{"id":"0dfebf8a-c4fc-4db5-a2d3-04c803179b1b","arxiv_id":"2505.15193","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Vapor-liquid equilibrium uncertainties, especially the medium boiler's vapor pressure, dominate purity losses in Pareto-optimized dividing wall columns, and losses increase with column stage count.","lead":"This study simulates how errors in thermodynamic models, such as vapor pressures and activity coefficients, affect the purity of products from energy-optimized dividing wall distillation columns. It identifies which model uncertainties matter most and shows that higher stage numbers can amplify purity losses.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim depends on the Table 3 uncertainty magnitudes, inferred mainly from System 1 and applied uniformly; if real pB° or γ errors differ across mixtures, the 'decisive' rankings and stage trends could shift.","rationale":"The reader's CONDITIONAL verdict is sound. I read the central claim as: for Pareto-optimized DWCs of ideal ternaries, VLE model errors dominate, pB° vapor-pressure uncertainty is the single most decisive property, and losses are amplified by higher stage numbers. What must be true for this claim is that the uncertainty ranges in Table 3 faithfully reflect actual model errors for these mixtures, at least in relative magnitude. The paper gives partial support (Fig. 5 for System 1; references 16 and 18), but does not demonstrate the same ranges for Systems 2 and 3. Since the conclusions are ranking claims, the relative sizes are load-bearing. This is not an accusation of incorrect numbers; it is a missing robustness check. The endpoint-only screening is a secondary but related issue: the worst case could lie inside the interval, and the qualitative ranking could depend on perturbation magnitude. The paper's own limitation statement ('no claim of general validity is made here') strengthens rather than weakens this concern, because the authors themselves recognize that the selection of relevant models and absolute values will deviate for other mixtures or different uncertainties. The proposed test—mixture-specific uncertainty quantification and intermediate perturbation levels—would resolve whether the highlighted claims survive. Because the concern is addressable with additional data and reruns rather than a conceptual flaw, I would keep the reader's CONDITIONAL verdict unchanged.","tokens_in":16143,"tokens_out":4875,"duration_ms":46947,"concrete_test":"Independently determine mixture-specific uncertainty intervals for all three systems: for each binary pair, compare NRTL/Aspen predictions against the full DDB/NIST data to obtain γij∞ deviations; for each pure component, compare extended-Antoine/DIPPR predictions to experimental vapor-pressure and enthalpy data over the relevant temperature range, as done in Fig. 5 only for System 1. Rerun the Section 3 worst-case scan using these mixture-specific intervals, and also include intermediate perturbations (±2.5%, ±5%, ±7.5%) for the leading properties. The central claim is settled if pB° still appears in all worst-case combinations and VLE-induced losses still exceed caloric-induced losses across all three mixtures.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline conclusions are conditional on the uncertainty magnitudes fixed in Table 3 (±10% infinite-dilution activity coefficients, ±2% pure vapor pressures, ±1% enthalpy of vaporization, ±2% vapor enthalpy). Section 2.5 shows the uncertainty quantification for System 1 only (Fig. 5), then states that the ranges are similar for all three mixtures and consistent with literature; no equivalent data are presented for Systems 2 or 3. Because the conclusions are comparative ('VLE decisive', 'vapor pressure of medium boiler decisive'), what matters is not just whether each range is individually plausible, but the relative sizes. If real pure-component vapor-pressure errors (especially pB°) are smaller than 2% relative to γ uncertainties, activity-coefficient perturbations would dominate and the pB° claim would not be 'decisive independent of mixture'; if larger, the quantitative loss estimates would increase. The paper's own caveat 'no claim of general validity' (Section 3.1) concedes sensitivity to this parameter choice. A related gap is that the parametric screen evaluates only the endpoints of each uncertainty interval, never intermediate perturbation levels; a non-monotonic response with a larger effect at, say, ±5% than at ±10% would be missed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates how uncertainties in thermodynamic models (activity coefficients, pure vapor pressures, and vapor/evaporation enthalpies) affect product purities in Pareto-optimized dividing wall columns (DWCs) for three near-ideal ternary mixtures at four different total stage counts. Using a stage-to-stage MESH simulation with the nominal optimized column specifications fixed, the authors screen individual and combined perturbations and report worst-case purity losses. They find that VLE-related uncertainties dominate over caloric ones, that losses increase with total stage number when the VLE or internal vapor/liquid ratio is perturbed, and that the medium-boiler vapor pressure is the most influential property. The results are interpreted using Vmin diagrams, and the joint effects of simultaneous uncertainties are presented as approximately additive superpositions.","tokens_in":16364,"tokens_out":6858,"duration_ms":58080,"significance":"The paper addresses a practically important gap: thermodynamic model uncertainty is rarely considered in DWC design optimization. The systematic screening across mixtures and stage numbers, and the connection to Vmin-diagram features, provides a useful framework for identifying critical properties and for understanding why extra stages can increase operational risk. The simulation setup is described in enough detail to re-implement, and the results are internally consistent. However, the headline conclusions are conditional on the assumed uncertainty ranges, which are mainly validated for one mixture and applied uniformly; this limits the generality of the \"decisive\" rankings. If the requested sensitivity checks are provided, the framework would be considerably more robust.","major_comments":[{"comment":"The uncertainty magnitudes (γ∞ ±10%, p° ±2%, Δhv ±1%, hv ±2%) are derived from comparisons that are shown mainly for System 1 (Fig. 5), then applied uniformly to all three mixtures without equivalent data for Systems 2 and 3. Because the conclusions are comparative (\"VLE decisive\", \"pB° decisive\"), the relative sizes of these ranges are load-bearing. If, for example, the real pure-component vapor-pressure error were significantly smaller than 2% for System 2, the pB°-dominance claim could fail. The caveat in Section 3.1 (\"no claim of general validity\") mitigates this, but the abstract and highlights present the conclusions unconditionally. Please either provide uncertainty estimates for all mixtures or perform a sensitivity analysis of the key results to the assumed ranges (e.g., ±5% γ, ±1% p°) to demonstrate robustness.","section":"Section 2.5, Table 3"},{"comment":"The combined scenarios use only the extreme perturbation levels (±10% γ, ±2% p°) and omit the nominal and any intermediate values. This screening design cannot detect non-monotonic responses in product purity, and it limits the support for the \"addition of individual influences\" statement in Section 3.2 (Fig. 8); the superposition claim is verified only for the extremes. Please test at least one intermediate perturbation level for the most critical combinations (e.g., the pB° + ABC worst cases), or provide a justification for why the extrema bound the losses, particularly given that the individual-scenario screening (Section 3.1) did include the nominal value.","section":"Section 3.2, Fig. 10"},{"comment":"The comparison between VLE and caloric properties is not on an equal footing: the assumed uncertainty for activity coefficients (±10%) is an order of magnitude larger than that for Δhv (±1%) and hv (±2%). The highlight \"Caloric properties are unsignificant\" is therefore a statement about these specific ranges, not about the models in general. Please either perform a symmetric comparison at equal perturbation levels (e.g., ±10% for all properties) or explicitly state in the abstract and conclusions that the ranking is conditional on the ranges identified as realistic. This is related to the first major comment but applies specifically to the VLE-versus-caloric claim.","section":"Section 3.1, Table 4"}],"minor_comments":[{"comment":"The text states that the mixtures have \"close-to-ideal behavior (i.e., activity coefficients γi ≈ γj ≈ 1)\", but Table 1 lists infinite-dilution activity coefficients up to 1.229 for System 2 (methanol–butanol). Please clarify the criterion used, or state that the activity coefficients are close to unity in the concentration ranges of interest.","section":"Section 2.2, Table 1"},{"comment":"The typeset equation for the Margules perturbation appears garbled (missing summation limits and unreadable bracket structure). Please provide a cleanly typeset version with correct indices and parentheses.","section":"Section 2.5, Eq. (3)"},{"comment":"The definition Δxi = 0.95 − xi should specify that losses are measured relative to the 95 mol% product specification, not necessarily to the actual nominal purity, which may be slightly above 0.95.","section":"Section 3.1, Table 4"},{"comment":"The Self-Organizing Patch Plots lack axis or colorbar labels. Adding a legend explaining the binary encoding (0 = negative deviation, 1 = positive deviation) and the color scale for purity loss would make the figures interpretable without referencing the text in detail.","section":"Section 3.2, Figs. 10–11"},{"comment":"The novelty claim \"For the first time\" is not substantiated relative to prior work. Please cite the closest works (e.g., [18] on simple columns) and specify what is new for DWCs, so the claim is verifiable.","section":"Abstract"},{"comment":"The highlight \"Vapor pressure of medium boiler is decisive independent of mixture or stages\" should be qualified with \"within the investigated uncertainty ranges\" to avoid overgeneralization, consistent with the caveat in Section 3.1.","section":"Highlights"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the simulation work appears technically sound. The main risk is that the strong comparative claims rest on a single set of uncertainty magnitudes, which are mainly validated for one mixture. The requested sensitivity checks and intermediate-value tests would substantially increase confidence in the conclusions. The paper relies heavily on the authors' own previous work [9] and on the perturbation method of Burger et al. [18], which is co-developed by one of the current authors; this is not a problem per se, but the incremental novelty should be clearly delineated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a solid, careful simulation study that fills a genuine gap: it gives the first systematic comparison of thermodynamic model uncertainties on Pareto-optimized dividing wall columns, with varying stage counts and three ideal mixtures. The headline claims—VLE uncertainties dominate, caloric properties are negligible in the tested range, and the vapor pressure of the medium boiler is decisive—are supported by the evidence presented. The stage-number dependence (larger purity losses at higher stage counts when the VLE is perturbed) and the correlation with the highest Vmin peak are useful, and the McCabe-Thiele analysis of why one product in System 2 is robust is a genuine insight.\n\nThe paper also does some things right. The perturbation method is not new (Margules-based from Burger et al. and Mathias), but applying it to optimized DWC designs is. The scenario enumeration is thorough: all 27 vapor-pressure combinations, all combinations of activity-coefficient factors, and 512 combined scenarios. The simulation setup is described in enough detail to re-implement, and the internal consistency checks (e.g., superposition of individual effects) are convincing.\n\nThe soft spots are real but manageable. The uncertainty magnitudes in Table 3 (±10% γ∞, ±2% p°, ±1% Δhv, ±2% hv) are point estimates, illustrated mainly for System 1, and then applied uniformly to the other mixtures. The paper acknowledges this and disclaims general validity. Still, the ranking of 'VLE decisive vs. caloric negligible' depends on the relative sizes of these ranges; if real vapor-pressure errors are much smaller than 2% for some mixtures, the activity-coefficient part would dominate even more, and if larger, the quantitative loss estimates would grow. The paper's qualitative conclusions are probably robust within a plausible range, but they are not stress-tested against alternative uncertainty ranges.\n\nA more concrete gap: the combined-scenario screen only samples the endpoints of each perturbation (the SOPPs show only 0 and 1 values), never intermediate perturbation levels or the nominal case. If the purity loss is non-monotonic in perturbation size—e.g., a resonance at ±5% not captured at ±10%—this screen would miss it. I don't see evidence of such non-monotonicity in the individual runs, but it is an untested assumption.\n\nNo code or data are provided; since the exact loss numbers are simulation outputs, this limits independent verification. The restriction to ideal mixtures and steady state is stated clearly and does not undermine the scope claims.\n\nOverall, this is a good, honest engineering science paper. It is not groundbreaking, but it gives process designers a useful ranking of where to spend data-quality effort, and it is internally consistent. It deserves a serious referee. My recommendation: engage with it, and ask the authors to (1) test intermediate perturbation values in the combined screen, (2) provide the simulation artifacts or at least tabulated results, and (3) include a short robustness check on the relative uncertainty magnitudes. Those are conditions, not fatal flaws.","headline":"Careful simulation study that fills a real gap; the central ranking holds within stated uncertainty ranges, but the combined screen should test intermediate perturbations before the numbers are used.","tokens_in":16929,"tokens_out":2858,"would_cite":true,"duration_ms":23729,"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":"This paper claims that in Pareto-optimized dividing wall columns, vapor-liquid equilibrium model uncertainty — not caloric property uncertainty — is what destroys product purity, and that the losses grow as the column gains theoretical…","keywords":["Dividing Wall Column","uncertainty","thermodynamic models","sensitivity","vapor-liquid equilibrium","Pareto optimization","Vmin diagram","distillation"],"falsifier":"Re-run the same combined 512-scenario screening on a ternary mixture whose experimentally calibrated uncertainty ranges differ from Table 3, for example with caloric errors near 5% or activity-coefficient errors below 5%, and check whether the ranking still holds: VLE over caloric, medium-boiler vapor pressure decisive, and higher-stage columns losing more purity than lower-stage columns. A single mixture where a light- or heavy-boiler vapor-pressure perturbation produces the largest loss, or where the 56-stage column is less sensitive than the 25-stage column, would bound the claim.","tokens_in":15899,"feed_emoji":"⚗️","tokens_out":7097,"duration_ms":62821,"temperature":0.7,"pith_summary":"This paper tries to establish that when a dividing-wall column is Pareto-optimized for energy and stage count, the dominant operational risk from thermodynamic model error sits in the vapor-liquid equilibrium: activity coefficients and pure-component vapor pressures. Across three near-ideal ternary mixtures and four stage counts each, worst-case purity losses grew with total stage number and reached roughly 3 to 6 mol% under combined VLE uncertainties, while perturbations of vapor enthalpy and enthalpy of vaporization stayed below about 0.5 mol%. The authors identify the vapor pressure of the medium boiler as decisive independent of mixture or column size, and the products of the binary split with the highest minimum vapor demand as the most affected. The practical stakes are a design rule: direct data-quality and model-refinement efforts at the difficult binary split and at low mid-boiler concentrations, and recognize that adding stages can enlarge worst-case purity losses instead of buying safety.","feed_headline":"VLE errors, not heat data, decide dividing wall column purity loss","feed_subtitle":"Worst-case losses grow with stage count, and the medium boiler's vapor pressure is the decisive property, across three ideal mixtures.","key_machinery":"The argument is carried by fixed Pareto-optimal DWC designs from the authors' earlier NQ-curve optimization, simulated rigorously stage-to-stage under perturbed property models. The perturbation machinery is a Margules-type term added to NRTL activity coefficients, anchored to ±10% changes in binary infinite-dilution activity coefficients with thermodynamic consistency, plus multiplicative ±2% scaling of pure-component vapor pressures and 1–2% scaling of enthalpy terms. The organizing identity is the Vmin diagram — a shortcut map of the minimum boil-up required for each sharp binary split — because the split with the highest Vmin peak sets the column's vapor demand, and its products are the ones whose operating lines sit closest to the VLE. Losses arise when a perturbed VLE shifts the equilibrium line toward or across the operating line, and McCabe-Thiele diagrams are used to show why one product can remain protected when its split is much easier than the difficult split.","core_discovery":"On its own terms, the paper claims that for steady-state Pareto-optimized dividing wall columns separating close-to-ideal ternary mixtures, model uncertainty in the vapor-liquid equilibrium is the dominant threat to product purity while caloric-property uncertainty is negligible. Screening all combinations of ±10% infinite-dilution activity coefficients and ±2% pure vapor pressures across four stage counts per mixture, the authors find worst-case purity losses that grow with total stage number and reach roughly 3 to 6 mol% when both VLE uncertainties act together; enthalpy of vaporization and vapor enthalpy perturbed by 1–2% cause losses below about 0.5 mol%. They identify the vapor pressure of the medium boiler as decisive independent of mixture and stage count, and the products belonging to the binary split with the highest minimum vapor demand, marked by the highest peak in the Vmin diagram, as the most affected. The side product B is the most sensitive product overall, except in the methanol/ethanol/butanol system where the medium/heavy-boiler split is easy and that heavy product stays nearly pure.","pith_inferences":["If the stage-dependence generalizes, common oversizing practice for DWCs should be reconsidered: adding stages to buy margin may increase worst-case sensitivity, and a natural extension is to optimize stage allocation against worst-case VLE rather than nominal VLE.","Because the Table 3 uncertainty ranges were calibrated mainly against one mixture, applying the ranking to a new system requires re-estimating those ranges; a system with larger caloric uncertainty could overturn the conclusion that caloric properties are insignificant.","The Vmin-peak rule suggests a cheap screening test for any candidate ternary system: compute the Vmin diagram, identify which binary split has the highest minimum boil-up, and concentrate VLE measurement on that binary's low-concentration range; this prediction can be tested on non-ideal mixtures.","In dynamic operation the vapor split is fixed by construction and cannot be re-tuned, so the same VLE perturbation could produce larger transient purity excursions than the steady-state losses reported here."],"forward_implications":["Design-stage thermodynamic data efforts should concentrate on VLE quantities, especially the medium boiler's vapor pressure and the infinite-dilution activity coefficients of the highest-Vmin binary split, rather than on caloric property models.","A Pareto-optimal DWC with more total stages is not automatically safer: if the VLE model is wrong, purity losses grow with stage number, so an extra-stages safety margin can widen worst-case product giveaway.","Combined VLE uncertainties act approximately additively in this setting, so screening the individual vapor-pressure and activity-coefficient corner cases separately and summing the worst cases gives a practical bound without running a full 512-scenario sweep.","Mixtures with one clearly easy binary split can shield the corresponding product from thermodynamic uncertainty, as seen for the heavy product in the methanol/ethanol/butanol system.","The same stage-dependent logic applies to any uncertainty that shifts the equilibrium line relative to the operating line, so process-variable and VLE uncertainties can be expected to compound in the same parts of the column."],"supporting_citations":[{"why":"Supplies the Pareto-optimized DWC designs, mixtures, stage allocations, and the earlier process-variable deviation results that this work extends to thermodynamic uncertainty.","marker":"[9]"},{"why":"Provides the thermodynamically consistent Margules perturbation scheme used to apply activity-coefficient uncertainties.","marker":"[18]"},{"why":"Supplies the method and precedent for perturbing activity coefficients via the Margules equation and the finding that mixture properties dominate VLE uncertainty effects.","marker":"[15]"},{"why":"Supports the mixture-dependence of VLE uncertainty effects and the observation that same-direction deviations have little influence.","marker":"[16]"},{"why":"Introduces the Vmin diagram used to identify the binary split with the highest minimum vapor demand, which organizes the mixture-dependent results.","marker":"[22]"},{"why":"Establishes the older observation that data-error effects can be negligible or enormous depending on separation factor and stage requirement.","marker":"[14]"},{"why":"Connects the vapor-split feasibility window to optimal operation, used in interpreting why near-optimal designs are sensitive.","marker":"[19]"},{"why":"Shows how stage allocation affects the feasible operating window, relevant to the stage-dependent sensitivity reported here.","marker":"[20]"}],"fun_headline_variants":["VLE uncertainty drives dividing wall column purity loss","Medium boiler's vapor pressure is the key VLE uncertainty","Worst-case purity loss grows with stage count, up to 6%","Thermo model errors dominate dividing wall column purity loss","VLE, not caloric uncertainty, sets dividing wall column limits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumed uncertainty ranges in Table 3 — ±10% for infinite-dilution activity coefficients, ±2% for pure vapor pressures, ±1% for enthalpy of vaporization, and ±2% for vapor enthalpy — are treated as representative for all three mixtures even though they were estimated mainly from one mixture, and the authors explicitly disclaim general validity.","fun_headline_variants_meta":{"raw":{"variants":["VLE uncertainty drives dividing wall column purity loss","Medium boiler's vapor pressure is the key VLE uncertainty","Worst-case purity loss grows with stage count, up to 6%","Thermo model errors dominate dividing wall column purity loss","VLE, not caloric uncertainty, sets dividing wall column limits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000702,"raw_usage":{"total_tokens":3156,"prompt_tokens":923,"completion_tokens":2233,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":2149}},"tokens_in":539,"tokens_out":2233,"duration_ms":14026,"temperature":1.0,"reasoning_tokens":2149,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:21:59.054829+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same combined 512-scenario screening on a ternary mixture whose experimentally calibrated uncertainty ranges differ from Table 3, for example with caloric errors near 5% or activity-coefficient errors below 5%, and check whether the ranking still holds: VLE over caloric, medium-boiler vapor pressure decisive, and higher-stage columns losing more purity than lower-stage columns. A single mixture where a light- or heavy-boiler vapor-pressure perturbation produces the largest loss, or where the 56-stage column is less sensitive than the 25-stage column, would bound the claim.","supporting_citations":[{"cited_title":"relevant","cited_arxiv_id":null,"evidence_quote":"Supplies the Pareto-optimized DWC designs, mixtures, stage allocations, and the earlier process-variable deviation results that this work extends to thermodynamic uncertainty."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the thermodynamically consistent Margules perturbation scheme used to apply activity-coefficient uncertainties."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the method and precedent for perturbing activity coefficients via the Margules equation and the finding that mixture properties dominate VLE uncertainty effects."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the mixture-dependence of VLE uncertainty effects and the observation that same-direction deviations have little influence."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Vmin diagram used to identify the binary split with the highest minimum vapor demand, which organizes the mixture-dependent results."},{"cited_title":"D., Henley, E","cited_arxiv_id":null,"evidence_quote":"Establishes the older observation that data-error effects can be negligible or enormous depending on separation factor and stage requirement."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Connects the vapor-split feasibility window to optimal operation, used in interpreting why near-optimal designs are sensitive."},{"cited_title":"J., & Skogestad, S","cited_arxiv_id":null,"evidence_quote":"Shows how stage allocation affects the feasible operating window, relevant to the stage-dependent sensitivity reported here."}],"review_version":1}