{"id":"68cf3106-a400-4a45-9678-37d3d010cced","arxiv_id":"2608.03257","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"NanoMorph-3D reconstructs 3D nanomaterial volumes from limited-angle electron tomography using a physics-driven unrolled network with projection-geometry attention, beating iterative and learned baselines on synthetic and real tilt series.","lead":"This paper presents a deep-learning system that reconstructs 3D nanomaterial structures from limited-angle electron microscope images, which normally lose details due to the missing wedge effect. It combines a physics-based projection model with trainable attention layers, and reports large gains over standard tomographic algorithms on simulated and real data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-data 'sim-to-real transfer' is untested: Table 2 evaluates the same seven tilt series used in unsupervised training with the same reprojection-consistency objective, so it measures training fit, not generalization.","rationale":"The paper's headline contribution is not merely a high-performing synthetic reconstruction; it is the claim that physics-driven unrolling transfers to real HAADF-STEM data (§3.6, §4.2). The only quantitative evidence for that claim is Table 2, and that table is methodologically circular: the seven real series are the same ones used to train the unsupervised branch, and the metric is the same reprojection-consistency objective. This is a textbook training-on-test leakage. Even if every physics detail in Eq. (1) were correct, Table 2 still could not distinguish a model that has memorized these tilt series from one that generalizes. The reader's weakest assumption focused on the Beer-Lambert/Radon forward model; that matters, but it is secondary because a proper held-out real evaluation would surface such mismatches empirically. The synthetic results, while strong, are self-consistent with the simulator and therefore do not resolve the external-validity question. I therefore agree with the reader's CONDITIONAL verdict, and the identified concern reinforces it; no verdict change is needed, but the acceptance condition should explicitly require held-out real-data evaluation and release of code/data.","tokens_in":12153,"tokens_out":5102,"duration_ms":53500,"concrete_test":"Conduct a leave-one-series-out experiment: retrain NanoMorph-3D and all baselines seven times, each time holding out one of the seven real tilt series from the unsupervised training stream, and report Table 2 metrics on the held-out series only. If NanoMorph-3D's margin over baselines collapses or reverses on held-out data, the 'robust transfer' claim fails; if the margin persists, the circularity concern is resolved. Additionally, release the code and the promised dataset so the forward model and synthetic test set can be independently checked.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in §4.2 — that NanoMorph-3D 'demonstrates successful simulation-to-reality transfer' — is supported by Table 2, which reports 2D reprojection PSNR/SSIM/LPIPS on real HAADF-STEM tilt series. However, §3.6 trains the unsupervised stream on unlabeled experimental projections Y_real and minimizes Eq. (7), L_unsup = ||A_{Θ_out}[V_real] − Y_{Θ_out}||_1 over held-out angles. Section 4.1 explicitly states that the seven real series used in the experiments are the same seven series used as unlabeled data in that unsupervised stream. Table 2 therefore evaluates the model on its own training data, using a metric (reprojection consistency over measured tilts) that is essentially the same objective the network was optimized to satisfy. This is a training-set fit, not evidence of generalization to new experimental data. The synthetic evaluations (Tables 1, 3, 4) are generated with the same physics simulator embedded in the network's Projection Consistency Unit, so they cannot provide external validation either. No held-out real series, no train/test split for the real data, no code, and no released dataset are provided. Consequently, the paper's real-world transfer claim is currently unsupported, even setting aside questions about the physical fidelity of Eq. (1).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"NanoMorph-3D proposes an end-to-end physics-driven unrolling network for limited-angle HAADF-STEM electron tomography. The forward model (Eq. 1) is a Beer-Lambert-linearized discrete Radon transform with additive noise. The network unrolls proximal gradient descent into ten learnable stages, alternating a Projection Consistency Unit (Eq. 3) with a Topology-Aware Proximal Mapping built from hierarchical volumetric attention, dual-domain sinusoidal attention, and physics normalization. A taxonomy-driven synthetic dataset is generated with nonlinear attenuation and instrumental noise, and an unsupervised dual-stream loss (Eq. 7) is applied to real tilt series. Experiments compare against SIRT, GENFIRE, RESIRE, AET-Net, LPD, and Denza-GS, reporting synthetic metrics in Tables 1, 3, and 4 and real-data reprojection metrics in Table 2.","tokens_in":12558,"tokens_out":5473,"duration_ms":55916,"significance":"If the claims were supported, the paper would make a useful contribution: it combines algorithmic unrolling with explicit projection-geometry attention, and it contributes a large synthetic dataset grounded in a morphological taxonomy. The architecture is a reasonable and timely synthesis of learned priors and physics constraints, and the sparse-view robustness results (Table 3) show a plausible direction for low-dose electron tomography. The main weakness is validation: the real-data evaluation is circular with the training objective, and the synthetic evaluation uses the same forward model embedded in the network, so the simulation-to-reality transfer claim is currently not demonstrated. The paper also contains no code or dataset release despite describing a substantial new dataset.","major_comments":[{"comment":"The real-data evaluation is not a generalization test. Section 4.1 states that the seven real series used in the experiments are the same seven series used as unlabeled data in the unsupervised stream. The Table 2 metric (reprojection PSNR/SSIM/LPIPS over measured tilts) is essentially the L_unsup objective of Eq. (7), which is minimized during training. Thus Table 2 measures fit to the training set, not simulation-to-reality transfer. The Sec. 4.2 claim of 'successful simulation-to-reality transfer' is unsupported. Please evaluate on held-out real series or clearly split the seven series into train/test subsets, and report per-series results. Using withheld tilt views as pseudo-ground truth does not escape this problem when the same views are used in the unsupervised objective.","section":"Sec. 3.6, Sec. 4.1, Table 2"},{"comment":"The synthetic test set is generated with the same discrete Radon operator A_Θ (after Beer-Lambert linearization) that is embedded in the network's Projection Consistency Unit (Eq. 3) and in the unsupervised loss (Eq. 7). Consequently, Tables 1, 3, and 4 validate inversion under an assumed model, not generalization to the real HAADF-STEM measurement process. The paper itself acknowledges in Sec. 5 a 'reliance on the Beer-Lambert law.' To support the central claim, add an independent validation: generate test data with a different ray-tracer or discretization, include misalignment or calibration errors, or use real volumes with independent ground truth (e.g., FIB-SEM correlated volumes or simulated inclusions embedded in experimental backgrounds).","section":"Sec. 3.5, Sec. 4.1, Eq. (1), Eq. (3)"},{"comment":"The real-data comparison reports single aggregate numbers over only seven series, without per-series breakdown, confidence intervals, or statistical significance. Several baselines (LPD, AET-Net) achieve PSNR around 15 dB, which is far below the others and suggests possible implementation or hyperparameter issues for those baselines. Since the real-data claim is load-bearing, please report per-series errors, standard deviations, and the exact hyperparameters used for each baseline (iterations, regularizers, learning rates).","section":"Sec. 4.2, Table 2"},{"comment":"The unrolled PGD formulation is underspecified. Eq. (3) uses operators F_Θ and B_Θ, while Eq. (1) defines A_Θ, and Eq. (7) again uses A_Θ. The paper does not state whether F_Θ = A_Θ, whether B_Θ is the exact adjoint or an approximation, or how the gradient step relates to the PGD step for the objective in Eq. (2). Since the method's name and motivation rest on PGD unrolling, please clarify these operators and provide a short derivation of the update rule.","section":"Sec. 3.3, Eq. (3)"}],"minor_comments":[{"comment":"The symbol λ is used both as a learnable scalar in Eq. (5) and as the loss-balancing weight in Eq. (7) and Sec. 4.1. Please use distinct symbols to avoid ambiguity.","section":"Sec. 3.4, Eq. (5)"},{"comment":"The text emphasizes 'non-linear exponential attenuation' and the Beer-Lambert law, but Eq. (1) is already in the linear logarithmic domain. Clarify how the nonlinear attenuation is simulated and how domain randomization of μ is applied in the synthetic pipeline.","section":"Sec. 3.5, Eq. (1)"},{"comment":"Baseline configurations are not reported. For SIRT, GENFIRE, RESIRE, LPD, and Denza-GS, specify the number of iterations, regularization parameters, and any preprocessing steps. This is necessary for a fair comparison.","section":"Sec. 4.1, Compared Methods"},{"comment":"The architecture description lacks key details: the number of attention heads, window sizes, feature dimensions, transformer depth, and the definition of the 'gradient-free block' used for initialization. These details are needed for reproducibility.","section":"Fig. 2, Sec. 3.3"},{"comment":"Denza-GS [42] is a self-cited arXiv preprint. Please state its publication status and clearly explain its relationship to the current method, especially since it is used as a strong baseline.","section":"References [42]"},{"comment":"The paper states that a large-scale dataset is constructed and provided for the community, but no code or data release link is given. Please add an availability statement.","section":"Sec. 4.1, Data Availability"}],"recommendation":"major_revision","confidential_remarks":"The core concern is validation: the real-data results in Table 2 are undermined by train/evaluation overlap, and the synthetic experiments share the forward model with the network. These issues can be addressed within the manuscript's scope by adding held-out real data, per-series reporting, and an independent synthetic forward model, so I recommend major revision rather than rejection. The dataset, if released, could be a valuable community resource, but the current manuscript does not provide it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper gets the unrolling machinery right and then undercuts itself on validation. The synthetic evaluation is solid but internally circular, and the real-data transfer claim is measured on the same seven tilt series used for unsupervised training, with the same reprojection objective. That's training fit, not generalization.\n\nWhat's actually new: the taxonomy-driven procedural generator for mesoscale nanomaterial volumes, the Beer-Lambert forward model with instrument noise, the PGD unrolling with Transformer-based proximal mapping, and the sinusoidal trajectory bias in attention. The ablations suggest each piece earns its keep, and the sparse-view results are impressive. The writing is unusually clear for an MM paper.\n\nWhere it falls down: Table 2. The paper never states that the real series are split into train/test. Section 4.1 says the seven series are used as unlabeled data in the unsupervised stream, and Table 2 reports reprojection consistency on 'experimental data'—same series, same L_unsup objective from Eq. 7. So 'successful simulation-to-reality transfer' is unsupported. The synthetic test (Table 1) also uses volumes generated by the same discrete Radon operator that sits inside the Projection Consistency Unit (Eq. 3). That doesn't invalidate the method, but it does mean the numeric gains over baselines could come from the network memorizing its own forward model rather than from physics. The authors acknowledge reliance on Beer-Lambert in the conclusion, so the physical model is a stated limitation, not hidden.\n\nThe paper would benefit from a real held-out experimental series, a train/validation split for real data, and code/data release. The promise of a benchmark dataset without release is also a mismatch.\n\nBottom line: worth a serious referee and likely worth building on, but the central transfer claim needs to be re-examined or re-framed. I wouldn't cite the real-data numbers as evidence until they show a held-out split.","headline":"A well-engineered unrolling framework with strong synthetic results, but the sim-to-real transfer claim is unsupported because the real-data evaluation uses training data and the same loss as the training objective.","tokens_in":12979,"tokens_out":1491,"would_cite":false,"duration_ms":15312,"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":"NanoMorph-3D claims that unrolling proximal gradient descent with trajectory-aware attention reconstructs 3D nanomaterials from ±60° tilt series with higher fidelity than six baselines, completing the missing wedge and transferring from sim","keywords":["electron tomography","missing wedge","deep unrolling","proximal gradient descent","HAADF-STEM","nanomaterials","sinogram consistency","physics-driven reconstruction"],"falsifier":"Reconstruct a nanomaterial from a dual-axis or full-tilt HAADF-STEM acquisition to obtain a ground-truth volume, then feed only the ±60° subset to NanoMorph-3D and compare its reconstruction and completed sinogram with the ground truth and the actually measured omitted views; if Fourier shell correlation in the wedge region and the withheld-view reprojection error are not clearly better than those of the best iterative baseline, the claimed resolution of missing-wedge hallucinations is not substantiated on real data.","tokens_in":12074,"feed_emoji":"🔬","tokens_out":12435,"duration_ms":105244,"temperature":0.7,"pith_summary":"NanoMorph-3D aims to fix the missing-wedge problem in electron tomography of nanomaterials, where hardware limits tilt angles to roughly ±60° and the unsampled wedge of Fourier space produces elongation and topological blur. The paper proposes an end-to-end network that unrolls proximal gradient descent into ten learnable stages, each alternating a projection-consistency gradient step (back-projecting the sinogram residual) with a deep topology-aware proximal map. A Dual-Domain Sinusoidal Attention injects the geometry of Radon projection trajectories into the attention logits, and a Physics-Normalization layer makes features scale-invariant across magnifications and densities. Trained on a large synthetic dataset built from a Nanomorphological Taxonomy with Beer-Lambert nonlinear attenuation, and adapted to real data by an unsupervised dual-stream reprojection loss, the framework reports the highest Fourier shell correlation (FSC, 0.6125) and lowest LPIPS among SIRT, GENFIRE, RESIRE, AET-Net, LPD, and Denza-GS, while completing missing-wedge sinograms and keeping fidelity down to 11 views.","feed_headline":"0.61 FSC: physics-aware network fills the missing wedge","feed_subtitle":"Physics-guided deep network reconstructs 3D nanomaterials from ±60° tilt series, beating six baselines.","key_machinery":"The key machinery is the unrolled proximal gradient step with the Projection Consistency Unit (PCU), $Z^{(k)} = \\mathrm{ReLU}\\left(V^{(k-1)} - \\frac{\\alpha^{(k)}}{M\\cdot N} B_\\Theta\\left(F_\\Theta(V^{(k-1)}) - Y\\right)\\right)$, followed by a deep Topology-Aware Proximal Mapping $H_\\theta$; the Dual-Domain Sinusoidal Attention adds the trajectory-compatibility bias $M_{\\text{geo}}$ into the attention logits, and Physics-Normalization (density-aware RMS norm) provides scale invariance. The Beer-Lambert-linearized forward model $Y = -\\frac{1}{\\mu s}\\ln\\left(1 - \\frac{I}{I_0}\\right) = A_\\Theta V + \\eta$ (Eq. 1) grounds both the simulator and the PCU, so the network's data-consistency path is the","core_discovery":"The central discovery is that unrolling proximal gradient descent into a learnable architecture with an explicit geometric bias removes missing-wedge hallucinations. Each of the ten unrolled stages performs a gradient step in the projection domain—back-projecting the residual between the forward projection of the current volume and the measured sinogram—followed by a deep proximal mapping that restores long-range 3D topology via hierarchical attention with Physics-Normalization. The geometric bias is a pairwise trajectory-compatibility prior $M_{\\text{geo}}$ computed from the sinusoidal locus $u(\\theta) = x\\cos\\theta + z\\sin\\theta$ of each voxel under the Radon transform; it is added to the","pith_inferences":["A direct extension the paper leaves implicit is that the same recipe—unrolled proximal steps plus a sinusoidal trajectory bias injected into attention—should transfer to any limited-angle Radon-type inverse problem (sparse-view X-ray CT, cryo-electron tomography, ground-penetrating radar), since the geometry prior is operator-specific but not microscope-specific.","The ablation results (FSC drops from 0.6125 to 0.354 when the projection-consistency gradient step is removed) suggest that many black-box 3D denoisers could be upgraded simply by alternating with a differentiable forward/back-projector, a change testable on existing networks without redesigning them.","A strong test of the physics-driven claim is to mask a real dual-axis or full-tilt dataset and compare the network's completed sinogram in the omitted wedge with the actually measured projections; if they disagree systematically, the sinogram consistency is an artifact of the shared simulator rather than the real forward model.","The reliance on Beer-Lambert implies a prediction: on real data with significant multiple scattering or detector nonlinearities, the method's margin over the best iterative baseline should shrink, and a version trained with a physics-perturbed simulator should show a corresponding drop."],"forward_implications":["Limited-angle HAADF-STEM tilt series can be reconstructed with a single forward pass of a 10-stage network, offering faster inference than iterative solvers and per-scene optimization used by neural rendering methods.","The dynamic-view-dropout curriculum lets the model complete withheld projection angles, enabling dose-friendly sparse-view acquisition down to 11 views while retaining SSIM 0.888 and FSC 0.495, both higher than full-41-view GENFIRE.","Because the unsupervised dual-stream loss uses unlabeled real tilt series as pseudo-targets via reprojection consistency, the method can adapt to lab-specific noise and vacuum artifacts without paired ground truth.","The taxonomy-driven synthetic dataset with Beer-Lambert nonlinear attenuation provides a reusable physics-grounded benchmark for evaluating generalizable nanomaterial reconstruction.","Physics-Normalization makes the learned priors invariant to magnification and mass-density scale, so a single trained model is intended to transfer across microscope settings."],"supporting_citations":[{"why":"Supplies the learned primal-dual unrolling paradigm that NanoMorph-3D extends and serves as the main unrolled-network baseline.","marker":"[1]"},{"why":"The image-domain post-processing baseline (AET-Net) whose physics-blind refinement the framework is designed to overcome.","marker":"[17]"},{"why":"Provides the public real HAADF-STEM tilt series used in the unsupervised dual-stream domain-adaptation training.","marker":"[18]"},{"why":"Iterative real-space reconstruction baseline (RESIRE) that the framework claims to outperform on 3D fidelity.","marker":"[24]"},{"why":"Iterative Fourier-space baseline (GENFIRE) that the framework claims to outperform, including at sparse views.","marker":"[25]"},{"why":"Defines the Beer-Lambert law that grounds the linearized forward model in Eq. (1), shared by the simulator and the Projection Consistency Unit.","marker":"[32]"},{"why":"Classical iterative reconstruction baseline (SIRT) against which fidelity and sparse-view robustness are measured.","marker":"[35]"},{"why":"3D Gaussian splatting method (Denza-GS) that the framework adapts to volumetric tomography and claims to beat.","marker":"[42]"}],"fun_headline_variants":["Unrolled physics net erases missing wedge artifacts","Proximal gradient unrolled to kill missing wedge","Dual-domain attention fills tomography blind spot","Learnable proximal steps beat wedge-limited tomography","Physics-driven unrolling reconstructs 3D nanomaterials"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The whole pipeline rests on Eq. (1): that after Beer-Lambert linearization, HAADF-STEM projections equal a discrete Radon transform of the density field plus noise, and that the same discretized operator appears in both the synthetic training simulator and the network's Projection Consistency Unit; if real beam-sample interaction deviates from this model—via multiple scattering, detector nonlinearity, or misalignment—the physics constraints and the test set are biased in the","fun_headline_variants_meta":{"raw":{"variants":["Unrolled physics net erases missing wedge artifacts","Proximal gradient unrolled to kill missing wedge","Dual-domain attention fills tomography blind spot","Learnable proximal steps beat wedge-limited tomography","Physics-driven unrolling reconstructs 3D nanomaterials"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000151,"raw_usage":{"total_tokens":1033,"prompt_tokens":739,"completion_tokens":294,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":233}},"tokens_in":483,"tokens_out":294,"duration_ms":3726,"temperature":1.0,"reasoning_tokens":233,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:21:54.903165+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reconstruct a nanomaterial from a dual-axis or full-tilt HAADF-STEM acquisition to obtain a ground-truth volume, then feed only the ±60° subset to NanoMorph-3D and compare its reconstruction and completed sinogram with the ground truth and the actually measured omitted views; if Fourier shell correlation in the wedge region and the withheld-view reprojection error are not clearly better than those of the best iterative baseline, the claimed resolution of missing-wedge hallucinations is not substantiated on real data.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the learned primal-dual unrolling paradigm that NanoMorph-3D extends and serves as the main unrolled-network baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The image-domain post-processing baseline (AET-Net) whose physics-blind refinement the framework is designed to overcome."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the public real HAADF-STEM tilt series used in the unsupervised dual-stream domain-adaptation training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Iterative real-space reconstruction baseline (RESIRE) that the framework claims to outperform on 3D fidelity."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Iterative Fourier-space baseline (GENFIRE) that the framework claims to outperform, including at sparse views."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Beer-Lambert law that grounds the linearized forward model in Eq. (1), shared by the simulator and the Projection Consistency Unit."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Classical iterative reconstruction baseline (SIRT) against which fidelity and sparse-view robustness are measured."}],"review_version":1}