{"id":"881e9cb1-da51-4c8b-a8ee-9e63c18a5f18","arxiv_id":"2412.02521","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A column generation heuristic with guided variable grouping finds better or comparable solutions to large space mission ConOps mixed-integer programs in less time than full solvers or metaheuristics.","lead":"This paper applies column generation, an optimization technique that starts with a small set of decision variables and adds the most promising ones step by step, to space mission planning with many discrete choices. It demonstrates on a Mars mission and a lunar logistics campaign that this approach finds good plans faster than solving the full mixed-integer program or using metaheuristics.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Case Study 1's reported optimum is infeasible as printed: Table 9 assigns Tank 2 a 5000 kg capacity yet reports 9830 kg of propellant drawn, violating Eq. 13/14; the paper's primary demonstration for Case Study 1 is therefore currently invalid.","rationale":"Reader's weakest assumption was the heuristic pricing step. I agree that is a genuine limitation, but the more load-bearing issue is that the paper's own reported Case Study 1 result violates its model. The paper's claim is an empirical 'can be useful' claim, so each demonstration matters. Table 9 is the only detailed Case Study 1 solution, and it contradicts Eqs. 13-14. That is a concrete, falsifiable inconsistency, not a judgement about problem-specific heuristics. The pricing issue is largely admitted by the authors in Section 3.4, which notes that alternative generated variable sets fail; this limits generality but does not by itself invalidate the two worked examples. The infeasibility does invalidate one worked example. I therefore keep the reader's CONDITIONAL verdict rather than upgrading to ACCEPT or downgrading to REJECT: the fix is clear and local, but until it is done the paper should not be read as having validated the method on Case Study 1. A secondary reason for keeping the verdict conditional is the retrospective selection of the generated variable set; a blind application to a third problem would strengthen the proposed guidelines.","tokens_in":20787,"tokens_out":17067,"duration_ms":185645,"concrete_test":"Run a feasibility check on the Section 5.1 solution: take the variable assignment implied by Table 9, substitute it into Eqs. 13-14 (Constraints 2A/2B), and verify whether Tank 2's total propellant draw (9830 kg) is less than or equal to its selected capacity (5000 kg). If the table contains a transcription error, correct the entry and re-solve Case Study 1; if the solver returned this point, the model or implementation has a bug. Report the corrected objective and the regenerated Figure 6 curve.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 defines Constraint 2A (Eq. 13) as sum_n m^prop_{t,n,v} ≤ sum_s x_{t,n=0,s,v} m_s^cap, and Constraint 2B (Eq. 14) as m^prop_{t,n,v} ≤ sum_s x_{t,n,s,v} m_s^cap. Table 9 in Section 5.1 reports the best-found Case Study 1 solution. The Tank 2 row gives 'Capacity = 5000 kg, Dry Mass = 780 kg' and propellant usage 9830 kg in manoeuvre 1, with zeros elsewhere. Under either constraint, this point is infeasible: the total draw from Tank 2 (9830 kg) exceeds its declared capacity (5000 kg), and the per-manoeuvre draw in manoeuvre 1 also exceeds it. The stated objective of 69,746 kg and the Figure 6 comparison of restricted versus full model therefore include an infeasible point. Since the paper's conclusion that the restricted problem can produce solutions of equal or greater quality relies on both case studies, this is not a cosmetic table typo; it is a current numerical counterexample to the authors' own model. The central claim cannot be credited for Case Study 1 until the table is corrected or the case is re-solved.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes using column generation to solve mixed-integer linear programs that arise in space mission concept-of-operations optimization. The authors introduce a terminology for grouping variables into generated sets, describe a reduced-cost pricing heuristic for integer variables, and present guidelines for selecting generated variable sets and grouping indices. The method is demonstrated on two case studies: a crewed Mars mission with discrete tank sizing, staging, launch vehicle, and parking orbit choices, and a lunar logistics scheduling problem for an extended Artemis campaign. The reported results indicate that restricted problems find solutions of similar or better objective value substantially faster than the full MILP, while the paper also openly discusses the heuristic nature of the pricing step and shows one alternative variable grouping that fails.","tokens_in":21098,"tokens_out":5317,"duration_ms":52465,"significance":"If the reported results are correct, the paper makes a useful practical contribution by showing how column generation can be applied to realistic space logistics MILPs without requiring problem-specific dynamic programming subproblems. The two case studies are nontrivial in size, and the authors are unusually candid about the heuristic status of the method and about a failure mode of the reduced-cost pricing approach. The paper does not claim a new theoretical guarantee, which is appropriate given that the pricing rule is based on LP relaxations of integer variables. The main value is as an engineering methodology demonstration; its scientific significance is limited by the absence of reproducible code or data and by the current inconsistency in the Case Study 1 results.","major_comments":[{"comment":"The reported best-found solution for Case Study 1 is infeasible as printed. Table 9 lists Tank 2 with capacity 5,000 kg and dry mass 780 kg, but the propellant usage row for Tank 2 shows 9,830 kg consumed in manoeuvre 1 (and zeros elsewhere). This violates both Constraint 2A (Eq. 13), which bounds total propellant drawn from a tank by its capacity, and Constraint 2B (Eq. 14), which bounds per-manoeuvre draw by the capacity of the tank present in that manoeuvre. The stated objective of 69,746 kg also cannot be reproduced from the table: summing the listed propellant usage and tank dry masses gives 70,160 kg. Because this infeasible point underlies the Case Study 1 objective and the comparison in Figure 6, the paper's central claim for Case Study 1 is currently unsupported. The authors should correct the table or re-solve the case study and verify all tank capacity constraints before the results can be accepted.","section":"§4.3, Table 8"},{"comment":"The comparison between the full MILP and restricted problems in Case Study 2 is presented as a comparison of best-found objectives, but the full MILP is reported not to have converged to a proven optimal solution. The table does not state the solver's termination criterion, the MIP gap at termination, or the hardware/solver settings used. Since the claim in the conclusion is that restricted problems can produce solutions 'of equal or greater quality' in less time, the comparison should explicitly report the final optimality gaps of all methods so that the reader can assess whether the improvement is meaningful relative to solution quality rather than just incumbent value.","section":"§4.3, Table 8"},{"comment":"The selection of the generated variable set and grouping index is made retrospectively after evaluating several alternatives. The paper does not provide a principled a priori rule for choosing the generated set, aside from the general observations in Section 3.2. This is not an internal inconsistency, and the authors are honest that the method is problem-specific, but it means the case studies demonstrate a successful retrospective application rather than a predictive guideline. The conclusions should be tempered to reflect that the method's success depends on a choice that, in the examples, was made after inspecting the alternative pricing behavior.","section":"§3.4"}],"minor_comments":[{"comment":"The captions of Figures 7-10 appear to have the case study labels swapped: Figure 7 is captioned 'Case Study 1' but describes B_{p,t} variables and time-index grouping from Case Study 2, while Figures 8-10 are captioned 'Case Study 2' but describe x_{t,n,s,v} and ℓ_{v,o} variables from Case Study 1. Please correct the captions.","section":"Figures 7-10"},{"comment":"The manuscript presents Section 4.3 after Sections 5.1 and 5.2, and Section 3.4 appears after Section 5.2. The numbering and placement should be reorganized so that the methodological discussion follows Section 3 and the case study performance discussion follows the case study sections.","section":"Section numbering and order"},{"comment":"Table 2 lists manoeuvre numbers 0, 1, 2, 4, and 5 with no manoeuvre 3, and the description 'DSM 2' is assigned to entry 4. Please clarify whether the numbering is intentionally non-contiguous or whether a manoeuvre is missing.","section":"Table 2"},{"comment":"The big-M coefficient ℳ in Constraint 1A is described as 'sufficiently large' but its value is never reported. To make the formulation reproducible and to rule out numerical issues, the authors should provide the value used or describe how it was computed from problem data.","section":"Equation 11"},{"comment":"The caption of Table 5 literally contains the placeholder text '[include refs]'. This must be completed with the actual references.","section":"Table 5 caption"},{"comment":"The symbol τ is used both for tank mass fraction in Case Study 1 and for real time-of-flight in Case Study 2. While the section-specific meanings are defined, the dual use is confusing in a paper that relies heavily on notation; consider renaming one of the two quantities.","section":"Nomenclature"},{"comment":"The tables in Section 4.2 appear out of numerical order: Table 5 is printed before Table 4, and Table 7 is printed before the constraints that refer to it. Please renumber the tables and place each near its first citation.","section":"Tables 3-5 ordering"}],"recommendation":"major_revision","confidential_remarks":"The infeasible Table 9 entry is likely a transcription error, as the reported objective is close to but not equal to the sum of the listed masses, but the authors must correct it or re-solve the case before the central claim can be credited. The paper is otherwise an honest engineering application of column generation to space mission MILPs. For a mathematical optimization journal, the methodological novelty is modest; the contribution is more of a case-study-driven guideline paper. This may be a scope consideration, though it does not by itself preclude publication in an application-oriented venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes a column generation recipe for MILPs with categorical decisions in space mission ConOps, and tests it on a Mars crewed mission and a lunar logistics model. The genuinely new bits are the concrete guidelines for choosing which variables to generate, how to group them, and how to build an initial feasible set, plus the two MILP formulations. The lunar case study is the strongest part: the restricted model finds a 576,259 kg solution, better than the full MILP's 590,992 kg in 2.7 days, and beats the metaheuristic baseline. The authors also deserve credit for being upfront that reduced-cost pricing on LP relaxations is a heuristic for integer variables, and Section 3.4 candidly reviews alternative generated-variable choices.\n\nThe soft spots are real but uneven. The biggest is Table 9 in Case Study 1: Tank 2 has a 5,000 kg capacity but is reported to deliver 9,830 kg of propellant in manoeuvre 1. That violates both constraints 2A and 2B. So the reported 69,746 kg objective and Figure 6's comparison rest on an infeasible solution. That's not cosmetic; it invalidates the current demonstration for Case Study 1 until re-solved or corrected.\n\nSecond, Section 3.4 admits the generated variable set and grouping index were selected after testing alternatives. That's honest, but it means the method's success isn't a blind prediction - it's a retrospective fit. A validation would need to fix the choice a priori or test on a third case.\n\nThird, the manuscript is unfinished: Table 5's caption literally says '[include refs]', 'formerly solved' should be 'formally solved', Figure captions 8-10 reference the wrong case study, and there are missing figure numbers. These are fixable but suggest a draft rather than a polished submission.\n\nThe central idea is plausible and the lunar results are encouraging. The infeasible table is the load-bearing issue; if it's a typo, the correction may be easy, but until then the primary claim for Case Study 1 doesn't stand. I'd send this to peer review with a request for major revision - fix the table, re-run Case Study 1, clarify the retrospective selection, and clean up the text. For a conference paper, I'd accept only with those changes.","headline":"Plausible column generation recipe for space ConOps, but Case Study 1's reported optimum violates its own capacity constraints.","tokens_in":21590,"tokens_out":2960,"would_cite":false,"duration_ms":28727,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C11","90C10","90C05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Column generation finds mission designs as good or better, faster","keywords":["mixed-integer linear programming","column generation","space mission design","concept of operations","lunar logistics","Mars mission","Dantzig-Wolfe decomposition","categorical decisions"],"falsifier":"Take a mission design instance where the generated variable set is chosen poorly by the paper's own criteria, such as the tank design variables grouped by sample index in Case Study 1; Section 3.4 indicates the relaxation will price only the largest, most mass-efficient design and miss smaller options, so a column generation run on that grouping should fail to reach the 69,746 kg solution found with the launch-vehicle grouping. Alternatively, run the lunar logistics model with more than five generated time-index groups and show that the restricted objective never improves beyond 576,259 kg while the full MILP eventually finds a better solution.","tokens_in":20594,"feed_emoji":"🚀","tokens_out":5808,"duration_ms":54848,"temperature":0.7,"pith_summary":"This paper argues that column generation, a method that solves a small 'restricted' version of a linear program and then selectively adds back promising variables, can make mixed-integer optimisation of space exploration concepts of operations tractable. The authors apply the method to two problems that are hard for ordinary MILP solvers because of their many categorical decisions: a crewed Mars mission with discrete choices about tank design, tank dropping, launch vehicles, and assembly orbit, and a lunar logistics campaign with awkward payload scheduling choices. In the lunar case the restricted problem with five generated time-index groups reaches an objective of 576,259 kg, beating the full model's 590,992 kg while using far less compute time. The paper's conclusion is that, when the generated variable set is chosen with attention to problem structure, restricted problems can produce solutions of equal or greater quality than the full problem in much less time.","feed_headline":"Column generation finds mission designs as good or better, faster","feed_subtitle":"A restricted lunar logistics problem reached 576,259 kg versus 590,992 kg for the full model, in far less compute time.","key_machinery":"The machinery is column generation built on Dantzig-Wolfe decomposition, specialised to mixed-integer programs. Variables are partitioned into an always-included set and a generated set; the restricted problem solves over a small subset of generated variable groups, then every unused group receives a price equal to the sum of reduced costs of its component variables from the continuous relaxation of the restricted problem. The variable group with the best price is added back, and an initial feasible variable set is produced either by solving a binary coverage problem or by solving the full problem with a simpler integer objective. This reduced-cost pricing is a heuristic when applied to integer variables, and the paper's guidelines for grouping indices determine which relaxations carry meaningful information.","core_discovery":"On the paper's own terms, the central discovery is that the Dantzig-Wolfe-style column generation process, treating the reduced costs of a continuous relaxation as a heuristic price for groups of integer variables, can solve realistic space mission ConOps optimisation problems that are too large for direct MILP. In Case Study 1, after generating eight additional launch-vehicle variable groups, the restricted MILP finds a 69,746 kg objective for the Mars mission, closing the gap to the best bound of the full problem. In Case Study 2, the minimum-feasible restricted lunar logistics problem gives 578,305 kg, and adding five generated time-index groups improves this to 576,259 kg, better than the full MILP's 590,992 kg, with the full model failing to prove optimality. The authors therefore claim that restricted problems can produce solutions of equal or greater quality to the full problem in a much shorter amount of time, provided the generated variable set is selected carefully.","pith_inferences":["The same price-and-regenerate loop could be applied to other combinatorial ConOps choices, such as orbital servicing schedules or constellation deployment orders, wherever an LP relaxation offers meaningful reduced costs.","The failure mode identified in Section 3.4 suggests a simple diagnostic: if all prices in a candidate grouping are identical or if the relaxation allows fractional versions of integral choices, the grouping should be abandoned before running the full column generation loop.","Because the method is heuristic on integer variables, its main value may be as a fast warm-start or incumbent-finding procedure inside a branch-and-bound tree, rather than a replacement for exact solvers on instances where optimality must be certified.","A direct comparison on a third mission design problem, with the same method and two different grouping strategies, would test how transferable the guidelines are."],"forward_implications":["Restricted MILPs equipped with a well-chosen generated variable set can close the optimality gap of large mission design problems without solving the full model.","Discrete tank sizing can be handled by sampling a nonlinear mass-capacity curve, avoiding a MINLP formulation while retaining solution quality.","For the lunar logistics campaign, allowing payload scheduling variables to be generated in pairs of even-odd time indices improves objective value beyond the minimum feasible set and beyond the full MILP's best found solution.","Column generation is problem-specific: the same algorithm needs a different generated-variable choice for each problem class, and some natural choices, such as grouping tank designs by sample index, fail because the relaxation admits fractional tanks.","Future work on problem-specific pricing subproblems could converge to good ConOps solutions even faster."],"supporting_citations":[{"why":"Supplies the Dantzig-Wolfe decomposition used to construct restricted problems.","marker":"[16]"},{"why":"Supports using LP relaxations and reduced costs for pricing in integer column generation.","marker":"[17]"},{"why":"Provides a technical review situating column generation among integer programming methods.","marker":"[20]"},{"why":"Serves as the metaheuristic baseline compared against in the lunar logistics case study.","marker":"[15]"},{"why":"Provides the time-expanded network flow model that underlies the lunar logistics formulation.","marker":"[4]"},{"why":"Gives the generalized multicommodity network flow foundation for space logistics modelling.","marker":"[3]"},{"why":"Offers an alternative dynamic-programming pricing approach that the paper contrasts with reduced-cost pricing.","marker":"[22]"}],"fun_headline_variants":["Column generation speeds up Mars and lunar mission planning","Faster categorical optimization for space exploration missions","Column generation outpaces full MILP in lunar logistics case","Mission ConOps optimized quicker with column generation","Column generation cuts compute time, matches or beats full model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pricing step assumes that reduced costs from the continuous relaxation of the restricted problem are a reliable guide to which integer variable groups are worth adding, even though fractional changes to those integer variables are not feasible in the original problem.","fun_headline_variants_meta":{"raw":{"variants":["Column generation speeds up Mars and lunar mission planning","Faster categorical optimization for space exploration missions","Column generation outpaces full MILP in lunar logistics case","Mission ConOps optimized quicker with column generation","Column generation cuts compute time, matches or beats full model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000275,"raw_usage":{"total_tokens":1691,"prompt_tokens":1041,"completion_tokens":650,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":657,"completion_tokens_details":{"reasoning_tokens":577}},"tokens_in":657,"tokens_out":650,"duration_ms":7310,"temperature":1.0,"reasoning_tokens":577,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:22:12.568323+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a mission design instance where the generated variable set is chosen poorly by the paper's own criteria, such as the tank design variables grouped by sample index in Case Study 1; Section 3.4 indicates the relaxation will price only the largest, most mass-efficient design and miss smaller options, so a column generation run on that grouping should fail to reach the 69,746 kg solution found with the launch-vehicle grouping. Alternatively, run the lunar logistics model with more than five generated time-index groups and show that the restricted objective never improves beyond 576,259 kg while the full MILP eventually finds a better solution.","supporting_citations":[{"cited_title":"Desaulniers, J","cited_arxiv_id":null,"evidence_quote":"Supports using LP relaxations and reduced costs for pricing in integer column generation."},{"cited_title":"A Technical Review of Column Generation in Integer Programming,","cited_arxiv_id":null,"evidence_quote":"Provides a technical review situating column generation among integer programming methods."},{"cited_title":"Hierarchical Framework for Space Exploration Campaign Schedule Optimization,","cited_arxiv_id":null,"evidence_quote":"Serves as the metaheuristic baseline compared against in the lunar logistics case study."},{"cited_title":"Dynamic modeling and optimization for space logistics using time -expanded networks,","cited_arxiv_id":null,"evidence_quote":"Provides the time-expanded network flow model that underlies the lunar logistics formulation."},{"cited_title":"A Column Generation Algorithm for a Rich Vehicle -Routing Problem,","cited_arxiv_id":null,"evidence_quote":"Offers an alternative dynamic-programming pricing approach that the paper contrasts with reduced-cost pricing."}],"review_version":1}