{"id":"0c6451bb-0161-4937-9692-ac16b925b779","arxiv_id":"2411.09499","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A reinforcement learning agent, guided by a neural-network surrogate, picks side-sill wall thicknesses that improve energy absorption and mass compared to a genetic algorithm baseline.","lead":"This paper uses reinforcement learning to choose wall thicknesses for a car side-sill so it absorbs crash energy while staying light. The method replaces expensive crash simulations with a neural network during training, then checks the final design with a real simulation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 3's Eq. 3.2 and Section 4.5's Eq. 4.3 define opposite objective directions for mass, so the reported RL-vs-GA comparison may reflect a sign error rather than the claimed optimisation advantage.","rationale":"The reader's conditional verdict is appropriate. The paper's biggest weakness is not, in my view, the surrogate accuracy per se: the final coupled run is validated against Abaqus, and even a poorly trained policy could in principle be corrected by the coupled loop. The more load-bearing problem is that the paper states one objective in Eq. 3.2 and implements a different one in Eq. 4.3. The sign of the mass term is reversed relative to the stated minimization, and the scaling is changed without comment. Since the only quantitative evidence for the central claim is Table 4's comparison with GA and NI, this inconsistency makes that evidence uninterpretable. If the GA was run with the literal Eq. 3.2, it was being asked to maximize mass, so the RL's lower mass is not a meaningful win. The authors may have implemented the reward correctly in code, but then Eq. 3.2 and the text are wrong; either way the paper needs a correction and re-run. This does not change the reader's conditional verdict; it sharpens the condition. The surrogate coverage concern raised by the reader is real but secondary; it would matter once the objective is fixed and the comparison is valid.","tokens_in":11681,"tokens_out":9080,"duration_ms":83546,"concrete_test":"Re-derive Eq. 3.2 with f1 = energy (to maximize), f2 = mass (to minimize). Express the same goal as a single minimization; the coefficient on f2 must be positive (or, equivalently, the reward must be -O). Then re-run the GA and network-inversion baselines and the coupled RL optimisation using exactly this corrected objective and the same computational budget. Specifically, check whether the GA, when forced to minimize O=(M1-f1)+0.5*f2 instead of the literal Eq. 3.2, produces a design with mass below 14.49 kg or energy above 1417.36 J. If it does, the reported advantage of RL over GA in Table 4 collapses and the central claim is not supported. Additionally, verify Eq. 4.3 by perturbing a single scaled mass and energy value and confirming that the reward changes with the intended signs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 defines the optimization problem as min_T O(T)=w1*(f1(T)-M1(T))+w2*f2(T), with w1=1, w2=-0.5 and f2=mass. Since O is minimized, the -0.5*f2 term rewards larger mass: a heavier design gives a more negative O. The text says 'mass has to be minimised as much as possible', but the sign does the opposite. Section 4.5 uses the same expression as a reward R=(f1-M1)-0.5*f2 and maximizes it; under the stated convention this is equivalent to minimizing O, so the implemented RL objective is inconsistent with the stated minimization problem. The correct minimization form for 'maximize energy, minimize mass' is O=(M1-f1)+c*f2 with c>0, or equivalently R=(f1-M1)-c*f2 with c>0. This is not a cosmetic typo: Table 4 compares the coupled RL result (1417.36 J, 14.49 kg) against a GA result (1370.35 J, 15.59 kg). If the GA baseline minimized Eq. 3.2 literally, it was rewarded for higher mass, so the comparison does not show that RL is better for the stated multi-objective problem; it may only show that the GA was given the wrong sign. The central 'encouraging results' claim therefore rests on an unresolved sign and scaling inconsistency between the mathematical problem and the implemented reward.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an inverse multi-objective optimization method for a seven-parameter multi-cell side sill design. It combines finite element (FE) simulations (310 Abaqus runs), a neural-network regression surrogate, and a reinforcement learning (A2C) agent to maximize energy absorption while minimizing mass. The agent is first trained on the surrogate and later coupled directly with Abaqus for final evaluation. Results are compared with a genetic algorithm and network inversion; the coupled RL reports 1417.36 J absorbed at 14.49 kg, versus 1370.35 J at 15.59 kg for the GA.","tokens_in":12057,"tokens_out":6840,"duration_ms":55440,"significance":"If the methodological issues are resolved, the paper would offer a useful demonstration of RL-based multi-objective crashworthiness optimization with FE-in-the-loop validation. The combination of a regression surrogate for initial training and an FE-coupled refinement is practical and relevant, and the residual analysis on a held-out test set is a reasonable first check. However, the current evidence does not yet support the central claim: the sign inconsistency in the objective function undermines the GA comparison, the surrogate's accuracy in unexplored regions of the 7-dimensional space is not assessed, and the RL results are based on a single 20-step episode. The paper does not provide code or data, though the FE setup is described in sufficient detail for partial reproducibility.","major_comments":[{"comment":"The optimization problem is stated as min_T O(T) = w1*(f1(T) - M1(T)) + w2*f2(T) with w1=1 and w2=-0.5. Since f2 is the mass and the problem is minimized, the negative coefficient on f2 makes larger mass reduce O, which contradicts the text's assertion that 'the mass has to be minimised as much as possible.' If the genetic algorithm baseline described in Section 4.3 uses this expression as its fitness, then the comparison in Table 4 is invalid because the GA would be rewarded for increasing mass. The sign convention should be corrected (e.g., O = (M1 - f1) + c*f2 with c > 0) and all baselines should be re-run with the same objective used for the RL reward.","section":"Section 3, Eq. (3.2)"},{"comment":"The reward R = (f1(T) - M1(T)) - 0.5*f2(T) is maximized by the RL agent, which is equivalent to minimizing (M1(T) - f1(T)) + 0.5*f2(T). This is not the same objective as Eq. (3.2), where O is minimized with a negative f2 coefficient. The manuscript does not reconcile these two formulations. If the GA and the RL use different scalarizations, the reported advantage of RL over GA in Table 4 may be an artifact of the inconsistency rather than a genuine property of the method. The paper should state explicitly how each method's objective is derived from the corrected Eq. (3.2).","section":"Section 4.5, Eq. (4.3)"},{"comment":"The regression surrogate is validated only through residuals within ±5% on a 20% hold-out (62 of 310 simulations). The RL agent, however, queries the surrogate at arbitrary points in the 7-dimensional thickness space, and the paper itself acknowledges in Section 4.5 that the agent 'cannot estimate better energy absorption values while reducing the mass' because of the extrapolation limitations of the regression model. No coverage assessment, uncertainty quantification, or sensitivity analysis is provided. This is load-bearing for the initial RL training, because if the surrogate misranks candidate designs, the agent optimizes the wrong function. Please add an evaluation of surrogate accuracy along the RL agent's exploration trajectories or an uncertainty-aware reward.","section":"Section 5.2 and Section 4.5"},{"comment":"The coupled RL results are identical for T2=3, 4, and 5, and the evaluation consists of a single 20-step episode. This does not establish convergence or robustness; identical outputs across different termination thresholds suggest that the agent follows a deterministic trajectory rather than optimizing a stochastic policy. Repeated runs with different random seeds, and a report of variance or a convergence criterion, are needed to support the claim that the RL environment 'provided encouraging results.'","section":"Section 4.5 and Table 4"}],"minor_comments":[{"comment":"The title contains a typo: 'Developement' should be 'Development', and the Section 2 heading 'optimsiation' should be 'optimisation'.","section":"Title and Section 2 heading"},{"comment":"The abstract contains a duplicated sentence: 'This challenge leads to the need for inverse multi-parameter multi-objective optimisation.' appears twice with slightly different wording.","section":"Abstract"},{"comment":"Section 5.3 uses 'KJ' and 'kJ' while Table 4 and the FE database use 'J'; please standardize the energy units throughout.","section":"Section 5.3"},{"comment":"Figure 4 axis labels and Figures 5-6 labels appear garbled (strings such as '/uni00000035/...' are visible); the figures should be regenerated with proper text rendering.","section":"Figures 4-6"},{"comment":"The ideal energy-mass arrays [800,600,13] and [825,625,14] are chosen without explaining how these values were derived; add a justification or a sensitivity study.","section":"Section 4.5"},{"comment":"The second termination condition T2 is described as 'the mean sum of the absolute difference'; it is unclear whether this is a mean or a sum, and the condition 'less than 5' should specify over which steps the mean is taken.","section":"Section 4.5"},{"comment":"The phrase 'the inverse network is then trained using the loss of the forward network' is confusing, since network inversion is implemented through gradient descent in the input space (Eq. 4.2); please revise the wording.","section":"Section 4.4"}],"recommendation":"major_revision","confidential_remarks":"The sign issue in Eq. (3.2) is the most serious concern; if the GA was implemented as written, the main comparison in Table 4 is invalid. I would ask the authors to correct the objective, re-run GA and network inversion with the same scalarization, and provide repeated runs and surrogate validation before resubmission. The paper is within the scope of the journal, but the current version requires substantial revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuine engineering application of an existing ML+RL idea, and the Abaqus coupling is real work. But the paper's headline claim — RL beats GA on multi-objective side-sill optimisation — is not supported as written because the objective and the reward have opposite signs for mass.\n\nThe genuinely new part is narrow: seven wall-thickness variables for a multi-cell side sill under oblique pole impact, optimized by an A2C agent in a custom RL environment, with a final loop that calls Abaqus directly. That coupling is nontrivial and the authors did it. The surrogate story is also mostly credible: 310 FE runs, a correlation check, and residuals within +/-5% on a held-out 20%. Good enough as a training aid, though not as proof of global optimum.\n\nNow the problem. Eq. 3.2 defines min_T O(T) = (f1-M1) - 0.5*f2 with f2=mass. Minimizing this rewards heavier design, because the negative mass term gets smaller as mass increases. The text says mass should be minimized. Section 4.5 then maximizes the same expression as reward. So either the minimization problem is mis-specified or the reward sign is wrong. If the GA baseline used Eq. 3.2 literally, GA was pushed toward high mass; Table 4's comparison (GA 1370 J/15.59 kg vs RL 1417 J/14.49 kg) then shows only that GA got the wrong sign, not that RL is better. This is load-bearing, not a typo.\n\nThe other weaknesses are less severe but real: no code or data, a single one-episode run for the coupled RL evaluation, identical output for three T2 thresholds (suggesting the agent hits the action/step limit rather than converging), no sensitivity analysis of the surrogate over the 7D discrete space, and an unsupported claim that conventional optimizers fail beyond three parameters. Minor slips: abstract repeats a sentence and Section 5.3 reports kJ where the table says J. The dependence on the authors' own prior framework [4] is fine; that is normal incremental research.\n\nWho gets value: engineers working on crashworthiness surrogates and RL-for-design will find the environment description and the Abaqus-coupling idea useful. Nobody should rely on the RL-vs-GA conclusion until the objective is corrected and both optimizers are rerun.\n\nMy recommendation: send to peer review, because this is a real application with a clearly fixable flaw and the reviewer time would be productive. Do not accept as is; ask for corrected objective, repeated runs with seeds, and the data or at least a detailed design table for the 310 runs.","headline":"A worthwhile engineering case study, but the central RL-vs-GA claim is undermined by a sign error in the objective/reward definition.","tokens_in":12532,"tokens_out":5267,"would_cite":false,"duration_ms":47884,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Reinforcement learning beats genetic algorithm for side-sill design","keywords":["reinforcement learning","crashworthiness","side sill","multi-objective optimization","surrogate model","finite element simulation","inverse design","wall thickness optimization"],"falsifier":"Take the trained regression surrogate and evaluate it on a grid of thickness combinations that were not part of the 310 training simulations, then run those exact designs through the finite-element solver: if the surrogate's ranking of candidate designs disagrees with the finite-element rankings often enough that the RL agent's chosen path changes, the claimed optimization gain would be an artifact of the surrogate rather than a real improvement.","tokens_in":11507,"feed_emoji":"🚗","tokens_out":7460,"duration_ms":66624,"temperature":0.7,"pith_summary":"This paper tries to establish that a reinforcement-learning optimizer can handle an inverse multi-parameter, multi-objective crashworthiness design problem: choosing seven wall thicknesses of a multi-cell side sill to maximize energy absorption while limiting mass. The authors train an advantage actor-critic (A2C) agent in a custom environment where the reward rewards higher combined energy absorption and penalizes mass, first using a regression surrogate trained on 310 finite-element simulations, then coupling the environment directly to the finite-element solver. If the approach is right, engineers could replace expensive population-based search over crash simulations with an agent that explores thickness parameters directly and adapts to new objective targets without restarting the whole optimization. The reported result is a design absorbing 1417 J at 14.49 kg, higher on energy than the genetic-algorithm baseline (1370 J at 15.59 kg) and comparable to network inversion.","feed_headline":"Reinforcement learning beats genetic algorithm for side-sill design","feed_subtitle":"Coupled to an FE solver, the RL agent absorbed 1,417 J versus 1,370 J for the genetic algorithm, at lower mass.","key_machinery":"The central mechanism is a custom reinforcement-learning environment in which the state is the seven wall thicknesses and each action changes one thickness by a fixed step. The reward function is $R=(\\bar f_1-M_1)-0.5\\bar f_2$, with both objectives scaled to [1,100] before the reward is computed, so the agent maximizes absorbed energy relative to a user-supplied ideal target while penalizing mass. The agent is an advantage actor-critic (A2C) policy-gradient method trained with a model-free reinforcement-learning implementation: first against a three-hidden-layer regression surrogate (trained on 310 finite-element simulations with the Adam optimizer), and later in a coupled mode where the environment calls the finite-element solver directly for each state, with the termination step limit reduced to 20 and the ideal objective array raised to [825, 625, 14].","core_discovery":"The central claim is that the reinforcement-learning environment, coupled with the explicit finite-element solver, can solve the side-sill optimization problem. In the environment, the state is the vector of seven wall thicknesses, each action increments or decrements one thickness by its manufacturing step size, and the reward is $R=(\\bar f_1-M_1)-0.5\\bar f_2$, where $\\bar f_1$ is the scaled combined energy absorbed by the inner sill and front part, $M_1$ is the scaled ideal energy target, and $\\bar f_2$ is the scaled side-sill mass. The A2C agent first learns from a regression surrogate whose prediction residuals on a held-out test set lie within $\\pm5\\%$, and then the same agent is evaluated with the environment calling the finite-element solver directly, with the termination step limit reduced to 20 and the ideal objective array raised to [825, 625, 14]. The final thickness vector [1.7, 2.0, 2.0, 1.0, 3.0, 3.4, 4.0] mm gives 1417.36 J absorbed and 14.49 kg mass, which the paper reports as better than the genetic-algorithm baseline and comparable to network inversion while giving energy absorption the higher weight.","pith_inferences":["The scalarized reward with fixed weights $w_1=1$ and $w_2=-0.5$ merges two objectives into one; a natural extension would be to run multiple weight vectors and trace a Pareto front of thickness designs, which the single run does not deliver.","The peak contact force is extracted from the simulations but never used as an objective; since occupant safety is the real goal, constraining or minimizing peak contact force alongside energy and mass would be a direct test of the method's practical value.","Several final thicknesses (t2, t3, t4, t7) sit at the boundary of their allowed ranges, so the true optimum may lie outside the studied ranges; extending those ranges and retraining would show whether the agent is converging or simply walking to a constraint boundary.","The 310 training simulations were chosen randomly without a coverage or design-of-experiments analysis; an active-learning or space-filling sampling plan could make the surrogate trustworthy with fewer simulations or reveal regions where the current agent is effectively guessing."],"forward_implications":["Because the RL environment separates the objective array from the agent, the same trained agent can be pointed at new energy and mass targets and re-evaluated without retraining the surrogate, as the paper demonstrates with different ideal objective arrays.","Coupling the environment to the finite-element solver removes the surrogate's extrapolation limit, allowing the optimizer to seek energy absorption above the roughly 1400 J average present in the training database.","The final thickness values were identical under three different termination tolerances (T2=3, 4, and 5), suggesting the coupled RL run settled on a stable point in the parameter space.","For similar multi-objective problems, the pipeline is not gradient-based and learns by exploring the parameter space, so it could be applied where evolutionary algorithms struggle with more than three parameters.","The paper states that only A2C was tested and that other model-free RL agents and reward functions could plausibly improve the results."],"supporting_citations":[{"why":"The previous machine-learning-based design optimization framework that this article extends, providing the surrogate-plus-RL methodology.","marker":"[4]"},{"why":"Supplies the weighted-sum multi-objective optimization formulation used to define the inverse optimization function.","marker":"[21]"},{"why":"The finite-element solver used to generate the simulation database and to run the coupled RL evaluation.","marker":"[22]"},{"why":"Provides the Keras sequential model used to build the regression surrogate.","marker":"[23]"},{"why":"The Adam optimizer used for training the regression surrogate and the RL agent.","marker":"[24]"},{"why":"Provides the StandardScaler used to standardize the surrogate outputs before training.","marker":"[25]"},{"why":"Provides the A2C agent and reinforcement-learning training loop used in the custom environment.","marker":"[27]"}],"fun_headline_variants":["RL optimization boosts side-sill crash energy absorption","Reinforcement learning improves side-sill design optimization","AI-driven side-sill design outperforms genetic algorithm","RL agent tunes side-sill thickness for better crash safety","Machine learning cracks side-sill multi-objective optimization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire optimization stands on the regression surrogate, trained on only 310 finite-element simulations with residuals checked only on a held-out 20%, being accurate enough to guide the RL agent's policy across the full seven-dimensional thickness range, and the paper provides no coverage or sensitivity analysis to show this.","fun_headline_variants_meta":{"raw":{"variants":["RL optimization boosts side-sill crash energy absorption","Reinforcement learning improves side-sill design optimization","AI-driven side-sill design outperforms genetic algorithm","RL agent tunes side-sill thickness for better crash safety","Machine learning cracks side-sill multi-objective optimization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000626,"raw_usage":{"total_tokens":2899,"prompt_tokens":950,"completion_tokens":1949,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":1873}},"tokens_in":566,"tokens_out":1949,"duration_ms":13650,"temperature":1.0,"reasoning_tokens":1873,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:34:21.460353+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained regression surrogate and evaluate it on a grid of thickness combinations that were not part of the 310 training simulations, then run those exact designs through the finite-element solver: if the surrogate's ranking of candidate designs disagrees with the finite-element rankings often enough that the RL agent's chosen path changes, the claimed optimization gain would be an artifact of the surrogate rather than a real improvement.","supporting_citations":[{"cited_title":"Smith, ABAQUS/ Standard User’s Manual, Version 6.9, Dassault Systèmes Simulia Corp, United States, 2009","cited_arxiv_id":null,"evidence_quote":"The finite-element solver used to generate the simulation database and to run the coupled RL evaluation."},{"cited_title":"Chollet, et al., Keras,https://keras.io (2015)","cited_arxiv_id":null,"evidence_quote":"Provides the Keras sequential model used to build the regression surrogate."},{"cited_title":"Pedregosa, G","cited_arxiv_id":null,"evidence_quote":"Provides the StandardScaler used to standardize the surrogate outputs before training."},{"cited_title":"Raffin, A","cited_arxiv_id":null,"evidence_quote":"Provides the A2C agent and reinforcement-learning training loop used in the custom environment."}],"review_version":1}