{"id":"2baf92c6-a534-4a3e-85e6-59ac640a5755","arxiv_id":"2506.12498","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SINDybrid uses a mixed-integer linear program over a library of candidate functions to locate and fit data-driven corrections for the uncertain equations in an ODE model.","lead":"SINDybrid is an algorithm that automatically finds which equations in a physics-based model are unreliable and patches them with data-driven correction terms learned from measurements. It lowers the expert barrier to building hybrid models that combine first-principles knowledge with machine learning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SINDybrid's location-identification claim assumes exact FPM structure and parameters; a parameter-misspecification test is needed before the abstract's unqualified wording is accepted.","rationale":"The reader's weakest_assumption correctly identifies the exactness of the mechanistic model and full-state measurement as the load-bearing premise. Section 3.4 confirms that the method assumes correctly identified FPM parameters and does not explore parameter-level uncertainty or partial observability. Since Eq. (2) defines the epistemic residual as the difference between the measured derivative and the FPM evaluation, any error in f is indistinguishable from epistemic uncertainty and will be absorbed by the MILP in whichever column most reduces the residual. The three case studies only inject known h functions into an otherwise perfect f, so the validation does not exercise this failure mode. Other concerns, such as unreported per-case function libraries and tuning on the same cases, are real but secondary because they affect reproducibility and strength of evidence rather than the logical dependence on FPM correctness. The paper's own report of a misidentification at 5 time samples in the continuous fermentation case (Section 3.2.3) also qualifies the abstract's 'minimum 5 samples' claim, but it is a narrower, case-specific limitation. The proposed parameter-perturbation test would directly settle whether the central 'automatic detection' claim survives when the FPM is not exact; if it fails, the method remains useful as a model-deviation locator only under strong assumptions, which is exactly the conditional status the reader assigned.","tokens_in":16725,"tokens_out":5861,"duration_ms":71330,"concrete_test":"Re-run the Lotka-Volterra or Meerwein validation with one FPM parameter deliberately misspecified (e.g., the Lotka-Volterra growth rate or k_M set 20% off its true value) while keeping the synthetic epistemic deviation unchanged, using the same data-generation and MILP settings. Compare the active-column set identified by SINDybrid with the true deviation set. If parameter errors cause the identified location to shift to a different equation, or if the R2 on validation data drops below the claimed 0.85 under ±20% parameter error, the central claim is conditional on exact FPM knowledge and the abstract's unqualified wording should be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the mechanistic model f(x,r) is structurally correct and has correctly identified parameters, so the residual h_exp = dx_exp/dt - f(x_exp, r_exp) in Eq. (2) contains only the equation-level epistemic correction. Section 3.4 states this explicitly: the method 'operates under the assumption that the FPM parameters have been correctly identified,' and parameter-level uncertainty and partial-state observation were not explored. In any realistic application with an imperfect mechanistic model, h_exp will contain structural and parametric error in addition to the epistemic correction. The MILP in Eq. (9) will then minimize the loss by fitting the wrong columns of Ξ, and the algorithm may identify an epistemic uncertainty in an equation that is actually correct, compensating for a parameter error with a spurious data-driven term. The validation protocol never perturbs f; all three case studies in the SI only add the known h deviations to an otherwise exact FPM, so the central 'automatic detection' claim is untested under the most common failure mode of mechanistic models. The full-state measurement requirement is a second aspect of the same assumption: if a state is unmeasured, h_exp cannot be formed for that column and the location search is ill-posed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SINDybrid, an algorithm for automatic hybrid model construction for systems described by ordinary differential equations. Given a first-principles mechanistic model f(x,r) and experimental data, the method computes the residual h_exp = dx_exp/dt - f(x_exp,r_exp) and solves a mixed-integer linear program (MILP) that selects, from a user-supplied library of basis functions, a sparse set of correction terms and the equations to which the corrections apply. The authors introduce binary variables for term and column selection, an L1 regularizer, and scaled basis columns, and validate the approach on three synthetic case studies: a Meerwein arylation reaction, a continuous fermentation process, and a Lotka-Volterra oscillator. For each case, a known deviation h is injected into one equation at a time, and the paper reports the success of the location identification and the accuracy of the resulting hybrid model under varying measurement noise, number of training batches, and number of time samples.","tokens_in":16986,"tokens_out":5917,"duration_ms":68483,"significance":"The MILP formulation is coherent and the idea of automatically identifying which mechanistic equations carry epistemic uncertainty is practically relevant, since hybrid model construction still relies heavily on expert judgment. The authors provide an open-source implementation and the synthetic experiments are described in enough detail to follow the pipeline. If the method is robust under realistic conditions, it could lower the barrier to adopting hybrid modelling in chemical engineering. However, the validation is entirely in silico with injected deviations, and the two most consequential assumptions of the method - that the first-principles model has exactly correct structure and parameters, and that the epistemic uncertainty is exactly expressible in the chosen function library - are not tested or even fully specified in the manuscript. These omissions leave the central 'automatic detection' claim weaker than the abstract suggests.","major_comments":[{"comment":"The central claim that SINDybrid automatically detects epistemic uncertainty is only valid when the mechanistic model f has the correct structure and parameters. If parameters are misspecified or the model structure is incomplete in ways not captured by the assumed h, the residual h_exp computed from Eq. (2) contains parametric or structural error in addition to the epistemic correction. The MILP in Eq. (9) will then minimise the loss by fitting the wrong columns of Xi, potentially attributing the error to an equation that is actually correct. The authors acknowledge this in Section 3.4 ('operates under the assumption that the FPM parameters have been correctly identified') and state that parameter-level uncertainty and partial-state observation were not explored, but the abstract and introduction make unqualified claims. The validation protocol never perturbs f: in all three case studies, the FPM is exact and only known h terms are added. A parameter-misspecification test, such as shifting a kinetic constant by 10-20% and checking whether the algorithm still locates the true deviation or instead activates a spurious column, is needed before the abstract's wording can be accepted. At minimum, the claims should be qualified to state the exactness assumption prominently.","section":"Sections 1, 3.1, and 3.4; Eq. (2)"},{"comment":"The candidate function libraries X_L used for the three case studies are not reported anywhere in the manuscript or the supplementary material. The true deviation functions h are given in the SI (equations s2, s4, s6), but the function libraries from which the algorithm selects are not. Since the success of SINDybrid depends on h being exactly representable as a sparse linear combination of the library columns, the absence of the library specification prevents reproducibility and makes it impossible to assess whether the positive results are an artefact of cherry-picked libraries that already contain the true deviation terms. The authors should report the full library for each case (e.g., the set of monomials, rational terms, and other candidate functions) and, ideally, test with libraries that do not contain the exact h, to demonstrate that the method does not rely on knowing the answer in advance.","section":"Section 3.1 and SI Section s1"},{"comment":"The robustness results are reported without error bars or any measure of variability. Each point in the figures is described as an average over multiple runs ('Each run sets the synthetic epistemic uncertainty location on a different model equation'), which corresponds to at most four runs for the Meerwein case, three for fermentation, and two for Lotka-Volterra. With such small numbers, the claims that the algorithm is 'robust' and that R^2 values are consistently above 0.85 are not statistically quantified. The authors should report the number of replicates, the standard deviation or confidence intervals, or show all individual runs in the figures. This is particularly important because the location identification results are binary (success/failure), and a single failure at the boundary significantly changes the reported success rates.","section":"Figures 1, 2, 3 and Section 3.2"}],"minor_comments":[{"comment":"The abstract claims R^2 scores above 0.85 on validation data, while the conclusion states 'test R^2 > 0.89 in all cases'; these threshold values should be reconciled.","section":"Abstract and Section 4"},{"comment":"There are typos: 'SINDyrid' instead of 'SINDybrid' in Section 3.2.3, and 'Loka-V olterra' instead of 'Lotka-Volterra' in Section 3.2.1.","section":"Section 3.2.1 and 3.2.3"},{"comment":"The sums use indices starting at 0 (e.g., Nt,NS sum over i,j=0), but the matrices are naturally indexed from 1; the convention should be stated or corrected.","section":"Equations (4), (5), (8), (9)"},{"comment":"After normalising the library columns with Eq. (10), the coefficients Xi are those for the scaled columns, not for the original library functions. The manuscript should state this explicitly so that the final correction h_pred = X_L * Xi can be interpreted correctly.","section":"Section 2.2, Eq. (10)"},{"comment":"The caption of Figure 2 mentions 'Plots a), b) and c)' but the panels are labelled a1/a2/a3, b1/b2/b3, and c1/c2/c3; the caption is confusing and should be aligned with the actual panel labels.","section":"Section 3.2.2"},{"comment":"The manuscript does not state which MILP solver was used (e.g., Gurobi, CPLEX, CBC) or give typical computation times. Since MILP scalability is a practical concern, a brief statement on solver and run times would help assess the method's usability.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper falls within the scope of the journal and addresses a relevant problem. The strongest risk is the unverified assumption of an exactly correct first-principles model; a parameter-misspecification experiment is the key missing piece that would make the central claim credible. The missing library specifications are a serious reproducibility issue that should be fixed. The novelty claim of being the 'first framework' should be checked against the closely related work of Willis and von Stosch [6] and the broader MILP-based sparse identification literature, since the authors build directly on that reference. If the authors add the requested experiments and full library details, and appropriately qualify the claims, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things up front. First, SINDybrid is a real, modest contribution: it formulates the problem of choosing which ODE equations in a first-principles model need data-driven corrections as a column-sparse MILP over a SINDy-style library, and it ships code. Second, the headline claim that it \"automatically detects epistemic uncertainty\" is broader than what the validation supports, because every test injects the missing term into an otherwise exact first-principles model. The paper itself concedes this in Section 3.4, where it says the method assumes correct FPM parameters and full-state measurement, and that neither was explored. The stress-test note is right on target here.\n\nWhat the paper does well: the MILP formulation in (8)-(9) is coherent and clearly explained; the regularizer linkage in the SI is a sensible touch. The validation is honest about its synthetic nature and covers a useful range of noise, batch counts, and time samples, and the authors report a real failure case (continuous fermentation at 5 time points) rather than sweeping it under the rug. The open-source code is a genuine plus. Section 3.3, which explains why the algorithm works in terms of the distribution shift in h_exp, is understated but useful.\n\nThe soft spots are real but fixable. No baseline comparisons against plain SINDy, standard hybrid approaches, or the Willis--von Stosch MILP formulation that the authors cite. The per-case function libraries are not reported, so the experiments are not fully reproducible as written. The regularizer lambda_1,xi is tuned on the same validation cases, which introduces a mild circularity. And the \"first framework\" claim is contradicted by their own citation of [6], which already does simultaneous mechanistic identification and nonparametric structure selection. But the biggest issue is the one the stress test flags: all three case studies add known h deviations to an exactly correct FPM. The algorithm's ability to locate the right equation is therefore only tested when the residual contains exactly what the algorithm is looking for. In practice, FPMs have parameter error, and the residual will be a mix of structural, parametric, and epistemic error. The MILP will happily absorb parameter error as a spurious data-driven term in the wrong equation. That is not a fatal flaw in the algorithm as a residual-correction tool, but it is a fatal gap in the paper's claim about automatic uncertainty localization.\n\nWho this is for: people working in hybrid modeling, especially in chemical engineering, who want a concrete MILP-based tool for selective correction. It deserves a serious referee. I would send it out and ask for baseline comparisons, full library specifications, and at least one experiment where the FPM parameters are deliberately misspecified. If those come back, this becomes a solid paper. As it stands, the algorithm is useful and the claims need trimming, but the core idea is sound.","headline":"A clear, honest MILP-based method for column-sparse SINDy correction terms, but the location-identification claim overreaches until FPM parameter errors are actually tested.","tokens_in":17492,"tokens_out":1251,"would_cite":false,"duration_ms":19034,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["37M05","90C11","93B30"],"pacs":[],"model":"deepseek-v4-flash","headline":"One optimization pass finds the unreliable equations and patches them automatically.","keywords":["hybrid modelling","SINDy","mixed-integer linear programming","epistemic uncertainty","sparse identification","ODE systems","symbolic regression","process systems engineering"],"falsifier":"Generate synthetic data from a first-principles model whose parameters are deliberately wrong (for example, set the catalytic arylation rate constant $k_M$ 20--30% off) with no added epistemic uncertainty; if SINDybrid still returns a sparse equation-level correction instead of flagging the parameter error, then the residual-as-uncertainty premise is refuted.","tokens_in":16518,"feed_emoji":"🔧","tokens_out":8633,"duration_ms":96199,"temperature":0.7,"pith_summary":"The paper tries to establish that hybrid modelling can be automated: given a mechanistic ODE model and measurement data, SINDybrid will locate the equations whose predictions are unreliable and add sparse, symbolic data-driven corrections to exactly those equations, so an expert does not have to decide where machine learning enters the model. This matters because hybrid modelling, which combines first-principles equations with data, is normally a bespoke task requiring modelling, statistics, and domain expertise; reducing it to a single optimisation problem lowers the barrier to using hybrid models in research and industry. The paper validates the claim on three dynamic systems — a catalytic batch reaction, a continuous fermentation process, and a predator-prey oscillator — reporting correct localisation of the uncertain equations and validation $R^2$ above 0.85 under measurement noise up to 20%, as few as 2 training batches, and 5 time samples per experiment. If the claim is right, a user can hand the algorithm a known model plus data and receive both a diagnosis of which equations carry epistemic uncertainty and a symbolic correction for each.","feed_headline":"Algorithm finds unreliable equations and patches them automatically","feed_subtitle":"It adds sparse, readable corrections only where needed, from as few as 2 training batches and 5 samples.","key_machinery":"The load-bearing object is the residual matrix $h_{\\rm exp} = dx_{\\rm exp}/dt - f(x_{\\rm exp}, r_{\\rm exp})$: the difference between numerically estimated time derivatives of the measured states and the first-principles model's prediction. A correct equation leaves a residual column scattered around zero, while an equation under epistemic uncertainty leaves a column with a structured, non-centred deviation, and that contrast is what the optimisation exploits. SINDybrid approximates the residual as $h_{\\rm pred} = X_L \\cdot \\Xi$, where $X_L$ is a library of candidate functions evaluated on the state data and $\\Xi$ is a sparse coefficient matrix; the MILP adds binary variables $\\delta_j$ that activate entire columns (equations) and $\\alpha_{ij}$ that bound individual coefficients, with regularisation terms penalising both coefficient magnitude and the number of active columns. Column-wise normalisation of $X_L$ prevents large-magnitude states from dominating the selection.","core_discovery":"On its own terms, the paper's central discovery is that constructing a hybrid model is equivalent to solving one mixed-integer linear program over a residual matrix. With a known mechanistic model $f$, the epistemic uncertainty $h$ is defined through the residual $h_{\\rm exp} = dx_{\\rm exp}/dt - f(x_{\\rm exp}, r_{\\rm exp})$, and the algorithm must decide which columns of this residual are nonzero (which state equations are wrong) and which functions from a user-supplied library best describe each nonzero column. The MILP formulation uses binary variables to force whole equations on or off, a sparsity penalty on the coefficient matrix, and an absolute-error objective, so the result is a hybrid model in which untouched equations remain exactly the mechanistic ones and corrected equations receive compact symbolic expressions. The paper further argues that the localisation works because an equation under epistemic uncertainty produces a residual column whose distribution is clearly shifted from the zero-centred Gaussian scatter produced by numerical differentiation and measurement noise.","pith_inferences":["A natural extension the paper leaves implicit is to use the column-wise distribution of $h_{\\rm exp}$ as a cheap pre-screen: the paper shows uncertain columns are shifted by roughly an order of magnitude more than noise, so a statistical test on column means could warm-start or independently verify the MILP's column selection.","Because the method operates in derivative space, its success depends on how derivatives are computed; replacing finite differences with a more robust differentiation scheme could plausibly remove the one reported failure at 5 time samples in the fermentation case, where the algorithm misassigns the deviation to a correlated equation.","The paper assumes every state is measured and all parameters are correct; if either assumption fails, the residual contains more than epistemic uncertainty, so a practical workflow would alternate parameter estimation and SINDybrid rather than applying it once.","For periodic or low-variance dynamics, the plateau behaviour suggests an experiment-design rule: allocate sampling effort where trajectory diversity is highest, since extra points in already-visited regions contribute little information."],"forward_implications":["A user can supply a mechanistic model and data and receive both the location of the unreliable equations and sparse symbolic corrections for them, with no expert pre-selection of where machine learning enters.","The approach keeps the trustworthy part of the model intact: equations that are not flagged are left exactly as the mechanistic model states them.","The reported robustness bounds — up to 20% measurement noise, as few as 2 training batches, and 5 time samples per experiment — imply the method can work on realistically limited experimental campaigns.","Systems with exponential or rational terms in the mechanistic model, such as the continuous fermentation case, are more sensitive to noise and may need de-noising or denser sampling to avoid overfitting and integration failures on test data.","Periodic systems benefit most from additional time samples, because their low output variance lets many candidate functions fit few points."],"supporting_citations":[{"why":"supplies the sparse identification framework and the candidate-function library construction that SINDybrid uses to represent deviations","marker":"[11]"},{"why":"supplies the MILP formulation for simultaneous parameter identification and nonparametric-structure discrimination that the optimisation problem is built on","marker":"[6]"},{"why":"provides the kinetic model and operating-condition bounds for the catalytic batch arylation case study","marker":"[24]"},{"why":"provides the fermentation kinetic model and parameter ranges used for the continuous bioethanol case study","marker":"[25]"},{"why":"provides the predator-prey model used as the third validation case","marker":"[26]"}],"fun_headline_variants":["SINDybrid auto-patches unreliable equations in dynamic models","MILP-driven hybrid modeling: find and fix equation uncertainty","SINDybrid learns which equations to fix, then fixes them automatically","Automated hybrid models: SINDybrid pinpoints and compensates epistemic gaps","SINDybrid: from sparse data, accurate hybrid models with auto-corrections"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the mechanistic model has the right structure and correctly identified parameters, and that every state is measured, so the leftover difference between the data and the model is entirely the missing correction rather than parameter error or unmeasured dynamics.","fun_headline_variants_meta":{"raw":{"variants":["SINDybrid auto-patches unreliable equations in dynamic models","MILP-driven hybrid modeling: find and fix equation uncertainty","SINDybrid learns which equations to fix, then fixes them automatically","Automated hybrid models: SINDybrid pinpoints and compensates epistemic gaps","SINDybrid: from sparse data, accurate hybrid models with auto-corrections"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0003,"raw_usage":{"total_tokens":1763,"prompt_tokens":1010,"completion_tokens":753,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":656}},"tokens_in":626,"tokens_out":753,"duration_ms":9010,"temperature":1.0,"reasoning_tokens":656,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:47:50.633251+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate synthetic data from a first-principles model whose parameters are deliberately wrong (for example, set the catalytic arylation rate constant $k_M$ 20--30% off) with no added epistemic uncertainty; if SINDybrid still returns a sparse equation-level correction instead of flagging the parameter error, then the residual-as-uncertainty premise is refuted.","supporting_citations":[{"cited_title":"Willis and Moritz von Stosch","cited_arxiv_id":null,"evidence_quote":"supplies the MILP formulation for simultaneous parameter identification and nonparametric-structure discrimination that the optimisation problem is built on"},{"cited_title":"Shukla, Ranjit S","cited_arxiv_id":null,"evidence_quote":"provides the kinetic model and operating-condition bounds for the catalytic batch arylation case study"},{"cited_title":"Geraldo, Carlos Eduardo Vaz Rossell, Rubens Maciel Filho, and Antonio Bonomi","cited_arxiv_id":null,"evidence_quote":"provides the fermentation kinetic model and parameter ranges used for the continuous bioethanol case study"},{"cited_title":"Lotka-volterra population models","cited_arxiv_id":null,"evidence_quote":"provides the predator-prey model used as the third validation case"}],"review_version":1}