{"id":"d458587a-5a7d-433e-bde5-826018aef651","arxiv_id":"2411.13366","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A graph neural network with separate encoders for tube, die, and stamp contacts predicts wall thickness changes in nosing forging from FEM training data, evaluated with a new ABTC metric.","lead":"This paper trains a graph neural network on finite element simulations to predict how tube wall thickness changes during cold nosing, a metal forming process. It also introduces a new error metric and reports that the network runs much faster than the simulation it mimics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central architectural claim—that separate edge-type encoders improve accuracy—is never tested against a shared-encoder baseline, leaving the paper's main novelty unsupported.","rationale":"The reader's stated weakest assumption is that the FEM simulation itself is unvalidated against physical experiments. That is a real limitation for the physical-process claim, but it is not the most decisive issue for the paper's technical contribution: a surrogate model can be valid relative to FEM without being experimentally validated. The more load-bearing weakness is internal and directly targets the claimed novelty: the separate edge-type encoders are asserted to enhance accuracy, yet no ablation removes or merges them. Section 4.3 is labeled 'Ablation studies' but varies step width, contact presence, and dataset coverage—never the encoder-sharing scheme. Thus the statement in Section 5 that separation 'resulted in more accurate models' has no supporting experiment. The paper does provide some genuine evidence: a mesh convergence study, a no-contact control showing the model does not hallucinate die forces, and a step-size study. These support the overall surrogate feasibility but do not substantiate the architecture-specific claim. I agree with the reader's overall CONDITIONAL verdict: the work is plausible and addressable, but a key claim needs a direct test. Since my concern reinforces the reader's conditional assessment rather than overturning it, the verdict should remain unchanged.","tokens_in":19548,"tokens_out":4458,"duration_ms":51420,"concrete_test":"Retrain the identical architecture with a single shared edge encoder for all three edge sets (using edge type as a one-hot feature) and compare ABTC and RMSE on the same nine held-out test configurations. If the shared-encoder model's total ABTC is not significantly worse than the 212.152 mm^2 reported for step=1 in Table 3, the claimed benefit of separate encoders is not supported. Report multiple random seeds and state whether the 0.8 relative-length truncation is applied to both models.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty is the claim that splitting message passing into three edge-type-specific encoders (Eqs. 9-10: tube-tube E_t, tube-die E_d, tube-stamp E_p) improves surrogate accuracy. The conclusion states this directly: 'The separation of edge types into their corresponding encoders resulted in more accurate models' (Section 5), and the abstract advertises this augmentation as accuracy-enhancing. However, Section 4.3, the only ablation section, tests (i) step width, (ii) removing die contact, and (iii) full-dataset rollouts; none compares the proposed multi-encoder architecture against the single-shared-encoder baseline of [35] or any equivalent. The reference to [35] is an architectural description, not an experimental comparison. Consequently, the distinguishing contribution of this work is untested. This is distinct from the FEM-validation concern: even granting FEM as ground truth, the paper does not establish that separate encoders cause the reported accuracy. The truncation of displayed thickness profiles at 0.8 relative tube length (Fig. 16) further suggests the worst collar region is excluded from headline results, but the missing baseline is the more direct threat to the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a graph neural network (GNN) surrogate for finite element method (FEM) simulations of tube nosing in cold forging. It first presents a domain analysis and a 2D axisymmetric FEM model to study how the deformation degree, die opening angle, and friction coefficient affect wall thickness. The GNN follows the encode-process-decode framework of Pfaff et al. but introduces separate edge encoders for three interaction types (tube–tube, tube–die, tube–stamp) with dynamically added contact edges, and a new evaluation metric called the area between thickness curves (ABTC). The model is trained on multi-step rollout targets from FEM and tested on nine held-out parameter configurations, showing visually plausible mesh predictions and thickness profiles that track FEM results, with a considerable computation-time reduction compared with FEM. The authors conclude that the separation of edge types into dedicated encoders improves accuracy and enables real-time process use.","tokens_in":19736,"tokens_out":6349,"duration_ms":69517,"significance":"If the central claims are substantiated, this work would be a useful contribution to surrogate modeling in metal forming, offering a domain-tuned GNN architecture and a concrete evaluation metric (ABTC) for comparing thickness profiles. The paper provides a reasonably thorough FEM preprocessing study, a pragmatic use of dynamic edges for contact, and a clear speed comparison. The ABTC metric is a sensible addition for this application. However, the main advertised novelty—the multi-encoder separation—is not tested against a shared-encoder baseline, and the evaluation omits variance measures, full-profile reporting, and experimental validation. The work is therefore promising but currently does not establish its central architectural claim.","major_comments":[{"comment":"The conclusion states that \"the separation of edge types into their corresponding encoders resulted in more accurate models,\" but no experiment in the paper tests this assertion against the single-encoder baseline of [35] or any equivalent shared-encoder architecture. The ablation studies in Section 4.3 cover step width, removal of die contact, and full-dataset rollouts, but none isolates the effect of the proposed multi-encoder design. Since this separation is the main architectural novelty advertised in the abstract and introduction, the central claim is currently unsupported. Please add a controlled comparison, e.g., the proposed architecture vs. a GraphNet with a single shared edge encoder, on the same training/testing split and metrics.","section":"Section 5 and Section 4.3"},{"comment":"The wall-thickness evaluation is truncated at 0.8 relative tube length, which is precisely where the model is reported to show its largest deviations (the collar/funnel region). The authors state that the prediction \"tends to deviate at the beginning of the tube\" and that the deviation occurs in the collar area, yet the quantitative ABTC results in Table 3 and Figure 20 appear to be computed on the truncated profiles. To avoid overstating accuracy, please present the full thickness profiles or explicitly justify why the excluded region is irrelevant to downstream process-monitoring tasks, and report ABTC on the full tube length as a secondary result.","section":"Section 4.2.1, Figure 16"},{"comment":"The execution-time comparison is not rigororous because the FEM and NN timings are measured on different hardware (FEM on an Intel Core i9-7980XE with GTX 1070; NN on an Intel Core i7-11700 with RTX 5000). The claimed speedup is therefore confounded by hardware differences. Please report timings on matched hardware, or at least provide per-simulation-step timings, CPU-only and GPU-only results, and state the assumptions that make the comparison meaningful.","section":"Table 4"},{"comment":"The test set consists of only nine hand-picked configurations, and all reported ABTC and RMSE values appear to be single deterministic rollouts. Because the model includes noise injection during training and random initialization, repeated runs will produce different results. Please report means and standard deviations over multiple training seeds and/or multiple rollouts for the test configurations, so the generalization claims in Section 4.3.3 are supported statistically.","section":"Section 2.2.3, Section 4.3.3"},{"comment":"The paper correctly lists comparison with real experimental setups as future work, but the abstract and introduction frame the surrogate as predicting wall thickness changes in nosing processes generally. The GNN is trained and evaluated only against FEM output, so its accuracy is contingent on the validity of the FEM model (2D axisymmetric assumption, Hollomon flow-curve parameters, Coulomb friction model, and material database). Please add an explicit scope statement that the surrogate is a model of the specific FEM setup, and discuss how FEM discretization and constitutive-model uncertainties might affect the predicted thickness profiles in real production.","section":"Section 5"}],"minor_comments":[{"comment":"In the sentence \"Within the processor, f d, f d, f p and f n are also MLPs,\" the second \"f d\" should be \"f t\"; the equation itself correctly uses f^d, f^t, f^p, and f^n.","section":"Section 3.2, Eq. (9)"},{"comment":"There are typos: \"rigit element\" should be \"rigid element,\" \"visibible\" should be \"visible,\" and \"paramters\" should be \"parameters.\"","section":"Section 2.2.1"},{"comment":"The legend labels \"ground truth, tube, die\" do not indicate which colors correspond to predicted versus ground-truth tube and die. Please clarify the color/line mapping in the caption.","section":"Figure 15"},{"comment":"The phrase \"the general different is still big enough\" should read \"the general difference is still big enough.\" Also, the hardware comparison is described only in the text; a supplementary table listing CPU/GPU models, memory, and software versions would improve reproducibility.","section":"Section 4.4"},{"comment":"The manuscript states \"Data availability: Not applicable\" and \"Code availability: Not applicable.\" For a machine-learning paper with a custom dataset and architecture, providing the FEM dataset, preprocessing scripts, and model code would greatly strengthen reproducibility and should be considered.","section":"Data and Code Availability"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of the Journal of Intelligent Manufacturing, but the main novelty (multi-edge-type encoders) is not validated against a baseline, and the evaluation has important gaps (truncated profiles, no error bars, unmatched hardware in timing). These issues are fixable with additional experiments. The data/code availability statement is unusual for this type of data-driven paper and may deserve editorial attention. I recommend major revision rather than rejection because the overall approach is sound and the missing experiments are within the scope of the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I agree with the reader's conditional verdict, and the stress-test note lands. The paper is not circular: it trains a GNN on FEM outputs and evaluates on held-out FEM outputs, which is supervised learning. The missing piece isn't logic, it's evidence.\n\nWhat's genuinely useful: the nosing process description is careful, the FEM setup includes a mesh convergence study, and Figure 13b comparing existing analytical models to FEM is a nice reality check showing the analytical formulas disagree and ignore friction. The die-contact ablation (Section 4.3.2) is a good control—it shows the model doesn't just hallucinate forming when the die is gone. The ABTC metric is simple but sensible for this application. Execution time numbers (301 min FEM vs 3.5 min GPU) make the practical motivation concrete.\n\nThe soft spots are real and concentrated in the evaluation. The central architectural claim—that splitting message passing into tube-tube, tube-die, and tube-stamp edge encoders improves accuracy—is stated in the abstract and Section 5 but never compared against the single-shared-encoder MeshGraphNet baseline. Section 4.3 tests step width, die removal, and full-dataset rollouts, but not the architecture itself. That's the one ablation that would justify the paper's main claim. Second, the thickness plots in Figure 16 stop at 0.8 relative tube length, which excludes exactly the collar region where the authors say the prediction deviates. That may be a defensible scope decision for process monitoring, but it needs to be stated as a deliberate exclusion with collar numbers reported separately, not just as a display choice. Third, every ABTC number is a single point—no repeated runs, no error bars, on a nine-case test set. With 2.88M parameters and noise injection, run-to-run variance could be nontrivial. Fourth, the FEM model itself is never checked against a physical experiment; the authors list that as future work, and the conclusion should carry the caveat that 'accuracy comparable to FEM' means comparable to an unvalidated simulation.\n\nNone of this is fatal to the engineering story. The surrogate looks plausible, the process analysis is informative, and the missing pieces are addressable. But the paper's most distinctive claim is the least supported one, and that needs fixing before publication.\n\nWho gets value: researchers working on mesh-based GNN surrogates for forming processes, and manufacturing engineers considering closed-loop control. I'd bring it to a reading group as a case study in evaluation design. I wouldn't cite it yet for the architecture claim, but I might cite the process analysis and ABTC if I were in that subfield.\n\nRecommendation: send to peer review, not desk reject, with a clear request for a shared-encoder baseline, full-range thickness curves including the collar, and repeated trials or uncertainty estimates. If those are added, the conditional acceptance is justified.","headline":"The paper has a real engineering use case and a plausible surrogate, but the headline claim—that separate edge encoders are better—is never tested against the shared-encoder baseline.","tokens_in":20335,"tokens_out":2648,"would_cite":false,"duration_ms":28044,"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 graph neural network predicts tube wall thickness changes in nosing as accurately as FEM, at a fraction of the cost.","keywords":["cold forging","nosing process","graph neural network","surrogate model","finite element method","wall thickness prediction","area between thickness curves","real-time process control"],"falsifier":"Run instrumented physical nosing trials on CuZn39Pb2 tubes spanning the parameter grid (φ from 0.05 to 0.35, α from 5° to 25°, μ = 0.05 and 0.2), measure wall-thickness profiles along the tube after forming, and compare them with both the FEM and the GNN rollouts using the ABTC metric; if the FEM itself deviates substantially from the measured profiles, or if the GNN's ABTC relative to the FEM is much larger than the values reported in Table 3 on the same scale, the surrogate's accuracy claim fails.","tokens_in":19310,"feed_emoji":"⚙️","tokens_out":4984,"duration_ms":51637,"temperature":0.7,"pith_summary":"The paper claims that a purpose-built graph neural network can replace slow finite-element simulations for predicting wall thickness changes in the cold nosing of tubes, with accuracy close to the FEM reference and runtime reduced from about 301 minutes to roughly 3.5 minutes on a GPU. It argues that because nosing involves three distinct contact mechanisms—deformation inside the tube, contact between tube and die, and contact between stamp and tube—the network should use a separate encoder for each interaction type, with edges dynamically added as parts come into contact. The authors also introduce a metric, the area between thickness curves (ABTC), that quantifies how well a predicted wall-thickness profile matches the FEM profile along the tube. If the claim holds, such surrogates could run inside closed-loop production control, where traditional FEM is too slow.","feed_headline":"Graph network matches FEM on tube forging, ~86x faster","feed_subtitle":"Separate encoders for tube, die, and stamp cut a 301-minute FEM run to 3.5 minutes on a GPU.","key_machinery":"The central mechanism is a graph neural network with three edge sets and corresponding encoders: internal tube edges model material deformation, dynamic edges between tube and die model forming contact, and dynamic edges between stamp and tube model axial feed. Node encoders include mesh connectivity, positions, and a solid/deformable flag; edge features include relative Cartesian coordinates, Euclidean distance, and the friction coefficient. A shared message-passing processor with k=15 steps updates node features, and a decoder outputs the per-node displacement Δp = $p^{{t+1}}$ − p^t; the trajectory is generated by forward integration. The paper extends the mesh-graph-network baseline by splitting the single edge encoder into three interaction-specific encoders, which is what makes the nosing-specific physics learnable.","core_discovery":"The core discovery is that a graph-neural-network surrogate—built on an encode-process-decode one-step predictor with three specialized edge encoders (tube-internal, tube-die, tube-stamp) and dynamic distance-based edge insertion—can reproduce the FEM-computed mesh deformation and wall-thickness evolution of a CuZn39Pb2 tube during die nosing. Trained on 207,778 FEM timesteps covering variations in forming degree φ, die opening angle α, and Coulomb friction μ, and tested on 48,045 held-out timesteps, the network rolls out 4,500 steps and matches the FEM thickness profiles well, with deviations concentrated at the collar region that leaves the die and with no catastrophic error growth. The network has 2.88 million trainable parameters and runs the full simulation in about 3.5 minutes on a GPU versus 301 minutes for FEM on a CPU. The authors additionally show that the model learns contact genuinely: when the die is moved out of contact, no forming occurs.","pith_inferences":["The same encoder-splitting architecture could be extended to other tube-forming operations such as expanding, necking with a mandrel, or flow forming by adding edge types for the new tool contacts, and to different materials by adding material parameters as node or edge features.","Because the model is trained only on FEM data, its practical value depends on FEM fidelity; a physical validation campaign comparing measured wall-thickness profiles to both FEM and GNN outputs would be the decisive test, and the ABTC metric could be reused as the evaluation measure.","The noisy predictions at high deformation degrees (φ = 0.35) suggest the dataset's edge-of-distribution region is a bottleneck; adding training simulations near process limits or using uncertainty-aware predictors could improve reliability where failures such as buckling are most likely."],"forward_implications":["If the model generalizes as claimed, real-time monitoring and closed-loop control of nosing processes become feasible, since a full 4,500-step rollout takes minutes on a GPU rather than hours on a CPU.","The separation of edge types suggests a design recipe for surrogate models of other forging and forming operations with multiple contacting tools: encode each distinct contact pair separately.","The ABTC metric gives a physically interpretable error measure for wall-thickness predictions and could become a standard for comparing forming surrogates.","The step-size ablation shows that larger time steps are only slightly worse, so the surrogate could be run with coarser steps, trading a little accuracy for additional speed.","The model's recovery from initial rollout deviations means short-horizon errors do not necessarily compound into failure over a full forming cycle."],"supporting_citations":[{"why":"Supplies the baseline encode-process-decode mesh-graph-network architecture and the dynamic-edge idea that the paper modifies with three specialized edge encoders.","marker":"[35]"},{"why":"Provides the graph-network simulation approach and the training-noise strategy used to stabilize multi-step rollouts.","marker":"[29]"},{"why":"Provides the empirical wall-thickness-change behavior and the thick-walled tube regime against which the FEM model is compared.","marker":"[3]"},{"why":"Provides the thin-walled tube formula and the retraction-ratio description of nosing that the paper builds on.","marker":"[2]"},{"why":"Supplies the empirical formula for final wall thickness and the parameter analysis (forming degree, die angle, friction) that motivates the FEM design.","marker":"[4]"},{"why":"Provides the mathematically derived wall-thickness formula used as a comparison baseline for the FEM results.","marker":"[44]"}],"fun_headline_variants":["Graph net matches FEM on tube forging, 86x faster","Neural surrogate predicts wall thickness, 86x speedup over FEM","Forging simulation cut from 301 min to 3.5 min via graph net","Specialized edge encoders power fast neural forging surrogate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the FEM simulation itself faithfully represents the real nosing process for CuZn39Pb2 tubes; the network is trained and evaluated only against FEM output, and the authors state that comparison with a real experimental setup is future work.","fun_headline_variants_meta":{"raw":{"variants":["Graph net matches FEM on tube forging, 86x faster","Neural surrogate predicts wall thickness, 86x speedup over FEM","Forging simulation cut from 301 min to 3.5 min via graph net","Specialized edge encoders power fast neural forging surrogate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000552,"raw_usage":{"total_tokens":2629,"prompt_tokens":936,"completion_tokens":1693,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":1617}},"tokens_in":552,"tokens_out":1693,"duration_ms":15216,"temperature":1.0,"reasoning_tokens":1617,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:30:20.925523+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run instrumented physical nosing trials on CuZn39Pb2 tubes spanning the parameter grid (φ from 0.05 to 0.35, α from 5° to 25°, μ = 0.05 and 0.2), measure wall-thickness profiles along the tube after forming, and compare them with both the FEM and the GNN rollouts using the ABTC metric; if the FEM itself deviates substantially from the measured profiles, or if the GNN's ABTC relative to the FEM is much larger than the values reported in Table 3 on the same scale, the surrogate's accuracy claim fails.","supporting_citations":[{"cited_title":"In: 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021","cited_arxiv_id":null,"evidence_quote":"Supplies the baseline encode-process-decode mesh-graph-network architecture and the dynamic-edge idea that the paper modifies with three specialized edge encoders."},{"cited_title":"In: III, H.D., Singh, A","cited_arxiv_id":null,"evidence_quote":"Provides the graph-network simulation approach and the training-noise strategy used to stabilize multi-step rollouts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the empirical wall-thickness-change behavior and the thick-walled tube regime against which the FEM model is compared."},{"cited_title":"teil i, ii und iii","cited_arxiv_id":null,"evidence_quote":"Provides the thin-walled tube formula and the retraction-ratio description of nosing that the paper builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the empirical formula for final wall thickness and the parameter analysis (forming degree, die angle, friction) that motivates the FEM design."},{"cited_title":"Berlin (1968) 38","cited_arxiv_id":null,"evidence_quote":"Provides the mathematically derived wall-thickness formula used as a comparison baseline for the FEM results."}],"review_version":1}