{"id":"cb3f0203-ea63-4120-9598-997c5d3f8fc9","arxiv_id":"2501.05869","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A U-Net trained on CFD-DEM simulations maps top-view dune images to per-grain force maps, but the experimental application lacks any ground-truth validation.","lead":"Researchers trained a neural network on computer simulations of sand dunes to guess the forces pushing on each grain from a top-down photo. They applied it to real water-tank dunes, but with no direct force measurements to check against, the real-dune numbers are unverified estimates.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Experimental force maps are unvalidated: the paper concedes there is no access to Y in the experimental domain, and the binarized preprocessing was adopted only after the original model gave unexpected experimental outputs.","rationale":"The reader's weakest assumption—that the morphology-to-force mapping learned in CFD-DEM is identical in the real experiments—is precisely the load-bearing issue. The paper's own language confirms the absence of experimental ground truth, so the experimental force maps are unvalidated projections, not measurements. I additionally note that the binarization step was chosen after seeing unsatisfactory outputs from the initially adversarially adapted model; this makes the one qualitative experimental 'check' circular rather than confirmatory. The numerical proof of concept may be a reasonable starting point, but it does not support the abstract's unconditional 'high accuracy' claim for real dunes. I agree with the reader's REJECT verdict and would not change it. The proposed concrete test is feasible with the authors' existing experimental setup and would either validate or refute the central claim.","tokens_in":13369,"tokens_out":3783,"duration_ms":43971,"concrete_test":"Use the existing high-speed top-view experimental recordings to track individual surface grains across consecutive frames (as in the group's earlier particle-tracking work, Refs. 15/16/19) and compute their accelerations. Compare the measured longitudinal component of m*a for a grain with the CNN-predicted resultant force for that grain. If the predicted and measured values do not agree within the expected uncertainty—e.g., no significant correlation or a systematic bias—then the claimed sim-to-real transfer is falsified. This directly tests the central claim without needing embedded force sensors.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that the CNN quantifies forces on grains of a real dune from images—requires the morphology-to-force mapping learned in CFD-DEM to transfer to the real water-tank experiments. The paper explicitly concedes this premise is unchecked: 'As there are no force measurements for experimental dunes (we have no access to Y in the experimental domain), one can only verify the capacity of the model in the numerical domain.' Consequently, the experimental force map in Fig. 2d is not a validated measurement. The only experimental check is qualitative and was made after the authors observed that the domain-adversarial model 'was not sufficient to infer correct results in experimental images' and then changed the input preprocessing to binarization. That is post-hoc selection on the target-domain outputs, so even the qualitative agreement is not an independent test. The numerical-domain results in Fig. 3 show only time-averaged forces and variances, not per-grain instantaneous errors, and they cannot certify cross-domain transfer. The abstract's 'high accuracy' and 'breakthrough' claims therefore outrun the evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a deep-learning method to estimate the longitudinal component of the resultant force on individual grains of subaqueous barchan dunes from top-view images. The authors train a U-Net on synthetic images generated by their own CFD-DEM simulations, where grain forces are known, and then apply the trained network to experimental images of water-tank barchans. They report qualitative agreement between predicted and simulated mean forces and variances on unseen simulated dunes, and they present an experimental force map (Fig. 2d) obtained from an experimental image. The central claim is that the method 'can quantify these forces with high accuracy' and constitutes a breakthrough in measuring grain-scale forces from images.","tokens_in":13502,"tokens_out":3303,"duration_ms":32974,"significance":"If the central claim were established, the proposed technique would be a genuinely novel measurement concept with potential applications beyond dune grains, including planetary remote sensing. The paper is commendable for making data, code, and model weights publicly available, and for explicitly discussing the domain-shift problem between simulations and experiments. However, the claimed result — quantifying forces on grains of real dunes — is not actually validated in the manuscript. The experimental predictions have no ground truth, the numerical validation is only qualitative, and the experimental protocol was adjusted after observing the target-domain outputs. The significance is therefore prospective rather than demonstrated.","major_comments":[{"comment":"The central quantitative claim about experimental dunes is unsupported. The authors explicitly write that 'there are no force measurements for experimental dunes (we have no access to Y in the experimental domain),' so the performance check is restricted to the numerical domain. The experiment shown in Fig. 2d is never compared against any measured force distribution, yet the abstract states that the method can 'quantify these forces with high accuracy.' This is a load-bearing gap: the title promises a way to measure forces on grains of a real dune, but the paper provides no evidence that the experimental force map is accurate.","section":"Estimating the resultant force on grains of real dunes"},{"comment":"The experimental evaluation is post-hoc and therefore not an independent test. The text reports that the domain-adversarial model 'was not sufficient to infer correct results in experimental images,' and that only after switching to binarized inputs did the outputs match the authors' expectation that the convex periphery should be red and the concave region predominantly blue. Because the preprocessing choice was selected after observing the experimental outputs, the subsequent qualitative agreement is not a valid confirmation of the method. The force map in Fig. 2d remains an unvalidated model projection rather than a measurement.","section":"Supplementary Materials, CNN model"},{"comment":"The numerical-domain generalization claim is supported only by qualitative visual comparison of time-averaged forces and variances. No quantitative error metrics, confidence intervals, or per-grain instantaneous errors are reported. The text describes the results with phrases such as 'well predicted' and 'reasonably estimate' without any numerical support. Given the abstract's 'high accuracy' claim, the absence of any quantitative assessment of prediction error is a major omission even within the simulation domain where ground truth exists.","section":"Generalization to dunes never seen by the CNN model, Fig. 3"},{"comment":"The training labels are exclusively force maps from the authors' own CFD-DEM simulations, and the paper's validation of those simulations refers to prior work (Refs. 21, 22) without showing that the force field itself was experimentally validated. Since the entire morphology-to-force mapping is learned from these simulated labels, any systematic error in the simulated forces would propagate into all predictions, including the experimental ones. The manuscript does not address this source of uncertainty.","section":"Supplementary Materials, Numerical simulations"}],"minor_comments":[{"comment":"The abstract overclaims by asserting 'high accuracy' and a 'breakthrough' while the body of the paper explicitly states that accuracy can only be verified in the numerical domain. These claims should be tempered to match the evidence.","section":"Abstract"},{"comment":"There is a grammatical slip: 'We also want aim to develop the capability' should read 'We also aim to develop the capability.'","section":"Main text, 'Estimating the resultant force' section"},{"comment":"The notation in the description of the mapping function, 'Φ : R → R3, g(x) ↦ f(x) | x ∈ Rn', is confusing; the domain and codomain are not written consistently and should be clarified.","section":"Supplementary Materials, CNN model"},{"comment":"The figure compares ground truth and model outputs only for time-averaged forces and variances; adding a panel with a quantitative error metric (e.g., mean absolute error or structural similarity) would make the comparison interpretable.","section":"Fig. 3"}],"recommendation":"reject","confidential_remarks":"The paper is well organized and the authors are transparent about the lack of experimental force ground truth, which is commendable, but that transparency directly contradicts the abstract's claims of high accuracy for real dunes. The experimental section is not a validation but an illustration, and the post-hoc binarization choice undermines even the qualitative check. Fixing this would require either new experiments with a force ground truth or a substantial reframing of the paper's claims as a simulation-only study with a speculative experimental extension. I do not see such a fix being achievable within the current manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is new and worth taking seriously: nobody has trained a U-Net to map top-view images of barchan dunes to per-grain longitudinal force maps, and the paper backs it with open code, data, and trained weights. That is real and reproducible. The simulation-to-simulation generalization, though only shown visually, does suggest the network can interpolate across Reynolds numbers and particle counts. The authors also write honestly about why domain adaptation is hard for regression and why they switched to binarized inputs.\n\nBut the soft spot is load-bearing. The paper explicitly concedes there is no access to the true forces in the experimental domain, so the experimental force maps in Fig. 2d are not validated measurements. The only check is qualitative: the authors expected simulation-like color patterns, and after the adversarial approach gave unexpected outputs, they changed the input preprocessing to binarization. That is post-hoc selection on the target-domain outputs, so even the qualitative agreement is not an independent test. The abstract's 'high accuracy' and 'breakthrough' claims outrun the evidence.\n\nThe simulation-domain validation also lacks quantitative error metrics. Figure 3 shows time-averaged forces and variances with no per-grain errors, no confidence intervals, no numbers. 'High accuracy' is asserted, not demonstrated.\n\nIf the paper were framed as a proof-of-concept—'we can train a network on CFD-DEM to predict forces on unseen simulated dunes, and we show it might transfer to experiments, pending validation'—it would be a solid methods contribution. As written, it overclaims measurement of real dune grain forces.\n\nThe citation pattern looks fine; prior work on grain trajectories and simulated forces is cited, and the new contribution is distinct. No invented entities, no obvious misconduct.\n\nWho gets value: granular physicists and remote-sensing folks interested in image-based inference of inaccessible quantities. A serious referee should see this, because the concept is promising and the code is public. But the authors need to either obtain independent force measurements, add quantitative sim-domain error metrics, and pre-specify the preprocessing, or substantially soften the claims. My recommendation: send it to peer review with the expectation of major revision.","headline":"A genuinely new sim-to-real force mapping idea with open code, but the experimental claim rests on unvalidated, post-hoc-adjusted outputs.","tokens_in":14082,"tokens_out":1626,"would_cite":false,"duration_ms":18003,"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":"A U-Net trained on simulated barchan dunes estimates the longitudinal force on individual grains from top-view images, and the authors apply it to real experimental dune images.","keywords":["barchan dunes","grain-scale forces","convolutional neural network","CFD-DEM simulation","image-to-image translation","subaqueous experiments","domain adaptation","deep learning"],"falsifier":"Run the trained network on experimental images of a subaqueous barchan while simultaneously measuring actual grain-scale forces, for example by tracking individual grains with a high-speed camera to compute their accelerations from the trajectories, and compare the predicted force map to these measured forces. Any systematic discrepancy between the two would falsify the claim that the CNN estimates real forces from images.","tokens_in":13138,"feed_emoji":"🏜️","tokens_out":9577,"duration_ms":72853,"temperature":0.7,"pith_summary":"This paper claims that a convolutional neural network can turn a plain top-view image of a sand dune into a quantitative map of the longitudinal component of the resultant force acting on each grain. The network is trained entirely on CFD-DEM simulations of subaqueous barchan dunes, where the true forces are known, and is then applied to high-speed camera images of real laboratory dunes. The authors report that the network predicts force distributions on simulated dunes it has never seen, and that it produces physically sensible force maps when applied to experimental images. If the approach holds, images from cameras or satellites could become a remote sensor for grain-scale forces on Earth and Mars.","feed_headline":"Dune images alone can reveal grain-scale forces","feed_subtitle":"CNN trained on simulated dunes predicts grain forces from top-view photos, no instrumented grains needed.","key_machinery":"The workhorse is a U-Net, an encoder-decoder convolutional network originally designed for image segmentation with limited training data. The authors treat force estimation as an image-to-image regression: the input is a binarized top-view image of the dune (binarization strips photometric noise and forces the network to focus on shape and contour), and the output is a color-coded map of the longitudinal component of the resultant force per grain. Training pairs come from CFD-DEM simulations that solve the fluid flow and the motion of individual grains, providing ground-truth forces. A domain-adversarial module with gradient reversal was tried to align simulation and experimental image statistics but was found insufficient; the final pipeline relies on input binarization, data augmentation (rotation, translation, zoom), and pooling of multiple simulation conditions in the training set.","core_discovery":"The central discovery is that the morphology of a barchan dune, as seen from above, carries enough information to recover the longitudinal component of the resultant force on its individual grains. By training a U-Net with paired inputs (binarized top-view dune images) and outputs (force-color maps from simulation), the network learns a mapping from shape to force that generalizes to dunes with different Reynolds numbers and particle counts. The same network, applied to experimental images captured in a water tank under matching flow conditions, yields force estimates on groups of grains, reproducing the expected pattern of strong downstream force at the dune's convex periphery and recirculation force in the lee. The authors are explicit that the experimental force maps cannot be directly verified because no force measurements exist for experimental dunes; validation is limited to the numerical domain.","pith_inferences":["The experimental force maps in the paper's Fig. 2d should be read as model projections rather than verified measurements, since the paper itself states that no experimental ground-truth forces exist.","A natural next step is to test the pipeline on eolian or Martian dune images, but the shift in image statistics and grain-scale physics will likely be far larger than the simulation-to-water-tank shift.","Because the final pipeline depends on binarization, the accuracy of any field application is bounded by the quality of grain segmentation; misclassified grains would propagate into force errors.","If the morphology-to-force mapping is as general as claimed, the same architecture could be retrained to output other grain-scale fields, such as grain velocity or erosion rate, from the same images."],"forward_implications":["Applied to experimental images under the same flow conditions, the network yields force estimates on groups of grains without any modification.","The CNN generalizes to simulated dunes never seen during training, including cases with different Reynolds numbers and particle counts.","The same approach could be extended to remote-sensing images to estimate forces on rocks, boulders, rovers, and man-made structures on Earth and Mars.","By providing grain-scale force information from images alone, the method avoids the need to embed accelerometers or other sensors in the dune."],"supporting_citations":[{"why":"The CFD-DEM simulations that produce the paired training images and ground-truth grain forces.","marker":"(21–23)"},{"why":"The U-Net architecture used for the image-to-image force regression.","marker":"(38)"},{"why":"The CFDEM coupling framework that links the CFD and DEM solvers.","marker":"(50)"},{"why":"The DEM solver used to compute individual grain motion and contact forces.","marker":"(51,52)"},{"why":"The gradient-reversal domain-adversarial method tried for aligning simulation and experimental image domains.","marker":"(70)"}],"fun_headline_variants":["Top-view photos estimate every dune grain's force","Sim-trained CNN maps dune images to grain forces","Dune shape encodes per-grain forces","No-touch force estimates for dune grains from photos"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the mapping from dune shape to grain forces learned from the authors' own CFD-DEM simulations remains valid for real experimental dunes once the images are binarized, even though no experimental force measurements exist to check it.","fun_headline_variants_meta":{"raw":{"variants":["Top-view photos estimate every dune grain's force","Sim-trained CNN maps dune images to grain forces","Dune shape encodes per-grain forces","No-touch force estimates for dune grains from photos"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001048,"raw_usage":{"total_tokens":4357,"prompt_tokens":849,"completion_tokens":3508,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":465,"completion_tokens_details":{"reasoning_tokens":3446}},"tokens_in":465,"tokens_out":3508,"duration_ms":25924,"temperature":1.0,"reasoning_tokens":3446,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:05:50.405683+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained network on experimental images of a subaqueous barchan while simultaneously measuring actual grain-scale forces, for example by tracking individual grains with a high-speed camera to compute their accelerations from the trajectories, and compare the predicted force map to these measured forces. Any systematic discrepancy between the two would falsify the claim that the CNN estimates real forces from images.","supporting_citations":[],"review_version":1}