{"id":"7117ea0d-f8e7-4879-b4be-1141dfeb7315","arxiv_id":"2608.11255","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Symbolic regression corrections added to Peng-Robinson equation-of-state predictions reduce vapor-liquid equilibrium errors for six nitrogen-n-alkane systems, with coefficients interpolated as functions of carbon number.","lead":"This paper uses symbolic machine learning to fit compact correction formulas on top of the Peng-Robinson equation of state, improving predictions of vapor-liquid equilibrium pressure and vapor composition for nitrogen-alkane mixtures. The corrected formulas match the measured data better than the standard equation and stay human-readable, but generalization to untested hydrocarbon chain lengths is not validated.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-system generalization is asserted but never tested at an unseen carbon number; reported test metrics only cover the six calibrated systems.","rationale":"The paper's central claim is that the two-level symbolic regression yields a unified correction model that generalizes across hydrocarbon chain lengths. That claim rests on the assumption that the system-specific coefficients alpha_m vary smoothly with carbon number, as implemented through the polynomial functions in Fig. 2 and Algorithm 2. The reported evaluation, however, uses a pointwise 80/20 split over the same six systems, so every test point comes from a system whose coefficients were directly calibrated. This does not test interpolation to untested chain lengths such as C8 or C11, nor does it test the shape of the coefficient curves between the six carbon numbers. The reader's weakest assumption identifies exactly this gap, and no passage in the manuscript supplies the missing leave-one-system-out evidence. The concern is therefore empirical rather than mathematical: the within-system accuracy improvements are plausible and the appendices provide explicit model forms, but the cross-system generalization claim is unverified. The proposed leave-one-system-out check would settle whether the concern lands, so the conditional verdict remains appropriate.","tokens_in":18512,"tokens_out":6369,"duration_ms":56228,"concrete_test":"Perform leave-one-system-out validation: for each of the six systems, fit the coefficient functions alpha_m^P(nC) and alpha_m^y(nC) using only the other five systems (including re-optimizing basis constants on the merged five-system training data), then evaluate MSE and R2 on the held-out system's full dataset. Repeat for all six systems. If held-out system errors are comparable to the in-sample test MSEs in Tables 4 and 5, the cross-system claim survives; if errors increase by an order of magnitude or R2 drops materially, the unified model is an interpolation artifact. As a stronger check, if N2 + n-C8 or n-C11 VLE data are available in the literature, fit on the six systems and predict those chain lengths directly, reporting the polynomial degree used in Fig. 2.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step for the paper's headline claim is the coefficient parameterization alpha_m(nC) in Algorithm 2, Fig. 2, and Tables A.8/A.9. The paper treats alpha_m as smooth polynomial functions of carbon number, but it never evaluates the model at a carbon number that was not used to determine those functions. Every reported test point (Tables 4 and 5) belongs to one of the same six systems C5, C6, C7, C9, C10, C12, so the 80/20 split only tests interpolation in (T_r, x_N2) within already-calibrated systems. If the 'polynomial interpolation' in Fig. 2 is an interpolant through the six system points, it reproduces the per-system coefficients exactly for those systems; the test metrics then cannot distinguish the unified model from a lookup table of system-specific coefficients. The claim that the model 'enables prediction across different hydrocarbon systems' is thus supported only by visual smoothness, not by out-of-system error metrics. No leave-one-carbon-number-out validation is reported, and the manuscript contains no limitation statement acknowledging this gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a two-level symbolic regression framework that learns interpretable correction terms for Peng–Robinson equation-of-state (PR-EOS) predictions of vapor–liquid equilibrium in binary N2 + n-alkane systems. In the first level, system-specific symbolic expressions for the pressure and vapor-composition residuals are discovered by symbolic regression for each of six alkanes (C5, C6, C7, C9, C10, C12). In the second level, recurring symbolic structures are collected into shared basis functions, and the basis coefficients are represented as polynomial functions of carbon number. The authors report that the resulting closed-form correction reduces pressure MSE by nearly two orders of magnitude relative to uncorrected PR-EOS and also improves vapor-composition prediction, while preserving the thermodynamic structure of the baseline model.","tokens_in":18755,"tokens_out":4632,"duration_ms":41265,"significance":"If the cross-system generalization claim were established, the paper would make a useful engineering contribution: it provides compact, interpretable, closed-form corrections to a widely used cubic equation of state, and the appendix gives explicit basis functions and system-specific coefficients that would allow the model to be reconstructed and applied. The within-system improvement over PR-EOS is clearly demonstrated in the reported tables and parity plots, and the symbolic-regression formulation is a sensible alternative to opaque neural-network corrections. The load-bearing weakness is that the headline claim of prediction across different hydrocarbon systems is not actually tested: the reported validation uses a pointwise 80/20 split within the six calibrated systems, so it only demonstrates interpolation in temperature and composition for systems whose data were used to determine the carbon-number coefficient curves.","major_comments":[{"comment":"The central claim that the unified model 'enables accurate prediction across different hydrocarbon systems' is not supported by the reported validation. Tables 4 and 5 use a common 80/20 pointwise split across all six systems, so every test point belongs to one of the six systems whose data determined the shared basis and the carbon-number coefficient functions. Because the polynomial interpolants in Fig. 2 pass through the six system coefficient values (Tables A.8/A.9), the test metrics only demonstrate interpolation in (Tr, xN2) within already calibrated systems; they cannot distinguish the unified model from a per-system lookup or interpolation of system-specific coefficients. The paper needs a leave-one-carbon-number-out evaluation, such as training on five systems and evaluating on the held-out alkane, or holding out an intermediate chain length, before the generalization claim can be assessed.","section":"3.1, Algorithm 2; Tables 4 and 5"},{"comment":"The coefficient parameterization alpha_m(nC), which is the mechanism that would enable prediction at new carbon numbers, is never written down. Fig. 2 shows only graphical interpolants, and Tables A.8/A.9 list system-specific coefficients but not the polynomial degrees, polynomial coefficients, or any regularization used for alpha_m(nC). Without these explicit expressions the model is not reproducible, and its behavior at unmeasured carbon numbers is unspecified. With only six points, the 'smooth' curves in Fig. 2 could be high-degree interpolants that oscillate between the fitted systems, so the visual smoothness is not evidence of predictive behavior at unmeasured chain lengths.","section":"Appendix B, Fig. 2, Tables A.8/A.9"},{"comment":"The PR-EOS baseline is incompletely specified: the binary interaction parameter k_ij in Eq. (7) is never assigned a numerical value and the manuscript does not state whether it was set to zero, taken from a literature correlation, or optimized per system. Because the paper's headline improvement is measured against PR-EOS, the numerical results are not reproducible without this information, and a fitted k_ij could absorb part of the error that the paper attributes to the uncorrected baseline.","section":"2, Eq. (7), Tables 4 and 5"},{"comment":"The normalization step in Algorithm 1, which normalizes Tr, xN2, Delta-P, and Delta-y within each system, is not integrated into the unified model. The per-system normalization constants are not reported, and Algorithm 2 does not state whether the merged optimization uses per-system normalized variables or a global normalization, nor how normalization would be defined for a new carbon number. A prediction for an untested alkane therefore requires these constants or an interpolated normalization rule; without that specification the unified model is incomplete.","section":"Algorithm 1, Algorithm 2"},{"comment":"The manuscript does not state whether the 80/20 split was applied before the system-specific symbolic regressions and the coefficient optimization, or only at the final reporting stage. If the symbolic basis and coefficients are fit to all 660 points and then split into train/test metrics, the reported 'test' numbers are not independent. The order of splitting and fitting must be clarified, and the split should be applied before any symbolic discovery or coefficient optimization.","section":"3.1, Tables 4 and 5"}],"minor_comments":[{"comment":"The caption in Fig. 10 calls the proposed model the 'lookup coefficient model,' while Figs. 9 and 11 call it the 'corrected' or 'two-level SR' model; the terminology should be made consistent, and the color legend text should be checked against the actual markers.","section":"Fig. 10 caption"},{"comment":"The subplot titles 'Pressure' and 'Vapor composition' are repeated in each panel and the legend order differs between the top, middle, and bottom rows; simplifying the layout would improve readability.","section":"Fig. 12"},{"comment":"The 'reference critical locus' is said to be included for visualization, but its source or method of construction is not given; a citation or a clear statement that it is a hand-drawn guide would be helpful.","section":"Fig. 13"},{"comment":"The appendix gives the optimized basis constants and system-specific coefficients, but it would be helpful to state explicitly how the shared basis functions in Appendix B relate to the symbolic forms identified in Tables 2 and 3, since the notation for the internal constants changes between the main text and the appendix.","section":"Appendix A and B"},{"comment":"The phrase 'after which their coefficients are represented as functions of carbon number' in the abstract could be more precise: the coefficients are represented by polynomial interpolants through six systems, not by a theoretically derived scaling law; this distinction should be stated in Section 1 so readers do not overinterpret the extrapolation claim.","section":"Abstract and Section 1"}],"recommendation":"major_revision","confidential_remarks":"The within-system accuracy results are convincing and the symbolic-correction idea is a good fit for the journal, but the manuscript's main contribution is presented as cross-system prediction, and that claim is not yet tested. If the authors add a leave-one-carbon-number-out analysis, provide the explicit polynomial coefficient functions, and clarify the PR-EOS baseline and normalization details, the paper could be suitable for publication. As submitted, the gap between the claim and the validation is too large for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a referee's time, but the headline claim needs one more experiment before I'd believe it. The two-level idea—learn system-specific symbolic corrections to PR-EOS, extract shared basis functions, then parameterize the coefficients by carbon number—is genuinely new relative to [34], which did symbolic regression for liquid density, not VLE. The appendices give the full basis functions and all coefficients, so the within-system model is explicit and reconstructable. Tables 4 and 5 show large, consistent gains over PR-EOS, with pressure MSE dropping by roughly two orders of magnitude, and the pooled symbolic regression baseline is a reasonable comparator. That part looks solid.\n\nThe soft spot is exactly where the stress-test note points. The claim that the unified model enables prediction across different hydrocarbon systems is supported only by the polynomial interpolation in Fig. 2. The six alkanes (C5, C6, C7, C9, C10, C12) have gaps at C8 and C11, and no leave-one-system-out or holdout-carbon-number evaluation is reported. The 80/20 split is pointwise within those six systems, so a test point still belongs to a system whose coefficients were used to build the interpolant. In that setting, the test metrics cannot distinguish the unified model from a lookup table of system-specific coefficients. The paper also does not state how the split interacts with the gradient-based coefficient fitting in Algorithm 2; if the basis constants and coefficient functions were fit on all six systems before splitting, the test errors are not independent. I'd want a clean leave-one-carbon-out test, or at minimum a model fit on C5–C10 and evaluated at C12.\n\nMinor but real: the PR-EOS binary interaction parameter k_ij is never specified. The magnitude of the residuals being corrected depends on it, and if it was fit to the same data, that is a free parameter not accounted for. No code or data are included; the appendices are enough to reconstruct the corrections for the listed systems, but not to reproduce the fitting or the split.\n\nThe citation pattern looks fine. [31] is the authors' own neural-network VLE work and is a legitimate comparison point.\n\nBottom line: as a method for correcting PR-EOS for these six systems, it works and is transparent. As a demonstration of cross-system prediction via carbon-number interpolation, it is not yet demonstrated. The fix is straightforward, so I would send it to review with the request for leave-one-system-out validation and a clear statement on k_ij and the split. I'd bring it to reading group as an example of interpretable ML for thermodynamic models, but I would not cite it as evidence for interpolation to untested chain lengths until that validation exists.","headline":"A real two-level symbolic correction for PR-EOS VLE residuals; the cross-system generalization claim rests on carbon-number interpolation that is never tested at an unseen chain length.","tokens_in":19282,"tokens_out":3466,"would_cite":false,"duration_ms":31345,"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":"The paper claims that a two-level symbolic regression can correct Peng–Robinson equation-of-state (PR-EOS) predictions for nitrogen–alkane vapor–liquid equilibrium by learning compact residual expressions and parameterizing their…","keywords":["Peng-Robinson equation of state","vapor-liquid equilibrium","symbolic regression","binary mixtures","nitrogen-alkane systems","carbon number","interpretable machine learning","equation-of-state correction"],"falsifier":"Train the two-level model on five of the six systems (for example, C5, C6, C7, C9, C10) and evaluate the interpolated coefficients on the held-out C12 data, or on published N2 + n-octane measurements not used in the paper; if the test error exceeds the within-system test error by more than the natural scatter, the carbon-number interpolation claim fails.","tokens_in":18298,"feed_emoji":"🧪","tokens_out":6501,"duration_ms":53687,"temperature":0.7,"pith_summary":"This paper seeks to make the Peng–Robinson equation of state (PR-EOS) accurate for nitrogen–alkane vapor–liquid equilibrium by adding small symbolic correction terms learned from measured data rather than replacing the equation. The authors first discover, separately for six binary N2 + n-alkane systems, explicit formulas for the residual errors in equilibrium pressure and vapor nitrogen fraction. They then find that the same few symbolic building blocks recur across systems and parameterize only the coefficients as polynomial functions of carbon number, producing one compact closed-form correction for all six systems. On a common 80/20 split, the corrected model lowers pressure mean-squared error by nearly two orders of magnitude relative to PR-EOS and improves vapor-composition accuracy substantially, with validation shown inside each fitted system.","feed_headline":"Learned formulas shrink an equation-of-state error ~100-fold","feed_subtitle":"Two-level symbolic regression converts PR-EOS residuals into carbon-number formulas for six N2-alkane systems.","key_machinery":"The load-bearing mechanism is post-processing residual correction with a shared symbolic basis. Specifically, the pressure correction is written $\\Delta P^{(s)}(T_r,x_{N_2}) = \\sum_{m=1}^5 \\alpha_m^{(P,s)} \\phi_m^{(P)}(T_r,x_{N_2})$ and the vapor-fraction correction as $\\Delta y^{(s)}(T_r,x_{N_2}) = \\sum_{m=1}^6 \\alpha_m^{(y,s)} \\phi_m^{(y)}(T_r,x_{N_2})$, with the $\\phi$ functions fixed across systems and the $\\alpha$ coefficients interpolated as polynomial functions of carbon number. This carries the argument because it converts a system-by-system data fit into a single formula whose only system-dependent input is the alkane chain length; the paper provides all basis constants and coefficients in appendices so the correction can be reconstructed directly.","core_discovery":"The central claim is that the systematic gap between PR-EOS predictions and measured VLE in N2 + n-alkane binaries can be captured by a compact two-level symbolic correction. In the first level, symbolic regression fits $\\Delta P = P_{\\rm exp}-P_{\\rm PR}$ and $\\Delta y_{N_2}=y_{\\rm exp}-y_{\\rm PR}$ as functions of reduced temperature $T_r$ and liquid nitrogen fraction $x_{N_2}$ for each of six systems (C5, C6, C7, C9, C10, C12). In the second level, recurring expressions—linear-exponential, coupled polynomial, exponential-nonlinear, and inverse-temperature forms—are assembled into a shared basis, and only the coefficients are allowed to vary, as polynomial interpolants in carbon number $n_C$. The paper reports that this unified model beats both the uncorrected PR-EOS and a single pooled symbolic-regression model on train and test splits for all six systems, with the pressure mean-squared error reduced by nearly two orders of magnitude and the vapor-composition error reduced by roughly an order of magnitude.","pith_inferences":["A natural next test is whether the carbon-number polynomials are smooth enough for interpolation to C8 and C11 and to heavier n-alkanes; if they hold, the model would become a parameter-light correlator for the whole homologous series.","The recurring exponential–linear basis forms resemble what a composition- and temperature-dependent binary interaction parameter $k_{ij}$ would do inside PR-EOS; rewriting the learned correction as a $k_{ij}(T_r, x, n_C)$ correction would make the physical content more direct and may improve extrapolation.","The same two-level recipe—residual fitting per system, shared basis, coefficient parameterization by a systematic descriptor—could transfer to other homologous families, with the descriptor possibly acentric factor instead of carbon number.","If validated on held-out chain lengths, the approach would give engineers a compact alternative to large training sets: six carefully chosen systems plus carbon-number interpolation, rather than thousands of deep-learning samples."],"forward_implications":["Pressure predictions on the six studied systems drop to roughly one-hundredth of the PR-EOS mean-squared error, so the corrected model reproduces measured VLE isotherms in regimes where the cubic equation of state deviates most.","The same shared basis with carbon-number-dependent coefficients yields an explicit, closed-form correction that can be written down and evaluated without retraining for any system covered by the coefficient interpolants.","The improvement is larger for heavier alkanes (C9, C10, C12), where PR-EOS errors are worst, suggesting the correction absorbs the chain-length-dependent interaction error that the mixing rule misses.","Because the corrected model keeps the PR-EOS structure intact, existing phase-equilibrium algorithms can substitute corrected pressure and composition outputs without changing fugacity or root-selection routines.","A pooled single-expression symbolic regression is consistently worse than the two-level model, indicating that factoring out carbon-number dependence helps, not just symbolic fitting itself."],"supporting_citations":[{"why":"supplies the high-pressure N2 + n-pentane VLE measurements used to fit the C5 correction.","marker":"[36]"},{"why":"supplies the high-pressure N2 + n-hexane VLE measurements used to fit the C6 correction.","marker":"[37]"},{"why":"supplies the high-pressure N2 + n-heptane VLE measurements used to fit the C7 correction.","marker":"[38]"},{"why":"supplies the high-pressure N2 + n-nonane VLE measurements used to fit the C9 correction.","marker":"[39]"},{"why":"supplies the N2 + n-decane VLE dataset used to fit the C10 correction.","marker":"[14]"},{"why":"supplies the N2 + n-dodecane VLE dataset used to fit the C12 correction.","marker":"[15]"},{"why":"defines the original Peng–Robinson equation of state whose residual errors are being corrected.","marker":"[6]"},{"why":"documents the known shortcomings and past modifications of PR-EOS that motivate a data-driven residual correction.","marker":"[7]"},{"why":"demonstrates the use of symbolic regression to modify a cubic equation of state, providing the methodological precedent.","marker":"[34]"},{"why":"provides the evolutionary symbolic-regression implementation used to discover the system-specific expressions.","marker":"[35]"}],"fun_headline_variants":["Symbolic ML finds formulas that fix N2-alkane VLE predictions","Two-level symbolic regression slashes N2-alkane EOS error","Interpretable formulas improve PR-EOS for N2-alkane equilibria","Machine-learned corrections cut N2-alkane VLE error ~100x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the six measured alkanes define smooth polynomial trends in carbon number for the correction coefficients, so interpolating these polynomials predicts untested chain lengths; the paper does not hold out a whole carbon number to check that.","fun_headline_variants_meta":{"raw":{"variants":["Symbolic ML finds formulas that fix N2-alkane VLE predictions","Two-level symbolic regression slashes N2-alkane EOS error","Interpretable formulas improve PR-EOS for N2-alkane equilibria","Machine-learned corrections cut N2-alkane VLE error ~100x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000342,"raw_usage":{"total_tokens":1871,"prompt_tokens":926,"completion_tokens":945,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":862}},"tokens_in":542,"tokens_out":945,"duration_ms":8222,"temperature":1.0,"reasoning_tokens":862,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:34:00.495860+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the two-level model on five of the six systems (for example, C5, C6, C7, C9, C10) and evaluate the interpolated coefficients on the held-out C12 data, or on published N2 + n-octane measurements not used in the paper; if the test error exceeds the within-system test error by more than the natural scatter, the carbon-number interpolation claim fails.","supporting_citations":[{"cited_title":"Avendaño-Gómez","cited_arxiv_id":null,"evidence_quote":"supplies the high-pressure N2 + n-pentane VLE measurements used to fit the C5 correction."},{"cited_title":"High-pressure vapor–liquid equilibria in the nitrogen + n-hexane system.J","cited_arxiv_id":null,"evidence_quote":"supplies the high-pressure N2 + n-hexane VLE measurements used to fit the C6 correction."},{"cited_title":"High-pressure (vapor + liquid) equilibria in the (nitrogen + n-heptane) system.J","cited_arxiv_id":null,"evidence_quote":"supplies the high-pressure N2 + n-heptane VLE measurements used to fit the C7 correction."},{"cited_title":"High-pressure vapor–liquid equilibria in the nitrogen–n-nonane system.J","cited_arxiv_id":null,"evidence_quote":"supplies the high-pressure N2 + n-nonane VLE measurements used to fit the C9 correction."},{"cited_title":"García-Sánchez, G","cited_arxiv_id":null,"evidence_quote":"supplies the N2 + n-decane VLE dataset used to fit the C10 correction."},{"cited_title":"Garcia-Cordova, D","cited_arxiv_id":null,"evidence_quote":"supplies the N2 + n-dodecane VLE dataset used to fit the C12 correction."},{"cited_title":"Anewtwo-constantequationofstate.Industrial & Engineering Chemistry Fundamentals, 15:59–64, 1976","cited_arxiv_id":null,"evidence_quote":"defines the original Peng–Robinson equation of state whose residual errors are being corrected."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"documents the known shortcomings and past modifications of PR-EOS that motivate a data-driven residual correction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"demonstrates the use of symbolic regression to modify a cubic equation of state, providing the methodological precedent."}],"review_version":1}