{"id":"b341d5d8-c377-4ae2-9900-02e49d7a2a79","arxiv_id":"1908.11001","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper estimates a template FTIR spectrum and decomposes post-treatment spectral changes into a fixed modification pattern and a per-treatment strength, applied to plasma-treated CFRP.","lead":"This paper builds a two-step statistical method that separates a surface treatment's effect on FTIR spectroscopy signals from measurement noise. It applies the method to plasma-treated carbon fiber composites and estimates how the treatment effect changes with plasma nozzle height.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The recovered modification pattern g is not identifiable: any unit vector in span{tilde g, 1, x0} gives the same post-treatment fit after reparameterizing delta and the alignment factors, and the SVD sign is arbitrary, so the chemical 'bonds created' claim is underdetermined.","rationale":"The reader's weakest assumption was the single-pattern rank-one structure (Eq. 3). My concern is distinct and more direct: even if that rank-one structure is exactly true, g itself is not identified from the data because of the free alignment factors and the arbitrary rotation/sign choices. The paper's own Sec. II-E concedes the rotation ambiguity, and the L1 heuristic does not resolve it; the reported selection is one of many equally fitting options, and the sign of g is never constrained. The simulation does not test this ambiguity because it starts from a known g and only checks shape recovery after the same arbitrary post-processing, so it cannot validate the identifiability of the final chemical interpretation. This undermines the strongest claim in the abstract and Sec. IV, although the estimation of the template and the qualitative decreasing trend in delta are less affected. The paper could be made acceptable by adding an explicit identifiability analysis, imposing physically motivated sign/interpretability constraints, or substantially weakening the chemical-bond conclusions; hence a conditional verdict is appropriate rather than outright rejection.","tokens_in":13639,"tokens_out":11097,"duration_ms":118768,"concrete_test":"Re-run Step 2 on the real post-treatment spectra with the SVD sign flipped: replace tilde g by -tilde g and delta by -delta in problem (10), repeat the Sec. II-E L1 selection over (theta, phi), and compare the optimized g, the residual sum of squares, and the resulting wavenumber peaks. If an equally fitting optimum has peaks of opposite sign in the 2700-3000 cm^-1 region, the 'bonds created' interpretation is an artifact of sign convention. Additionally, enumerate all local minima of G(theta, phi) in Fig. 15 and map each optimized g to the IR absorption table; if different equally fitting minima produce different chemical-bond lists, the conclusion is not identified. A unique, sign- and rotation-invariant bond list would settle the concern; otherwise the chemical claims must be dropped or constrained by independent measurements such as XPS.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that g identifies specific chemical bonds created by plasma is not supported even under the stated model. In the post-treatment model (Eq. 3), only the products delta_i * g enter the mean, and Step 2 (Sec. II-C/D) leaves the per-sample alignment factors c_i, d_i free. As the paper itself acknowledges in Sec. II-E, g can be any linear combination of tilde g, the constant vector 1, and the template x0 without changing the fit. The subsequent L1 minimization over (theta, phi) is a non-convex search over equally fitting representations, not an identification of g from the data. Moreover, the SVD in problem (10) determines tilde g only up to sign; flipping tilde g and delta simultaneously leaves the objective unchanged. Therefore, whether the FTIR peaks in Sec. IV correspond to bonds being created or destroyed is a sign convention, not a data-driven inference. The reported choice is also internally inconsistent: the text says phi near 0 or pi is preferred, but the reported real-data optimum is phi* = 0.5053 (Sec. IV-B, Fig. 15). Because the same spectra can be represented with a g of opposite sign or with different mixing of x0 and 1, the chemical-bond list and even the statement that g 'matches existing engineering knowledge' are not robust consequences of the experiment.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-step statistical procedure for analyzing FTIR spectra collected before and after a surface treatment. In Step 1, a template spectrum is estimated from pre-treatment signals by solving a constrained least-squares problem that accounts for multiplicative and offset uncertainties. In Step 2, the post-treatment deviation from the template is decomposed into a modification pattern g and a per-sample effect magnitude δ through an alternating coordinate-descent algorithm and a rank-one SVD approximation. The method is validated on synthetic data and applied to a plasma-treatment case study on CFRP coupons, where the estimated δ is used to infer the effective plasma-height range and the estimated g is used to identify chemical bonds created by plasma exposure.","tokens_in":13983,"tokens_out":5713,"duration_ms":61374,"significance":"The paper addresses a real and practically important problem: quantifying treatment effects on spectroscopic measurements that are corrupted by multiplicative errors and baseline shifts. The statistical model in Eqs. (1) and (3) is reasonable, and the two-step estimation scheme is computationally tractable. If the identifiability issues were resolved, the δ estimate could provide useful engineering guidance on treatment-strength thresholds, and the preprocessing step could be valuable for handheld spectral metrology. However, as presented, the central claim that g identifies specific chemical bonds created by plasma exposure is not supported by the data because g is identifiable only up to a subspace and up to sign, and the final g is selected by a non-convex post-processing criterion among equally fitting representations. The simulation validation is also only qualitative, comparing shapes rather than magnitudes. The work is therefore best viewed as a method for estimating an equivalence class of modification patterns and a relative effect trend, not as a tool for chemical identification.","major_comments":[{"comment":"The pattern g is not identifiable from the data: any g' = cosφ·g̃ + (1/√p)cosθ·sinφ·1 + sinθ·sinφ·x0 gives exactly the same post-treatment fit after reparameterizing δ and the alignment factors c_i, d_i. The L1 minimization over (θ, φ) selects one member of this equivalence family based on an interpretability heuristic, not on information in the spectra. Consequently, the chemical-bond claims in Section IV (O=C=O, N=C=O, N=N=N, N=C=N, N=C=S, C–H, O–H, N–H are created) are not consequences of the data alone. Please either provide additional assumptions that identify g, or explicitly reframe the claims as describing a representative of an equivalence class rather than a uniquely recovered pattern.","section":"Section II-E, Eqs. (8) and (10)"},{"comment":"The reported selection of (θ*, φ*) is internally inconsistent with the stated preference. The text says the authors select a local optimum with φ≈0 or φ≈π so that g is mainly determined by g̃, but the real-data optimum is φ*=0.5053 and the simulation optimum is φ*=0.8741, neither of which is close to 0 or π. Additionally, the SVD solution to problem (10) determines g̃ only up to sign; flipping g̃ and δ simultaneously leaves the objective unchanged. The reported sign of δ (positive for small plasma heights) and the interpretation that bonds are 'created' rather than 'destroyed' therefore depend on an unstated sign convention. Please state the convention explicitly and discuss whether the conclusions survive the sign flip.","section":"Section IV-B, Figs. 15 and 16"},{"comment":"The definition H_i = x0,i x0,i^T (x0,i^T x0,i)^{-1} [I - 11^T/p] is not the projection matrix onto span{x0,i, 1} that the least-squares solution actually requires. The correct projection is [x0,i, 1]([x0,i, 1]^T[x0,i, 1])^{-1}[x0,i, 1]^T. As written, the reduction to f(x0) = Σ ||H_i x0 - x0||^2 and the resulting eigenproblem (6) do not follow. Please correct the formula or show the intermediate algebra in detail.","section":"Section II-B, around Eq. (6)"},{"comment":"The simulation validation is only qualitative. The text acknowledges that the estimated δ is 'significantly different' from the ground truth and that the magnitude of the estimated g is 'significantly different' from the true g, with only the shapes being similar. Since the paper claims to 'quantify' the treatment effect, please report quantitative error metrics such as correlation, normalized RMSE, or bias after accounting for the non-identifiability, and clarify which aspects of δ and g are recoverable and which are not.","section":"Section III, Figs. 8 and 9"},{"comment":"The block-wise coordinate descent algorithm is not shown to converge to a global optimum of the non-convex problem (8), and no convergence criterion or initialization strategy is specified beyond δ←0 and arbitrary g. Because the objective landscape in Fig. 10 is highly non-convex, the final estimate may depend on initialization. Please provide a convergence analysis or, failing that, a sensitivity analysis over multiple random initializations.","section":"Section II-D, Algorithm 1"}],"minor_comments":[{"comment":"The phrase 'the objective of (3)' appears to refer to the wrong equation; it should likely be 'the objective of (5)' or 'the objective in (5)'.","section":"Section II-B, text after Eq. (5)"},{"comment":"The notation (c1⊤) ⊙ X is not clearly defined; please state the dimensions of c1, d1, and X and explain the elementwise product operation in a way that a reader can verify Eq. (9) without guessing.","section":"Section II-D, Eq. (9)"},{"comment":"There are several typographical errors, including 'Ramen spectra' (should be 'Raman spectra') and 'outfit error' (should likely be 'offset error'). These should be corrected in revision.","section":"Throughout"},{"comment":"The elimination of the six post-exposure signals from the 2 mm and 4 mm coupons is described in one sentence; please provide more detail on the criterion used to decide that these signals should be excluded, since this choice directly affects the estimated δ and g.","section":"Section IV-A"}],"recommendation":"major_revision","confidential_remarks":"The skeptical reading of the non-identifiability of g is warranted and should be addressed head-on. The paper's most striking claim—that the recovered g identifies specific chemical bonds created by plasma exposure—is not supported under the stated model, and the text itself acknowledges the equivalence class in Section II-E. I would advise the editor that the paper can be made publishable if the authors substantially reframe the contributions: (i) present the method as estimating a template and a rank-one treatment effect up to a well-defined equivalence class and sign convention, (ii) replace the chemical 'bonds created' language with a statement about a representative modification pattern, and (iii) provide a quantitative simulation assessment. The engineering insight from δ (effective plasma height range) is more defensible and could carry the application value."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid two-step statistical method for quantifying treatment effects on FTIR spectra that suffer from offset and multiplicative noise. Step 1's MLE-based template estimation that lets noise scale with signal magnitude is a genuine improvement over standard MSC, and Step 2's rank-one decomposition of the post-treatment change into a pattern g and a strength vector δ is a sensible way to summarize treatment effects. The simulation shows the template and δ trend are recovered well, and the plasma-height case study is a relevant engineering application. If the goal is a process-control tool for setting plasma parameters, the δ vector alone gives a plausible threshold around 10 mm.\n\nThe weaknesses are concentrated in the interpretability of g, and the stress-test note is right: g is only identified up to the span of the estimated component, the constant vector, and the template. The L1 search over (θ, φ) is a non-convex heuristic, not a data-driven identification—the paper essentially admits this in Section II-E. Worse, the SVD gives g only up to sign, so the claim that specific bonds are 'created' is a sign convention, not an inference. The reported optimum φ* = 0.5053 also sits awkwardly with the stated preference for φ near 0 or π. The chemical-bond list in Section IV should be softened or defended with independent evidence.\n\nOther soft spots: the derivation of H_i in Step 1 is garbled, the constrained eigen solution is cited rather than shown, and the simulation validates shape but explicitly not magnitude. The post-hoc exclusion of the two strongest treatment groups, while explained by carbonization, further narrows the generality of the real-data conclusions. The single-pattern assumption is acknowledged in the conclusion, which is fair.\n\nNone of this sinks the method for its primary intended use—estimating the relative magnitude of treatment effect—and the δ estimates are not undermined by the g ambiguity. But the paper's own claim to have 'identified affected chemical bonds' is not supported. A serious referee should ask for an identifiability analysis, a treatment of sign ambiguity, and either a validated criterion for choosing g or a more cautious discussion of what g can and cannot mean.\n\nI would send this to peer review. It is a worthwhile statistical framework with a real application, and the flaws are fixable in revision. I would not cite the chemical-bond identification, but I might cite the template estimation or the rank-one treatment decomposition.","headline":"A useful engineering statistics paper whose chemical-interpretation claim outruns the identifiability of the model; worth refereeing after the authors confront the sign and basis ambiguity of g.","tokens_in":14444,"tokens_out":1376,"would_cite":true,"duration_ms":17266,"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":"A two-step statistical procedure estimates a pre-treatment FTIR template and decomposes the post-treatment change into a single modification pattern and a per-sample strength, giving engineers a quantitative readout of plasma treatment…","keywords":["FTIR spectroscopy","treatment effect","multiplicative error","offset shift","template estimation","plasma surface treatment","spectral decomposition","coordinate descent"],"falsifier":"Fit the model separately to low-, medium-, and high-dose post-treatment spectra and compare the estimated patterns: if g changes shape across dose ranges, or if a model allowing two patterns g_1 and g_2 reduces the held-out residual variance substantially compared with the single-g model, then the single-pattern assumption fails and δ and g do not directly represent the underlying bond changes.","tokens_in":13476,"feed_emoji":"🧪","tokens_out":4069,"duration_ms":45522,"temperature":0.7,"pith_summary":"The paper proposes a two-step statistical procedure for quantifying how a surface treatment changes FTIR spectra when measurements are corrupted by offset shift and multiplicative error. Step one estimates a template pre-treatment spectrum by aligning repeated pre-treatment measurements; step two represents the post-treatment change as one fixed pattern g multiplied by a per-sample strength δ, estimated by a constrained rank-one fit. The authors validate the approach on simulated data and on a plasma-exposure experiment on carbon-fiber-reinforced polymer, where δ reveals that plasma heights above roughly 10 mm have little effect and g points to specific chemical bonds created by plasma. The central claim is that this decomposition turns noisy handheld-spectrometer data into interpretable information about the treatment effect and the underlying chemistry.","feed_headline":"Plasma chemistry read from noisy FTIR spectra in two steps","feed_subtitle":"A template-plus-decomposition fit separates measurement drift from treatment change and finds a 10 mm plasma-height cutoff.","key_machinery":"The central object is the rank-one treatment-effect model x0 + δ_i g, where x0 is a normalized template spectrum, g is the pattern of modification satisfying g^T x0 = 0, g^T 1 = 0, and ||g|| = 1, and δ_i is the scalar effect strength. The machinery consists of two optimization stages: a constrained eigenvalue problem for the template x0, and an alternating coordinate-descent algorithm for (δ, g) whose inner step is a rank-one singular value decomposition of the corrected post-treatment residuals; a final ℓ1-minimizing rotation g = g̃ cos φ + (1/√p) cos θ sin φ + x0 sin θ sin φ makes the pattern interpretable. This machinery carries the argument because it reduces a high-dimensional spectral comparison to two small identifiable quantities, one describing the shape of the treatment effect and one describing its magnitude at each treatment level.","core_discovery":"The paper claims that, despite offset shift and multiplicative error, the pre-treatment FTIR spectra share a common template x0 that can be recovered by a constrained optimization problem, and that every post-treatment spectrum can be written as the template plus δ_i g, where g is a single modification pattern and δ_i is the treatment strength. Step one minimizes the sum of squared distances between aligned pre-treatment spectra and the template, reducing to a constrained eigenvalue problem. Step two, after projecting the corrected post-treatment residuals onto the space orthogonal to the template and the constant vector, obtains g as the leading right singular vector and δ as the corresponding left singular vector times the singular value. The paper further shows that a re-rotation of g within the space spanned by the template, the constant vector, and the initial estimate g̃ can make g sparse and hence interpretable. The authors assert that in the CFRP plasma case study the estimated δ marks the range of effective plasma height and the estimated g matches known chemical bond changes, including O=C=O, N=C=O, N=C=N, C–H, O–H, and N–H.","pith_inferences":["Extension: if the single-pattern assumption is only approximately true, δ can still serve as a treatment-ranking score but should not be read as a direct measure of chemical concentration; a natural extension is to allow several patterns g_1, g_2, ... with treatment-dependent weights.","Extension: the same template-and-decomposition pipeline could test whether treatment effects are additive across repeated or sequential exposures by checking whether total δ adds linearly over passes.","Extension: the ℓ1 re-rotation step is a heuristic for interpretability; an objective comparison against a quantitative reference library, rather than visual inspection of an absorption table, would sharpen the bond assignments.","Extension: since δ and g are identified only up to scaling and rotation within the span of the template and constant vector, cross-experiment comparisons would need external anchoring, for example to X-ray photoelectron spectroscopy measurements of surface composition."],"forward_implications":["The estimated vector of effects δ can be used to identify the effective range of plasma height: in the case study, heights beyond about 10 mm produce little further chemical change.","The pattern of modification g provides a map of treatment-induced chemical changes across frequency bands and identifies plausible bonds created by plasma exposure, including O=C=O, N=C=O, C–H, O–H, and N–H.","Because offset and scale are treated as nuisance parameters, the framework extends to other spectroscopic measurements with similar uncertainties, including UV-Vis, XRD, and Raman spectroscopy.","The method replaces visual, subjective inspection of FTIR signals with a quantitative decomposition that can be used to guide plasma parameter selection in composite surface preparation.","Simulation results indicate that both the template and the modification pattern can be recovered in shape and that δ trends match the simulated ground truth."],"supporting_citations":[{"why":"Provides the extended multiplicative signal correction model that also handles offset and multiplicative error, serving as the baseline the proposed template model improves on.","marker":"[5]"},{"why":"Reviews preprocessing procedures for FTIR and Raman spectra and motivates the need for a systematic template-estimation step.","marker":"[15]"},{"why":"Supplies a heteroscedastic noise model for Raman spectra in which noise depends on signal level, informing the noise structure assumed for FTIR measurements.","marker":"[19]"},{"why":"Provides the constrained eigenvalue result used to solve the template-extraction optimization in Step 1.","marker":"[20]"},{"why":"Gives the singular value decomposition theorem used to obtain δ and g as the rank-one approximation of the corrected post-treatment residual matrix.","marker":"[21]"},{"why":"Supplies the infrared absorption table used to assign the discovered spectral peaks to specific chemical bonds.","marker":"[23]"},{"why":"Reports that plasma grafting increases the oxygen concentration on carbon-fiber surfaces, corroborating the chemical interpretation of the recovered pattern g.","marker":"[7]"},{"why":"Documents that plasma-modified carbon fibers show increased oxygen and nitrogen concentration, supporting the claim that the identified bonds are created by plasma exposure.","marker":"[6]"}],"fun_headline_variants":["Two-step FTIR fit isolates plasma effect","Template decomposition separates drift from treatment","FTIR noise-proofed: plasma effect mapped","Align, then decompose: plasma effect from messy FTIR","Plasma height range read from FTIR two-step fit"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire decomposition rests on assuming that every post-treatment spectrum is the pre-treatment template plus one fixed shape of change g multiplied by a number δ_i, so the chemistry of the treatment is the same at every strength and only its magnitude changes.","fun_headline_variants_meta":{"raw":{"variants":["Two-step FTIR fit isolates plasma effect","Template decomposition separates drift from treatment","FTIR noise-proofed: plasma effect mapped","Align, then decompose: plasma effect from messy FTIR","Plasma height range read from FTIR two-step fit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000215,"raw_usage":{"total_tokens":1447,"prompt_tokens":985,"completion_tokens":462,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":391}},"tokens_in":601,"tokens_out":462,"duration_ms":5795,"temperature":1.0,"reasoning_tokens":391,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:28:16.843730+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit the model separately to low-, medium-, and high-dose post-treatment spectra and compare the estimated patterns: if g changes shape across dose ranges, or if a model allowing two patterns g_1 and g_2 reduces the held-out residual variance substantially compared with the single-g model, then the single-pattern assumption fails and δ and g do not directly represent the underlying bond changes.","supporting_citations":[{"cited_title":"Extended multiplicative signal correction in vibrational spectroscopy, a tutorial,","cited_arxiv_id":null,"evidence_quote":"Provides the extended multiplicative signal correction model that also handles offset and multiplicative error, serving as the baseline the proposed template model improves on."},{"cited_title":"Quantitative application of in situ ATR-FTIR and raman spectroscopy in crystallization processes,","cited_arxiv_id":null,"evidence_quote":"Reviews preprocessing procedures for FTIR and Raman spectra and motivates the need for a systematic template-estimation step."},{"cited_title":"Generalized wavelet shrinkage of inline raman spectroscopy for quality monitoring of continuous manufacturing of carbon nanotube buckypaper,","cited_arxiv_id":null,"evidence_quote":"Supplies a heteroscedastic noise model for Raman spectra in which noise depends on signal level, informing the noise structure assumed for FTIR measurements."},{"cited_title":"Some modified matrix eigenvalue problems,","cited_arxiv_id":null,"evidence_quote":"Provides the constrained eigenvalue result used to solve the template-extraction optimization in Step 1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the singular value decomposition theorem used to obtain δ and g as the rank-one approximation of the corrected post-treatment residual matrix."},{"cited_title":"spectra data online (FTIR/Ramen search)","cited_arxiv_id":null,"evidence_quote":"Supplies the infrared absorption table used to assign the discovered spectral peaks to specific chemical bonds."},{"cited_title":"Carbon fiber surfaces and composite interphases,","cited_arxiv_id":null,"evidence_quote":"Reports that plasma grafting increases the oxygen concentration on carbon-fiber surfaces, corroborating the chemical interpretation of the recovered pattern g."},{"cited_title":"Evaluation of fiber surface treatment on the interfacial behavior of carbon fiber-reinforced polypropylene composites,","cited_arxiv_id":null,"evidence_quote":"Documents that plasma-modified carbon fibers show increased oxygen and nitrogen concentration, supporting the claim that the identified bonds are created by plasma exposure."}],"review_version":1}