{"id":"b05b87c7-abe0-4b5f-94c2-f97018517446","arxiv_id":"2608.00593","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A deep-ensemble active-learning surrogate predicts per-triangle impact velocities and particle flux in stream finishing with Spearman correlations up to 0.93 from 13% of DEM simulations.","lead":"This paper trains a neural network surrogate to predict how an abrasive finishing process erodes a workpiece, using only 13% of the possible workpiece orientations. If the method holds up, manufacturers could plan finishing steps for complex parts at a fraction of the current simulation cost.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Active-learning benefit is unverified: no random-sampling baseline, so the 13%-data accuracy cannot be attributed to uncertainty guidance.","rationale":"The reader's weakest_assumption focuses on whether the per-triangle fields are smooth, deterministic functions of the 11 geometric features. For the fixed cuboidal geometry, this concern is less acute than it appears: p0 identifies the facet, and the rotated position/normal encode the global orientation, so the mapping is in principle deterministic; the open question is learnability, which the paper's held-out results address empirically. The more load-bearing gap is the absence of any control for the active-learning rule. The paper's contribution is precisely the uncertainty-guided acquisition loop, and the experimental design never compares it to random or diversity-only sampling. Thus the observed improvement across iterations could be driven simply by adding more training data. The reader's rationale does mention this missing baseline, but it is not the stated weakest_assumption, hence 'partial' agreement. A randomized control is cheap, decisive, and directly tests the central claim; if it fails, the method reduces to a generic surrogate with no demonstrated advantage. The verdict should remain CONDITIONAL until this control is run, which is unchanged from the reader's assessment.","tokens_in":13604,"tokens_out":14293,"duration_ms":191379,"concrete_test":"Re-run the active-learning pipeline exactly as in §2.2.6 with the same 52 farthest-point seeds, M=12 ensemble, batch size 10, and 5 iterations, but replace the acquisition function with random selection from the remaining feasible orientations; repeat with 5 or more random seeds and report mean±SD of the held-out R2 and ρs trajectories and the low-uncertainty band fractions at each iteration. Also include a farthest-point-only acquisition baseline (add the 10 most distant remaining orientations each iteration) to separate diversity from uncertainty. If uncertainty-guided acquisition is not significantly better than random or farthest-point on both metrics, the central active-learning claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central contribution is the uncertainty-guided active-learning loop: ensemble disagreement selects which DEM orientations to simulate, and the headline result is that 92 simulations (~13% of 696) suffice. The load-bearing assumption is not merely that the per-triangle fields are learnable from the 11 geometric features—for a fixed geometry those features encode both the facet identity (p0) and the orientation (rotated position and normal), so the mapping is in principle deterministic—but that the acquisition rule is actually better than random or diversity-only sampling. This is never tested. The paper reports monotone gains on a fixed 14-orientation held-out set and growth of the low-uncertainty band on the remaining feasible set, but any rule that adds 10 diverse orientations per iteration would produce some improvement because the training set grows. The seed set already covers the feasible set by the d* threshold (§2.2.5), so random additions within the covered space may be nearly as informative. Without a control arm, the observed improvement cannot be attributed to the epistemic-uncertainty signal, and the '13%' claim is not evidence for the active-learning method. Secondary evaluation weaknesses—only 14 hand-picked validation orientations, uncertainty thresholds fit on the same 14, no code/data—compound this but do not substitute for the missing baseline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a deep-ensemble surrogate to predict per-triangle normal impact velocity, tangential impact velocity, and particle impact flux for a cuboidal workpiece in stream finishing, using 11 geometric features per triangle. A farthest-point seeding scheme selects 52 initial DEM orientations, followed by five active-learning iterations that add the 10 orientations with the largest ensemble disagreement. The surrogate is evaluated on 14 held-out orientations and on uncertainty-band growth over the 696 feasible orientations. The authors report final-iteration Spearman rank correlations of 0.93, 0.89, and 0.93 for the three fields, claim the uncertainty is well calibrated, and reconstruct wear fields via the Finnie model matching DEM with Spearman correlation up to 0.97 at low uncertainty, using only 13% of the feasible orientations. The core idea is plausible, but the evidence presented does not yet establish the central claims about active-learning benefit and calibration.","tokens_in":13960,"tokens_out":6785,"duration_ms":83699,"significance":"The paper has several genuine strengths. The decision to predict smooth primary impact fields and to confine the nonlinear Finnie wear model to an analytical reconstruction is physically sensible and likely improves learnability. The use of an applicability-domain condition to avoid extrapolation is a thoughtful safeguard. The problem is practically relevant: orientation-dependent stream-finishing wear maps are expensive to obtain with DEM, and a reliable surrogate could substantially reduce cost. If the active-learning gains were supported by an appropriate control, the framework would be a useful template for geometry-to-wear surrogates. However, the manuscript currently stops short of demonstrating the central contribution: that uncertainty-guided acquisition outperforms simpler sampling strategies under an equal simulation budget.","major_comments":[{"comment":"No baseline against random or diversity-only sampling is provided. The active-learning loop in §2.2.6 adds the 10 orientations with the largest σ_e and the abstract and §4 attribute the 13%-budget result to this uncertainty-guided strategy. But because the seed set already covers the feasible set by the d* criterion (§2.2.5), and because the training pool grows by 10 orientations per iteration, any reasonable rule that adds orientations within the covered region could plausibly produce similar gains. The reported monotone growth of the low-uncertainty band and the held-out accuracy is therefore not evidence for the value of the uncertainty signal. A control arm with the same seed set, batch size, retraining schedule, and either random or farthest-point additions is required to support the paper's central claim.","section":"§2.2.6, §3.2"},{"comment":"The uncertainty calibration and its validation use the same 14 held-out orientations. Figure 5 reports Spearman correlations between σ_e and R²/ρ_s on those 14 orientations, and the σ_e thresholds for the low/medium/high bands are then fitted to the same 14-point σ_e–ρ_s relationship. This conflates calibration with validation, so the 'truthful uncertainty' claim and the band-growth figures are not independent evidence. A separate calibration split, or nested cross-validation, should be used, and the threshold values should be reported with uncertainty. With N=14, the rank correlations also have wide confidence intervals; point estimates of −0.77 and −0.82 should be accompanied by confidence intervals. In addition, §2.2.6 does not describe how the 14 validation orientations were selected; if hand-picked, this introduces another potential bias.","section":"§3.1"},{"comment":"No error bars or repeated realizations are reported for any main result. The deep ensemble uses random initializations and bootstrap subsets (§2.2.2), so the selected batches and final metrics are stochastic. Table 2.2 and Figures 6-7 present point estimates of R², ρ_s, and band fractions; the claims of monotone improvement and saturation at iteration 3 could reflect one realization. Re-running the loop with different ensemble seeds, or at least reporting variance over repeated training runs, is needed to support the quantitative 13%-data claim.","section":"§3.2, Table 2.2, Figs. 6-7"},{"comment":"The band thresholds fitted at iteration 5 are applied to earlier iterations ('the thresholds are fixed and applied unchanged to every iteration'), yet the σ_e distribution changes substantially across iterations, as shown in Table 2.2. The claim that a fixed threshold denotes the same accuracy level is reasonable only if the σ_e–ρ_s relationship is stable across iterations; this is asserted but not demonstrated. Reporting the σ_e–ρ_s relationship at each iteration, or fitting thresholds on a separate calibration set that is not also used for the Fig. 5 truthfulness test, would address this concern.","section":"§3.1"}],"minor_comments":[{"comment":"Typos and copyediting: 'uncertainity' in the Introduction, 'presenrlty' in the Summary, and the formatting of 'Ansys Rocky 2024R2' should be corrected.","section":"Abstract/Introduction"},{"comment":"The relationship between d*, defined as the 95th percentile of nearest-neighbor distances within F, and its later use as a covering-radius stopping criterion needs clarification. Please state explicitly whether farthest-point sampling continued until the maximum distance to the nearest seed was below d*, and confirm that no feasible orientation lies beyond d* from a seed.","section":"§2.2.5"},{"comment":"The paper uses orientation-averaged R² and ρ_s, but notes that R² is unbounded below and one bad orientation can dominate an average. Reporting medians or per-orientation distributions instead of only averaged values would give a more complete picture, especially with only 14 validation orientations.","section":"§3.1"},{"comment":"No code or data availability statement is provided. Given the reproducibility value of the DEM surrogate pipeline, the authors should state whether the trained models, generated data, and scripts will be released.","section":"Global"},{"comment":"The 'irreducible limit' from unresolved stochastic collisions is acknowledged, but its size is never estimated. A comparison between ensemble error and the run-to-run variability of repeated DEM simulations at the same orientation would help contextualize the achievable accuracy of the surrogate.","section":"§2.2.3"}],"recommendation":"major_revision","confidential_remarks":"The missing random-sampling baseline is the key gap; I would not accept the paper until it is addressed. The paper is otherwise technically sound and within the journal's scope. If the authors add a control arm and an independent calibration split, I would expect the revised manuscript to be publishable. No concerns about attribution or novelty disclosure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper does something worth doing, and does it mostly competently, but the headline claim that uncertainty-guided active learning is what gets them to 13% data is not supported. There is no comparison to random or diversity-only acquisition, so the 13% result shows that a decent surrogate can be trained on 92 diverse simulations, not that ensemble disagreement is the right acquisition signal.\n\nThe genuinely new bit is the domain formulation: predict the smooth per-triangle impact fields (normal/tangential velocity, flux) and push the sharp nonlinearity into the Finnie wear model, instead of regressing wear directly. That is a smart, physically motivated choice. The input features (rotated position, normal, baseline coordinates, radial distance, flow alignment) are all computable from geometry, which makes the surrogate applicable to arbitrary STL geometries. The demonstration on a cuboid in stream finishing is concrete and the results are plausible: rank correlations ~0.93, wear reconstructions near 0.97 in the low-uncertainty regime, and a believable monotone relationship between ensemble disagreement and realized error.\n\nThe soft spots are real, though. The biggest one, as the stress-test says, is the missing baseline. The seed set is already a farthest-point cover of the feasible set, so every orientation is within d* of a training example; adding 10 diverse orientations per iteration would likely improve any surrogate. Without a random-sampling arm or a diversity-only acquisition, you cannot attribute the observed improvement to the epistemic-uncertainty signal. The '13%' headline is therefore about data efficiency in general, not about active learning specifically. The paper should either add the baseline or soften the claim.\n\nThe evaluation is also thin in places: 14 held-out orientations total, thresholds for the uncertainty bands fit on those same 14, no error bars on the main numbers, and no code or data. Those are fixable but they matter, especially because the uncertainty-band growth plots are the main evidence that the loop is working.\n\nOn the plus side, the paper is honest about its limits — it explicitly notes the irreducible stochastic-collision floor in Sec 2.2.3 and does not oversell extrapolation. The physics-informed features are sensible, and the idea of predicting smooth fields rather than wear itself is something other surrogate-model papers could reuse. I'd send it to a serious referee, but the revision needs to address the baseline and tighten the evaluation before I'd trust the active-learning claim.","headline":"Useful surrogate for stream finishing wear, but the active-learning benefit is not actually demonstrated without a random-sampling baseline.","tokens_in":14351,"tokens_out":2172,"would_cite":false,"duration_ms":25958,"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":"The paper shows that an uncertainty-guided surrogate can map stream-finishing wear over all 696 feasible orientations from just 92 DEM simulations, with calibrated uncertainty that predicts where the reconstruction can be trusted.","keywords":["stream finishing","discrete element method","surrogate model","deep ensemble","active learning","epistemic uncertainty","wear field prediction","Finnie erosion model"],"falsifier":"Run DEM on 30 orientations that the final surrogate labels low-uncertainty but were never simulated, then compare predicted and DEM wear fields. The central claim fails if a substantial fraction of these orientations show Spearman rank correlations below about 0.90, or if the predicted uncertainty no longer orders the orientations by realized error.","tokens_in":1571,"feed_emoji":"⚙️","tokens_out":1639,"duration_ms":74319,"temperature":0.7,"pith_summary":"Stream finishing wears a workpiece differently at each orientation, but simulating every feasible orientation with the discrete element method (DEM) is too expensive. This paper argues that a surrogate network can predict the three per-facet fields that determine erosion—normal impact velocity, tangential impact velocity, and particle flux—directly from geometric features, and that feeding these predictions through the Finnie wear formula reconstructs the entire wear-rate map. The surrogate is a deep ensemble: disagreement among its twelve members estimates where the model is uncertain, and that uncertainty selects which orientations to simulate next. Trained on 92 of 696 feasible orientations (13%), the surrogate reaches hold-out Spearman rank correlations of 0.93, 0.89, and 0.93 for the three fields, and reconstructed wear fields match DEM with rank correlation up to 0.97 for low-uncertainty orientations. If correct, orientation planning for new parts becomes affordable and comes with a trustworthy signal for where more simulation is needed.","feed_headline":"92 simulations replace 696 for stream-finishing wear maps","feed_subtitle":"Deep-ensemble uncertainty picks the next simulations, reconstructing wear with rank correlation up to 0.97.","key_machinery":"The argument is carried by combining a deep ensemble with the Finnie wear relation. Rather than regressing wear rates directly, the surrogate predicts three smooth per-triangle fields from an 11-dimensional physical feature vector: rotated centroid position, surface normal, baseline coordinates, radial distance from the container axis, and alignment of the normal with the azimuthal flow. Wear is then reconstructed analytically through the Finnie formula, which confines the sharp angular and quadratic nonlinearity to an equation. The deep ensemble's disagreement provides the epistemic uncertainty that classifies orientations into accuracy bands and selects the next DEM simulations. A farthest","core_discovery":"The central claim is that the wear-rate field of a workpiece in stream finishing is a learnable function of geometry alone, and that its smooth constituents—per-facet normal impact velocity, tangential impact velocity, and particle impact flux—can be learned from a small fraction of all orientations. The paper shows that a deep ensemble seeded with 52 farthest-point orientations and refined through five rounds of ten uncertainty-selected DEM runs predicts these fields with held-out Spearman rank correlations of 0.93, 0.89, and 0.93. The ensemble's between-member standard deviation is a truthful pre-DEM estimate of realized error, with rank correlations of -0.77 and -0.82 against R² across he","pith_inferences":["A natural but untested extension: the uncertainty thresholds are fit to the flat-wall workpiece, so applying the same pipeline to a new geometry would require recalibration of the σ_e-to-accuracy mapping with a small validation set.","The 11 features are local and static; for parts with strong mutual shielding or internal passages, a facet's impact statistics may depend on the surrounding geometry, which could shrink the learnable smooth part and require more active-learning iterations.","Because the DEM model uses a reduced Young's modulus, the predicted wear magnitudes are relative; converting surrogate outputs to absolute wear would still require experimental calibration.","Choosing the next simulations by mean field-level uncertainty may be less efficient than targeting uncertainty in the reconstructed wear field or in the rank ordering of hotspots; that is a testable modification the paper does not explore."],"forward_implications":["A new workpiece's full orientation-dependent wear map can be obtained from roughly a tenth of the exhaustive DEM programme, making orientation-sequence design feasible in practice.","The input features are computable from any triangulated mesh, so the same pipeline transfers to arbitrary complex part shapes without new input engineering.","Low-uncertainty orientations can be trusted without running DEM; only high-uncertainty orientations need additional simulation, and the uncertainty itself tells the user which ones.","Rank-based agreement remains high even where absolute wear magnitudes degrade, so decisions based on relative wear across orientations remain usable at medium uncertainty.","The low-uncertainty population is still growing at iteration 5, so a modest number of further batches is expected to bring nearly the whole feasible set into confident coverage."],"supporting_citations":[{"why":"Supplies the discrete element method used to generate the per-impact training data and the ground truth for validation.","marker":"[4]"},{"why":"Provides the Finnie erosion formula that combines the predicted velocity and flux fields into reconstructed wear rates.","marker":"[5]"},{"why":"Introduces the deep-ensemble method that provides the epistemic uncertainty estimate driving the active-learning loop.","marker":"[22]"},{"why":"Justifies the reduced Young's modulus that makes each DEM simulation computationally affordable while preserving bulk flow behaviour.","marker":"[26]"},{"why":"Supplies the k-nearest-neighbour applicability-domain approach used to set the interpolation threshold for seed selection.","marker":"[30]"},{"why":"Extends the nearest-neighbour applicability idea to deep models, supporting the in-distribution requirement for the surrogate.","marker":"[31]"},{"why":"Provides the farthest-point/core-set sampling strategy used to build the initial seed set from geometry alone.","marker":"[33]"}],"fun_headline_variants":["Uncertainty-guided AI predicts wear with 92 DEM runs, not 696","Deep ensemble maps stream-finishing wear from 13% of orientations","Active learning cuts DEM simulations from 696 to 92 for wear","13% of orientations predict wear fields at 0.97 correlation"],"cache_read_input_tokens":16128,"weakest_assumption_plain":"The load-bearing premise is that each surface facet's impact velocity and flux are smooth, deterministic functions of the 11 geometric features; if unresolved many-body collision noise dominates the variation, the surrogate's sparse-sample accuracy and its uncertainty calibration will not transfer to new geometries or operating conditions.","fun_headline_variants_meta":{"raw":{"variants":["Uncertainty-guided AI predicts wear with 92 DEM runs, not 696","Deep ensemble maps stream-finishing wear from 13% of orientations","Active learning cuts DEM simulations from 696 to 92 for wear","13% of orientations predict wear fields at 0.97 correlation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000423,"raw_usage":{"total_tokens":2019,"prompt_tokens":765,"completion_tokens":1254,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":1188}},"tokens_in":509,"tokens_out":1254,"duration_ms":12354,"temperature":1.0,"reasoning_tokens":1188,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T00:31:02.812839+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DEM on 30 orientations that the final surrogate labels low-uncertainty but were never simulated, then compare predicted and DEM wear fields. The central claim fails if a substantial fraction of these orientations show Spearman rank correlations below about 0.90, or if the predicted uncertainty no longer orders the orientations by realized error.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the discrete element method used to generate the per-impact training data and the ground truth for validation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Finnie erosion formula that combines the predicted velocity and flux fields into reconstructed wear rates."},{"cited_title":"Dem speedup: Stiffness effects on behavior of bulk material.Particuology, 12:107–112, 2014","cited_arxiv_id":null,"evidence_quote":"Justifies the reduced Young's modulus that makes each DEM simulation computationally affordable while preserving bulk flow behaviour."},{"cited_title":"Defining a novel k-nearest neighbours approach to assess the applicability domain of a qsar model for reliable predictions.Journal of cheminformatics, 5(1):27, 2013","cited_arxiv_id":null,"evidence_quote":"Supplies the k-nearest-neighbour applicability-domain approach used to set the interpolation threshold for seed selection."}],"review_version":1}