{"id":"859caecc-3e0e-4715-8154-5b4e2fb64282","arxiv_id":"2412.11035","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MLP surrogate models accelerate multi-objective optimization of LPG thermal cracking to under one minute, and five MCDM methods select a single recommended operating point.","lead":"This paper trains a single-hidden-layer neural network on a thermal cracking simulator and uses the network as a fast objective function for multi-objective particle swarm optimization. The workflow cuts optimization time from roughly two days to under one minute and picks a recommended operating point by combining five ranking methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The speedup is credible only if the surrogate remains accurate at optimizer-selected extremes; test-set R² and four spot-checks in Case A do not establish that the MCDM-recommended points are truly optimal for the original model.","rationale":"The reader's weakest assumption combines two things: the mathematical model's real-world fidelity and the surrogate's fidelity to that model. I focus on the second, which is the condition that must hold for the paper's own speedup and Pareto-front claims to be internally valid. The concern is sharpened by noting that surrogate-assisted optimization suffers from an optimizer's curse: a test R² measured on random model-generated data does not measure accuracy at the extremes and corners that MOPSO systematically selects. The only direct check, Table 2, tests four Case-A points that are not the final recommended solution, so it does not cover the points that would actually be implemented. This does not force rejection, because the missing validation is straightforward to perform and there is no evidence that the surrogate is wrong; it does make the central practical claim conditional on that validation. Hence the reader's CONDITIONAL verdict remains appropriate.","tokens_in":18269,"tokens_out":4835,"duration_ms":53429,"concrete_test":"Run the original mathematical model at the MCDM-recommended operating points in Tables 3–8 and at 100 additional sampled points from each surrogate Pareto front; compare all six predicted objectives with full-model values. The claim is invalidated if any recommended point deviates beyond the reported test RMSE, or if a neighboring full-model point dominates it under the original model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 builds one single-hidden-layer MLP per objective from data generated by \"the mathematical model\" (not identified in the main text) and reports test-set R² values of at least 0.995. The load-bearing step is not the average R² but the accuracy of those MLPs at the operating points that MOPSO actually selects. The decision-variable plots (Figs. S2–S8) show selected regions pinned to range boundaries — e.g., Fin = 14, Tin = 703, COT = 853, COP = 1.5, SR = 0.4 — and MOPSO maximizes or minimizes MLP outputs, so any small systematic overprediction in those regions is precisely what the optimizer will exploit. Section 3.1 validates only four arbitrary Case-A front points (Table 2) against the full model; these are not the MCDM-recommended solution from Table 3, and Cases B–F are not checked at all. Thus the central practical claim—that the recommended operating conditions are Pareto-optimal for the actual process model—rests on untested surrogate accuracy in optimizer-selected regions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a surrogate-based multi-objective optimization workflow for the LPG thermal cracking process. A complete factorial design over five operating variables is run on an (unidentified) mathematical model to generate training data; single-hidden-layer MLPs are trained for six objectives and report test-set R² values of at least 0.995; MLP-aided MOPSO is then applied in six case studies to produce Pareto fronts; CRITIC weights and five MCDM methods (MABAC, PROBID, SAW, sPROBID, TOPSIS) are used to recommend a single operating point. The paper's central practical claims are that the MLP-aided optimization completes in under one minute versus about two days for direct simulation-based MOPSO, and that the recommended points are valid Pareto-optimal solutions for the thermal cracking process.","tokens_in":18499,"tokens_out":6645,"duration_ms":62177,"significance":"If the surrogate fidelity at the optimizer-selected points and the validity of the underlying process model are established, the workflow is a useful template for accelerating MOO of cracking furnaces: it combines factorial data generation, MLP surrogates, MOPSO, CRITIC weighting, and a transparent MCDM selection stage, and the six case studies illustrate plausible trade-offs among ethylene, propylene, heat duty, run length, and selectivity indices. The four-point spot check in Table 2 is a useful start, and the majority-voting procedure across five MCDM methods is a reasonable safeguard against method-specific ranking artifacts. However, the current evidence does not yet establish that the recommended operating conditions are Pareto-optimal for the actual mathematical model, so the significance is conditional on additional validation.","major_comments":[{"comment":"The 'mathematical model' used to generate all training data is never identified or cited in the main text, and its accuracy against plant or experimental data is not addressed. Since every surrogate prediction inherits any error in this model, the absence of model identification and validation is load-bearing rather than stylistic. The reference list contains Nabavi et al. (2009), which appears uncited in the text and may be the intended source; the authors should cite it explicitly and provide the model equations or a clear pointer to them.","section":"Section 3, first paragraph"},{"comment":"The only full-model validation is a set of four points (I-IV) from Case A, and these are not the MCDM-recommended solution reported in Table 3; Cases B-F have no full-model checks at all. Because the optimizer drives decision variables to the boundaries of their ranges in the recommended solutions (e.g., Fin=14, Tin=703, COT=853, COP=1.5, SR=0.4 in Tables 3-8), a global test-set R² over a factorial design does not establish surrogate accuracy in the regions where the optimizer actually selects solutions. The authors should evaluate the full mathematical model at the recommended solutions of Tables 3-8 and report the resulting errors, and ideally evaluate a sample of nondominated solutions from each case to show that the surrogate Pareto front is not an artifact of MLP approximation.","section":"Section 3.1, Table 2"},{"comment":"The headline speedup ('within one minute' versus 'an average of two days') is not substantiated. The manuscript provides no hardware details, no measured wall-clock times, no number of function evaluations for either approach, no repetitions, and no description of the conventional MOPSO implementation used as the baseline. Since this acceleration claim appears in the abstract and conclusions, it should either be supported by a reproducible benchmark or removed and replaced by a more modest qualitative statement.","section":"Section 3.1, first paragraph; Section 4"}],"minor_comments":[{"comment":"There are typographical inconsistencies in the MCDM method names: 'PRBOID' and 'sPRBOID' in Sections 2.5 and 2.7 should be 'PROBID' and 'sPROBID', and Table 7 contains 'sPROBOD' and 'TOPISIS'.","section":"Sections 2.5 and 2.7; Table 7"},{"comment":"The summation in the formula for c_j is not clearly typeset; please state explicitly that the sum runs over k=1 to n and clarify how the j=k term is treated, since this directly affects the CRITIC weights.","section":"Section 2.3, Eq. (7)"},{"comment":"The factorial design is described only qualitatively; the number of levels per decision variable and the total dataset size should be reported so that the reported R² values can be interpreted and the coverage of the boundary regions can be assessed.","section":"Section 3.1"},{"comment":"There are several language slips that should be corrected, including 'taken removed from the production line', 'heat fraction process', and 'speed of heat fractions'; in context these appear to refer to cracking rather than 'fraction' processes.","section":"Sections 3.4 and 3.5"},{"comment":"The definition of the severity index C-3/C3= should be double-checked: as written, it includes propylene in both the numerator and the denominator, which makes the index's behavior and interpretation unclear.","section":"Section 3.5"},{"comment":"For a data-driven workflow, providing the trained MLP weights or the code used for surrogate training, MOPSO, and MCDM would substantially improve reproducibility; currently only architecture-level information is indicated.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a process-systems/chemical-engineering audience, and the proposed integration of MLP surrogates with MOPSO and MCDM is potentially valuable. The main risk is not the idea but the evidence: the central claims require either stronger validation against the full process model or appropriate qualification. The reference list includes Nabavi et al. (2009), which appears to be the likely process-model source but is not cited in the text; this should be corrected during revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read on this one: it's a straightforward application of an existing surrogate-based MOO+MCDM pipeline to LPG thermal cracking. What's new is the case study: six objective pairings, Pareto fronts, MCDM recommendations, and the three-orders-of-magnitude speedup claim. The MLP surrogate itself is a single hidden layer, so calling it 'deep learning' oversells it, but that's a framing issue, not a technical one.\n\nThe paper does the engineering well enough. The data generation via factorial design is sensible, the test R2 values are reported (at least 0.995), and they did spot-check four Case A points against the mathematical model. Those checks agree within ~1%, which is real evidence for interpolation quality. The MCDM part is clearly explained, and the majority-voting logic is transparent.\n\nThe soft spots are exactly where the reader and stress-test notes land. First, the mathematical model is never identified or cited. That matters because the surrogate is only as good as the data it learns from; if the model is not independently accessible, the results can't be reproduced or evaluated. Second, the four spot-checks in Case A are not the MCDM-recommended point, and Cases B-F have no validation at all. Since the optimizer is expected to push toward corner regions of the input space, test-set R2 is not enough to guarantee the MLP is accurate exactly where MOPSO selects. The recommended operating points are therefore not actually shown to be Pareto-optimal for the real process model. That is the load-bearing gap.\n\nA third, lesser concern: there is no uncertainty quantification around the surrogate predictions or the MCDM rankings. Given that the whole point is to support a single operating recommendation, a sensitivity analysis would help.\n\nMy take: the central claim—that this workflow can cut optimization time from days to under a minute—is credible for this process, but the practical value depends on demonstrating that the surrogate-optimized solutions are faithful to the full model at the recommended points. That's an addressable problem. I'd send it to review with a request for the simulator identity, validation of the recommended points for each case (or at least a representative high-dimensional case), and ideally code/data availability.\n\nThis paper is for practitioners who want to apply this pipeline to their own cracking furnaces. It earns a serious referee, not a desk reject, but it needs a revision focused on validation.","headline":"A credible application of an existing surrogate-based MOO+MCDM pipeline to LPG cracking, with a real speedup claim but weak validation of the recommended points.","tokens_in":19037,"tokens_out":1520,"would_cite":false,"duration_ms":13555,"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":"Trained multilayer perceptron surrogates let MOPSO map Pareto fronts for an LPG thermal cracking furnace in under a minute, roughly three orders of magnitude faster than optimizing with the full mathematical model, and a five-method MCDM…","keywords":["liquefied petroleum gas","thermal cracking","multilayer perceptron","surrogate modeling","multi-objective particle swarm optimization","Pareto-optimal front","multi-criteria decision making","CRITIC weighting"],"falsifier":"Take every recommended operating point from the six case studies, re-evaluate the six objectives with the full mathematical model, and compare them with the MLP predictions; if the errors are not small relative to the spread of the Pareto front, or if full-model MOPSO finds clearly better points, the surrogate-optimality claim collapses. The paper's own spot-check covers only four points in Case A, so this full check is the direct experiment that would settle the question.","tokens_in":18052,"feed_emoji":"🏭","tokens_out":12939,"duration_ms":104433,"temperature":0.7,"pith_summary":"This paper aims to establish that a multilayer perceptron (MLP) neural network, trained on data generated by a mathematical model of an LPG thermal cracking furnace, can stand in for that model during multi-objective particle swarm optimization (MOPSO) and cut the time to compute a Pareto-optimal front from about two days to under one minute. After the Pareto front is found, the paper adds a decision stage: CRITIC weighting derives objective weights from the front itself, five multi-criteria decision making (MCDM) methods—MABAC, PROBID, SAW, sPROBID, and TOPSIS—rank the solutions, and majority voting picks one implementable operating condition. Across six case studies the surrogate models reach test $R^2 \\geq 0.995$, and the fronts show the expected trade-offs between ethylene production, propylene production, heat duty, run length, and selectivity. If the acceleration is genuine, the approach makes multi-objective optimization practical for process models whose two-day runs currently discourage routine trade-off studies.","feed_headline":"Neural-net shortcut turns 2-day olefin optimization into 1 minute","feed_subtitle":"A neural network trained on the cracking model lets engineers weigh output, energy, and run length in under a minute.","key_machinery":"The load-bearing mechanism is the trained MLP surrogate: each of the six process objectives is predicted almost instantly from five decision variables (feed flow rate, coil outlet temperature, coil outlet pressure, steam-to-feed ratio, and coil inlet temperature), so the swarm can evaluate candidate operating conditions hundreds of times per minute instead of waiting on the slow mathematical model. On top of the surrogate, the Pareto-optimal archive produced by MOPSO is processed by CRITIC weighting, which assigns objective weights from the standard deviation and pairwise correlations of the normalized objective matrix, and then by five MCDM methods that rank the Pareto solutions; the paper's final selection rule is majority voting supplemented by similarity among the recommended solutions.","core_discovery":"The central claim is that replacing the full mathematical model with MLP surrogates inside MOPSO preserves the essential optimum structure of the LPG cracking process while making each optimization tractable in under a minute on the same computer that would need about two days for the conventional approach. The paper trains one single-hidden-layer MLP for each of six outputs—annual ethylene and propylene production, heat duty, run length, ethylene selectivity, and severity index—from a complete factorial dataset over five operating variables, then runs MOPSO for 100 generations with 60 or 70 particles in six case studies. It reports test $R^2 \\geq 0.995$ for all surrogate models, and in Case A four surrogate-optimal points are shown to give objective values close to those obtained by re-evaluating the mathematical model. The resulting Pareto fronts align with cracking chemistry, and the CRITIC plus five-MCDM procedure with majority voting yields a single recommended solution in each case.","pith_inferences":["A consequence the paper leaves implicit is that the one-minute runtime makes near-real-time re-optimization possible: when feed quality or product prices shift, the operating point can be updated within the hour rather than after a multi-day study.","Because the paper validates surrogate predictions against the full model at only four points in Case A, a direct test would be to feed every MCDM-recommended solution from all six cases back into the full model and compare objective values.","In this reader's view, the agreement among several MCDM methods is evidence of stability but not a formal guarantee; changing the CRITIC weighting or using other weight schemes could still move the majority-vote winner.","The success of a single-hidden-layer MLP suggests the true input–output map of this furnace is smooth, so simpler response-surface surrogates might achieve much of the same speedup; the paper does not compare against them."],"forward_implications":["In all six case studies the surrogate-based optimization completes in under one minute, so the paper's claim implies that routine re-optimization of cracking conditions becomes cheap enough for industrial practice.","The Pareto fronts quantify the trade-offs among ethylene, propylene, heat duty, run length, and selectivity, giving operators a map of how much of one objective is sacrificed for another.","Because CRITIC weights come from the front itself and several MCDM methods often converge on the same solution, the workflow converts the Pareto set into one defensible setpoint rather than leaving the choice to the engineer.","The consistently high test $R^2 \\geq 0.995$ implies the MLP models are accurate enough to serve as objective functions for optimization screening, with full-model verification reserved for the final candidate."],"supporting_citations":[{"why":"Provides the ML-aided MOO/MCDM framework and two chemical-process demonstrations that the present LPG cracking study transplants and extends.","marker":"Wang et al., 2022"},{"why":"Introduces the CRITIC standardization and weighting steps used to assign objective weights from the Pareto matrix.","marker":"Diakoulaki et al., 1995"},{"why":"Introduces the MABAC border-approximation-area ranking method used in the MCDM comparison.","marker":"Pamučar & Ćirović, 2015"},{"why":"Introduces PROBID and its simpler variant sPROBID, the ideal-average-distance ranking methods used here.","marker":"Wang et al., 2021"},{"why":"Provides the SAW equations and its application to selecting from the Pareto-optimal front.","marker":"Wang & Rangaiah, 2017"},{"why":"Introduces TOPSIS, the distance-to-ideal ranking method included in all six MCDM comparisons.","marker":"Hwang & Yoon, 1981"},{"why":"Supplies the heat-fraction kinetics expectation that ethylene and propylene production move in opposite directions, used to validate the Case A Pareto front.","marker":"Towfighi et al., 2006"},{"why":"Supplies the review-level relationships among coil temperature, steam ratio, pressure, coking, and selectivity used to explain decision-variable trends.","marker":"Sadrameli, 2015"}],"fun_headline_variants":["Olefin optimization: 2 days to 1 minute via neural nets","MLP surrogates slash olefin cracking optimization time","Deep learning cuts thermal cracking optimization from days to minutes","Neural nets speed up olefin plant optimization 2000x","One-minute olefin optimization with neural net surrogates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The scheme rests on the assumption that the mathematical model used to generate the training data faithfully represents the real LPG cracking furnace, so that a point that is best for the neural-network stand-in is also best for the actual process.","fun_headline_variants_meta":{"raw":{"variants":["Olefin optimization: 2 days to 1 minute via neural nets","MLP surrogates slash olefin cracking optimization time","Deep learning cuts thermal cracking optimization from days to minutes","Neural nets speed up olefin plant optimization 2000x","One-minute olefin optimization with neural net surrogates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000678,"raw_usage":{"total_tokens":3085,"prompt_tokens":949,"completion_tokens":2136,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":2050}},"tokens_in":565,"tokens_out":2136,"duration_ms":14901,"temperature":1.0,"reasoning_tokens":2050,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:22:04.398672+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take every recommended operating point from the six case studies, re-evaluate the six objectives with the full mathematical model, and compare them with the MLP predictions; if the errors are not small relative to the spread of the Pareto front, or if full-model MOPSO finds clearly better points, the surrogate-optimality claim collapses. The paper's own spot-check covers only four points in Case A, so this full check is the direct experiment that would settle the question.","supporting_citations":[],"review_version":1}