{"id":"627aa7e3-fa80-4a54-ab9f-aaaa68424686","arxiv_id":"2501.00923","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A new adjoint shape optimization framework using the BGK equation and a body-fitted mesh accurately computes sensitivities and efficiently optimizes airfoils across continuum to free-molecular flow regimes.","lead":"This paper develops an adjoint-based method to optimize the shape of solid bodies in gas flows that range from normal to extremely rarefied, using the kinetic BGK equation. The method computes shape sensitivities efficiently and finds optimal 2D airfoil shapes for drag reduction in about a dozen optimization steps.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The accuracy claim hinges on the deferred multiscale BGK solver from [18]: all in-paper validations (FDM and topology comparison) share that solver, so a systematic solver bias would invalidate the reported sensitivities and optima without any independent signal.","rationale":"The reader's weakest assumption—that the method's accuracy rests on the unstated correctness of the multiscale scheme from [18]—is also the most load-bearing concern here. The paper's strongest independent evidence is the finite-difference validation of the adjoint sensitivity, which is genuinely useful: it checks that the continuous/discrete hybrid sensitivity is consistent with the discretized objective. However, that validation cannot detect a systematic error in the underlying primal BGK solver, because the FDM objective evaluations use the same solver. The comparisons to topology optimization are likewise internal to the authors' solver family, and the admitted non-convergence of the Re=200 topology result further weakens the external check. Thus the central claim that the method is accurate 'in all flow regimes' and converges to optimal shapes is conditional on solver correctness that is not independently demonstrated in this manuscript. I do not see a separate fatal flaw in the adjoint equations or the optimization procedure; the agreement with FDM and the objective decreases support those parts. The appropriate verdict remains CONDITIONAL, which is unchanged from the reader's verdict.","tokens_in":96,"tokens_out":15227,"duration_ms":219867,"concrete_test":"Recompute the Table 2 channel-airfoil cases with independent solvers: DSMC for Kn=10 and a validated Navier-Stokes solver for Kn=0.001, evaluating the drag coefficient Cd for the published initial and optimized airfoil shapes under the same channel geometry and freestream conditions. Accept the central claim if the independent drag reductions agree with Table 2 within the estimated discretization and statistical error (e.g., within 10% of the reported Delta-Cd); if the independent Delta-Cd differs by more than that or an independent shape search improves on the reported drag beyond the SLSQP tolerance, the deferred solver is the weak link and the method's accuracy across regimes is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the adjoint-based method yields accurate shape sensitivities and near-optimal airfoils from Kn=0.001 to 10. The new adjoint sensitivity itself has real support: the finite-difference checks in Section 4.1 directly test the discrete consistency of the adjoint equation (22) and the sensitivity formulas (33)-(36), and they agree. The load-bearing step is therefore not the adjoint derivation but the multiscale BGK solver used for both the primal and adjoint equations. Section 3.1 explicitly omits the discretization details, stating only that they are 'generally the same' as those in [18]. Every validation in the paper is internal to that solver family: the FDM sensitivities perturb the same solver, and the topology-optimization comparisons in Sections 4.2 and 4.3 use the same underlying gas-kinetic scheme from [18]. The one ostensibly external comparison is weakened by the paper's own admission that the Re=200 topology optimization in [18] did not fully converge, and by the choice to initialize the chord length at the [18] optimum. Consequently, if the deferred solver has a bias in its diffuse-boundary treatment, continuum-limit acceleration, or velocity-space resolution, the sensitivities, optimized shapes, and drag reductions all inherit that bias, and no independent DSMC or Navier-Stokes benchmark in this paper would reveal it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an adjoint-based shape optimization method for gas flows described by the BGK equation with diffuse-reflection boundary conditions. A continuous adjoint equation is stated and solved with a multiscale discrete-velocity finite-volume scheme taken from the authors' prior work [18], and the sensitivity of boundary functionals is obtained through a discrete-adjoint-like differentiation of the discretized boundary integrals. The sensitivities are validated against finite differences for an elliptic cylinder at Kn = 10, 0.1, and 0.001, and the method is applied to drag reduction of NACA0012-type airfoils in subsonic channel flow and supersonic flow, reporting drag reductions up to 13.36% and convergence in roughly a dozen optimization steps.","tokens_in":21743,"tokens_out":6788,"duration_ms":69107,"significance":"The main contribution is a shape-optimization framework that avoids the high design-variable count and lower boundary accuracy of topology optimization while covering the whole Knudsen regime. The finite-difference validation of sensitivities at three Knudsen numbers is a genuine strength and provides direct evidence that the adjoint implementation is discretely consistent. The observation that CST parameterization suppresses velocity-discretization-induced sensitivity oscillations is practically useful. However, the paper defers the core numerical scheme to a prior publication, and the derivation of the adjoint equation is omitted; these gaps currently limit the reproducibility and self-containedness of the central method. If those details are supplied and the topology-optimization comparison is appropriately qualified, the method would be a valuable contribution to rarefied-flow design.","major_comments":[{"comment":"The discretization and solution procedure for both the primal and adjoint equations are deferred to the authors' prior work [18], with the statement that the procedures are 'generally the same' as in that reference. Since the accuracy of the adjoint sensitivities and all downstream optimization results depends directly on this multiscale scheme, and since the adjoint equation (22) involves terms (negative advection, boundary integrals coupling to phi_w, the phi_tau term) that are specific to this problem, the paper should include either the full derivation and discretization of the adjoint system or a substantial appendix describing the numerical scheme. Without this, the central claim is not reproducible from the manuscript alone.","section":"Section 3.1, Eqs. (26)-(27)"},{"comment":"The adjoint governing equation is stated without derivation. The treatment of the diffuse boundary condition, especially the coupling through the wall density rho_w and the associated integral boundary condition, is nontrivial; the finite-difference comparisons in Section 4.1 test the final sensitivity but do not demonstrate that the derivation is correct for other objectives or flow conditions. A derivation, or at least a detailed outline of the variational steps leading to Eq. (22), should be added.","section":"Section 2.5, Eq. (22)"},{"comment":"The agreement with the topology-optimization results of [18] is presented as validation, but it is not independent: both methods use the same multiscale gas-kinetic solver, and the shape optimization is initialized with the chord length set to the [18] optimum, as the paper itself states in Section 4.2. The Re = 200 topology result in [18] is also admitted to be unconverged. This comparison should be reframed as a consistency check between two geometry representations rather than as independent verification of physical accuracy, and the text should be adjusted accordingly.","section":"Section 4.2 and 4.3, Figs. 8 and 13"},{"comment":"The node-coordinate sensitivities show persistent oscillations that do not disappear even at a 300x300 velocity-space discretization, and the paper asserts that CST parameterization removes these oscillations as a 'low-pass filter'. It is not demonstrated, however, whether this filtering removes numerical noise or also removes physically meaningful high-frequency sensitivity that could bias the optimizer. A quantitative convergence study (for example, CST sensitivities at increasing Bernstein-polynomial order, or a comparison against explicitly filtered node sensitivities) would support the claim that the parameterization does not alter the converged optimum.","section":"Section 4.1, Fig. 2"}],"minor_comments":[{"comment":"The finite-difference step size epsilon in Eq. (46) is never specified; please provide the value used and, ideally, a short step-size sensitivity check.","section":"Section 4.1, Eq. (46)"},{"comment":"The FDM results are reported only for the 60x60 velocity discretization; please clarify whether finer FDM runs were attempted and, if not, why the 60x60 comparison was considered sufficient.","section":"Section 4.1, Fig. 2 caption"},{"comment":"The sentence 'It is shown that the our adjoint results agree well with the FDM results' contains a grammatical error and should read 'our adjoint results'.","section":"Section 4.1, text near Fig. 4"},{"comment":"The claim that extension to 3D is 'straightforward' is plausible but not substantiated; at minimum, the 3D analogue of Eq. (37) should be outlined or explicitly referenced.","section":"Section 5"},{"comment":"The convergence criterion (44) depends only on the change in the objective; reporting the constraint feasibility at each iteration would be useful to demonstrate that the volume constraint is actually satisfied at convergence.","section":"Section 3.4, Eq. (44)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the finite-difference sensitivity validation is a strong point. The main risk is not the correctness of the results but the heavy reliance on the authors' own prior paper [18] for the solver details and the absence of an adjoint derivation; both should be addressed for a self-contained methods paper. I would also encourage the editor to consider whether a code/data availability statement is needed, given that no implementation details are otherwise provided."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest take: this is the first adjoint-based shape optimization for the BGK equation with diffuse reflection, and the core sensitivity claim checks out against finite differences at Kn=10, 0.1, and 0.001. I believe the method works. The paper deserves a serious referee, but it is not self-contained: both the adjoint derivation and the multiscale solver are deferred to the authors' earlier work [18], and the validations are all internal to that solver family.\n\nWhat's genuinely new: the continuous adjoint formulation for the diffuse boundary condition, the observation that velocity-space discretization produces sensitivity oscillations that CST parameterization filters out, and demonstration that shape (not topology) optimization converges in a dozen iterations. The finite-difference verification of the sensitivity is real evidence, especially the agreement for CST design variables where the 60x60 velocity mesh already matches FDM. The optimization results—sharp leading edges in rarefied flow, bow shock to oblique shock in the supersonic case, 13.36% drag reduction—are plausible and match the topology optimization from [18] where the comparison is meaningful.\n\nSoft spots, in rough order of importance: 1. The paper states the adjoint equation (22) without derivation, and the numerical schemes are 'generally the same' as [18]. For a methods paper this is a real gap. It may be acceptable if [18] is solid, but the present paper should at least supply the adjoint derivation in an appendix or make clear what is being reused. Also, the FDM comparison validates the consistency of the discrete adjoint with the primal solver, not the absolute accuracy of the BGK solver. An independent benchmark (DSMC or NS in the continuum limit) would make the physical accuracy claim stronger. 2. The topology-optimization comparison is weaker than it looks: the authors initialize the chord length at the [18] optimum and admit the Re=200 baseline never fully converged. So the agreement is suggestive but not conclusive. 3. The velocity-space sensitivity oscillations are observed and fixed, but the fact that raw sensitivities still oscillate at 300x300 makes me want a better explanation than 'white noise.' It doesn't invalidate the method, but it's a loose end.\n\nThe citation pattern is honest: the heavy reuse of [18] is declared, and the admitted non-convergence of the baseline is the kind of transparency you want. No code or data is shipped, which in a numerical methods paper is a missed opportunity.\n\nWho this is for: anyone doing adjoint or gradient-based design for rarefied gas flows. It's a legitimate advance, not a revolution. If I were the editor, I would send it to review with a request to make the derivation and numerical discretization self-contained.","headline":"First adjoint shape-optimization for diffuse-boundary BGK, with real FDM validation, but leaning heavily on [18] and worth refereeing with a request for self-contained derivation.","tokens_in":22225,"tokens_out":2432,"would_cite":true,"duration_ms":21684,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper develops an adjoint shape-optimization method for gas flows that computes accurate drag sensitivities and converges to optimal airfoils in roughly a dozen iterations across Knudsen numbers from 0.001 to 10.","keywords":["shape optimization","adjoint method","rarefied gas flow","discrete velocity method","BGK equation","diffuse boundary condition","CST parameterization","Knudsen number"],"falsifier":"Take one of the optimized airfoils (e.g., the Kn = 0.01 supersonic case) and recompute its drag with an independent, high-accuracy kinetic solver such as direct simulation Monte Carlo under the same Mach number, wall temperature, and Knudsen number; if the recomputed drag is not lower than that of the initial NACA0012 airfoil, or if significant differences appear when the same comparison is done for the other optimized shapes, the adjoint gradient that drove the optimization is wrong.","tokens_in":21254,"feed_emoji":"✈️","tokens_out":7347,"duration_ms":62772,"temperature":0.7,"pith_summary":"The paper tries to establish that shape optimization of solid bodies in gas flows can be done uniformly from the continuum regime to the free-molecular regime by deriving shape sensitivities from the BGK kinetic equation with diffuse-reflection boundary conditions. Previous adjoint optimization for rarefied gases used topology optimization, which requires one design variable per mesh cell and a non-body-fitted boundary; the proposed method uses a body-fitted mesh and parameterizes the boundary with a small number of design variables. The paper claims that the calculated sensitivities match finite-difference sensitivities at Knudsen numbers 0.001, 0.1, and 10, that the CST parameterization removes the sensitivity oscillations caused by velocity-space discretization, and that airfoil drag optimizations converge in roughly a dozen steps in 5 to 20 minutes on 40 to 160 cores. In a Mach-2 case the drag is reduced by 13.36%. If these claims hold, gradient-based design becomes practical for rarefied and multiscale gas flows without solving expensive surrogate problems.","feed_headline":"Airfoil drag cut 13% by adjoint method for all gas regimes","feed_subtitle":"Shape optimization that spans Knudsen numbers from 0.001 to 10 and converges in about a dozen steps.","key_machinery":"The load-bearing object is the adjoint BGK system, specifically the adjoint distribution function $\\varphi$ and its wall flux $\\phi_w$; the adjoint equation propagates sensitivity information backwards through the flow, and the wall boundary condition converts the objective into a constant flux source $m_J$. Around this, the method combines continuous adjoint analysis (to derive that equation) with discrete-adjoint-style differentiation (to convert derivatives with respect to surface element area, centroid, and normal into derivatives with respect to boundary nodes), and then a chain rule maps those node derivatives onto CST parameters. The CST parameterization is more than a geometry representation: it acts as a low-pass filter that removes the high-frequency sensitivity oscillations induced by discretizing the molecular velocity space, making the gradient usable by the optimizer.","core_discovery":"The central discovery is that the adjoint of the BGK equation, together with the diffuse-reflection boundary condition, provides shape sensitivities for gas flows at every Knudsen number, and that the velocity-space discretization noise in those sensitivities is eliminated by parameterizing the boundary with CST shape functions. The paper derives a Lagrangian for the objective (drag or heat) constrained by the steady BGK equation, obtains a continuous adjoint equation whose wall boundary condition contains the objective moment $m_J$ as a flux source, and then computes the final sensitivity by discrete-like differentiation of the discretized objective and adjoint boundary terms with respect to the boundary element's area, centroid position, and outward normal. This sensitivity is transferred through a chain rule to CST design variables and fed to an SLSQP quasi-Newton optimizer. Numerically, the paper shows that the adjoint sensitivity agrees with central finite differences under the same velocity-space discretization, that the CST parameterization yields smooth and mesh-converged sensitivity profiles, and that the optimized airfoil shapes match those from a previous topology-optimization method.","pith_inferences":["Beyond the paper: the same hybrid continuous/discrete adjoint structure could be applied to other kinetic models, such as the Shakhov or ellipsoidal BGK equations, with extra collision terms appearing in the adjoint equation; nothing in the derivation is specific to the relaxation-time model.","Beyond the paper: if CST smoothing is understood as a low-pass filter on a noisy gradient, then coarser velocity grids might be usable in optimization without losing accuracy, at the price of reparameterizing the boundary; this could cut cost further in three dimensions.","Beyond the paper: the paper's claim that extension to 3D only requires the 3D analogue of the geometric derivative (equation 37) is plausible, but the velocity-space oscillations that had to be smoothed in 2D may be more severe in 3D because the velocity grid is much larger.","Beyond the paper: the non-monotonic optimal-thickness trend with Knudsen number (maximum thickness at Kn = 0.1) is a prediction about rarefied aerodynamics that could be tested experimentally or with high-fidelity direct simulation Monte Carlo in a controlled flow."],"forward_implications":["Drag-reduction optimizations of airfoils in a channel converge in 9 to 18 optimization steps at Kn = 0.001, 0.1, and 10, with total runtimes of 6 to 20 minutes on 40 to 160 cores.","Under Mach 2 flow, the method reduces drag by 13.36% at Kn = 0.01 and 4.22% at Kn = 0.5, with the optimized airfoil converting a detached bow shock into two oblique shocks.","The method reproduces the optimal shapes obtained by the earlier topology-optimization approach of Reference [18] for several flow conditions, despite using a completely different geometry representation.","Because the boundary is body-fitted and the design space is low-dimensional (18 CST parameters in the tested setup), quasi-Newton optimization is feasible without the hundreds of design variables that hinder topology optimization.","The sensitivity validation against finite differences across Kn = 0.001 to 10 indicates that the adjoint derivation, including the diffuse boundary condition, is algebraically correct in all flow regimes."],"supporting_citations":[{"why":"Supplies the implicit multiscale gas-kinetic schemes used to solve the primal and adjoint BGK equations, and the topology-optimization results that the present optimized shapes are compared against.","marker":"[18]"},{"why":"Provides the interface-flux treatment of the discrete unified gas-kinetic scheme that keeps the solver accurate across Knudsen regimes.","marker":"[28]"},{"why":"Introduces the class/shape-function transformation used to parameterize the airfoil boundary and to suppress sensitivity oscillations.","marker":"[32]"},{"why":"Supplies the radial-basis-function interpolation used to deform the volume mesh after each shape update.","marker":"[34]"},{"why":"Provides the SLSQP quasi-Newton optimizer that consumes the computed sensitivities and drives the design updates.","marker":"[35]"},{"why":"Describes the sequential quadratic programming method on which the optimizer's update and constraint handling rest.","marker":"[36]"},{"why":"Earlier adjoint-based topology optimization for the Boltzmann equation, the approach the paper extends to shape optimization on body-fitted meshes.","marker":"[16]"},{"why":"Earlier adjoint-based topology optimization using adjoint IP-DSMC, showing the need for a shape-based alternative with fewer design variables.","marker":"[17]"},{"why":"Defines the BGK relaxation model that serves as the governing equation for both the primal and adjoint systems.","marker":"[22]"}],"fun_headline_variants":["Adjoint shape optimization for all gas regimes converges fast","Airfoil drag reduced via adjoint BGK in a dozen steps","Smooth adjoint sensitivities enable rapid airfoil shape design","Continuum to free-molecular flow shape optimization in 12 iterations","Adjoint method cuts airfoil drag across Knudsen numbers 0.001–10"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The governing assumption is that the implicit multiscale gas-kinetic scheme adopted from Reference [18] works correctly for the adjoint BGK system; the paper states that the detailed computation procedures are omitted because they are 'generally the same' as in [18], so if that solver is inaccurate for the adjoint equation, the sensitivities and therefore all optimized shapes would be wrong.","fun_headline_variants_meta":{"raw":{"variants":["Adjoint shape optimization for all gas regimes converges fast","Airfoil drag reduced via adjoint BGK in a dozen steps","Smooth adjoint sensitivities enable rapid airfoil shape design","Continuum to free-molecular flow shape optimization in 12 iterations","Adjoint method cuts airfoil drag across Knudsen numbers 0.001–10"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000266,"raw_usage":{"total_tokens":1624,"prompt_tokens":974,"completion_tokens":650,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":555}},"tokens_in":590,"tokens_out":650,"duration_ms":6835,"temperature":1.0,"reasoning_tokens":555,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:39:37.099702+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one of the optimized airfoils (e.g., the Kn = 0.01 supersonic case) and recompute its drag with an independent, high-accuracy kinetic solver such as direct simulation Monte Carlo under the same Mach number, wall temperature, and Knudsen number; if the recomputed drag is not lower than that of the initial NACA0012 airfoil, or if significant differences appear when the same comparison is done for the other optimized shapes, the adjoint gradient that drove the optimization is wrong.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the implicit multiscale gas-kinetic schemes used to solve the primal and adjoint BGK equations, and the topology-optimization results that the present optimized shapes are compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the interface-flux treatment of the discrete unified gas-kinetic scheme that keeps the solver accurate across Knudsen regimes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the class/shape-function transformation used to parameterize the airfoil boundary and to suppress sensitivity oscillations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the radial-basis-function interpolation used to deform the volume mesh after each shape update."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the SLSQP quasi-Newton optimizer that consumes the computed sensitivities and drives the design updates."},{"cited_title":"Kraft, A software package for sequential quadratic prog ramming, Forschungsbericht- Deutsche Forschungs- und Versuchsanstalt fur Luft- und Raumfahrt (19 88)","cited_arxiv_id":null,"evidence_quote":"Describes the sequential quadratic programming method on which the optimizer's update and constraint handling rest."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier adjoint-based topology optimization for the Boltzmann equation, the approach the paper extends to shape optimization on body-fitted meshes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier adjoint-based topology optimization using adjoint IP-DSMC, showing the need for a shape-based alternative with fewer design variables."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the BGK relaxation model that serves as the governing equation for both the primal and adjoint systems."}],"review_version":1}