list, the energy threshold per crystal, and any noise or digitization effects, since these directly affect the realism of the training data.","section":"Section 2, simulation setup"},{"comment":"There are several typographical and formatting issues, including the author name 'V enelin' with an unwanted space and the figure caption 'Right:. ' in Figure 3; a careful proofread is needed.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is a short methods-oriented manuscript that needs a moderately substantial revision before it can be accepted. The central technical gap is the absence of cluster-level efficiency and fake-rate metrics: the reported accuracy is only for matched clusters under a permissive window, which undermines the abstract's general claim. The abstract's overlap-separation claim should also be toned down unless supported by new analysis. In its current form, the contribution is a promising pilot study rather than a fully established reconstruction method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nI've read the paper and the stress-test note. I think the note is right: the reported accuracy is conditional on a 5x5-bin matching window, and the missing match efficiency and fake rate are the biggest holes.\n\nWhat's actually new: the group applied their labeled-autoencoder approach, previously used for 1D pulse traces, to 2D shower clusters in a BGO calorimeter. The idea is simple and legitimate: upsample the labels to 0.5 cm bins, train a CNN to map crystal energies to a sparse position-energy map, then post-process with a moving 5x5 sum. The main results—sub-centimeter position residuals and a near-zero mean energy bias for one of the two architectures—are believable as far as the simulation goes. The comparison between a 4-layer and a 5-layer encoder is a concrete check, and the paper is honestly written: it doesn't pretend cluster separation is solved, and it flags mismatched events as a possible cause of large energy tails.\n\nThat said, the soft spots are real but not fatal. First, the matching criterion is generous: any non-zero prediction inside a 2.5x2.5 cm window counts as a match, and all position/energy distributions are built only from matched clusters. Without the match efficiency and fake rate, you can't tell how often the method actually finds a photon versus hallucinating or merging nearby showers. This is internal to the analysis, so it matters even assuming perfect simulation. Second, there is no baseline—no energy-weighted centroid, no nearest-neighbor clustering, no classical algorithm. For a reconstruction paper that is a big gap, because the reader can't judge whether the CNN has any real advantage. Third, there are no error bars from repeated training runs and no code or data release, so the stability of the sigma values is unknown. Fourth, simulation-to-real transfer is untested, which is a standard caveat for this kind of study but still worth stating. Also, the abstract claims the method can be applied where separation of overlapping signals is key, but the paper itself says that is future work; that overreach should be fixed.\n\nNet: this is a modest, honest contribution that could be made solid with a moderately sized revision. It deserves peer review—a referee would give the authors a clear checklist—and I'd expect it to be acceptable after adding the missing efficiency numbers and at least one classical comparison. I don't think it's important enough to cite in my own work unless I happen to work on calorimeter clustering, but I'd bring it to a reading group as an example of how evaluation choices can shape ML reconstruction claims.","headline":"A credible simulation-only proof-of-concept for CNN calorimeter clustering, but the reported metrics are conditional on a generous matching window and there is no baseline.","tokens_in":5625,"tokens_out":3095,"would_cite":false,"duration_ms":35403,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["29.40.Vj"],"model":"deepseek-v4-flash","headline":"A modified convolutional autoencoder trained on simulated energy matrices reconstructs photon impact points in a segmented BGO calorimeter to within 1 cm and yields a near-zero mean energy bias.","keywords":["convolutional neural network","autoencoder","electromagnetic calorimeter","cluster reconstruction","BGO crystal calorimeter","impact point reconstruction","energy reconstruction","Monte Carlo simulation"],"falsifier":"A beam test that sends photons of known energy at known positions into a 29-by-29 BGO crystal matrix would settle it: if the network's predicted impact points are systematically off by more than 1 cm, or if the mean energy residual moves far from 3 MeV, the simulation-to-experiment transfer fails.","tokens_in":4642,"feed_emoji":"🎯","tokens_out":10393,"duration_ms":109502,"temperature":0.7,"pith_summary":"The paper is trying to show that a convolutional autoencoder, trained on labeled Monte Carlo events, can reconstruct individual photon hits in a segmented electromagnetic calorimeter directly from the pattern of energy deposited in its crystals. Unlike a standard autoencoder, the network is not asked to reproduce its input; it is trained to output an upsampled grid in which only the bin at each true photon impact point carries that photon's energy. On a held-out simulated dataset, most predicted impact points fall within 1 cm of the true positions, and the deeper of the two models reconstructs energies with a mean bias of 3.15 MeV and a spread of 75.46 MeV. The authors see this as a step toward reliable automatic separation of overlapping showers, which is the main obstacle to accurate event reconstruction in high-rate calorimeter data.","feed_headline":"CNN autoencoder locates calorimeter photon hits to under 1 cm","feed_subtitle":"A label-trained network also reads photon energies with a mean bias of 3 MeV, a step toward full single-network event reconstruction.","key_machinery":"The carrying object is a convolutional autoencoder: an encoder that compresses the 29-by-29 energy map through successive 2D convolution layers, and a decoder of transposed convolution layers that expands it back. The modification that makes reconstruction work is the use of labels as the training target rather than the input, together with two upsampling layers that produce a 116-by-116 output. This upsampling is what gives sub-crystal spatial resolution. A final post-processing step scans the output with a 5-by-5 window, merges neighboring non-zero values into the local maximum, and reads off one position and one energy per recognized cluster. The optimization minimizes the chi-squared difference between the label array and the predicted array.","core_discovery":"The paper's central claim is that a modified CNN autoencoder can map the 29-by-29 matrix of accumulated crystal energies directly to the positions and energies of all photons that hit the detector during an event. The training target is a 116-by-116 array obtained by dividing each 2 cm by 2 cm crystal into 0.5 cm by 0.5 cm bins and placing each photon's energy in the bin that contains its true impact point. At inference, the position of the maximum value in each predicted cluster is read as the impact point, and a 5-by-5 summation window around that maximum gives the cluster energy. On independent simulated events with up to 50 photons per event, the method places most reconstructed clusters within 1 cm of the true hit point, and the deeper model achieves an energy residual with mean 3.15 MeV and standard deviation 75.46 MeV, close to zero bias.","pith_inferences":["Editorial inference: a stress test with showers arriving closer together than one crystal width would reveal the actual two-photon separation limit, which the paper leaves for future work.","Editorial inference: the same label-up-sampling idea could be applied to other segmented detectors whose readout is a 2D grid, such as hadron calorimeters or imaging time-projection chambers.","Editorial inference: the small systematic x-coordinate bias of about 2 mm suggests a shift in the encoder's feature maps; training with mirrored or rotated events could show whether it is a simulation artifact or a model bias."],"forward_implications":["A single trained network could replace hand-written clustering and position-fitting code for calorimeters of this type.","The near-zero mean energy bias of the deeper model means reconstructed energies need little or no additional offset correction after the network output.","Because each crystal is divided into 0.5 cm bins in the labels, position precision is tied to the label bin size rather than to the physical crystal size.","The method targets high-occupancy events with many overlapping showers, where conventional cluster separation is most likely to fail.","Deeper architectures with dropout improve energy reconstruction without changing position reconstruction, so the two tasks can be tuned separately."],"supporting_citations":[{"why":"Supplies the simulated training and test datasets; every reported accuracy is measured on its output.","marker":"[7]"},{"why":"Provides the convolutional autoencoder architecture that the modified, label-based training scheme extends.","marker":"[2]"},{"why":"Earlier label-based reconstruction for time-series pulses in individual crystals, adapted here to image-shaped calorimeter data.","marker":"[8]"}],"fun_headline_variants":["CNN autoencoder maps calorimeter hits to photon positions","ML reconstructs calorimeter clusters within 1 cm","Autoencoder locates photon impacts and estimates energies","Neural network resolves overlapping calorimeter clusters","CNN reads crystal energies to find photon hit points"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simulated detector behaves like the real BGO calorimeter, so the sub-centimeter positions and near-zero energy bias learned from simulated showers will survive contact with experimental data.","fun_headline_variants_meta":{"raw":{"variants":["CNN autoencoder maps calorimeter hits to photon positions","ML reconstructs calorimeter clusters within 1 cm","Autoencoder locates photon impacts and estimates energies","Neural network resolves overlapping calorimeter clusters","CNN reads crystal energies to find photon hit points"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1398,"prompt_tokens":841,"completion_tokens":557,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":457,"completion_tokens_details":{"reasoning_tokens":485}},"tokens_in":457,"tokens_out":557,"duration_ms":5945,"temperature":1.0,"reasoning_tokens":485,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:13:52.990236+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A beam test that sends photons of known energy at known positions into a 29-by-29 BGO crystal matrix would settle it: if the network's predicted impact points are systematically off by more than 1 cm, or if the mean energy residual moves far from 3 MeV, the simulation-to-experiment transfer fails.","supporting_citations":[{"cited_title":"(2020) Visualkeras","cited_arxiv_id":null,"evidence_quote":"Supplies the simulated training and test datasets; every reported accuracy is measured on its output."},{"cited_title":"(2017) A better autoencoder for image: Convolutional autoencoder In Proceedings of the ICONIP17-DCEC, Guangzhou, China, 14--18 October 2017","cited_arxiv_id":null,"evidence_quote":"Provides the convolutional autoencoder architecture that the modified, label-based training scheme extends."},{"cited_title":"(2003) GEANT4--a simulation toolkit Nucl","cited_arxiv_id":null,"evidence_quote":"Earlier label-based reconstruction for time-series pulses in individual crystals, adapted here to image-shaped calorimeter data."}],"review_version":1}