{"id":"c39f8ee1-c255-468a-a037-1022ff25c0ce","arxiv_id":"2505.05664","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"This paper introduces a 'synthetic representation' of calorimeter hits based on simulated optical-photon tracks as an intermediate target for ML reconstruction, and qualitatively shows a U-Net can predict these tracks from realistic hits.","lead":"This paper proposes training a neural network to reconstruct a synthetic, simulation-only view of particle showers, the paths of light photons inside a calorimeter, as an intermediate step before physics reconstruction. The hope is that this makes the mapping from detector data to particle properties less lossy and more interpretable, but the paper demonstrates only the first step and does so qualitatively.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'less lossy' claim is information-theoretically impossible as stated; any advantage must be inductive bias, which is never tested.","rationale":"The reader correctly identified the untested premise that S is 'closer' to D and better-conditioned. My stress-test sharpens this into a stronger objection: if the paper's language is read information-theoretically, the claim is not merely unverified but impossible, because the deployed composite g∘f is a function of D and therefore cannot convey more mutual information about T than D itself. The paper explicitly says 'less susceptible to information loss' and 'less lossy compression and recovery,' which invites exactly this reading. The only way to salvage the claim is to reinterpret it as an inductive-bias or sample-efficiency statement, but then the paper still provides no comparison against a direct D->T baseline, no trained NN2, and no quantitative reconstruction metric. The qualitative U-Net images in Section VII are suggestive but do not test the central hypothesis. I do not find a reason to move the reader's REJECT verdict; the paper is clearly written and honest about being a first implementation, but the load-bearing assertion remains unsupported in either of its possible interpretations.","tokens_in":7391,"tokens_out":5518,"duration_ms":62674,"concrete_test":"Train NN2 on simulated (S,T) pairs, then evaluate the full pipeline g(f(D)) on the held-out 1,000-event test set and compare label accuracy or energy regression against a direct D->T network with matched architecture, parameter count, and training budget. Also estimate an upper bound on achievable T-information from D (for example, by training a high-capacity direct classifier and measuring test accuracy); if the two-stage pipeline does not outperform the direct baseline, the less-lossy claim is falsified. To separate the information story from the conditioning story, additionally measure classification accuracy using oracle S (from simulation) as input: if oracle S does not beat D, the premise that S is richer in task-relevant information fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (abstract, Section I, Section IV) is that factorizing D->S->T is 'less lossy' than direct D->T. As stated, this cannot be true for information content. In the deployed pipeline, Section II computes S1=f(D1) and T1=g(S1), so the final label is a deterministic function g∘f of D. The data processing inequality gives I(T; g(f(D))) ≤ I(T; D): no two-stage composition through S can recover information about T that is absent from D, and no choice of S can make the composite less lossy than the best direct D->T map. The only defensible version is an inductive-bias/regularization claim: S may make optimization better-conditioned at finite data or model capacity. That version is not tested. NN2 (S->T) is never trained, and no quantitative comparison of D->S->T vs D->T is reported. Section VII offers only qualitative U-Net images and an interpretation that the Cerenkov signal is learned first; this does not establish that S is more informative than D, nor that the two-stage map is better. Therefore the manuscript's headline claim is either internally inconsistent (if 'less lossy' means more mutual information) or empirically unsupported (if it means better-conditioned learning).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes an intermediate 'synthetic representation' space S, constructed from full-simulation optical-photon track hits in a dual-readout crystal calorimeter, to factorize the reconstruction mapping from detector space D to truth space T into two stages, D→S→T. The authors define two networks, NN1 (D→S) and NN2 (S→T), and implement NN1 as a 3-level U-Net trained on 50 GeV single-particle events to infer the synthetic channels from realistic detector channels. The presented results are qualitative inference images (Figure 7), along with an interpretation that the network learns the Cerenkov signal first and thereby implicitly performs a dual-readout correction. NN2 is never trained, and no quantitative reconstruction accuracy or comparison with a direct D→T mapping is reported.","tokens_in":7663,"tokens_out":3096,"duration_ms":37479,"significance":"If the central hypothesis were validated, the idea of using hardware-motivated synthetic observables as an interpretable intermediate representation for ML-based reconstruction would be genuinely valuable, both for future collider detectors and for other experimental domains facing similar domain-gap problems. The paper also provides a reasonably detailed description of a reconfigurable DD4hep/Geant4 dual-readout calorimeter simulation with custom optical-photon readout, and the channel encoding in Table I is clearly documented. However, the paper's central claim—that the factorization D→S→T is 'less lossy' than direct D→T—is never tested in any quantitative sense, and as literally stated it is in tension with the data processing inequality. The qualitative observation that a U-Net can learn to produce plausible synthetic Cerenkov and scintillation track images is a useful proof-of-concept detail but does not by itself establish the proposed method's advantage.","major_comments":[{"comment":"The claim that factorizing the mapping as g∘f : D→T through S is 'less lossy' than a direct D→T mapping is not supported and, if understood information-theoretically, is contradicted by the data processing inequality: for any deterministic functions f and g, I(T; g(f(D))) ≤ I(T; D). The paper explicitly computes S1=f(D1) and T1=g(S1) in Section II, so the composite is a deterministic function of D and cannot recover information about T that is absent from D. If the intended claim is instead that the two-stage pipeline provides a useful inductive bias or better optimization conditioning at finite data or model capacity, that claim requires explicit empirical testing, which is not provided.","section":"Section I and Section II (Eq. 1-3)"},{"comment":"NN2, the S→T mapping, is defined in Section II but never trained or evaluated in this work. Section VI trains only the U-Net for the image-to-image task D→S, and Section VII shows only qualitative inference images. Consequently, no reconstruction performance metrics (energy resolution, particle identification efficiency, or classification accuracy) are reported for the full D→S→T pipeline, so the paper cannot substantiate any statement about the quality of the final reconstruction.","section":"Section II and Section VI"},{"comment":"There is no quantitative comparison between the proposed D→S→T pipeline and a direct D→T baseline. The only empirical evidence is a set of qualitative U-Net inference images (Figure 7), which show that the model learns features resembling Cerenkov and scintillation tracks. The interpretation that this 'implicitly performs the dual-readout correction' is plausible but is not tied to any measurable reconstruction quantity, and no comparison to a standard dual-readout calibration or to a direct mapping is attempted.","section":"Section VII and Figure 7"},{"comment":"The construction of the synthetic space S depends on several manually chosen parameters: the wavelength range 200–600 nm and the scaling/offset factors se=2, sp=5, we=wp=1. The paper provides no sensitivity or ablation study showing how the results (or the information content of S) depend on these choices. Since the central argument is that S is a particularly informative representation, the dependence of the method on these free parameters must be characterized before the claim can be evaluated.","section":"Table I and Section IV"}],"minor_comments":[{"comment":"The paper uses the spelling 'Cerenkov' throughout; the standard physics spelling is 'Cherenkov'. Please use one consistent spelling, ideally 'Cherenkov'.","section":"Throughout"},{"comment":"The last sentence of Section VI says 'Then we run inference to generate images for a test set of 1000 images for each particle type and classify them,' but no classification task or classification result is described anywhere in the paper. This sentence should be clarified or connected to an actual reported analysis.","section":"Section VI"},{"comment":"The caption of Figure 7 says 'various training epochs and batch sizes' but the individual panels are not labeled with the exact epoch and batch size. Adding explicit per-panel labels would make the qualitative comparison much easier to follow.","section":"Figure 7"},{"comment":"The angular acceptance is described only as 'a uniform angular distribution in the barrel of the detector only.' Please specify the exact polar-angle range used so the study is reproducible.","section":"Section V"},{"comment":"The term 'machinistic interpretability' is unusual. Consider defining it or replacing it with a more standard term such as 'mechanical interpretability' or 'architectural interpretability'.","section":"Section I and VII"},{"comment":"The assertion that this is 'the first fully dynamic and reconfigurable geometry written for a collider detector' is a strong novelty claim and would benefit from a supporting reference or a clearer qualification of what 'fully dynamic and reconfigurable' means in this context.","section":"Section III"}],"recommendation":"reject","confidential_remarks":"The manuscript's central claim is not empirically supported: the second-stage network is never trained, no quantitative reconstruction metric is reported, and no direct D→T baseline is provided. The information-theoretic framing of the claim is also problematic as stated. These issues are load-bearing rather than cosmetic, so I do not see how a standard minor revision could resolve them. The simulation infrastructure and the qualitative observations may be a useful starting point, but the paper would need substantial additional quantitative work before it could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core concept here is new and worth taking seriously. Training a network to map realistic calorimeter hits to simulated optical-photon tracks, then using that synthetic representation as an intermediate for downstream inference, is a real departure from the latent-space and unfolding literature. The author has built an actual full simulation with a dual-readout crystal calorimeter, defined a concrete synthetic target space, and shown that a U-Net can plausibly infer the missing synthetic channels from realistic-only inputs. That is a legitimate proof-of-concept, clearly written and honest about its preliminary nature.\n\nThe soft spot is exactly where the reader and the stress test put it. The headline claim—that D→S→T is \"less lossy\" than direct D→T—is never tested. No NN2 is trained, no reconstruction accuracy reported, no direct comparison to a one-stage mapping. Worse, the claim as stated cannot be true in the information-theoretic sense: the composite g(f(D)) is a deterministic function of D, so I(T; g(f(D))) ≤ I(T; D). Any advantage must come from inductive bias—better-conditioned optimization, easier targets, or regularization at finite data/compute—and that version is not tested either. The qualitative images in Fig. 7 are suggestive, and the interpretation that the U-Net picks up the sparse Cerenkov signal first is plausible, but that does not establish that S is more informative than D or that the two-stage pipeline beats the direct one.\n\nI also note the \"first fully dynamic and reconfigurable geometry\" is supported by a self-citation (ref. 7). That is not a flaw by itself, but it is a reminder that the detector construction is the author's own prior contribution, so the novelty sits in the synthetic-target idea, not the simulation.\n\nWho is this for? Researchers working on ML-based reconstruction in HEP, especially those interested in physically interpretable intermediate representations or in the interplay between differentiable simulation and reconstruction. They will find the idea stimulating, but they will not find the evidence to adopt it yet.\n\nMy recommendation: send it to peer review, because the idea deserves referee time, but do not accept as is. The revision should either reframe the \"less lossy\" language as an inductive-bias hypothesis and then test it, or drop the claim. A quantitative comparison D→T vs. D→S→T with the same architecture would be the obvious next step.","headline":"A genuinely new idea for reconstruction—training on simulated optical-photon tracks as an intermediate target—but the 'less lossy' claim is untested and, as stated, information-theoretically wrong; deserves a serious referee, not acceptance.","tokens_in":8156,"tokens_out":1861,"would_cite":false,"duration_ms":22733,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes that reconstructing particle properties from detector data should be factorized through a synthetic intermediate representation of detector physics, because the two-step mapping loses less information and is more…","keywords":["synthetic representations","domain gap","dual-readout calorimetry","image-to-image translation","U-Net","detector simulation","interpretable machine learning","calorimeter reconstruction"],"falsifier":"Train the full factorized pipeline $D\\to S\\to T$ and a direct $D\\to T$ regressor or classifier on the same simulated events with the same training budget, and compare truth-reconstruction accuracy on a held-out set. If the direct mapping matches or beats the factorized pipeline, the central claim is false; a second check is to test whether $S$ carries information about the truth label $T$ that is not already in $D$, since without such extra information no factorization can reduce information loss.","tokens_in":7182,"feed_emoji":"⚛️","tokens_out":7564,"duration_ms":79783,"temperature":0.7,"pith_summary":"This paper proposes that reconstructing particle properties from detector data should be factorized as $D\\to S\\to T$, passing through a synthetic intermediate space $S$ that encodes detector-specific physics available only in full simulation. The claim is that because $S$ is closer to the detector space $D$ in dimensionality and information content, the two-step mapping loses less information than the direct map $D\\to T$, and it anchors the machine-learning model to a known physical method. As a first demonstration, the paper trains a U-Net to recover simulated scintillation and Cerenkov optical-photon track images from realistic calorimeter hits for 50 GeV single particles in a dual-readout crystal calorimeter. The first results are qualitative: the network resolves the sparse Cerenkov structure first, which the author interprets as the model implicitly performing the dual-readout correction. If the hypothesis holds, ML-based reconstruction and anomaly detection would become more accurate and interpretable across experiments that rely on high-statistics Monte Carlo simulation.","feed_headline":"The network learns the Cerenkov track first","feed_subtitle":"Recovering synthetic S/C photon tracks from calorimeter hits anchors ML to dual-readout physics.","key_machinery":"The load-bearing machinery is the synthetic representation space $S$, instantiated as per-crystal track hits of optical scintillation and Cerenkov photons saved in full simulation with a track-level wavelength cut (200–600 nm) instead of an energy-deposit threshold. This turns the intermediate representation into a simulated observable: it would never appear in a real detector, yet it encodes a real physical process tied to detector hardware. The U-Net's image-to-image translation $D\\to S$ then anchors the network's latent space to dual-readout physics, and the eventual $S\\to T$ mapping is argued to be better conditioned because $S$ is closer to $D$ in dimensionality and information content.","core_discovery":"The central claim, stated as a hypothesis in the introduction, is that factorizing inference through a synthetic representation of truth reduces information loss compared with a direct mapping from truth to experimental data. The synthetic space $S$ is engineered from detector hardware capabilities and real physical processes, here the track-level trajectories of scintillation and Cerenkov photons in a longitudinally segmented dual-readout crystal calorimeter, so that the mapping from realistic hits $D$ to $S$ is less degenerate, and the subsequent mapping $S\\to T$ operates on a richer, better-conditioned feature space. The paper's demonstration trains a three-level U-Net to infer the synthetic channels from the realistic channels of event images, and finds that the Cerenkov signal is resolved first, which it reads as the network implicitly learning the dual-readout correction.","pith_inferences":["A direct quantitative test is implied but not performed: comparing the information content or reconstruction accuracy of $g\\circ f$ against a direct $D\\to T$ network on identical data would settle whether the factorization actually reduces information loss.","The claim that the Cerenkov signal is learned first because it is sparse could be checked against a control network trained on a non-physical intermediate target; if a random sparse target produces the same learning order, the physics grounding is not what drives the observed behavior.","The synthetic-space idea suggests a new anomaly-detection strategy: use residuals in $S$-space, predicted versus true synthetic tracks, as a physics-anchored anomaly score, potentially flagging new particles or detector effects as deviations in known physical structure.","Because $S$ is defined by the detector's hardware capabilities, geometrically reconfigurable simulation could be used to optimize detector design jointly with the reconstruction network, treating the synthetic representation as a tunable inductive bias."],"forward_implications":["A trained $D\\to S$ mapping can recover fine-grained shower structure, such as skin depth and hair-like track protrusions, that is invisible in realistic detector hits, giving reconstruction access to hidden physics information.","Because the network locks onto the sparse Cerenkov signal before the scintillation signal, the method offers a built-in interpretability axis: the order of feature learning mirrors the physics of dual-readout correction.","The same construction, an intermediate synthetic response grounded in hardware physics, can be applied to other detectors or to cosmological surveys that map low-dimensional truth to high-dimensional observations.","If the factorization is trained end-to-end with the combined loss, the pipeline yields truth labels $T$ directly from detector hits while keeping the synthetic anchor during training.","Anomalous detector signals flagged by networks trained this way would be more likely to correspond to genuine physical anomalies rather than artifacts of the latent space."],"supporting_citations":[{"why":"Baseline fast-simulation approach (deep generative photon shower simulation) that the paper's intermediate representation is meant to improve upon.","marker":"[1]"},{"why":"CaloGAN, the GAN-based fast calorimeter simulation baseline whose direct truth-to-shower mapping motivates the factorization claim.","marker":"[2]"},{"why":"Another GAN fast-simulation baseline for particle detectors, contrasted with the proposed physics-grounded intermediate space.","marker":"[3]"},{"why":"OmniFold, the invertible unfolding approach that maps detector data back to truth; the paper contrasts it with the hardware-grounded synthetic space.","marker":"[4]"},{"why":"Survey of machine-learning unfolding methods, situating the direct inversion approach that $S$ is designed to improve upon.","marker":"[5]"},{"why":"Prior study establishing the dual-readout segmented crystal calorimeter concept and its jet-resolution performance, motivating the detector choice.","marker":"[6]"},{"why":"The differentiable full detector simulation used to generate both realistic and synthetic hit images for this study.","marker":"[7]"}],"fun_headline_variants":["Synthetic tracks bridge detector domain gap","Cerenkov first: ML learns dual-readout physics","Less lossy: synthetic truth anchors ML reconstruction","Detector-aware synthetic training reduces lossy mapping"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the untested premise that the synthetic space $S$ is genuinely closer to the detector space $D$ in information content, making the two-stage map $D\\to S\\to T$ better conditioned and less lossy than the direct $D\\to T$; if $S$ does not actually capture the relevant physics, or the factorization is no better conditioned, the central advantage of the proposal collapses.","fun_headline_variants_meta":{"raw":{"variants":["Synthetic tracks bridge detector domain gap","Cerenkov first: ML learns dual-readout physics","Less lossy: synthetic truth anchors ML reconstruction","Detector-aware synthetic training reduces lossy mapping"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000481,"raw_usage":{"total_tokens":2401,"prompt_tokens":992,"completion_tokens":1409,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":1350}},"tokens_in":608,"tokens_out":1409,"duration_ms":10616,"temperature":1.0,"reasoning_tokens":1350,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:59:30.396947+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the full factorized pipeline $D\\to S\\to T$ and a direct $D\\to T$ regressor or classifier on the same simulated events with the same training budget, and compare truth-reconstruction accuracy on a held-out set. If the direct mapping matches or beats the factorized pipeline, the central claim is false; a second check is to test whether $S$ carries information about the truth label $T$ that is not already in $D$, since without such extra information no factorization can reduce information loss.","supporting_citations":[{"cited_title":"Andreassen , author P","cited_arxiv_id":null,"evidence_quote":"OmniFold, the invertible unfolding approach that maps detector data back to truth; the paper contrasts it with the hardware-grounded synthetic space."},{"cited_title":"Lucchini , author W","cited_arxiv_id":null,"evidence_quote":"Prior study establishing the dual-readout segmented crystal calorimeter concept and its jet-resolution performance, motivating the detector choice."}],"review_version":1}