{"id":"d73be7ac-4429-47a6-b8a8-543c1535fe62","arxiv_id":"2411.13869","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Filtered local member-combination features and a convolution-based transfer let a neural-net surrogate reduce structural analyses in simulated-annealing topology optimization of periodic lattice structures.","lead":"This paper trains neural networks to predict how periodic lattice frames deform, using local combinations of neighboring bars as inputs. The surrogate speeds up simulated-annealing topology optimization by skipping expensive structural analyses, though the reported designs are slightly worse than those from plain simulated annealing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Transfer equivalence fails: the 8x8 duplication of the 4x4 optimum (Fig. 10) has compliance 0.81299 vs 0.71656 Nm, but convolution maps it to the same input vector, so the NN surrogate is ~13% biased at the start point of the 8x8 local search.","rationale":"The reader's weakest assumption identified the 4x4-to-8x8 mechanical equivalence as a load-bearing point; I agree that this is the right area, but the concern can be sharpened into an internal inconsistency using numbers already in the paper. The 8x8 initial structure obtained by duplicating the 4x4 optimum has compliance 0.81299 Nm, while the 4x4 optimum has compliance about 0.71656 Nm (Table 3), a 13% difference. Because the convolution with weight 0.25 averages four identical fine members back to the coarse indicator, the convolved input to the NN is identical to the 4x4 optimum's input. Therefore the NN's prediction for the 8x8 initial structure is essentially the prediction for the 4x4 structure, which is far from the actual 8x8 compliance. This directly undermines the claim that the scaled cross-sections make 4x4 and 8x8 units mechanically equivalent for the sparse, optimized topologies that actually occur, rather than for dense ground structures (Table 2 ground-structure compliance values are close: 0.53748 vs 0.54031). The local search still found near-optimal solutions in Table 4, so the paper is not fatally flawed, but the transfer mechanism is less supported than claimed. The appropriate verdict remains CONDITIONAL: the method is promising and reimplementable, but this specific transfer bias needs to be either fixed, quantified, or removed from the claims. I therefore leave the reader's verdict unchanged, with a partial agreement because the reader flagged the equivalence assumption but not the direct contradiction from the paper's own initial-structure data.","tokens_in":7839,"tokens_out":10628,"duration_ms":108353,"concrete_test":"Recompute the compliance and the NN prediction for the 8x8 initial structure of Fig. 10 before any local-search iterations. Convolve the 8x8 binary member arrays with the same 2x2 kernel (weight 0.25) and apply the sigmoid filter of Eq. (5), then feed the result into the trained 4x4 NN. Compare the predicted objective value with the actual 0.81299 Nm and with the NN's prediction for the corresponding 4x4 optimum. If the prediction tracks the 4x4 value (about 0.71656) rather than the actual 8x8 value, the surrogate has a bias of roughly 13% at the starting design, and the transfer-based acceleration is unreliable in the low-compliance regime where optimization operates.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the pre-trained 4x4 model can be applied to 8x8 units via convolution (Sec. 3.3, 5.1.3) rests on the assertion that the cross-section scaling in Table 1 makes 4x4 and 8x8 units mechanically equivalent, so averaging 2x2 blocks with weight 0.25 yields a faithful coarse representation. The paper's own data contradict this equivalence for the structures of interest. In Sec. 5.2.2, the 8x8 initial structure is generated by locally duplicating each member of the best 4x4 structure into four members; the resulting 8x8 structure has compliance 0.81299 Nm (Fig. 10), whereas the original 4x4 optimum has compliance about 0.71656 Nm (Table 3, red entry). After the proposed 2x2 convolution with weight 0.25, every coarse member position is exactly the average of four identical fine-member indicators, so the convolved input equals the 4x4 optimum's input; the pre-trained NN therefore predicts roughly the 4x4 optimum's objective value for a structure whose true compliance is ~13% higher. This is not a small distribution-shift detail: it is a systematic bias exactly at the initial point of the optimization the method is supposed to accelerate. The local search in Table 4 still recovers solutions within about 1% of pure SA, but that recovery is achieved despite the surrogate being wrong at the starting design, not because of the transfer. If the convolution representation cannot preserve compliance under the paper's own scaling, then the 'sufficient accuracy' claim for the transfer is unsupported, and the 8x8 optimization results lose their stated foundation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a machine-learning-assisted topology optimization method for periodic lattice structures. It introduces a filtering process that encodes local member connectivity as input features, uses f_regression-based feature selection, and proposes a convolution operation to transfer a network trained on small (4x4) units to larger (8x8) units. The resulting neural network predicts compliance and is embedded into a simulated-annealing framework that skips structural analyses for predicted-inferior candidates. Numerical examples for 4x4 and 8x8 grids compare the NN-assisted SA with pure SA in terms of obtained compliance and number of structural analyses.","tokens_in":8227,"tokens_out":6040,"duration_ms":48106,"significance":"If validated, the proposed surrogate-based acceleration could reduce the cost of heuristic topology optimization for periodic lattices. The paper has several strengths: the final optimized structures are evaluated by finite-element analysis, so the optimization result is not based on the surrogate; the filtering and feature-selection pipeline is clearly described; and the convolution transfer idea is interesting and potentially useful. However, the evidence supporting the central claims is incomplete: accuracy comparisons are based on training loss curves, the convolution weight is selected on the target dataset, and the optimization results show pure SA achieving lower compliance, which conflicts with the abstract's wording. These issues currently prevent the paper from establishing its claims.","major_comments":[{"comment":"The abstract states that the proposed method 'can find the optimal structure in less computation time than the pure simulated annealing,' but Tables 3 and 4 show the opposite in terms of solution quality: pure SA yields lower average compliance in both cases (0.77588 vs 0.79047 for 4x4; 0.78699 vs 0.79318 for 8x8) and a lower best compliance in both cases. The wording should be revised to say that the NN-assisted method gives comparable but slightly worse solutions while reducing the number of structural analyses. The 'less computation time' claim is also not directly supported because wall-clock time is not measured, and the overhead of NN prediction and data preprocessing is not accounted for.","section":"Abstract and Sec. 5.2 (Tables 3 and 4)"},{"comment":"The comparison of filtering types (n_m=2 vs n_m=3 vs no filtering) and of different feature counts is based only on training MSE curves, not on held-out test error or repeated runs. Since the paper's central claim is that filtered input 'improves prediction accuracy' and that 340 features preserve accuracy, the authors should report test-set metrics (for example, MSE or R^2 on a held-out set) with standard deviations over multiple random seeds. Without these, the superiority of the chosen configuration is not established.","section":"Sec. 5.1.2 and Figs. 6-7"},{"comment":"The convolution weight of 0.25 is selected by minimizing prediction error on the 8x8 dataset in Fig. 8 and is then used for the 8x8 optimization in Sec. 5.2.2. This is a form of selection on the target data, so the reported transfer accuracy is not an independent validation. Moreover, the assumed mechanical equivalence between 4x4 and 8x8 units is contradicted by the paper's own numbers: the 8x8 structure obtained by duplicating the best 4x4 structure (Fig. 10) has compliance 0.81299 Nm, whereas the original 4x4 optimum has compliance about 0.71656 Nm (Table 3, red entry). After the proposed 2x2 convolution with weight 0.25, the duplicated structure maps to the same input vector as the 4x4 optimum, so the NN predicts an objective value close to the 4x4 optimum for a structure whose true compliance is about 13% higher. This systematic bias at the initial design undermines the claim that the convolution representation provides a sufficiently accurate transfer to larger structures.","section":"Sec. 5.1.3 and Sec. 5.2.2"},{"comment":"The NN is trained to predict compliance (Sec. 3.1), but Step 3 of the SA algorithm compares the NN prediction against a threshold c that is initialized as the penalized objective g(x0; lambda) and updated with increments w_d and w_i. The paper does not explain how the compliance prediction is combined with the volume penalty when evaluating y_p. If only compliance is predicted, then the comparison in Step 3 is not with the penalized objective used in Steps 1 and 5, making the acceleration criterion inconsistent. Please clarify the prediction target or amend the algorithm.","section":"Sec. 4, Steps 1 and 3"}],"minor_comments":[{"comment":"There are two subsections numbered 3.2: 'Filtering process' and 'Feature selection'; the latter should be numbered 3.3, and the subsequent 'Neural networks' subsection should be renumbered accordingly.","section":"Section numbering"},{"comment":"The F-statistic in Eq. (3) appears inconsistent with the f_regression function in scikit-learn, which tests each feature individually; for a univariate correlation the formula should be F = r^2/(1-r^2)*(n-2)/1. Please correct the formula or clarify the degrees of freedom used.","section":"Eq. (3)"},{"comment":"The phrase 'Rectifier Linear Units (ReLU)' should be 'Rectified Linear Units'.","section":"Sec. 3.3"},{"comment":"Figure 8 reports 'prediction error' without specifying the metric (e.g., MSE vs MAE), the axis scale, or whether the error is computed on a held-out set; please define the metric in the caption or text and include variability over multiple runs.","section":"Sec. 5.1.3 and Fig. 8"},{"comment":"The 'average number of structural analyses' is a proxy for computational cost; to support the 'less computation time' claim, the authors should report wall-clock times or at least estimate the overhead of NN prediction and feature extraction.","section":"Sec. 5.2"}],"recommendation":"major_revision","confidential_remarks":"The main weaknesses are the selection of the convolution weight on the same dataset used for evaluation, the use of training loss curves for accuracy claims, and the overstatement of the optimization results relative to the abstract. These are fixable with additional experiments and careful rewording. The contribution is incremental but within the journal's scope. If the authors provide out-of-sample test errors, a proper validation protocol for the convolution weight, and a revised comparison statement, the paper could be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The authors propose a practical way to accelerate simulated annealing for periodic lattice topology optimization by training an ML surrogate on filtered member-connectivity features and transferring a 4x4-trained model to 8x8 via a 4-channel 2x2 convolution. The genuinely new bit is that transfer: no one in the cited literature does it, and anyone building surrogates for discrete topology problems should know about it.\n\nThe paper does several things well. The filtering representation from Tamura et al. is applied cleanly, and the feature-selection step (f_regression) is a sensible way to cut input dimension. The NN-assisted SA procedure is clearly specified, and the reduction in structural analyses is real: roughly 7x fewer for 4x4 and 4.5x for 8x8. I also credit the authors for evaluating final designs with FEM rather than with the surrogate, so the optimization results are not definitionally rigged.\n\nThe soft spots are in the claims. The abstract says the proposed method can find the optimal structure in less time, but Tables 3 and 4 show pure SA finds lower average and best compliance in both test cases. 'Optimal' is too strong; 'near-optimal' is accurate. The accuracy comparisons in Figs. 6-8 are training-loss curves with no test error, no variability, and no code or data, so the accuracy claims are not independently checkable. The speedup numbers also exclude the cost of generating 150k samples and training the NN, which is a real cost even if it amortizes.\n\nThe more specific concern is the transfer equivalence. The paper says the Table 1 scaling makes 4x4 and 8x8 units mechanically equivalent. The paper's own data contradict this. The 8x8 initial structure in Fig. 10, formed by locally duplicating the best 4x4 structure, has compliance 0.81299 Nm, while the 4x4 optimum is 0.71656 Nm. After the 0.25-weight convolution, that 8x8 structure maps to the same input as the 4x4 optimum, so the NN's prediction is about 13% biased at the exact starting point of the 8x8 local search. That is not a subtle distribution shift; it is a systematic offset in the transfer. The local search still gets within about 1% of pure SA, so the bias is not fatal, but the mechanical-equivalence justification is wrong and the transfer claim needs to be reframed as an approximation with a known bias.\n\nWho is this for? Readers working on ML-assisted structural optimization or surrogate-based discrete topology optimization will find the transfer idea useful and the experimental setup clearly described. It deserves a serious referee, but it needs a revision that fixes the claims, reports test error and variability, and addresses the transfer bias directly.","headline":"A plausible ML-assisted topology optimization pipeline whose transfer claim is undercut by the paper's own numbers, but whose local search still works—worth a serious look and a careful revision.","tokens_in":8779,"tokens_out":4178,"would_cite":false,"duration_ms":38926,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["74P15","90C27"],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural network trained on filtered member-connectivity data can replace most structural analyses inside simulated annealing for periodic lattice topology optimization, giving comparable designs in a fraction of the time.","keywords":["lattice structures","topology optimization","machine learning","filtering process","feature selection","neural network","simulated annealing","periodic structures"],"falsifier":"Take an independent set of 8x8 lattice layouts, compute their compliance by finite-element analysis, and compare those values with the transferred network's predictions while sweeping the convolution weight; if the lowest-error weight is not near 0.25, or if NN-assisted simulated annealing consistently finds worse feasible designs than pure SA under the same volume constraint, the central claim is refuted.","tokens_in":7614,"feed_emoji":"🏗️","tokens_out":9187,"duration_ms":77651,"temperature":0.7,"pith_summary":"This paper claims that a neural network predicts the horizontal-load compliance of periodic lattice structures more accurately when its input encodes local connectivity of adjacent members (via a filtering process) than when it receives only the 0-1 presence or absence of each member. It further claims that feature selection can reduce the filtered input from 448 to 340 features with little loss of accuracy, and that a network trained on 4x4-grid units can be transferred to 8x8-grid units through a convolution-based conversion with kernel weight 0.25. If these claims hold, periodic lattice design can be accelerated by using the network as a cheap surrogate inside simulated annealing: the assisted search performs far fewer structural analyses while finding designs with compliance close to that of pure simulated annealing.","feed_headline":"Filtered data speeds lattice design by cutting analyses 7x","feed_subtitle":"Pairwise member encodings let a neural surrogate skip most structural analyses in simulated annealing.","key_machinery":"The load-bearing mechanism is the filtering representation of member connectivity: each node of the lattice has eight incident members forming a subregion, and the filter enumerates all combinations of choosing $n_m$ of those members, converting existence data into indicators of local patterns. Feature selection via the F-statistic then picks the 340 most correlated filtered features, and a four-channel convolution with kernel weight 0.25 converts a fine 8x8 layout into an equivalent coarse 4x4 layout so the pre-trained network can be applied without retraining. The prediction network itself is a three-hidden-layer fully-connected regressor with 900-600-300 units and ReLU activation, trained on mean squared error with the Adam optimizer; inside the SA loop it acts as a gate, using an adaptive threshold to decide which candidate solutions deserve a full structural analysis.","core_discovery":"The central discovery is that filtered local member combinations make a fully-connected neural network a much better compliance predictor for lattice structures, and that this predictor can be reused across scales. For a 4x4-grid unit, filtering with $n_m=2$ (pairwise combinations of the eight members around each node) gave the lowest test error, and keeping only 340 of the 448 filtered features retained most of that accuracy. The paper shows that a network trained only on 4x4 units can predict 8x8-unit responses by splitting members into four orientation channels and applying a 2x2 convolution with kernel weight 0.25, justified by the cross-section scaling in Table 1 that makes the two unit sizes mechanically equivalent. Embedding this network in simulated annealing with an adaptive threshold reduces the average number of structural analyses from 64,000 to 8,853 for 4x4 units and from 12,800 to 2,835 for 8x8 units, while producing designs whose compliance is close to what pure SA finds.","pith_inferences":["A natural extension, not tested here, is to apply the same filtering-and-feature-selection pipeline to other periodic unit cells, load cases, or three-dimensional lattices; the expected benefit is largest when local connectivity strongly controls stiffness.","The convolution-transfer idea implies a hierarchy of scales: repeated 2x2 pooling could let one base-scale network cover many unit sizes, but the weight 0.25 is tied to the specific cross-section scaling and would need recalibration for other aspect ratios.","Using a surrogate to skip structural analyses inside SA introduces a systematic risk: if the network's prediction bias aligns with the search direction, the threshold update could prune good solutions; a safeguard would be to occasionally evaluate skipped candidates.","Because the transfer claim rests on mechanical equivalence under Table 1 scaling, a cheaper check would be to compare the transferred network's predictions with finite-element results on a few hundred random 8x8 layouts before trusting the optimizer."],"forward_implications":["Filtered connectivity inputs are worth their higher dimension: pairwise member-encoding beats raw binary inputs for compliance prediction, so future response models for frames should encode local combinations rather than only member existence.","Feature selection can compress the filtered representation from 448 to 340 features with little accuracy loss, lowering the cost of training and evaluating the surrogate.","A network trained on small units transfers to larger periodic structures through 2x2 convolution when cross-sections are scaled for mechanical equivalence, so retraining is not required for each unit size.","NN-assisted simulated annealing with an adaptive prediction threshold cuts structural analyses by roughly a factor of 7 for 4x4 units and 4.5 for 8x8 units while producing comparable compliance values.","The 4x4 optimal topology is a good starting point for an 8x8 local search, yielding near-optimal periodic designs with similar lattice patterns."],"supporting_citations":[{"why":"Introduces the filtering process for brace placement that this paper adapts to encode member connectivity in lattice units.","marker":"[5]"},{"why":"Supplies the f_regression statistic used for feature selection on the filtered input data.","marker":"[6]"},{"why":"Provides the deep-learning implementation used to build and train the neural network models.","marker":"[7]"},{"why":"Supplies the Adam optimizer used for training the neural networks.","marker":"[8]"}],"fun_headline_variants":["Filtered connectivity data cuts lattice design analyses 7x","Pairwise member filters shrink lattice optimization cost 7x","ML surrogate with connectivity reuse scales lattice design","Connectivity-filtered features make lattice ML 7x faster","Scale-shared ML reduces lattice optimization analyses 7x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that an 8x8-grid unit with the Table 1 cross-section scaling is mechanically equivalent to a 4x4-grid unit, so averaging 2x2 member blocks with weight 0.25 preserves the stiffness behavior the network learned; if that equivalence fails, the transferred predictions and the 8x8 optimization results lose their foundation.","fun_headline_variants_meta":{"raw":{"variants":["Filtered connectivity data cuts lattice design analyses 7x","Pairwise member filters shrink lattice optimization cost 7x","ML surrogate with connectivity reuse scales lattice design","Connectivity-filtered features make lattice ML 7x faster","Scale-shared ML reduces lattice optimization analyses 7x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000231,"raw_usage":{"total_tokens":1501,"prompt_tokens":973,"completion_tokens":528,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":449}},"tokens_in":589,"tokens_out":528,"duration_ms":69445,"temperature":1.0,"reasoning_tokens":449,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:47:52.315205+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an independent set of 8x8 lattice layouts, compute their compliance by finite-element analysis, and compare those values with the transferred network's predictions while sweeping the convolution weight; if the lowest-error weight is not near 0.25, or if NN-assisted simulated annealing consistently finds worse feasible designs than pure SA under the same volume constraint, the central claim is refuted.","supporting_citations":[{"cited_title":"Jpn Archit Rev, 2018;1, 419-430","cited_arxiv_id":null,"evidence_quote":"Introduces the filtering process for brace placement that this paper adapts to encode member connectivity in lattice units."},{"cited_title":", Gramfort, A., Michel, V., Thirion, B","cited_arxiv_id":null,"evidence_quote":"Supplies the f_regression statistic used for feature selection on the filtered input data."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the deep-learning implementation used to build and train the neural network models."},{"cited_title":"P., & Ba, J., (2015)","cited_arxiv_id":null,"evidence_quote":"Supplies the Adam optimizer used for training the neural networks."}],"review_version":1}