{"id":"9e9477a8-1b66-4bf9-ad4f-0b1366a4dbde","arxiv_id":"2411.16887","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MGCA uses convex combinations of previously found near-optimal solutions to rapidly generate new feasible energy portfolios with user-imposed constraints and objectives.","lead":"This paper introduces MGCA, a fast post-processing method that lets users generate new near-optimal energy system portfolios inside the space mapped by previous optimization runs. It can create new portfolios, add constraints, compare tradeoffs, and estimate costs and emissions in fractions of a second on a laptop.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on conv(Z) approximating the near-optimal feasible region, but no coverage evidence is provided; all exploration and the 50-point accuracy test are confined to the hull of 200 iterates (4 vertices for the accuracy sample), so omitted regions remain invisible.","rationale":"The paper's mathematical core is sound: convex combinations of feasible LP solutions are feasible and within the original budget slack, and Eq. 7 for affine capacity metrics is correct. The issue is not internal logic but whether the finite hull is an adequate surrogate for the near-optimal feasible region. The reader identified the same weakest assumption, namely that Z must adequately span the near-optimal region, and that is exactly where the central claim is least secure. All downstream capabilities, including real-time objective optimization, constraint imposition, Pareto approximation, and lower-budget exploration, inherit whatever coverage gap conv(Z) has. The accuracy experiment in Section 3.6 cannot detect this gap because it samples a 3-simplex formed by 4 of the 200 vertices and compares against the re-solved dispatch of those same interpolates rather than against the true optima for user-defined objectives. The budget interpolation in Section 3.3 is a concrete case where the gap materializes: pairwise interpolation with the least-cost solution yields a scaled subset of the original hull, not the full lower-budget feasible region, so the claim of effectively performing a new MGA run is an overstatement. I keep the CONDITIONAL verdict because the method is a useful and fast inner-approximation tool, but the paper should either provide coverage diagnostics or qualify claims about exploring the full range of near-optimal space and about budget interpolation.","tokens_in":13,"tokens_out":11204,"duration_ms":168435,"concrete_test":"Solve the original GenX LP directly with the 10% budget and a user objective (e.g., minimize total natural gas capacity), obtaining the true best portfolio. Then solve Eq. 6 with the same objective over conv(Z) using the paper's 200 iterates, and again using 50, 100, and 400 iterates. Compare objective values and capacity vectors; also repeat for a lower budget (6%) by solving the full model with the 6% cost constraint and compare against the Section 3.3 pairwise interpolates. If the hull-constrained optimum converges to the true optimum as the iterate count grows, the inner-approximation concern is minor; if the gap stays large (e.g., more than 10% in objective value) or the pairwise 6% set misses feasible points that the full 6% solve finds, the central real-time feasible portfolio claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the finite set of MGA iterates Z spans the part of the near-optimal feasible region users care about. Section 2.2 only instructs that MGA should \"explore as much of the near-optimal feasible space as possible\"; it gives no check that conv(Z) covers the region relevant to a user's objective or added constraints. Equation 6 optimizes over conv(Z), so if the true optimum for a user objective lies outside conv(Z), MGCA returns the best hull point, not the best near-optimal portfolio. The paper acknowledges this in Section 3.5 (\"likely to not include the true optimal solution\") and recommends \"bracketing runs\" in Section 3.2, confirming that conv(Z) is generally an inner approximation. The headline accuracy claim does not test coverage: the n=50 interpolates in Sections 3.2 and 3.6 are random convex combinations of 4 exterior points, i.e., a 3-simplex slice of the 12-dimensional hull, and are compared against their own full-CEM re-solve, not against the true optimum of any stakeholder objective. A related overclaim is Section 3.3's budget interpolation: pairwise interpolation with the least-cost solution only produces a homothetic subset of the 10% hull and cannot in general represent the convex hull of the 6% feasible region, because other constraints besides the cost budget shape the lower-budget slice. A user can therefore be unaware that better portfolios satisfying their objective and constraints exist outside the generated set.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 'Modelling to Generate Continuous Alternatives' (MGCA), a post-processing method for linear-programming capacity expansion models. After an MGA exploration has produced a finite set of near-optimal capacity solutions, MGCA projects those solutions onto a subspace of capacity decisions and outcome metrics, then treats the resulting points as vertices of a convex hull. The method (i) generates new portfolios as convex combinations of these vertices, (ii) solves a small optimization problem (Eq. 6) over this hull so that users can impose objectives and additional constraints, (iii) computes affine capacity metrics exactly and operational metrics by convex interpolation, and (iv) exports chosen portfolios back to the full CEM for least-cost dispatch. The demonstrations on a three-zone GenX ISONE case show sub-second solve times for the reduced problems, including budget interpolation, constrained exploration, Pareto-front approximation, and re-dispatch of 50 interpolated portfolios. The paper also reports that interpolated operational metrics are generally within about 5-10% of the full-CEM re-solve values for the tested points.","tokens_in":2110,"tokens_out":2162,"duration_ms":79480,"significance":"If the claims are properly qualified, MGCA is a useful and practical methods contribution. The central convexity argument is correct for LP planning models: convex combinations of feasible, within-budget iterates are feasible and within budget, and affine capacity metrics are exactly interpolated. The dimensionality reduction to capacity decisions and metrics is simple but effective, and the reported speedups (sub-second post-processing versus hundreds of seconds per MGA iterate) are credible and potentially valuable for interactive stakeholder tools. The paper uses an open-source model (GenX) and openly describes the demonstration case, which supports reproducibility. The main weaknesses are overstatements about the convex hull of computed iterates representing the near-optimal feasible region, the claim in Section 3.3 that pairwise budget interpolation represents the lower-budget convex hull, and the limited scope of the accuracy test that underlies the abstract's 'within 10%' claim. These issues are fixable in revision and do not invalidate the core method.","major_comments":[{"comment":"The paper does not consistently state that all MGCA outputs live in conv(Z), the convex hull of the computed MGA iterates, and that conv(Z) is generally an inner approximation of the full near-optimal feasible region. Section 2.2 says the projected points 'can then be used as vertices for the convex hull of the near-optimal feasible region of the planning problem,' but this is only true if the MGA iterates actually span that region. Equation (6) optimizes over conv(Z), so if the user's true optimum lies outside this hull, the returned point is the best hull point, not the best near-optimal portfolio. The paper does acknowledge this later in Section 3.5 ('likely to not include the true optimal solution') and recommends bracketing runs in Section 3.2, but the abstract and introduction would mislead readers who do not reach those sections. Please add an explicit statement in the abstract and in Section 2 that the method operates on the convex hull of computed iterates and that coverage of the near-optimal region is an input assumption that must be verified.","section":"Section 2.2 and 2.3"},{"comment":"The budget-interpolation demonstration is overstated. The procedure of combining each MGA iterate with the least-cost solution at weight 0.6 produces only the union of line segments between the least-cost point and each iterate, a homothetic subset of the 10% hull. This does not, in general, 'represent the convex hull of the feasible region within this tighter constraint,' because other constraints besides the cost budget shape the lower-budget slice of conv(Z). A point with cost at 6% slack can lie outside the union of these segments. To explore the full 6% budget slice, one should solve Eq. (6) with the additional cost constraint; the pairwise interpolation should be described as a rapid approximation or a partial sample, not as the convex hull of the lower-budget feasible region.","section":"Section 3.3, Figures 5-6"},{"comment":"The headline accuracy claim is not supported by the experiment as described. The n=50 interpolates in Section 3.2 are random convex combinations of only 4 exterior points, i.e., points in a 3-simplex inside the 14-dimensional reduced space. Section 3.6 compares the metric values of these interpolates against the full-CEM re-solve of the same portfolios, which measures interpolation error for those specific points but does not compare against the true optimum of any user objective, nor does it test the hull's coverage of the near-optimal region. The abstract's phrase 'within 10% of the optimal value' should be replaced with a more precise statement, for example 'within 10% of the full-CEM re-solve value for the tested interpolated portfolios,' and the text should note that the test is confined to a small simplex and cannot bound errors elsewhere in the hull.","section":"Section 3.6 and Abstract"},{"comment":"Equation (6) contains undefined notation: the term 'lambda^T Z X' uses 'X' without definition, and the preceding text says 'z = lambda^T Z' while earlier Z is defined as a subspace. The model size stated ('2m variables and 3n+d+1 constraints') also does not match the displayed problem, which has n+m variables and n+1+d+m+n constraints (depending on how variable bounds are counted). Please correct the equation, define all symbols, and fix the complexity statement so readers can rely on the claimed dimensionality reduction.","section":"Section 2.3, Eq. (6)"},{"comment":"The statement that an interpolated operational metric is 'a feasible value within the convex hull of those found through the full model with economic dispatch' is imprecise. A convex combination of scalar metric values is not itself a feasible decision in the original model; it is simply an interpolation of the vertex values. The paper already notes that operational metrics 'will not match economic dispatch' and that 'no guarantee of lower overall system metric values is possible,' which is the correct caveat. The wording about feasibility of the interpolated metric should be revised to avoid implying that the interpolated scalar is a feasible outcome of the optimization model.","section":"Section 2.4"}],"minor_comments":[{"comment":"Typo: 'constraint coefficient matricies' should be 'matrices'.","section":"Section 2.1"},{"comment":"Typo: 'analagous' should be 'analogous'.","section":"Section 3.2"},{"comment":"Please state explicitly the dimension of the reduced space used in the demonstration (12 technology capacity aggregates plus 2 system metrics = 14) and clarify that the convex hull and all interpolations are performed in this full 14-dimensional space, while Figures 2-7 show only pairwise 2D projections.","section":"Section 3.1"},{"comment":"The text says the budget interpolation 'of all solutions computed in 0.35 seconds'; for reproducibility, please clarify how many solutions were interpolated and whether this includes the time to generate Figure 5 and Figure 6.","section":"Section 3.3"},{"comment":"The vertical axes are labeled 'Percent Difference' and 'Difference in Zonal Percent Share'; please state clearly whether the latter is in percentage points, and report the n=50 sample size in the figure captions.","section":"Section 3.6, Figures 10-11"},{"comment":"The claim that the exploration problem 'will, by definition, always find an exterior solution' is acceptable for a linear objective over a polytope if 'exterior' means a vertex, but the wording is confusing because vertices of conv(Z) are the original MGA iterates only if the iterates are not redundant; please clarify the intended meaning.","section":"Section 2.3"},{"comment":"The sentence 'This Pareto frontier is may not be identical' contains a grammatical error; please rephrase as 'This Pareto frontier may not be identical to a similar Pareto frontier calculated on the original space.'","section":"Section 3.5"},{"comment":"The paper relies on the companion paper Lau, Patankar, and Jenkins (2024) for the quality of the MGA exploration; please cite a published version if one exists, or add a note that the companion is under review, since the coverage assumptions in the main paper depend on it.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The core idea is sound and the method is a reasonable extension of Pedersen et al. (2021), with the practical advantage of avoiding Delaunay triangulation and allowing user-driven optimization over the hull. The paper would be acceptable after the overclaims in the abstract and Sections 2.2 and 3.3 are corrected and the accuracy claim in Section 3.6 is qualified. The evaluation uses a single test case, which is fine for a methods paper but should be stated as a limitation. The companion preprint [11] is important to the coverage assumption; the authors should confirm its status or include a version reference."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is simple and the paper executes it cleanly: take the convex hull of MGA iterates, project to capacity and metric space, and let users solve a tiny LP over convex combinations with custom objectives and constraints. The convexity argument is correct for LP planning models—convex combinations of feasible, within-budget iterates stay feasible and within budget, affine capacity metrics interpolate exactly, and operational metric estimates are feasible but not exact. The speed gains are real: 0.6 ms per interpolate and 0.01 s per constrained iterate on a laptop is genuinely useful for interactive decision support. It also improves on Pedersen's MAA by sidestepping Delaunay triangulation, which breaks down above ~10 dimensions, and by allowing user-directed rather than purely random sampling.\n\nThe soft spots are mostly about scope of claims. The set conv(Z) is an inner approximation of the near-optimal feasible region, and the paper acknowledges this in Section 3.5 and in its bracketing-run advice, but elsewhere it writes as if the hull were the region itself. No coverage diagnostic is offered—no check that the 200 iterates span the parts of the space a user might care about. The accuracy test is also narrow: the 50 interpolates are random convex combinations of just 4 exterior vertices, so they occupy a 3-simplex slice of a 12-dimensional hull. That supports 'interpolates are feasible and roughly within 10%' for that slice, but not for the hull generally. The budget-interpolation trick in Section 3.3 is pairwise homothetic scaling with the least-cost solution; it is a convenience for sketching lower-budget slices, not 'effectively performing an entirely new MGA problem,' as the text claims. Missing code and data make the runtime and accuracy numbers hard to verify independently, and there is no baseline comparison against MAA or a random-sampling alternative.\n\nThose issues are fixable. The central method stands, and the paper is honest about the main limitation (inner approximation) even if some promotional sentences overshoot. For the MGA decision-support community this is a useful advance. A serious referee should see it, though the revision needs coverage diagnostics, a more representative accuracy evaluation, code and data, and toned-down claims about budget interpolation and full-region coverage. I would accept it for peer review and engage with the revised version.","headline":"A sound convex-combination post-processing method for MGA, but several claims about coverage and budget interpolation outrun the evidence.","tokens_in":19538,"tokens_out":1413,"would_cite":true,"duration_ms":15188,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C05","90C25","90C90"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper shows that previously computed MGA results can be turned into a continuous, live-explorable space of feasible portfolios by interpolating over the convex hull of the iterates.","keywords":["modelling to generate alternatives","convex combination","convex planning models","capacity expansion","near-optimal feasible space","Pareto frontier approximation","dimensionality reduction","interactive decision support"],"falsifier":"Take a convex capacity-planning LP, run MGA once with a deliberately small or one-sided set of iterates that provably miss a known portion of the near-optimal region, then use MGCA to generate portfolios and Pareto frontiers there; if the omitted portfolios never appear or the frontiers stop short of known feasible extremes, the method is shown to be limited by the seed set. Alternatively, export the 50 interpolated portfolios to the full economic-dispatch model and tally errors: if substantially more than a small fraction of operational-metric interpolations exceed the reported 10 percent gap, the accuracy claim as stated fails.","tokens_in":75,"feed_emoji":"⚡","tokens_out":6404,"duration_ms":137432,"temperature":0.7,"pith_summary":"Capacity-expansion models that optimize for least cost return one plan, but real planning needs many viable alternatives. Modelling to Generate Alternatives (MGA) produces a scatter of near-optimal plans, and this paper adds a post-processing layer that turns that scatter into a continuous space users can explore live. The method, MGCA, projects MGA solutions onto capacity decisions and outcome metrics, forms their convex hull, and lets a user generate any interior portfolio by specifying weights, extra constraints, or a new objective. Because the resulting optimization problem is tiny, it solves in milliseconds to seconds on a laptop. The paper shows that capacity-related metrics are reproduced exactly by interpolation, operational metrics stay within feasible bounds and generally within 10 percent of an economic-dispatch calculation, and exported portfolios can be re-solved by the full model to recover exact operational details.","feed_headline":"Generate feasible energy portfolios in milliseconds from old MGA runs","feed_subtitle":"Drop a constraint or objective and the convex hull of past results yields a viable new plan with metrics within about 10 percent.","key_machinery":"The central object is the convex hull exploration problem (Eq. 6), a small linear or convex program over the MGA-iterate weights $\\lambda$ that produces a point $z_k = \\lambda^T Z$ inside the hull, subject to user-added inequalities $Az_k \\le b$. Its companion identity is Eq. 7, the scalar-product property $c^T(\\sum_i \\lambda_i x_i) = \\sum_i \\lambda_i (c^T x_i)$, which lets affine capacity metrics be evaluated exactly by the same weights. Dimensionality reduction first projects each MGA solution onto selected capacity decisions and linear metrics, shrinking a model with $O(N \\times T)$ operational variables to one with $O(N)$ entries, which is what makes interactive solve times possible.","core_discovery":"MGCA's central claim is that the near-optimal feasible region of a convex capacity-planning model, once approximated by MGA iterates, can be treated as the convex hull of those iterates in a reduced capacity-and-metric subspace. Inside that hull, any convex combination of stored vertices is a feasible, within-budget portfolio, so a user can generate new plans anywhere in the space by solving a small optimization problem over the weights, with user-specified objectives and constraints, rather than re-running the large model. Affine functions of capacity decisions, including many outcome metrics, are then exactly recoverable from the weights, while generation-based operational metrics are interpolated to feasible values that a full economic dispatch would improve or match; the paper reports interpolated metric values generally within 10 percent of the optimal value. The method also supports synthesizing tighter-budget results, approximating Pareto frontiers between metrics, and exporting chosen capacity portfolios back to the full capacity-expansion model to obtain least-cost operational details.","pith_inferences":["In our reading, the same convex-hull machinery transfers to any linear program with a large near-optimal set, such as water-resource or supply-chain planning, where stakeholders also need to see alternatives rather than a single optimum.","If the interpolation accuracy holds at scale, interactive decision-support dashboards could let opposing stakeholders co-design compromises directly, replacing the current workflow of pre-computing and hand-picking discrete MGA results.","A natural stress test would be to seed MGCA with deliberately adversarial or sparse MGA iterates and measure how much of the true near-optimal region is missed; the paper's guarantees are conditional on the seed set spanning the space.","The 10 percent accuracy figure is a demonstration on one test case, not a proven bound, so users would need case-specific validation before relying on interpolated operational metrics for regulatory decisions."],"forward_implications":["Stakeholders can explore near-optimal energy plans interactively, imposing constraints such as a cap on gas capacity or a farmland-preservation objective and seeing the feasible remainder in under a second.","Budget analysis at multiple cost slacks can be produced from one MGA run by interpolating each iterate toward the least-cost point, instead of re-running the planning model at each budget.","Pareto-frontier-style trade-off curves between any pair of outcome metrics can be approximated in post-processing with an epsilon-constraint method over the hull.","Any generated portfolio can be exported and re-solved with the full capacity-expansion model at fixed capacities, so operational metrics reflect least-cost dispatch rather than the interpolation estimate.","Because the exploration problem has only a few hundred variables, nonlinear convex objectives can be used in the live search."],"supporting_citations":[{"why":"Defines modelling to generate alternatives and the budget-over-cost framing that MGCA post-processes.","marker":"[3]"},{"why":"Supplies the MGA methods and guidance for generating diverse iterate sets that seed the hull.","marker":"[11]"},{"why":"Establishes convex combinations as a way to sample interior alternatives, the approach MGCA generalizes.","marker":"[12]"},{"why":"Cited as the reason Delaunay triangulation becomes intractable above about ten dimensions, motivating MGCA's direct use of the iterate set as vertices.","marker":"[24]"},{"why":"Provides the definitions of convex combination, convex hull, and the affine geometry that Eqs. 6 and 7 rely on.","marker":"[27]"},{"why":"Supplies the open-source capacity expansion model used for the 8760-hour demonstration and the export/re-solve test.","marker":"[29]"}],"fun_headline_variants":["New energy plans in milliseconds from old MGA results","Convex hull of past runs yields instant feasible portfolios","Sub-second portfolio generation via convex combination","Real-time exploration of energy design space from stored solutions"],"cache_read_input_tokens":21504,"weakest_assumption_plain":"The load-bearing premise is that the set of MGA iterates collected as vertices spans the near-optimal feasible region well enough that its convex hull stands in for the true region; if large parts of the region are missed, every generated portfolio, budget interpolant, and Pareto frontier is confined to an inner approximation.","fun_headline_variants_meta":{"raw":{"variants":["New energy plans in milliseconds from old MGA results","Convex hull of past runs yields instant feasible portfolios","Sub-second portfolio generation via convex combination","Real-time exploration of energy design space from stored solutions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000313,"raw_usage":{"total_tokens":1820,"prompt_tokens":1025,"completion_tokens":795,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":735}},"tokens_in":641,"tokens_out":795,"duration_ms":8359,"temperature":1.0,"reasoning_tokens":735,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:46:51.944370+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a convex capacity-planning LP, run MGA once with a deliberately small or one-sided set of iterates that provably miss a known portion of the near-optimal region, then use MGCA to generate portfolios and Pareto frontiers there; if the omitted portfolios never appear or the frontiers stop short of known feasible extremes, the method is shown to be limited by the seed set. Alternatively, export the 50 interpolated portfolios to the full economic-dispatch model and tally errors: if substantially more than a small fraction of operational-metric interpolations exceed the reported 10 percent gap, the accuracy claim as stated fails.","supporting_citations":[{"cited_title":"Using modeling to generate alternatives (MGA) to expand our thinking on energy futures","cited_arxiv_id":null,"evidence_quote":"Defines modelling to generate alternatives and the budget-over-cost framing that MGCA post-processes."},{"cited_title":"Context-Specific Energy Strategies: Coupling En- ergy System Visions with Feasible Implementation Scenarios","cited_arxiv_id":null,"evidence_quote":"Establishes convex combinations as a way to sample interior alternatives, the approach MGCA generalizes."},{"cited_title":"Perspectives of Informed Citizen Panel on Low-Carbon Electricity Portfolios in Switzer- land and Longer-Term Evaluation of Informational Materials","cited_arxiv_id":null,"evidence_quote":"Cited as the reason Delaunay triangulation becomes intractable above about ten dimensions, motivating MGCA's direct use of the iterate set as vertices."},{"cited_title":"Policy Decision Support for Renewables Deploy- ment through Spatially Explicit Practically Optimal Alternatives","cited_arxiv_id":null,"evidence_quote":"Provides the definitions of convex combination, convex hull, and the affine geometry that Eqs. 6 and 7 rely on."},{"cited_title":"Tsitsiklis.Introduction to linear optimization","cited_arxiv_id":null,"evidence_quote":"Supplies the open-source capacity expansion model used for the 8760-hour demonstration and the export/re-solve test."}],"review_version":1}