{"id":"deb6ba15-f8b8-4b70-b511-bbc80525e5ec","arxiv_id":"2607.23790","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Clamped B-spline shaping of cylindrical coordinates plus differential flatness cuts median ΔV by ~22–45% versus high-order hodographic shaping on fine-grid low-thrust rendezvous campaigns.","lead":"A new shape-based design method parameterizes low-thrust interplanetary paths with clamped B-splines and recovers thrust from the path algebraically. On large Earth-to-Mars/Mercury/asteroid/comet grids it beats a standard hodographic benchmark on fuel cost and often on speed.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The ΔV-superiority claim rests on an asymmetric comparison: B-spline NLPs get a gradient-based interior-point solver with a physics-informed MEE initial guess, while the hodographic benchmark gets derivative-free Nelder–Mead capped at 1,000 evaluations — so the measured gap may partly reflect optimi","rationale":"The reader identified exactly this concern as the weakest assumption, and my independent read lands in the same place: §4.1 vs §4.2 is the least secure load-bearing point in the paper. I would add two sharpenings. First, the 1,000-evaluation cap on Nelder–Mead for a 6-D problem is a specific, quantifiable reason to suspect the benchmark is under-optimized, not just \"a different optimizer.\" Second, there is a dimensionality confound layered on top (18 vs 6 free variables), so even a matched-optimizer ablation would leave the interpretation \"richer shape space wins\" available — though that is still a legitimate, if weaker, contribution. I checked whether other concerns outrank this one: (a) the \"finite solutions retained\" selection filter could bias statistics if failure rates differ across methods, but failure counts are not reported and the claim is explicitly hedged to retained solutions, so it is secondary; (b) the unconstrained-acceleration/no-mass modeling limits practical relevance but is openly acknowledged and does not threaten the internal comparison; (c) the ε-smoothing is deactivated for reported costs and quadrature is harmonized in postprocessing, which are genuine strengths. The method itself (flatness + clamped B-spline reduction to a static NLP) is mathematically sound as presented, and code is available, which makes the proposed ablation directly executable. Because the reader already conditioned the verdict on exactly this fix, and my analysis confirms rather than escalates the concern, the verdict should remain CONDITIONAL: the matched-optimizer ablation is the single experiment that determines whether the 22–45% median improvements are a property of B-spline shaping or of the numerical machinery around it.","tokens_in":23801,"tokens_out":1755,"duration_ms":76016,"concrete_test":"Re-solve the hodographic benchmark's six-coefficient NLP (§4.2, Table 1 bases) with IPOPT through CasADi using analytic gradients of the same smoothed ΔV objective (Eq. 23), the same continuation/warm-start rules, and no 1,000-evaluation cap, on the four fine grids. Recompute Table 3's hodographic rows. If the hodographic mean/median/min statistics move materially toward the B-spline values (e.g., the median gap for Mercury or 1989 ML shrinks by more than ~50%), the ΔV advantage is substantially optimizer-driven and the headline percentages must be qualified; if the statistics barely move, the shape-space attribution stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is a head-to-head empirical statement: across fine grids, B-spline configurations beat high-order hodographic shaping on mean/median/min ΔV (Table 3) and on runtime (Table 4). For that claim to attribute the advantage to the B-spline shape space, the two methods must be given comparably competent numerical treatment. They are not. Per §4.1, the B-spline problem (20) is solved with IPOPT via CasADi — exact analytic gradients, a bounded NLP, and an initial guess fitted to a quintic modified-equinoctial-element profile (Eq. 22), which is itself a smooth, dynamically plausible interpolating trajectory. Per §4.2, the hodographic six-free-coefficient problem is solved with SciPy Nelder–Mead, an initial simplex built from 10^-2 coordinate perturbations, hard-capped at 1,000 objective evaluations, with 10^-6 tolerances. Nelder–Mead on a 6-D nonconvex objective with a 1,000-evaluation budget is frequently under-converged; continuation warm-starting helps only along grid directions where a usable predecessor exists. Meanwhile the B-spline side has 18 decision variables (nc=10 → 3(10−4)) versus the hodographic's 6, so the comparison also conflates shape-space dimension with shape-space type. The paper partially mitigates this by re-evaluating both methods' ΔV with the same high-accuracy Gauss–Legendre quadrature in postprocessing (good — that removes the quadrature confound), and by hedging the claim to \"tested fine-grid campaigns and among the finite solutions retained.\" But nothing in the manuscript isolates the optimizer/initialization contribution. If a gradient-based solve of the hodographic problem (its integrand is analytic and differentiable, so gradients are cheap) closed even half of the reported 22–45% median gap, the central quantitative claim would shrink to \"a richer parameterization with a better optimizer does better,\" which is substantially weaker than the framing. This is a correctness-risk concern about attribution, not an internal incons","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript proposes a shape-based method for fixed-time, minimum-ΔV low-thrust interplanetary rendezvous in which the heliocentric cylindrical coordinates are parameterized by clamped B-splines. Exploiting the differential flatness of the fully actuated cylindrical two-body dynamics, the control acceleration is recovered algebraically from the shaped trajectory and its first two derivatives (Eq. 3), and the clamped knot structure enforces endpoint position and velocity conditions analytically (Eq. 14). The continuous problem (1) is thereby reduced to a bounded finite-dimensional NLP (20) in the 3(nc−4) free interior control points, with quadrature needed only for the ΔV objective. The method is benchmarked against high-order hodographic shaping [27] over large fine-grid campaigns (≈1.3×10^5–1.5×10^5 attempts per method per target) for Earth-to-Mars, Mercury, 1989 ML, and Tempel 1 transfers. The headline claim (Table 3) is that all tested B-spline configurations reduce mean, median, and minimum ΔV relative to the hodographic benchmark among the finite solutions retained, with a quintic 10-control-point spline cutting median ΔV by ~22–45% at ~30–33% lower per-attempt runtime (Table 4).","tokens_in":24216,"tokens_out":2806,"duration_ms":124156,"significance":"If the comparative claim holds under matched numerics, the contribution is a genuinely useful preliminary-design tool: a single, target-independent architecture (standard clamped B-splines of the position coordinates) that requires no hand-crafted per-target basis sets, satisfies dynamics and boundary conditions by construction, and is fast enough for full porkchop-scale screening. The mathematics in §3 (flatness of cylindrical dynamics, clamped endpoint encoding, derivative recursions) is standard and correctly applied. The paper also deserves explicit credit for several practice-level strengths: very large campaigns on identical grids for both methods; re-evaluation of both methods' ΔV with the same high-accuracy Gauss–Legendre quadrature in postprocessing, which cleanly removes the quadrature confound; public source code (FASTTRANSFER repository), making the comparison reproducible in principle; honest hedging of the claim to \"tested fine-grid campaigns and among the finite solutions retained\"; and a candid limitations statement (unconstrained u, no mass model, local optimality within the shape space). These make the manuscript a solid candidate for the journal provided the one,","major_comments":[{"comment":"§4.1 vs §4.2 (central to Tables 3–4): the two methods are given asymmetric numerical treatment. The B-spline NLP (20) is solved with IPOPT via CasADi — exact gradients, box bounds, and a strong physics-informed initial guess obtained by least-squares fitting the control polygon to a quintic MEE interpolating profile (Eq. 22) — while the hodographic six-free-coefficient problem is solved with SciPy Nelder–Mead, an initial simplex built from 10^-2 coordinate perturbations, hard-capped at 1,000 objective evaluations. Nelder–Mead on a 6-D nonconvex objective under a 1,000-evaluation budget is frequently under-converged, so part of the reported ΔV gap (Table 3) and runtime gap (Table 4) may reflect optimizer quality rather than the shape spaces. Since the central claim is a head-to-head attribution ('all B-spline configurations reduce... relative to high-order hodographic shaping'), a matched","section":"§4.1–4.2, Tables 3–4"},{"comment":"§5.2, Table 3: the statistics are computed 'among the finite solutions retained,' but the manuscript nowhere reports per-method failure/retention rates. If IPOPT and Nelder–Mead produce non-finite or non-converged results at different rates (plausible, especially for Mercury where the hodographic mean is 73.5 km/s against a median of 49.8 km/s — a heavy upper tail that could be either genuine or convergence failure), then mean and median comparisons in Table 3 are computed over different implicit subsamples for the two methods. Please report, per method and target: number of attempts, number retained, and the retention criterion; and recompute Table 3 on the intersection of attempts where both methods return finite solutions, or state explicitly that retention rates were effectively equal.","section":"§5.2, Table 3"},{"comment":"§3.2.2/§5: the comparison conflates shape-space type with shape-space dimension. The recommended B-spline configuration (nc=10) has 3(10−4)=18 decision variables versus 6 free coefficients for the hodographic benchmark. A six-degree-of-freedom B-spline configuration (e.g., quintic with nc=6, or cubic with nc=6) would isolate the contribution of the basis type from that of dimension. The nc=40 results partially address flexibility scaling on the B-spline side, but there is no analogous higher-dimensional hodographic run (e.g., adding further basis functions to match ~18 free coefficients). At least one matched-dimension experiment, even on a subset of one campaign, would substantially strengthen the attribution claim.","section":"§5, Table 3"},{"comment":"§5.1 and §6: no solution from either method is validated against an independently converged reference (e.g., a direct collocation or indirect solution) for even a single transfer. The paper's own framing is comparative within shape spaces, and the conclusions are appropriately hedged, so this does not block the comparative claim — but the aggregate claim that B-spline solutions are 'better' preliminary designs would be materially grounded by one or two spot checks showing that the lower-ΔV B-spline solutions (e.g., Fig. 3, Fig. 7) remain superior after refinement by a standard transcription method, and that the near-coast profiles of Fig. 18 are not artifacts of the unconstrained-acceleration formulation that would disappear under any thrust limit.","section":"§5.1, §6"}],"minor_comments":[{"comment":"§4.1: IPOPT convergence tolerances, maximum iterations, and the handling of failed/restore-mode exits are not reported; these matter for interpreting per-attempt runtimes in Table 4 and the retention issue above.","section":"§4.1"},{"comment":"§4.1, Eq. (23): the smoothing parameter ε=10^-6 is stated in 'canonical acceleration units'; for reproducibility, give its equivalent in physical units and a brief sensitivity note (does the ranking in Table 3 change for ε=10^-5 or 10^-7?).","section":"§4.1, Eq. (23)"},{"comment":"§4.2: the paper notes [27] used O'Neill's Nelder–Mead implementation while this work uses SciPy's; a one-line justification (or a spot check that both implementations agree on a few grid points) would preempt questions about benchmark fidelity, since the comparison is against a reimplementation rather than published numbers.","section":"§4.2"},{"comment":"Rendering artifacts: the MEE vector appears as 'ααα' (triple alpha) in Eq. (22) and surrounding text; 'AU' is used both for the astronomical unit and for the canonical acceleration unit in Eq. (25) ('A U = 4π² AU/yr²') — rename the latter to avoid confusion.","section":"§4.1, Eqs. (22), (25)"},{"comment":"Fig. 1: basis-function indices in the legend run from B_{0,3} while the text indexes control points and basis functions from i=1 (Eqs. 4–5); align the indexing convention.","section":"Fig. 1"},{"comment":"§5.1.1, Fig. 3: the lowest-ΔV solutions compared occur at different departure dates, transfer times, and (for Mercury/Tempel 1) different N branches; the text is clear about this, but the captions should state it explicitly so the per-case percentages (1.4%, 4.8%, 9.3%, 10.3%) are not misread as same-attempt comparisons.","section":"Figs. 3, 7, 11, 15"},{"comment":"Table 4: report the hardware-parallelism caveat (parallel campaign, serial-equivalent timing) in the table caption itself, and state the IPOPT linear solver used (e.g., MA27/MUMPS), which can materially affect per-attempt times.","section":"Table 4"},{"comment":"The GitHub repository is a strong point; please add a tagged release or commit hash in the Code Availability statement so the exact benchmarked version is citable.","section":"Code availability"}],"recommendation":"major_revision","confidential_remarks":"The formulation is competent, the campaigns are large, and the open code is welcome. My one substantive reservation is that the headline comparison may partly measure optimizer quality (IPOPT + MEE warm start vs capped Nelder–Mead) rather than shape-space quality. This is fixable within the paper's existing codebase — the author has all the machinery to run a matched-optimizer control — so I recommend major revision rather than rejection. If the matched control shows the gap largely closes, the paper's contribution would need to be repositioned, and I would want to see that outcome reported honestly rather than the control omitted."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: Sanchez gives a clean clamped B-spline parameterization of heliocentric cylindrical position, recovers thrust from differential flatness, enforces the rendezvous BCs by construction, and runs large fine-grid campaigns (Earth–Mars/Mercury/1989 ML/Tempel 1) against a reimplementation of Gondelach–Noomen high-order hodographic shaping. Code is linked. That is real preliminary-design tooling, not a restatement.\n\nWhat is new is the specific combination and the systematic hyperparameter study (degree and control-point count), not the ingredients. Flatness of fully actuated cylindrical two-body dynamics, clamped B-splines, Bézier cylindrical shaping, and MEE cubic splines are all in the cited literature. The paper is honest about that and about its own limits: unconstrained acceleration, no mass, local optima inside a shape space. Dynamics and BCs are satisfied by construction; the objective is only quadrature. Tables 3–4 and the porkchops are clear, and post-processing both methods with the same Gauss–Legendre quadrature removes one confound.\n\nThe soft spot that matters is attribution, not internal math. B-spline NLPs get CasADi/IPOPT with analytic gradients and a physics-informed MEE initial guess; the six-coefficient hodographic problems get SciPy Nelder–Mead, 10^{-2} simplex, 1000-eval cap. Dimension also differs (18 free vars for nc=10 vs 6). The paper hedges to “tested grids / finite solutions retained,” which is fair, but nothing isolates shape space from solver/init. If a gradient-based hodographic solve closed a large fraction of the 22–45% median gap, the claim shrinks to “richer parameterization + better optimizer.” That is a real correctness-risk on the strongest quantitative sentence; it is not a reason to dismiss the formulation.\n\nMath and citations look solid. Who it is for: people who actually run LT porkchops and need fast initial guesses or screening. I would send it to peer review, ask for a matched-optimizer ablation or a clearer framing of the comparison, and treat the method as citeable engineering progress either way.","headline":"Solid, usable B-spline + flatness shaping with large multi-target grids and public code; the headline ΔV gains vs hodographic are real on the reported setup but partly confounded by unequal optimizers.","tokens_in":24963,"tokens_out":548,"would_cite":true,"duration_ms":12037,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Clamped B-splines of heliocentric cylindrical coordinates cut low-thrust rendezvous delta-v versus high-order hodographic shaping while needing no dynamics propagation.","keywords":["low-thrust trajectory optimization","B-spline shaping","shape-based methods","interplanetary rendezvous","differential flatness","hodographic shaping","cylindrical coordinates"],"falsifier":"Re-solve the same fine grids with one shared optimizer, identical initialization and continuation, and matched quadrature; if the median and minimum delta-v gaps between ten-control-point quintic B-splines and the hodographic bases largely close, the central performance claim fails.","tokens_in":24685,"feed_emoji":"🚀","tokens_out":1016,"duration_ms":21517,"temperature":0.7,"pith_summary":"Preliminary design of continuous-thrust interplanetary rendezvous needs fast trajectory shapes that still meet position and velocity at both ends and keep propellant cost low. This paper parameterizes the spacecraft path in heliocentric cylindrical coordinates with clamped B-splines, then recovers the required thrust acceleration algebraically from the path and its derivatives by using differential flatness of the dynamics. Endpoint conditions are built into the first and last control points, so the fixed-time minimum-delta-v problem collapses to a bounded nonlinear program over only the free interior control points, with numerical integration used solely for the cost. Fine-grid Earth-to-Mars, Mercury, 1989 ML, and Tempel 1 campaigns show that every tested B-spline setup lowers mean, median, and minimum accumulated delta-v relative to a high-order hodographic benchmark; a ten-control-point quintic is offered as the practical default, and denser control polygons as a refinement tool. A sympathetic reader cares because the same architecture applies across targets without hand-built basis functions and yields usable porkchop maps at lower serial cost per attempt.","feed_headline":"B-spline paths cut low-thrust delta-v vs hodographic shaping","feed_subtitle":"Ten control points beat the benchmark on four targets and run about a third faster per attempt.","key_machinery":"Clamped B-spline cylindrical-position shaping plus differential flatness: the path q=(r,θ,z) is a clamped B-spline whose endpoint control points enforce position and velocity by construction; control acceleration is recovered algebraically from q and its first two derivatives, so dynamics need not be propagated and only interior control points are optimized.","core_discovery":"Across the tested fine-grid campaigns and among the finite solutions retained, all clamped B-spline configurations of cylindrical coordinates reduce mean, median, and minimum accumulated delta-v relative to high-order hodographic shaping. A quintic with ten control points is identified as the favorable compromise: it cuts median delta-v by roughly 22–45% depending on target and uses about 30–33% less computation time per transfer attempt, while forty-control-point shapes further lower the best-found minimum when cost is secondary.","pith_inferences":["If mass depletion and hard thrust bounds were folded into the same flat B-spline NLP, the method could move from unconstrained screening toward propulsion-aware preliminary design without leaving the shape-based setting.","The local support of B-splines suggests natural multi-leg or gravity-assist extensions by splicing segments while keeping endpoint clamps.","A matched-optimizer ablation against other shape families (Fourier, Bézier, equinoctial splines) would isolate how much of the gain is basis geometry versus solver technology."],"forward_implications":["A single clamped B-spline cylindrical architecture can screen multi-revolution Earth departures to Mars, Mercury, NEAs, and comets without target-specific analytical bases.","Ten-control-point quintic shapes are a default for broad porkchop campaigns when both cost and serial time matter.","Forty-control-point shapes are reserved to refine selected opportunities when the lowest delta-v inside the shape space is the priority.","Shaped trajectories can seed higher-fidelity direct or indirect optimal-control solvers because dynamics and boundaries already hold by construction.","Raising spline degree and control-point count localizes thrust into shorter arcs separated by longer coasts, changing the qualitative control history available to the designer."],"fun_headline_variants":["Clamped B-splines cut low-thrust delta-v vs hodographic shaping","Ten-point quintic B-splines beat hodographic benchmark on four targets","B-spline cylindrical paths reduce median delta-v 22-45% vs hodographic","All tested B-spline configs lower mean median min delta-v vs hodographic","Quintic 10-control-point B-splines balance delta-v cuts and runtime"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The reported delta-v edge is treated as coming from the B-spline shape space itself, even though the B-spline problems are solved with a gradient interior-point solver and the hodographic problems with a derivative-free simplex method under different sampling and warm-start rules.","fun_headline_variants_meta":{"raw":{"variants":["Clamped B-splines cut low-thrust delta-v vs hodographic shaping","Ten-point quintic B-splines beat hodographic benchmark on four targets","B-spline cylindrical paths reduce median delta-v 22-45% vs hodographic","All tested B-spline configs lower mean median min delta-v vs hodographic","Quintic 10-control-point B-splines balance delta-v cuts and runtime"]},"model":"grok-4.5","effort":"low","cost_usd":0.004239,"raw_usage":{"total_tokens":1331,"prompt_tokens":828,"num_sources_used":0,"completion_tokens":96,"cost_in_usd_ticks":42388000,"prompt_tokens_details":{"text_tokens":828,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":407,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":828,"tokens_out":96,"duration_ms":6943,"temperature":1.0,"reasoning_tokens":407,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T12:20:46.014939+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-solve the same fine grids with one shared optimizer, identical initialization and continuation, and matched quadrature; if the median and minimum delta-v gaps between ten-control-point quintic B-splines and the hodographic bases largely close, the central performance claim fails.","supporting_citations":[],"review_version":1}