{"id":"6bd9b398-ddf6-4be2-845f-a3a4e60f1bd5","arxiv_id":"2507.18222","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Using maximum-likelihood estimation on integral-based pulse-shape parameters reduces simulated 3D position-reconstruction bias in strip HPGe detectors from about 0.4 mm to below 0.05 mm at 1 keV RMS noise.","lead":"This paper simulates an orthogonal-strip germanium detector and shows that combining several pulse-shape measurements with maximum-likelihood estimation cuts position-reconstruction bias roughly tenfold compared with standard methods. The result is a more accurate, noise-resistant way to locate gamma-ray interactions, which matters for medical imaging and astronomy instruments.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline bias/resolution numbers are in-sample: evaluation uses the same 0.1 mm grid points that define the likelihood mapping, so near-zero bias at grid points does not establish accurate continuous 3D reconstruction.","rationale":"The reader's verdict flags lack of experimental validation and in-sample evaluation. I go further: the in-sample grid is the most load-bearing concern because it undermines the quantitative claim even before experimental validation. If the evaluation is only at the grid points used to define the likelihood, then the near-zero bias is nearly tautological: the MLE is a nearest-grid-point classifier trained and tested on identical coordinates. The simulation-accuracy concern is real but secondary; the same in-sample flaw would persist with a perfect simulator. The paper never states that test positions are held out or that off-grid positions are reconstructed, and Eq. 2.8 gives no interpolation rule. The reported X/Y maximum bias, 0.016 mm, is less than one-sixth of the 0.1 mm grid spacing, which cannot hold for a continuous position unless the mapping is interpolated; the paper does not describe such interpolation. A simple offset-grid cross-check would settle whether the method generalizes. Until then, the conclusion 'maximum bias reduced from 0.4 mm to 0.016 mm' overstates what has been demonstrated. Because the paper is explicitly a simulation study and the method is plausible, CONDITIONAL remains appropriate, requiring an off-grid/cross-validated evaluation and ideally code release.","tokens_in":8017,"tokens_out":5226,"duration_ms":60636,"concrete_test":"Generate an independent validation set at off-grid positions offset by half a grid cell (e.g., add 0.05 mm to x, y, and z of every training point), simulate their pulses with the same SolidStateDetectors.jl setup and noise model, and apply the existing MLE mapping built on the original 0.1 mm grid. Recompute the maximum and mean absolute bias and resolution separately for these off-grid events. If the maximum X/Y bias exceeds roughly 0.05 mm (half the grid spacing) or the Z bias degrades beyond the reported 0.02 mm, the in-sample evaluation is the reason for the headline numbers, and the central claim needs an interpolation or continuous-likelihood step plus cross-validation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 defines the training set as a 0.1 mm grid of simulated interaction points; Section 2.4 builds the likelihood lookup (Eq. 2.8) at exactly those points, and Section 3 evaluates bias/resolution using 'all interaction events in the corresponding X-Y plane' — i.e., the same grid points. The MLE is therefore tested on its training points, with no off-grid or held-out validation. At a training point, the expected likelihood is maximized at the true grid coordinate by construction, so the reported <0.02 mm bias mostly reflects Monte Carlo noise, and the 'resolution' measures how often noise causes a jump to a neighboring grid point. The method's practical claim is continuous 3D reconstruction; a real event will land between grid points. If the arg max is restricted to the training grid (as written), off-grid positions are quantized to the nearest grid point, producing bias up to half the 0.1 mm spacing and a floor on resolution; if an interpolation scheme is used, it is neither described nor validated. The reported X/Y bias of 0.016 mm — far below half the grid spacing — is only achievable when true positions lie exactly on the grid. This is a load-bearing gap independent of simulation fidelity: even with a perfect simulator, the paper's numbers do not yet support the claimed near-zero bias for arbitrary positions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a maximum-likelihood-based three-dimensional position reconstruction method for orthogonal-strip planar HPGe detectors. The authors simulate pulse shapes from a central pixel on a 0.1 mm grid using SolidStateDetectors.jl, extract four integral-based parameters (S_C, S_I, S_AX, S_AY), model their per-position multivariate normal distribution from repeated noise realizations, and reconstruct positions by maximizing the likelihood. The paper reports that, under 1 keV RMS electronic noise for 100 keV events, the maximum Z bias drops from 0.4 mm to 0.02 mm centrally and from 2 mm to 0.15 mm near electrodes, maximum X/Y bias drops to 0.016 mm, and position resolution is 0.07-0.16 mm in Z and 0.07-0.44 mm in X/Y, with improved noise robustness relative to amplitude-based parameters. The central methodological claim is that joint modeling of multiple pulse-shape parameters removes the systematic bias caused by cross-dimensional interference.","tokens_in":8207,"tokens_out":3945,"duration_ms":44664,"significance":"If the reported performance holds for continuous, off-grid interaction positions and for a real detector, the method would be a useful advance for HPGe position reconstruction, particularly because it replaces noise-sensitive amplitude parameters with integral-based ones and addresses cross-dimensional interference. The paper benefits from a coherent simulation pipeline that includes realistic physical effects (drift, diffusion, self-repulsion, RC response, Gaussian noise), and from explicit comparisons against a conventional independent reconstruction baseline. However, the current evidence is entirely in-sample: both the likelihood model and the evaluation use the same 0.1 mm simulation grid, so the headline bias and resolution numbers do not yet establish the continuous-reconstruction claim. Experimental validation or at least a perturbed-simulation cross-check would be needed to transfer the method from simulation to practice.","major_comments":[{"comment":"The manuscript does not specify how the arg max in Eq. (2.8) is computed: is it restricted to the discrete set of 0.1 mm grid points, or is it optimized over continuous coordinates? If the latter, the likelihood function is only defined at grid points and the interpolation method is missing. This is load-bearing because the central claim is three-dimensional position reconstruction for arbitrary interaction sites, not only for the sampled grid. Please clarify the implementation and, if grid-only, add an off-grid evaluation or an interpolation scheme.","section":"Section 2.2, Section 2.4, Section 3.1"},{"comment":"While the paper is presented as a simulation study, the abstract and introduction suggest practical applicability to medical imaging and gamma-ray astronomy. The absence of any experimental data or model-perturbation study is a limitation that should be stated more prominently, and a concrete cross-validation test is necessary to support the extrapolation from simulated to real detector performance.","section":"Section 2.2 and Conclusion"},{"comment":"Several implementation parameters are not reported, which prevents reproducibility. The number of noise realizations N used for estimating the mean vector and covariance matrix in Eqs. (2.6)-(2.7) is not stated, nor is the number of repeated reconstructions M in Eq. (2.9). The mismatch of indices in Eq. (2.9) (the summation runs to N while the normalization uses M-1) also makes the exact definition ambiguous. Please specify N and M, correct the indices, and report the statistical uncertainty (e.g., the standard error of the mean bias) so that differences such as 0.02 mm versus 0.03 mm are meaningful.","section":"Section 2.4, Eq. (2.8), and Section 3"}],"minor_comments":[{"comment":"Equation (2.9) contains a typo: the sum over j runs from 1 to N while the normalization is 1/(M-1), and the definition of \\bar{\\hat{x}} also uses N. Both should be indexed by M to match the text describing M repeated reconstructions.","section":"Eq. (2.9)"},{"comment":"The caption reads \"S_AX (left) as a function of the X postion\" and \"S_AY (right) as a function of the Y postion\"; \"postion\" should be \"position.\"","section":"Figure 10 caption"},{"comment":"The manuscript assigns an energy of 100 keV to each simulated interaction event but does not clarify whether these are single-site energy deposits or full photoelectric absorption events. Since HPGe photon interactions at 100 keV typically involve a single photoelectron, this is likely fine, but a sentence defining the simulated event type (point-like single energy deposit) would avoid ambiguity.","section":"Section 2.2"},{"comment":"The integration window is fixed at 300 ns (900-1200 ns) for all events. The authors state it covers the entire rise edge, but drift times vary with interaction depth and can approach or exceed this window near the cathode under the adopted field. A justification or a position-dependent window would strengthen the parameter definitions.","section":"Section 2.2 and Figure 6"},{"comment":"The impurity type is not specified (n-type vs p-type). The sign of the impurity gradient affects the electric field profile and hence the pulse shapes; stating the polarity would improve reproducibility.","section":"Section 2.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a simulation study; the reviewers should judge it as such. The main concern is that the evaluation is in-sample on the training grid, which is a fixable but load-bearing issue. The authors should either add off-grid validation or adjust the claims to explicitly scope them to grid-point reconstruction. I would not recommend rejection because the method is plausible and the simulation pipeline is detailed; the required changes (off-grid tests, implementation clarification, parameter reporting) are within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I’ll cut to it: the paper tests a good idea — joint MLE over four integral-based pulse-shape parameters for 3D position reconstruction in orthogonal-strip HPGe — but the headline numbers are in-sample, and the practical claim of continuous sub-100-micron reconstruction is not supported by the evidence as presented.\n\nWhat’s genuinely new: prior work reconstructs X/Y and Z independently from amplitude-based parameters (CTD, A_asy). Here the authors build a single multivariate Gaussian likelihood over S_C, S_I, S_AX, S_AY, learning the parameter-position mapping from simulated pulses. The integral parameters are well-motivated, and the noise-robustness comparison against amplitude-based parameters is useful and physically credible. Figures 12 and 14 show real cross-dimensional interference, so the joint treatment should indeed reduce systematic bias in simulation. The authors are also honest about the simulation-only scope and explicitly flag that the method relies on simulation accuracy.\n\nThe soft spot is structural and independent of simulator fidelity. The likelihood is defined on a 0.1 mm grid (Section 2.4), and Section 3 evaluates bias and resolution on that same grid. The arg max in Eq. 2.8 is over the discrete grid, so at a training point the estimate can sit on the true grid point even if the model is completely wrong for any off-grid location. Reported X/Y bias of 0.016 mm and resolution of 0.07 mm are then essentially measures of grid quantization and the frequency of jumps to neighboring points — not evidence of unbiased continuous reconstruction. The paper never describes an interpolation scheme or any off-grid test. This is a load-bearing gap: for a real event, the interaction lands between grid points, and the method as written would quantize to the closest grid point, giving a bias floor of at least half the spacing and a resolution floor that is not captured by the reported numbers.\n\nMissing details (N, M, maximization scheme) and no code/data release make the numbers hard to verify, though these are minor and fixable.\n\nWho benefits: this is for detector simulation groups and HPGe instrument teams, particularly those working on Compton telescopes or small-animal SPECT. It deserves a serious referee — the method is plausible and the paper is clear — but the revision needs to address off-grid evaluation (e.g., leave-one-out on the grid or adding a jittered test set) and describe the arg max procedure. I’d accept to review, not desk-reject.","headline":"Sound simulation study of a jointly-reconstructed multi-parameter MLE for HPGe strip detectors, but the headline bias/resolution are in-sample and the method is unproven for off-grid positions.","tokens_in":8853,"tokens_out":3362,"would_cite":false,"duration_ms":33871,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["29.40.Wk"],"model":"deepseek-v4-flash","headline":"A maximum-likelihood fit over four integral pulse parameters reconstructs 3D interaction positions with near-zero bias and sub-mm resolution in simulated strip HPGe detectors.","keywords":["HPGe detectors","position reconstruction","pulse shape analysis","maximum likelihood estimation","integral-based parameters","orthogonal-strip detectors","gamma-ray imaging","Compton telescope"],"falsifier":"Place a collimated 100 keV source on a real orthogonal-strip planar HPGe detector of the same geometry, acquire pulses with known beam positions, extract the four integral parameters, and run the paper's MLE map; if the measured biases exceed roughly 0.5 mm or the resolutions exceed roughly 0.5 mm at 1 keV equivalent noise—well above the simulated 0.02–0.16 mm—then the simulation-to-detector transfer is the failure point, not the estimator.","tokens_in":7717,"feed_emoji":"📐","tokens_out":9306,"duration_ms":94228,"temperature":0.7,"pith_summary":"An orthogonal-strip planar HPGe detector can localize a photon interaction in three dimensions from the shapes of the charge signals it induces on crossed strip electrodes. The paper claims that the standard approach—reconstructing depth from collection-time difference and lateral position from image-signal amplitudes, each axis independently—carries systematic bias because each parameter also depends on the other two coordinates. It replaces this with a maximum-likelihood estimator that treats four integral-based pulse parameters jointly, modeling their noise-corrupted values as a multivariate normal whose mean and covariance are mapped over the pixel volume. In pulse-shape simulations of 100 keV interactions under 1 keV RMS electronic noise, the claimed result is near-zero bias (Z bias reduced from 0.4 mm to 0.02 mm centrally and from 2 mm to 0.15 mm near electrodes; X/Y bias from 0.4 mm to 0.016 mm) and sub-millimeter resolution (0.07–0.16 mm in Z, 0.07–0.44 mm in X/Y). If these numbers hold on a real detector, the method would give Compton imagers and gamma-ray telescopes precise interaction locations without finer pixel segmentation.","feed_headline":"Four pulse integrals plus a likelihood fit cut 3D position error to 0.02 mm","feed_subtitle":"Jointly fitting integral-based pulse parameters removes cross-dimensional bias that single-axis amplitude methods cannot.","key_machinery":"The load-bearing mechanism is a maximum-likelihood position estimator built from a precomputed parameter-to-position map. For every simulated interaction site, $N$ noisy realizations of the four integral parameters give a sample mean $\\hat{\\mu}(x,y,z)$ and covariance $\\hat{\\Sigma}(x,y,z)$; the estimated position is the grid point maximizing the multivariate Gaussian density $P(A \\mid \\hat{\\mu},\\hat{\\Sigma})$. The parameters themselves matter: replacing collection-time difference by the net areas $S_C$ (collection signals) and $S_I$ (image signals) keeps depth information while averaging out high-frequency noise, and replacing image-signal amplitudes by the normalized absolute-area asymmetries $S_{AX}$, $S_{AY}$ does the same for the lateral coordinates. The joint likelihood is what removes the systematic cross-dimensional bias, because each candidate position is scored against all four parameters simultaneously instead of one axis at a time.","core_discovery":"The central discovery is that cross-dimensional interference—the fact that a depth parameter like collection-time difference also shifts with the lateral position of the interaction—can be absorbed, rather than averaged away, by fitting all coordinates at once. The paper constructs, on a 0.1 mm grid inside a central pixel, a position-dependent multivariate normal model for the parameter set $\\{S_C, S_I, S_{AX}, S_{AY}\\}$, where $S_C$ and $S_I$ are net-area integrals of collection and image signals and $S_{AX},S_{AY}$ are normalized absolute-area asymmetries of neighboring image signals. A measured parameter vector is assigned the position that maximizes this likelihood. In simulation, this reduces maximum Z bias from 0.4 mm to 0.02 mm in the central region and from 2 mm to 0.15 mm near the electrodes, reduces X/Y maximum bias from 0.4 mm to 0.016 mm, and yields position resolutions of 0.07–0.16 mm (Z) and 0.07–0.44 mm (X/Y) at 1 keV RMS noise. It also shows that the integral-based parameters degrade approximately linearly with increasing noise, whereas the amplitude-based parameters degrade exponentially.","pith_inferences":["Beyond the paper: the multivariate likelihood's covariance structure could be used to detect simulation mismatch in experiment—if residuals between measured and simulated parameter vectors are correlated across $S_C, S_I, S_{AX}, S_{AY}$, that is a signature that the simulated electric field or drift model is off, not just random noise.","Beyond the paper: a self-calibrating variant is testable—use Compton-scatter kinematics or a collimated beam to label a subset of real events, then fit the parameter-to-position map from data instead of simulation, turning the simulator into a prior rather than the source of truth.","Beyond the paper: the near-cathode X/Y resolution loss (0.44 mm) suggests that adding a second-neighbor or corner image signal to the parameter set would recover lateral information in that zone; the likelihood formalism makes the information gain of such an addition directly measurable."],"forward_implications":["Sub-millimeter 3D interaction positions from a standard orthogonal-strip HPGe detector would sharpen gamma-ray track reconstruction, improving Compton imaging angular resolution without smaller strips.","Because the integral-based parameters keep resolution nearly linear in electronic noise, the method should maintain usable positioning at higher noise levels or lower signal thresholds than amplitude-based methods.","The same joint-likelihood mapping can be rebuilt for other segmented semiconductor detectors such as CdZnTe, extending unbiased 3D positioning beyond germanium.","The near-electrode Z degradation (0.15 mm bias, 0.16 mm resolution) remains the main limitation inside the pixel volume, showing where additional parameters or field shaping would help most."],"supporting_citations":[{"why":"Supplies the detector simulation that generates all training and evaluation waveforms with known interaction positions.","marker":"[15]"},{"why":"Gives the induced-charge theorem used to compute the signal on each electrode as charge carriers drift.","marker":"[16]"},{"why":"Provides the companion induced-current relation that, with weighting potentials, forms the simulated pulse shapes.","marker":"[18]"},{"why":"Reviews how the Shockley-Ramo theorem applies to semiconductor gamma-ray detectors, supporting the signal-formation model.","marker":"[17]"},{"why":"Defines the collection-time-difference depth measurement that serves as the conventional Z-reconstruction baseline.","marker":"[10]"},{"why":"Establishes three-dimensional position sensing in orthogonal-strip germanium detectors, the detector concept and conventional reconstruction framework.","marker":"[11]"},{"why":"Uses neighboring image-signal amplitudes for lateral positioning, the conventional X/Y baseline the joint MLE replaces.","marker":"[12]"}],"fun_headline_variants":["ML joint fit on pulse integrals cuts HPGe 3D error to 0.02 mm","Integral-based likelihood shrinks 3D position bias 20-fold in HPGe detectors","Joint likelihood on four pulse integrals pinpoints gamma hits to 0.02 mm","MLE on charge integrals removes cross-axis bias, 0.02 mm accuracy in HPGe","Fit four pulse integrals together: HPGe 3D positions accurate to 0.02 mm"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simulated pulse shapes—computed from the assumed impurity profile, electric field, charge-carrier drift, and RC response—match the real detector at the sub-0.1 mm level; if they do not, the tiny reported biases are biases of the fit to the simulator rather than of the detector.","fun_headline_variants_meta":{"raw":{"variants":["ML joint fit on pulse integrals cuts HPGe 3D error to 0.02 mm","Integral-based likelihood shrinks 3D position bias 20-fold in HPGe detectors","Joint likelihood on four pulse integrals pinpoints gamma hits to 0.02 mm","MLE on charge integrals removes cross-axis bias, 0.02 mm accuracy in HPGe","Fit four pulse integrals together: HPGe 3D positions accurate to 0.02 mm"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00059,"raw_usage":{"total_tokens":2831,"prompt_tokens":1069,"completion_tokens":1762,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":685,"completion_tokens_details":{"reasoning_tokens":1645}},"tokens_in":685,"tokens_out":1762,"duration_ms":15583,"temperature":1.0,"reasoning_tokens":1645,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:17:41.957872+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Place a collimated 100 keV source on a real orthogonal-strip planar HPGe detector of the same geometry, acquire pulses with known beam positions, extract the four integral parameters, and run the paper's MLE map; if the measured biases exceed roughly 0.5 mm or the resolutions exceed roughly 0.5 mm at 1 keV equivalent noise—well above the simulated 0.02–0.16 mm—then the simulation-to-detector transfer is the failure point, not the estimator.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the detector simulation that generates all training and evaluation waveforms with known interaction positions."},{"cited_title":"Shockley,Currents to conductors induced by a moving point charge,J","cited_arxiv_id":null,"evidence_quote":"Gives the induced-charge theorem used to compute the signal on each electrode as charge carriers drift."},{"cited_title":"Ramo,Currents induced by electron motion,Proceedings of the IRE27(1939) 584","cited_arxiv_id":null,"evidence_quote":"Provides the companion induced-current relation that, with weighting potentials, forms the simulated pulse shapes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Reviews how the Shockley-Ramo theorem applies to semiconductor gamma-ray detectors, supporting the signal-formation model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the collection-time-difference depth measurement that serves as the conventional Z-reconstruction baseline."},{"cited_title":"Amman and P","cited_arxiv_id":null,"evidence_quote":"Establishes three-dimensional position sensing in orthogonal-strip germanium detectors, the detector concept and conventional reconstruction framework."},{"cited_title":"Vetter, M","cited_arxiv_id":null,"evidence_quote":"Uses neighboring image-signal amplitudes for lateral positioning, the conventional X/Y baseline the joint MLE replaces."}],"review_version":1}