{"id":"475d5922-556f-413d-b0a0-439131f7bd23","arxiv_id":"2412.08286","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A neural network fitted to 34 experimental samples reports 95.24% accuracy for bolted-joint load capacity and friction coefficients, but the targets are derived from the input measurements and only 7 test samples are used.","lead":"This paper trains a small feed-forward neural network on 34 measurements of bolted joints to predict load capacity and friction coefficients, reporting 95.24% accuracy on a small test set. It shows a preliminary ML application to bolted joint design, but the tiny, homogeneous dataset and circularly derived targets limit its value.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 95.24% accuracy is not evidence of predictive skill because Section 3.1 indicates the output targets were empirically estimated from the same torque and preload time series that form the input features, making the learned mapping a potential identity.","rationale":"The reader's weakest assumption is the same as the load-bearing concern I identify: the output targets must be genuine independent quantities, not deterministic functions of the input features. Section 3.1 explicitly says the outputs were 'further empirically estimated' from the measured torque and preload time series, and the input layer contains exactly those torque and preload values plus bolt size and strength grade. If friction coefficients and load capacity are computed from these inputs using standard mechanical relations, the network's near-perfect test accuracy is the expected result of fitting a known formula, and it carries no information about capturing nonlinear bolted-joint behavior. This concern is decisive for the central claim independent of the secondary issues of the 7-sample test set, hyperparameter selection on the test set, and the 5% threshold accuracy metric. Because the reader already rejected the paper on essentially this basis, my verdict is unchanged. The proposed concrete test — recomputing the Section 3.1 targets from the input features and checking for exact identity — would settle the matter analytically without requiring new experiments.","tokens_in":8873,"tokens_out":3540,"duration_ms":41741,"concrete_test":"Reconstruct the target-generation procedure from Section 3.1 using the original time-series data and the empirical formulas for head friction coefficient, thread friction coefficient, and remaining load capacity. Write each output as an explicit function of the six input features (bolt size, strength grade, tightening torque, head torque, thread torque, preload force). If every output is algebraically determined by the inputs — e.g., friction coefficient as torque divided by the product of preload force and a geometric radius, with bolt size fixing the radius — then the model is learning identities. The decisive check is to recompute the target values from the input features alone and compare them with the targets used in training Model 4; an exact match demonstrates that the 95.24% accuracy reflects formula reproduction, not independent predictive capability.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a feed-forward neural network 'effectively captures nonlinear relationships' and achieves 95.24% predictive accuracy for load capacity and friction coefficients. For that claim to be meaningful, the output targets must be independent measurements or at least not deterministic functions of the input features. Section 3.1 states: 'From these values, the friction coefficients of the head and threads and the remaining load capacity were further empirically estimated,' where 'these values' are the measured preload force, tightening torque, head torque, and thread torque. The model inputs include bolt size, strength grade, tightening torque, head torque, thread torque, and preload force. Under standard bolted-joint mechanics, friction coefficients are ratios of torque to preload times geometry, and load capacity is similarly derived from the same measured quantities. If the targets were computed from the inputs, then the network is fitting known algebraic identities rather than learning physical joint behavior, and the 95.24% test accuracy — essentially 20 of 21 individual output values within a 5% threshold — is a statement about curve fitting, not prediction. This concern is reinforced by the tiny test set (7 samples from the same M6/M10 configurations) and the explicit hyperparameter tuning against test accuracy in Section 3.3. The Discussion's own caveat about overfitting and limited generalizability does not repair the core issue: the headline accuracy is uninterpretable unless the target-generation formulas are shown to be independent of the inputs.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a feed-forward neural network that takes bolt size, strength grade, tightening torque, head torque, thread torque, and preload force as inputs and predicts three outputs: remaining load capacity and head and thread friction coefficients. The data consist of 34 samples from M6 and M10 8.8 bolted joints, split 80/20 into training and test sets. After experimenting with 136 models, the authors report that their best model (Model 4) reaches 95.24% test accuracy, defined as predictions within 5% of target values. The paper emphasizes the integration of empirical data with ML and acknowledges that the small dataset limits generalizability.","tokens_in":9133,"tokens_out":4471,"duration_ms":49913,"significance":"If the reported accuracy reflected true predictive skill on independently measured targets, the contribution would be a useful engineering convenience: a fast surrogate for derived bolted-joint parameters. The paper is transparent about its architecture, hyperparameters, training pseudocode, and limitations, which is commendable. However, the central claim is not supported by the evidence as presented: the target values are estimated from the same measured quantities that are used as input features, the test set is tiny and drawn from configurations represented in training, and hyperparameters were tuned against test accuracy. The paper therefore demonstrates curve fitting of a nearly deterministic transformation rather than predictive skill for unseen joint designs.","major_comments":[{"comment":"The output targets are not independent measurements. Section 3.1 states that 'From these values, the friction coefficients of the head and threads and the remaining load capacity were further empirically estimated,' where 'these values' are the measured preload force, tightening torque, head torque, and thread torque—exactly the quantities that, together with bolt size and strength grade, form the six input features listed in Sections 1.2 and 3.1. Under standard bolted-joint mechanics (Section 1.1), friction coefficients are derived from torque-preload ratios and geometry, and remaining load capacity is computed from the same measured preload and geometry. The mapping from inputs to targets is therefore nearly deterministic, and the network's high accuracy largely reflects reproducing the estimation equations. To support the 'predictive' claim, the authors should report the exact estimation formulas, quantify the uncertainty of the estimated targets, and validate against independent measurements (e.g., direct friction tests) or at least demonstrate that the outputs are not algebraic functions of the inputs.","section":"Section 3.1"},{"comment":"Hyperparameters were tuned using the test set, so the reported 95.24% accuracy is an in-sample selection result. Section 3.3 states that 'The evaluation and hyperparameter tuning process involved analyzing the accuracy and loss curves from training, as well as the error metrics and accuracy curves from testing, to assess the model's alignment with the data' and that 'Hyperparameters were iteratively adjusted until satisfactory accuracy was achieved.' This means the test set was not held out; it was used for model selection. The reported accuracy cannot be taken as an unbiased estimate of generalization. The authors should use nested cross-validation or an untouched test set for the final evaluation.","section":"Section 3.3"},{"comment":"The test set is far too small to support the central generalization claim. A 20% split of 34 samples yields roughly 7 test samples, all drawn from the same M6/M10 configurations and preload levels present in training. The overall 95.24% figure corresponds to 20 of 21 individual output values being within the 5% threshold, and the head friction coefficient accuracy is 85.71%, i.e., 6 of 7 values. No confidence intervals, per-output error bars, or per-configuration breakdowns are provided. This sample size cannot justify the assertion that the model 'effectively captures nonlinear relationships' across different bolt configurations. The authors should report per-output MAE/RMSE with bootstrap confidence intervals and explicitly separate interpolation from extrapolation performance.","section":"Section 4 / Table 3"},{"comment":"The accuracy metric is not sufficiently defined or justified. The authors state that 'Predictions were deemed accurate if they deviated by no more than 5% from the target values,' but it is not specified how the three outputs are aggregated—whether each output value is counted separately (which the 20/21 arithmetic implies) or whether each test sample must have all three outputs within tolerance. The 5% threshold is arbitrary, and a threshold-based accuracy on a single small test set is not a complete performance characterization. The authors should report continuous error metrics per output and include the uncertainty in the target values themselves.","section":"Section 4"}],"minor_comments":[{"comment":"The phrase 'rule-based formulations of bolt behavior like VDI VDI 2230' contains a duplicated 'VDI'; the reference should also be disambiguated (VDI 2230 Part 1, 2, or 3).","section":"Section 5"},{"comment":"The claim that the model 'effectively captures nonlinear relationships' is not supported by any comparison to linear regression, an analytical baseline, or a physics-based formula; adding such a baseline would help calibrate what the neural network actually adds.","section":"Abstract / Section 1"},{"comment":"No equations are given for the 'empirically estimated' friction coefficients and remaining load capacity; the authors should either provide the formulas or cite specific equations from Wettstein and Matthiesen (2020) and VDI 2230 so that the target construction is reproducible.","section":"Section 3.1"},{"comment":"The manuscript does not state whether the code, data, or trained model weights are publicly available, which prevents an independent check of the reported 95.24% accuracy.","section":"General"},{"comment":"The architecture figure does not show activation functions, bias initialization, or scaling layers, and Table 2 does not specify how normalization parameters were applied to the output for Model 4; providing this information would improve reproducibility.","section":"Figure 2 / Table 2"}],"recommendation":"reject","confidential_remarks":"The circularity issue is not a routine fix: the paper's central claim is built on a regression whose targets are derived from the input features. A standard revision cannot repair this without new independent target data or a fundamental reframing of the contribution as a curve-fitting demonstration rather than a predictive model."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline number — 95.24% predictive accuracy for load capacity and friction coefficients — doesn't survive contact with the methods. In Section 3.1 the friction coefficients and remaining load capacity are \"further empirically estimated\" from the same torque and preload time series that serve as the network's inputs. Under standard bolted-joint mechanics, those targets are algebraic functions of the inputs. So the network is fitting known identities, not learning physical behavior. The stress-test note is right: the accuracy claim is uninterpretable.\n\nThe paper does some things well. It is a clear write-up of a small feed-forward NN regression exercise: proper scaling, a sensible loss function, and an honest Discussion that flags overfitting risk and limited generalizability. The related work is adequate and actually shows prior NN models for load capacity and torque loss, so the novelty is minimal — this is a preliminary case study, not a new method.\n\nThe soft spots are substantive. The test set has 7 samples from the same M6/M10 configurations used in training, so it doesn't measure generalization to new designs. Hyperparameters were tuned against test accuracy, which inflates the reported number. The 5% accuracy tolerance is loose — hitting 20 of 21 output values within 5% is not impressive when the targets are derived from the inputs. The authors' own caveats do not repair the circularity.\n\nWho is this for? Readers new to ML in mechanical design might learn the workflow, and someone teaching pitfalls could use it as an example of target leakage. But as a research result, it doesn't support the conclusion that NNs offer a reliable design-time surrogate for these parameters.\n\nMy recommendation: don't send this to peer review in its current form. The fix is straightforward in principle — measure the friction coefficients and load capacity independently, or at least show they are not deterministic functions of the inputs. Without that, the central claim is unsupported. A serious referee would reject on those grounds.","headline":"The 95.24% accuracy is not a predictive result: the output targets are derived from the input features, so the network is fitting known relations rather than learning joint behavior.","tokens_in":9689,"tokens_out":2323,"would_cite":false,"duration_ms":24255,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A feed-forward neural network trained on 34 experimental runs predicts bolted-joint load capacity and friction coefficients to 95.24% accuracy.","keywords":["Bolted Joints","Data-Driven Design","Predictive Modeling","Neural Network","Artificial Intelligence","Feed-Forward Neural Network","Friction Coefficient","Load Capacity"],"falsifier":"The decisive check is whether a much simpler model, such as linear regression on the same six inputs and the same 34 records, matches the 95% accuracy on the same holdout; if it does, the neural network offers no evidence of learning beyond the estimation formula used to create the targets.","tokens_in":8670,"feed_emoji":"🔩","tokens_out":9155,"duration_ms":86362,"temperature":0.7,"pith_summary":"The paper tries to show that a small feed-forward neural network can serve as a fast, accurate substitute for slow analytical or finite-element work in bolted-joint design. Trained on 34 experimental records of M6 and M10 bolted joints, the network takes bolt size, strength grade, tightening torque, head torque, thread torque, and preload force as inputs and predicts remaining load capacity plus head and thread friction coefficients. The authors report 95.24% predictive accuracy, where a prediction counts as correct if it deviates by no more than 5% from the target; load capacity and thread friction reached 100% on the test split and head friction 85.71%. The result matters because, if it holds, routine friction and load-capacity checks could be done in seconds without advanced mechanical expertise. The authors themselves caution that the tiny dataset limits generalizability, framing the work as a preliminary demonstration.","feed_headline":"Neural net hits 95% on bolted-joint predictions","feed_subtitle":"Small feed-forward model trained on torque and preload data forecasts load capacity and friction coefficients in seconds.","key_machinery":"The mechanism carrying the argument is a fully connected feed-forward neural network with six input nodes (bolt size, strength grade, tightening torque, head torque, thread torque, preload force), two hidden layers, and three output nodes (load capacity, head friction coefficient, thread friction coefficient). Training uses stochastic gradient descent at learning rate 0.01 with Huber loss and batch size 4; activation and initialization vary across the four reported models, with the best model using sigmoid activation, Xavier initialization, normalization, 34 samples, and preload/load measured in kN/MN. The load-bearing preprocessing step is rescaling load capacity to MN so that output scales are comparable, which the discussion credits with reducing noise in the loss and stabilising training. Accuracy is scored as the fraction of test predictions within 5% of the target, yielding the headline 95.24% figure.","core_discovery":"On the paper's own terms, the central claim is that a supervised feed-forward neural network with two hidden layers, trained on empirical tightening data, captures the nonlinear relationship between tightening variables and functional behaviour of bolted joints well enough to predict the remaining load capacity and the head and thread friction coefficients simultaneously to 95.24% accuracy. The decisive configuration detail is that rescaling load capacity to meganewtons, so that its numerical range matches the friction coefficients, was what stabilised training: with the raw newton scale the loss was dominated by load capacity. The network's accuracy is defined relative to a 5% tolerance band, and the test-set breakdown is 100% accuracy for load capacity and thread friction and 85.71% for head friction, with one head-friction point deviating. The paper presents this as evidence that data-driven prediction can complement or replace idealized analytical assumptions and computationally heavy numerical models for the tested configurations, while explicitly limiting the claim to this dataset and loading condition.","pith_inferences":["The paper does not test whether a simpler baseline, such as linear regression on the same six inputs, achieves comparable accuracy on the same holdout; because the targets were empirically estimated from the torque and preload time series, such a test would show whether the network is learning new physics or an estimation formula.","With 34 samples and a 20% holdout, the test set contains roughly seven points, so a single head-friction outlier is the difference between 100% and 85.71% on that output; resampling the train/test split would likely move the headline accuracy by several points.","A natural extension the authors do not run is pretraining the same architecture on synthetic data from the VDI 2230 calculation rules and fine-tuning on the experimental records; improved out-of-configuration accuracy would support hybrid modeling more strongly than the current experiment.","The usability claim that the model can be operated without mechanical expertise could be checked by an independent design study in which engineers unfamiliar with the test rig size bolts using the tool; accuracy alone does not establish usability."],"forward_implications":["Within the tested M6 and M10 configurations, a trained network produces load capacity and both friction coefficients in about 90 seconds, making interactive design iteration feasible.","Under the paper's scoring rule, a prediction is correct only if it falls within 5% of the target; on the test split, load capacity and thread friction hit 100% and head friction 85.71%.","The rescaling of load capacity to MN was the preprocessing change that stabilised training, so similar unit alignment should be used when extending the method to outputs with mixed scales.","Expanding the dataset to more bolt sizes, strength grades, and load cases is the stated precondition for interpreting 95.24% as a general capability rather than a configuration-specific result."],"supporting_citations":[{"why":"Supplies the experimental torque/preload time series for the M10 and M6 bolted joints that form the training and test data.","marker":"(Wettstein and Matthiesen, 2020)"},{"why":"Provides the engineering relations for preload, load capacity, and friction that justify the selected input and output features.","marker":"(VDI 2230, 2015)"},{"why":"Documents how tightening torque is split between head and thread friction and preload, grounding the friction-coefficient outputs.","marker":"(VDI 2230, 2024)"},{"why":"Prior neural-network ensemble prediction of bolted lap-joint load capacity that this study extends to simultaneous friction and capacity prediction.","marker":"(Fernández-Ceniceros et al., 2012)"},{"why":"Methodological framework for combining empirical bolted-joint data with data-driven functional modeling.","marker":"(Afifi et al., 2024)"},{"why":"General rationale for data-driven modeling of nonlinear engineering behaviour from data.","marker":"(Montáns et al., 2019)"}],"fun_headline_variants":["Rescaled neural net predicts joint strength at 95%","Bolted joints: neural net forecasts friction and load","Machine learning predicts bolted-joint parameters with 95% accuracy","Neural net rescaling enables 95% joint prediction accuracy","Feed-forward net predicts joint load and friction to 95%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the load capacity and friction coefficients used as training targets are independent, reliable values; if those targets are merely recomputed from the same torque and preload inputs by a known estimation rule, the reported 95.24% accuracy mostly reflects formula inversion rather than predictive skill for new joint designs.","fun_headline_variants_meta":{"raw":{"variants":["Rescaled neural net predicts joint strength at 95%","Bolted joints: neural net forecasts friction and load","Machine learning predicts bolted-joint parameters with 95% accuracy","Neural net rescaling enables 95% joint prediction accuracy","Feed-forward net predicts joint load and friction to 95%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000415,"raw_usage":{"total_tokens":2107,"prompt_tokens":870,"completion_tokens":1237,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":1152}},"tokens_in":486,"tokens_out":1237,"duration_ms":11072,"temperature":1.0,"reasoning_tokens":1152,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:59:59.902223+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The decisive check is whether a much simpler model, such as linear regression on the same six inputs and the same 34 records, matches the 95% accuracy on the same holdout; if it does, the neural network offers no evidence of learning beyond the estimation formula used to create the targets.","supporting_citations":[{"cited_title":"and Matthiesen, S","cited_arxiv_id":null,"evidence_quote":"Supplies the experimental torque/preload time series for the M10 and M6 bolted joints that form the training and test data."},{"cited_title":"Blatt 1 -- systematische berechnung hochbeanspruchter schraubenverbindungen - zylindrische einschraubenverbindungen","cited_arxiv_id":null,"evidence_quote":"Provides the engineering relations for preload, load capacity, and friction that justify the selected input and output features."},{"cited_title":"Blatt 3 -- systematische berechnung hochbeanspruchter schraubenverbindungen - zylindrische einschraubenverbindungen","cited_arxiv_id":null,"evidence_quote":"Documents how tightening torque is split between head and thread friction and preload, grounding the friction-coefficient outputs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior neural-network ensemble prediction of bolted lap-joint load capacity that this study extends to simultaneous friction and capacity prediction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Methodological framework for combining empirical bolted-joint data with data-driven functional modeling."}],"review_version":1}