{"id":"09d8e350-80b7-4808-a458-206d86eb44a3","arxiv_id":"2506.20129","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A JAX-based automatic differentiation framework, Vibrojet, constructs Taylor-expanded rovibrational Hamiltonians in sum-of-products form and demonstrates convergent vibrational energies for H2CO and NH3.","lead":"This paper introduces Vibrojet, a Python framework that uses automatic differentiation in JAX to build Taylor series expansions of molecular kinetic and potential energy operators in internal coordinates. The tool aims to automate a step that currently requires manual symbolic algebra and system-specific code in rovibrational spectroscopy calculations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'arbitrary molecules' claim is not supported: the method needs user-supplied coordinate transformations that make Taylor expansions converge, and the only nontrivial example (NH3) relies on hand-chosen sin/cos and Morse maps (Eq. 10).","rationale":"I evaluated the strongest claim as stated: an automated framework for arbitrary molecules. The code and derivations in Sections III-IV are credible; the Eckart derivative recursion (Eqs. 7-9) is explicit, and the public repository is a real artifact. The vulnerability is not in the AD implementation but in the scope claim. The NH3 large-amplitude treatment (Eq. 10 and surrounding text) is essential to the demonstration, and it is a hand-crafted coordinate transformation. The paper nowhere says how a user would obtain such a transformation for a new molecule or detect when a naive coordinate choice will fail. This is exactly the gap the reader flagged as the weakest assumption, and I agree it is load-bearing because the abstract's 'arbitrary molecules' is the headline contribution. I do not see an internal inconsistency or a mathematical error in the presented examples. The appropriate disposition remains the reader's CONDITIONAL judgment: the framework is promising and apparently works for the shown cases, but the generality claim needs either significant additional demonstration or explicit caveats about user-supplied large-amplitude coordinates. Hence verdict_should_be UNCHANGED.","tokens_in":12874,"tokens_out":7114,"duration_ms":86208,"concrete_test":"Take a molecule with a documented large-amplitude coordinate and no known special coordinate trick, e.g., hydrogen peroxide (HO-OH torsion). Define internal coordinates as the two O-O-H angles and the dihedral angle without sin/cos or Morse substitution, use Vibrojet to build KEO/PES Taylor expansions of orders 4, 6, 8, and 10, and compare the resulting vibrational (torsional) energies against a grid-based exact-KEO calculation with the same PES. If the Taylor energies do not converge to the grid result (or to the literature torsional splittings), the 'arbitrary molecules' claim fails for generic coordinate choices. If they do converge, the concern is substantially weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract promises automated Taylor-series KEO/PES construction for 'arbitrary molecules, internal coordinate systems, and molecular frame embedding conditions.' The machinery itself is concrete and checkable: given a smooth, invertible internal-to-Cartesian map, Sections II-III turn derivatives of that map into polynomial coefficients, and Figures 2-3 verify fourth-order expansions against exact G-matrix/PES slices for the two examples. What would have to be true for the central claim is that such a map, or an automated way to find one, exists for any molecule and coordinate choice. That is where the support stops. For NH3, the paper abandons the natural coordinates and introduces Eq. (10): y6^(G)=cos rho for the KEO, y6^(V)=sin rho for the PES, plus Morse functions for stretches. These transformations are chosen ad hoc; no algorithm, criterion, or even heuristic is given for discovering similar transformations for another large-amplitude system. The energy convergence plots (Figs. 6-7) compare expansion orders D and D-2 against each other, so a uniformly wrong-but-convergent Hamiltonian in badly chosen coordinates would not be flagged. Consequently, the demonstrated result is weaker than the abstract: Taylor-mode AD can automate the bookkeeping once a user already knows a suitable coordinate transformation, but the 'arbitrary molecules' part of the central claim has no general procedure behind it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Vibrojet, a Python/JAX framework that uses Taylor-mode automatic differentiation to construct truncated Taylor series expansions of molecular rovibrational kinetic energy operators (KEOs) and potential energy surfaces (PESs) in sum-of-products form. The KEO formalism follows standard Wilson-type treatment with vibrational, rotational, and translational coordinates, and the frame embedding (Eckart or principal axes) is incorporated into the internal-to-Cartesian mapping. The implementation extends JAX's jet module with additional linear algebra primitives and custom routines for Eckart-frame differentiation. The method is demonstrated on H2CO and NH3: one-dimensional G-matrix and PES slices are compared with exact values, and vibrational energy levels are computed with varying N-mode truncation and Taylor expansion orders to show convergence. The code is publicly available on GitHub.","tokens_in":13177,"tokens_out":3327,"duration_ms":38802,"significance":"If substantiated, this is a useful methodological contribution: it replaces tedious symbolic derivation and hand-coding of rovibrational operators with a general numerical pipeline, and it extends earlier Fortran-based AD work by leveraging JAX's Taylor-mode jet and the Griewank-Utke-Walther directional-derivative interpolation for efficient N-mode expansions. The direct validation of fourth-order Taylor expansions against exact G-matrix and PES slices in Figs. 2-3 is a genuine strength, as is the open-source implementation. The paper also addresses a real technical issue by providing a derivative-friendly iterative solution of the Eckart equations. However, the advertised generality to 'arbitrary molecules' rests on unstated user expertise in choosing coordinate transformations, and the vibrational-energy validation is only self-consistency between truncation orders rather than a comparison with an unexpanded Hamiltonian or reference data.","major_comments":[{"comment":"The central claim of an automated framework for 'arbitrary molecules, internal coordinate systems, and molecular frame embedding conditions' is not fully supported. The only nontrivial example, NH3, relies on hand-selected transformations y6^(G)=cos ρ, y6^(V)=sin ρ, and Morse functions for the stretches in Eq. (10). No algorithm, criterion, or heuristic is provided for discovering such transformations for other large-amplitude coordinates, yet the convergence properties of the Taylor expansion depend critically on these choices. The paper should either supply a systematic procedure for selecting/constructing suitable coordinates or explicitly qualify the claim to 'given a suitable user-supplied coordinate transformation.'","section":"§V, Eq. (10) and §VI"},{"comment":"The vibrational-energy convergence study compares energies computed at Taylor orders D and D−2 (and N-mode orders N and N−1) within the same expanded-Hamiltonian family. This demonstrates internal consistency but not absolute accuracy: a systematically wrong but convergent Hamiltonian would not be detected. A benchmark against a calculation using the exact, unexpanded G-matrix and PES, or against established reference vibrational energies, is needed to validate both the KEO and PES expansions. This is especially important for NH3 because the PES expansion is performed around one inversion minimum in transformed coordinates, and the tunneling splitting is not checked.","section":"§V, Figs. 6-7"},{"comment":"The N-mode convergence analysis fixes the PES at 8th order when testing the KEO, and vice versa, but the combined truncation error is not assessed. The statement that 'N=4 is sufficient to achieve sub-wavenumber accuracy' refers only to differences between N and N−1 truncations, not to accuracy relative to the exact N=6 result. Reporting the maximum absolute deviation from the N=6 calculation (rather than pairwise differences) would make the convergence claim more quantitative and directly interpretable.","section":"§V, Figs. 4-5"}],"minor_comments":[{"comment":"The argument order in the deriv_list example is inconsistent: the interface is defined as deriv_list(func, x0, deriv_ind, if_taylor), but the example call is deriv_list(func, deriv_ind, q0, if_taylor=True).","section":"§IV"},{"comment":"Panel (a) of Fig. 7 is labeled 'N = 4, N = 6, N = 8' but the panel shows Taylor expansion order convergence and should read 'D = 4, D = 6, D = 8' to match the text and panel (b).","section":"Fig. 7 caption"},{"comment":"There is a typo: 'For example, in can be used to expand PESs' should read 'it can be used to expand PESs.'","section":"§IV"},{"comment":"The linear system for κ in Eq. (6) would benefit from an explicit definition of the matrix entries and the right-hand side in terms of u and λ; currently the reader must infer the arrangement from the preceding text and Eq. (5).","section":"§II, Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper is a serious piece of tool-building, not a breakthrough in theory. The core advance over the group's 2015 Fortran work is a Python/JAX implementation of Taylor-mode automatic differentiation for rovibrational KEOs, with new linear-algebra primitives (det, eig, matrix exponential) and Eckart/PAS frame support. The code is public, the interface is clean, and the convergence plots for H2CO and NH3 are internally consistent. The side-by-side comparison of fourth-order Taylor expansions against exact G-matrix and PES slices (Figs. 2–3) is real evidence that the expansion machinery works. I believe the package will be useful to people who already know what coordinates they want and just need the bookkeeping automated.\n\nThe soft spots are exactly where the reader's report and the stress-test note point. First, the 'arbitrary molecules' claim in the abstract is not supported. The method assumes a smooth, invertible internal-to-Cartesian map, and the NH3 example shows that the natural coordinates fail—you need the hand-chosen transformations in Eq. (10): cos rho for the KEO, sin rho for the PES, plus Morse functions for stretches. Those are ad hoc, and no general procedure is given for finding them. So the demonstrated claim is: given a good coordinate transformation, the Taylor expansion is automated. That is still valuable, but it is not 'arbitrary molecules'.\n\nSecond, the energy validation is self-referential. Figures 6 and 7 compare expansion orders D and D−2 against each other. A uniformly wrong-but-convergent Hamiltonian would not be caught. The paper would be stronger with a single comparison to a full unexpanded calculation or an established line list. The absence of that is a gap, not a fatal flaw.\n\nThe two free parameters (Morse exponent and the rho expansion point) are not fitted to the target energies, so the circularity burden is genuinely low. I don't share the harshest reading that this is just a rehash—the Python implementation, new primitives, and N-mode coordinate transformations are concrete additions.\n\nWho is this for? Spectroscopists who build variational codes and want a maintainable, flexible KEO constructor. I'd cite it if I were setting up a new molecule. The paper deserves peer review; it's technically sound in what it does, but the authors should be pushed to soften the generality claim and add one external benchmark for the energies. That's a revise-and-resubmit, not a rejection.","headline":"Solid, well-documented tool paper that converts clever AD machinery into a usable Python package, but it oversells 'arbitrary molecules' and needs an external energy benchmark before the convergence claims carry full weight.","tokens_in":13720,"tokens_out":1033,"would_cite":true,"duration_ms":13441,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single user-defined coordinate mapping is enough for the framework to produce Taylor-series kinetic and potential operators that yield converged vibrational energies for formaldehyde and ammonia.","keywords":["automatic differentiation","Taylor series","kinetic energy operator","rovibrational Hamiltonian","sum-of-products representation","Eckart frame","vibrational energy levels","large-amplitude motion"],"falsifier":"Compute the ammonia inversion tunneling splitting with an eighth-order Taylor expansion of the kinetic and potential operators in the transformed coordinates of Eq. (10) and compare with a converged grid calculation; if the splitting or the lowest levels deviate beyond the sub-wavenumber convergence claimed in the paper, the single-expansion strategy for large-amplitude coordinates is falsified. A stronger test would be to run the framework on a second double-well molecule using only raw internal coordinates—no Morse or trigonometric transformations—and check whether the energy differences continue to shrink with increasing Taylor order.","tokens_in":12683,"feed_emoji":"⚛️","tokens_out":9448,"duration_ms":94945,"temperature":0.7,"pith_summary":"This paper presents an automated framework for constructing the kinetic and potential energy operators that appear in a molecule's rovibrational Hamiltonian. The framework takes a user-supplied function that maps internal coordinates (bond lengths, angles, and similar) to the Cartesian positions of the atoms, plus an optional condition for orienting the molecular frame, and returns the operators as truncated Taylor series in a sum-of-products form. That form makes subsequent quantum calculations efficient because matrix elements factor into products of one-dimensional integrals. The authors demonstrate the approach on formaldehyde and ammonia, showing that computed vibrational energy levels converge as the Taylor order and the N-mode coupling order are increased, with fourth-order coupling giving sub-wavenumber accuracy for the lowest 100 states.","feed_headline":"Taylor-mode AD automates vibration Hamiltonians","feed_subtitle":"Users supply one coordinate mapping; kinetic and potential operators follow, and vibrational energies converge.","key_machinery":"The load-bearing mechanism is Taylor-mode automatic differentiation: arithmetic operations, intrinsic functions, and linear algebra routines are redefined to act on truncated Taylor polynomials, so derivatives of any composition follow from the chain rule (Faà di Bruno) and the Leibniz product rule. For a molecule, the input is the map $r_i = f_i(\\xi)$ from internal coordinates $\\xi$ to Cartesian atomic positions, and the output is the contravariant metric tensor $G$ and pseudopotential $U$ defining the kinetic operator. Frame embeddings are handled by parametrizing the orientation rotation as $d = e^{-\\kappa}$ with skew-symmetric $\\kappa$; differentiating the Eckart conditions $l$ times yields a linear system for the derivative $\\kappa_l$ whose matrix is the same at every order, so its inverse is computed once and reused. Multivariate partial derivatives are obtained by propagating families of univariate Taylor series and interpolating, which avoids building full high-dimensional derivative tensors. The final expansions are organized in an N-mode representation that keeps only terms coupling up to $N$ coordinates at a time.","core_discovery":"The central discovery is that Taylor-mode automatic differentiation—propagating truncated Taylor polynomials through every operation in the coordinate transformation—can replace hand derivation of kinetic energy operators. The implementation extends the Taylor-mode differentiation module of a Python automatic-differentiation library with the missing linear algebra primitives (matrix determinant, inverse, exponential, and eigendecomposition) and with routines for differentiating the Eckart-frame rotation matrix, parametrized as $d = e^{-\\kappa}$ with skew-symmetric $\\kappa$. Given a user-supplied internal-to-Cartesian function, the framework computes the kinetic metric tensor, the pseudopotential, and the potential energy surface at an expansion point and returns Taylor coefficients for any requested multi-indices. For ammonia, the paper shows that expanding in specially transformed coordinates—Morse functions for bond stretches and $\\sin \\rho$ for the inversion angle—lets a single Taylor series reproduce the double-well potential and converge much faster than expansion in raw internal coordinates. Variational calculations for both molecules show the resulting vibrational energies converging with Taylor order and with N-mode coupling order.","pith_inferences":["As an extension beyond the paper, the framework could automatically produce dipole moment surface expansions and therefore vibrational transition intensities, since the paper notes its derivative routine is not specific to the kinetic operator.","A natural stress test is to apply the framework to a molecule with two coupled large-amplitude coordinates without hand-picked transformations; the convergence rate would show how much of the ammonia success comes from the framework itself and how much from the chosen coordinates.","The paper's recursive Eckart differentiation sidesteps eigenvalue-based frame solvers, which the authors note fail for highly symmetric molecules such as methane; testing on such symmetric cases would clarify the practical range of the 'arbitrary molecules' claim."],"forward_implications":["For any molecule with a smooth, invertible internal-coordinate mapping, the same code produces a sum-of-products kinetic and potential operator without system-specific symbolic derivation.","Because matrix elements of the resulting operators factor into one-dimensional integrals, variational vibrational calculations can be driven directly by the Taylor coefficients.","Large-amplitude motions such as ammonia inversion can be described by a single Taylor expansion, without a grid over the large-amplitude coordinate, provided suitable transformed coordinates are chosen.","The Eckart frame, which is usually too algebraically complex to hand-derive, becomes routine because its rotation matrix is differentiated numerically through the exponential parametrization.","The same expansion machinery applies to any multivariate function, so potential energy surfaces and dipole moment surfaces can be expanded alongside the kinetic operator."],"supporting_citations":[{"why":"Establishes the automatic-differentiation route to Eckart-frame kinetic energy operator power series that the present framework reimplements and extends.","marker":"[17]"},{"why":"Supplies the univariate-Taylor propagation and interpolation method used to compute multivariate partial derivatives.","marker":"[21]"},{"why":"Provides the Python library with the Taylor-mode differentiation module at the core of the implementation.","marker":"[33]"},{"why":"Gives Taylor-series formulas for parametric eigenvalue problems used to differentiate eigenvalue-based primitives.","marker":"[34]"},{"why":"Shows that rectilinear-coordinate Taylor expansions around a planar geometry can describe ammonia inversion, supporting the single-expansion strategy.","marker":"[37]"},{"why":"Supplies the formaldehyde potential energy surface used in the convergence demonstration.","marker":"[38]"},{"why":"Supplies the ammonia potential energy surface used in the convergence demonstration.","marker":"[39]"}],"fun_headline_variants":["Taylor-mode AD automatically generates rovibrational Hamiltonians","Replace hand derivation with Taylor-mode AD for Hamiltonians","JAX Taylor-mode AD constructs molecular Hamiltonian operators","Automated Taylor expansion of rovibrational operators","Taylor-mode AD: from coordinates to Hamiltonian automatically"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework presumes the user can supply a smooth, invertible mapping from internal coordinates to Cartesian positions, and presumes that for large-amplitude vibrations a coordinate transformation exists that makes a low-order Taylor series faithfully represent the operators; the paper gives no general recipe for finding that transformation.","fun_headline_variants_meta":{"raw":{"variants":["Taylor-mode AD automatically generates rovibrational Hamiltonians","Replace hand derivation with Taylor-mode AD for Hamiltonians","JAX Taylor-mode AD constructs molecular Hamiltonian operators","Automated Taylor expansion of rovibrational operators","Taylor-mode AD: from coordinates to Hamiltonian automatically"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00021,"raw_usage":{"total_tokens":1352,"prompt_tokens":826,"completion_tokens":526,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":442,"completion_tokens_details":{"reasoning_tokens":453}},"tokens_in":442,"tokens_out":526,"duration_ms":19142,"temperature":1.0,"reasoning_tokens":453,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:55:25.162483+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the ammonia inversion tunneling splitting with an eighth-order Taylor expansion of the kinetic and potential operators in the transformed coordinates of Eq. (10) and compare with a converged grid calculation; if the splitting or the lowest levels deviate beyond the sub-wavenumber convergence claimed in the paper, the single-expansion strategy for large-amplitude coordinates is falsified. A stronger test would be to run the framework on a second double-well molecule using only raw internal coordinates—no Morse or trigonometric transformations—and check whether the energy differences continue to shrink with increasing Taylor order.","supporting_citations":[{"cited_title":"Yachmenev and S","cited_arxiv_id":null,"evidence_quote":"Establishes the automatic-differentiation route to Eckart-frame kinetic energy operator power series that the present framework reimplements and extends."},{"cited_title":"Griewank, J","cited_arxiv_id":null,"evidence_quote":"Supplies the univariate-Taylor propagation and interpolation method used to compute multivariate partial derivatives."},{"cited_title":"Solving the Parametric Eigenvalue Problem by Taylor Series and Chebyshev Expansion","cited_arxiv_id":"2302.03661","evidence_quote":"Gives Taylor-series formulas for parametric eigenvalue problems used to differentiate eigenvalue-based primitives."},{"cited_title":"Neff and G","cited_arxiv_id":null,"evidence_quote":"Shows that rectilinear-coordinate Taylor expansions around a planar geometry can describe ammonia inversion, supporting the single-expansion strategy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the formaldehyde potential energy surface used in the convergence demonstration."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ammonia potential energy surface used in the convergence demonstration."}],"review_version":1}