{"id":"b85201d2-b1a8-442d-b77e-e4cd103c8419","arxiv_id":"2411.19109","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A level-set plus adjoint-state optimization framework, implemented in the NeuralMag micromagnetic solver, designs magnonic devices such as a frequency-selective YIG demultiplexer with constant memory cost.","lead":"This paper combines a level-set geometry parameterization with an adjoint-state gradient method to optimize the shape of nanoscale magnetic devices inside a GPU micromagnetic solver. It demonstrates the approach by designing a spin-wave demultiplexer that routes 2.6 GHz and 2.8 GHz signals to separate outputs, and by shaping a nanoparticle to a target hysteresis curve.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The adjoint equations in Eqs. (6)-(7) describe a terminal-cost adjoint, but the FFT and field-sweep objectives in Eqs. (10) and (8) are distributed in time, so the printed gradient formula omits the necessary source term.","rationale":"The reader's weakest assumption was that the adjoint equations yield the exact gradient of the objective, including damping and smoothing, and that this remains unverified. My reading sharpens that into a specific mathematical omission: the printed adjoint is for a terminal cost only, while both demonstrated objectives depend on the full time evolution (FFT windows in Eq. (10), field-ramp sums in Eq. (8)). This is more concrete than a general lack of finite-difference verification, and it directly targets the 'universal tool' claim. I do not claim the numerical implementation is fraudulent; the discrepancy may be an undocumented state augmentation or a shorthand in the derivation. The paper earns credit for two convergent demonstrations, a frequency-sweep robustness check, and multiple initial designs, but those results cannot certify the gradient method if the published equations are inconsistent with the objective. A single gradient check on a coarse design would settle the matter, so the appropriate disposition remains conditional rather than a flat rejection or acceptance. The verdict is therefore unchanged from the reader's CONDITIONAL, with the concrete test above as the condition.","tokens_in":12661,"tokens_out":8196,"duration_ms":75303,"concrete_test":"At one early optimization step of the 4×4-hole demultiplexer setup, compute ∂J/∂s_i for a subset (e.g., 8–16) of the 400 RBF amplitudes by three routes: (i) central finite differences of the full time-domain FFT objective in Eq. (10) with epsilon≈1e−4 in normalized s, (ii) the adjoint defined by Eqs. (6)-(7), and (iii) the augmented adjoint with source term −∂g/∂m, where g is the windowed FFT integrand. If route (ii) matches (i) to about 1% relative error, the concern is resolved and the paper should state how the FFT source term was included. If route (iii) matches (i) but route (ii) does not, the published adjoint equations are incorrect and must be corrected before the method can be reproduced. Repeat once for the field-sweep objective in Eq. (8).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of a memory-efficient level-set/adjoint optimizer depends on the gradients used by Eq. (5) being gradients of the stated objective. As written, Eq. (6) is the adjoint for a terminal cost: a(T)=∂J/∂m(T) and a′=−a^T ∂L/∂m. The objectives are not terminal. In the demultiplexer, J in Eq. (10) is a sum of FFT amplitudes over frequency windows, so ∂J/∂m(t) is nonzero throughout the simulation time window; in the particle example, Eq. (8) sums over field-ramp steps. The correct continuous adjoint for J=∫ g(t,m,s)dt contains a source term −∂g/∂m and has a(T)=0 (plus any terminal contribution), but no such term appears in Eq. (6), and Eq. (7) integrates only a^T ∂L/∂s. If the implementation literally follows Eqs. (6)-(7), the optimizer is not descending the stated objective, so the demultiplexer separation in Fig. 6 would not validate the claimed method. If instead the implementation augments the state or uses a stored trajectory to backpropagate through the FFT, that mechanism is undocumented, and the 'constant memory in simulation time' claim would need qualification because the FFT time series must be accumulated. The missing source term is an internal inconsistency, not a matter of external consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a level-set method for inverse topology optimization of magnonic devices, parameterizing the level-set function by radial basis functions and mapping it to magnetic/nonmagnetic regions through a smooth sigmoid. The design parameters (RBF amplitudes) are updated by gradient descent, with gradients computed via an adjoint-state solution of the Landau-Lifshitz-Gilbert equation, implemented in the GPU-accelerated NeuralMag solver with torchdiffeq. The method is validated on two tasks: optimizing a nanoparticle's hysteresis curve to match a hard-axis linear response, and designing a 300-nm-wide YIG demultiplexer that separates 2.6 GHz and 2.8 GHz spin waves. The paper reports smooth convergence, topology-splitting behavior, frequency-selective output separation, robustness to different initial hole configurations, and a frequency operating window of about 70 MHz.","tokens_in":12922,"tokens_out":5570,"duration_ms":53742,"significance":"If the adjoint-gradient computation is correct, the work is a valuable contribution to inverse design in magnonics: it introduces a level-set parameterization that handles topology changes, combines it with a micromagnetic solver, and demonstrates application to a realistic nanoscale device. The multiple-initialization study and the frequency-sweep robustness check strengthen the demultiplexer claim. The demultiplexer separation is, of course, a direct consequence of the chosen objective and thus validates the optimizer rather than predicting new physics; this is normal for inverse-design papers and is not itself a flaw. The main value would be as a reusable computational framework, but the paper as written does not make the code available and, more importantly, contains an inconsistency in the adjoint equations relative to the stated objectives.","major_comments":[{"comment":"The adjoint system in Eq. (6) is written for a terminal-cost objective: the backward equation has no source term, and the terminal condition is given as a(T)=∂J/∂m(T). The objectives actually used in the paper are distributed in time: Eq. (8) sums over field-ramp steps j=0...N, and Eq. (10) sums FFT amplitudes over frequency windows, so ∂J/∂m(t) is nonzero throughout the integration interval. The correct continuous adjoint for an objective J=∫g(t,m,s)dt contains a source term -(∂g/∂m)^T in the adjoint equation and has a(T)=0 (plus any terminal contribution), and the gradient formula must include the corresponding source contribution. As printed, Eq. (7) omits the running-cost contribution to ∂J/∂s. This is an internal inconsistency: if the implementation literally follows Eqs. (6)-(7), the optimizer is not descending the stated objectives, and the demultiplexer results would not validate the claimed method. The authors should either correct the adjoint equations to include the source terms, or document the actual augmented-state/checkpointing/automatic-differentiation procedure used, and provide a finite-difference gradient check for at least one small problem.","section":"Optimisation with adjoint method (Eqs. 6-7)"},{"comment":"The memory-efficiency claim that the adjoint approach ensures constant memory as a function of simulation time is not demonstrated for the FFT-based objective of Eq. (10). Computing the FFT of the magnetization at the output cells requires either storing the time series over the simulation window or checkpointing the state, so the memory footprint is not trivially constant in simulation time unless a specific streaming or partial-FFT strategy is used. The paper should quantify the memory scaling with the number of time steps and describe how the FFT objective is incorporated in the adjoint computation. This point is load-bearing for the 'versatile and universal tool' claim in the Discussion.","section":"Demultiplexer example (Eq. (10), Discussion)"},{"comment":"No numerical verification of the adjoint gradient against brute-force finite differences or forward-mode automatic differentiation is reported. Because the paper introduces an optimization framework and the gradient formula is at issue (see previous comment), a gradient check for a small problem (e.g., the particle case with a reduced number of RBFs) should be included. Such a check would resolve whether the smooth convergence and optimized geometries in Figs. 3-7 are genuine minima of the stated objectives or artifacts of the optimizer operating on a different functional.","section":"Methods / all examples"}],"minor_comments":[{"comment":"In the paragraph after Eq. (10), the sentence 'These frequency windows are illustrated in Fig. 6(a, b)' should presumably refer to Fig. 6(f, g), since the colored frequency windows appear in the FFT spectra panels; Fig. 6(a, b) show the objective evolution and propagation maps.","section":"Results, Demultiplexer example"},{"comment":"The summation is written from j=0 to N, which gives N+1 terms, while the text says the summation is performed over N simulation steps; please clarify the indexing.","section":"Eq. (8)"},{"comment":"The statements that p=90 and a=50 introduce negligible error are not quantified. A brief sensitivity study or convergence check with respect to p and a would support this assertion.","section":"Eqs. (3)-(4)"},{"comment":"The statement 'No datasets were generated or analysed during the current study' is inconsistent with the reported simulation results; please clarify what simulation data, parameter files, and post-processing scripts are available, especially given that the code is not public.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The missing source term in the printed adjoint equations is the central issue. If the implementation truly follows Eqs. (6)-(7), the reported optimized designs may correspond to minima of different functionals, which would undermine the central claims. This is fixable in principle by correcting the adjoint derivation or documenting the actual implementation, and the framework itself seems promising, so I recommend major revision rather than rejection. A gradient check against finite differences is essential and should be required."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this for the method integration: level-set parameterization with RBFs and p-norm, combined with adjoint-state gradients in NeuralMag, is a sensible and useful addition to micromagnetic inverse design. The two test problems—particle shape optimization and a spin-wave demultiplexer—show the optimizer can handle topological changes (hole nucleation, merging) and converge from different initial designs. The demultiplexer separation is exactly what the objective asks for, which is fine, and the 70 MHz operating window and the multiple initializations are legitimate robustness checks.\n\nThe load-bearing soft spot is in the adjoint derivation. Equations (6)-(7) describe an adjoint for a terminal cost: a(T)=∂J/∂m(T) and a'=-a^T ∂L/∂m. But the objectives used are not terminal. The demultiplexer objective integrates FFT amplitudes over the whole time window, and the particle objective sums over field steps. A correct adjoint for those needs a source term involving ∂J/∂m(t), or an augmented-state formulation. The paper shows neither. If the implementation literally follows the printed equation, the gradients are not for the stated J; if it uses torchdiffeq's built-in adjoint with augmentation, that mechanism is undocumented. Either way, the math as presented is internally inconsistent, and the constant-memory claim depends on the missing details. A finite-difference gradient check would settle this.\n\nReproducibility is limited: no code or data, and several key parameters (damping, excitation amplitude, FFT window widths, learning rate schedule) are not given. The paper's closing claim of being a 'universal tool' is stronger than the evidence supports.\n\nThis deserves a serious referee, but revision is needed. The integration is valuable, the results are plausible, but the adjoint equations must be corrected or clarified, and the gradient should be verified. After that, it would be a solid methods paper for the magnonics community.","headline":"A useful integration of level-set topology optimization with adjoint gradients for magnonics, but the printed adjoint equations don't match the time-distributed objectives—needs a gradient check and better documentation.","tokens_in":13518,"tokens_out":4334,"would_cite":false,"duration_ms":40403,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper demonstrates a level-set and adjoint-state optimization framework that designs nanoscale magnonic devices—including a spin-wave demultiplexer separating 2.6 and 2.8 GHz signals—with memory cost independent of simulation time.","keywords":["inverse design","level-set method","adjoint-state method","topology optimization","magnonics","spin-wave demultiplexer","micromagnetic simulation","yttrium iron garnet"],"falsifier":"Take the 30×30 particle design task, leave the discretization and objective unchanged, and compare the adjoint gradient against a central finite-difference gradient for a handful of radial-basis amplitudes at the first optimization step; any relative discrepancy well above the solver's own tolerance would falsify the gradient premise. A second, independent check: fix the design region, double the simulated time for the demultiplexer, and monitor peak GPU memory—a noticeable rise with time would falsify the constant-memory claim.","tokens_in":12400,"feed_emoji":"🧲","tokens_out":10002,"duration_ms":89600,"temperature":0.7,"pith_summary":"The paper aims to make inverse design practical for magnonic devices by combining a level-set description of material boundaries with an adjoint-state method for computing gradients, and it claims this combination removes the main memory bottleneck that limited earlier approaches. The boundary between magnetic and non-magnetic regions is written as the zero contour of a smooth function built from radial basis functions, so the geometry can change topology freely while remaining differentiable. Gradients of a device-performance objective are obtained by solving an adjoint equation backward in time, which keeps memory use constant as simulation time grows. The method is demonstrated on two tasks: shaping a magnetic particle to follow a target hysteresis curve, and carving air holes in a yttrium-iron-garnet strip so that 2.6 GHz and 2.8 GHz spin waves separate into different output conduits. The paper concludes that the framework is a versatile and universal tool for inverse design in magnonics.","feed_headline":"In 100 steps, optimizer designs a spin-wave frequency splitter","feed_subtitle":"Backward-in-time gradient math keeps memory flat while the optimizer carves a magnetic chip that separates two close spin-wave frequencies.","key_machinery":"The load-bearing machinery is the pair formed by the level-set parameterization and the adjoint-state gradient formula. The level-set function $\\Phi(x,y)$ is built from $n$ radial basis functions $g_i(x,y)$ such that $\\Phi = (\\sum_i g_i^p + \\Delta\\phi)^{1/p} - \\Delta\\phi$ with $p=90$; its zero contour defines the boundary, and a sigmoid $\\psi = 1/(1+\\exp(-a\\Phi))$ with $a=50$ converts it into smooth material parameters. The adjoint-state method solves the system $\\partial m/\\partial t = L(t,m,s)$ and $\\partial a/\\partial t = -a^T \\partial L/\\partial m$ backward in time, then forms $\\partial J/\\partial s = -\\int_T^{t_0} a(t)^T \\partial L/\\partial s\\,dt$; this replaces step-by-step backpropagation with a second time integration, so intermediate magnetization states need not be stored. That integral and the backward integration are evaluated by automatic differentiation, and the amplitudes are updated by a gradient-descent optimizer with adaptive moments. Together these pieces let the algorithm nucleate, merge, and remove holes while keeping memory independent of simulation time.","core_discovery":"The central claim is that level-set topology optimization combined with the adjoint-state method can design functional nanoscale magnonic devices with a memory cost that does not grow with simulation time. The level-set function is assembled from radial basis functions through a differentiable p-norm approximation of the maximum, and a sigmoid maps it to material parameters, making the geometry a smooth function of optimizable amplitudes. The adjoint-state approach computes the gradient of the objective by integrating a second dynamics equation backward in time and an integral formula, rather than backpropagating through every step of the forward simulation. On the two test problems, the framework converges smoothly: the particle optimization reaches a stable objective value within 100 steps and can split a single shape into multiple particles, and the demultiplexer optimization separates spin waves at the two target frequencies with output amplitudes differing by an order of magnitude and remains functional over a roughly 70 MHz window. The authors state that these results validate the software as a versatile and universal tool for future inverse-design studies in magnonics.","pith_inferences":["The adjoint-gradient premise can be tested cheaply on the 30×30 particle case by comparing against central finite differences; if it holds, the method should transfer to any objective built from local magnetization measurements.","Because the optimizer finds distinct local minima from different initial hole patterns, pairing the level-set descent with a global search (e.g., multi-start or restart with random initial amplitudes) could systematically find better-performing geometries.","The paper's extension list—3D structures, chiral exchange interactions, and finite-temperature stochastic dynamics—suggests the same machinery could be applied to design devices based on topological spin textures or thermally stable memory elements, though these are not demonstrated.","A direct benchmark against the earlier binary-search demultiplexer on identical geometry, tracking number of simulations to convergence and final output contrast, would quantify the claimed efficiency advantage."],"forward_implications":["Because only the objective function is device-specific, the same optimizer can be pointed at other magnonic functions—mode multiplexing, directional coupling, or logic gates—without changing the machinery.","Devices designed this way tolerate fabrication errors: the demultiplexer keeps separating spin waves when frequencies deviate by up to roughly 70 MHz from the design values.","Longer or more detailed spin-wave simulations become affordable, since the adjoint method holds memory use constant in simulation time.","Topology changes such as hole nucleation and merging are handled automatically, so the optimizer can start from a minimal design and grow only the features that are needed.","The resulting geometries have smoother, rounder edges than binary-search designs, which is advantageous for lithographic fabrication."],"supporting_citations":[{"why":"Establishes the demultiplexer task and provides the binary-search baseline this work replaces.","marker":"[17]"},{"why":"Introduces level-set front propagation, the geometric representation used throughout.","marker":"[22]"},{"why":"Reviews level-set methods for structural topology optimization, including renormalization and handling of topological changes.","marker":"[23]"},{"why":"Supplies the adjoint-method tutorial and equation framework used for gradient computation.","marker":"[26]"},{"why":"Provides the adjoint-state method review whose backward-equation and integral formula are used here.","marker":"[31]"},{"why":"Gives the neural-ODE adjoint technique and its implementation for constant-memory backward integration and gradient evaluation.","marker":"[32]"},{"why":"Describes the micromagnetic solver used for all simulations and gradients.","marker":"[33]"},{"why":"Source of the parametric level-set approach whose p-norm maximum approximation is used in Eq. (3).","marker":"[38]"}],"fun_headline_variants":["Level-set optimizer carves spin-wave splitter in 100 steps","Memory-light adjoint-state design for magnonic logic","100-step topology optimization yields spin-wave demux","Backward-time gradients enable efficient magnonic design","Inverse design of magnonic devices with level-set"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the adjoint-state equations, solved backward in time, return the exact gradient of the objective with respect to the level-set parameters for the damped magnetization dynamics and the smoothed sigmoid mapping; the paper provides no comparison with a brute-force finite-difference gradient, so if the adjoint gradient is biased, the reported convergence and geometries could be optimizer artifacts.","fun_headline_variants_meta":{"raw":{"variants":["Level-set optimizer carves spin-wave splitter in 100 steps","Memory-light adjoint-state design for magnonic logic","100-step topology optimization yields spin-wave demux","Backward-time gradients enable efficient magnonic design","Inverse design of magnonic devices with level-set"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000751,"raw_usage":{"total_tokens":3328,"prompt_tokens":918,"completion_tokens":2410,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":2333}},"tokens_in":534,"tokens_out":2410,"duration_ms":38811,"temperature":1.0,"reasoning_tokens":2333,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:31:17.424646+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the 30×30 particle design task, leave the discretization and objective unchanged, and compare the adjoint gradient against a central finite-difference gradient for a handful of radial-basis amplitudes at the first optimization step; any relative discrepancy well above the solver's own tolerance would falsify the gradient premise. A second, independent check: fix the design region, double the simulated time for the demultiplexer, and monitor peak GPU memory—a noticeable rise with time would falsify the constant-memory claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the demultiplexer task and provides the binary-search baseline this work replaces."},{"cited_title":"& Sethian, J","cited_arxiv_id":null,"evidence_quote":"Introduces level-set front propagation, the geometric representation used throughout."},{"cited_title":"P., Maute, K., Langelaar, M","cited_arxiv_id":null,"evidence_quote":"Reviews level-set methods for structural topology optimization, including renormalization and handling of topological changes."},{"cited_title":"A tutorial on the adjoint method for inverse problems","cited_arxiv_id":null,"evidence_quote":"Supplies the adjoint-method tutorial and equation framework used for gradient computation."},{"cited_title":"T., Rubanova, Y., Bettencourt, J","cited_arxiv_id":null,"evidence_quote":"Gives the neural-ODE adjoint technique and its implementation for constant-memory backward integration and gradient evaluation."},{"cited_title":"NeuralMag: an open-source nodal finite-difference code for inverse micromagnetics","cited_arxiv_id":"2411.11725","evidence_quote":"Describes the micromagnetic solver used for all simulations and gradients."},{"cited_title":"& Maute, K","cited_arxiv_id":null,"evidence_quote":"Source of the parametric level-set approach whose p-norm maximum approximation is used in Eq. (3)."}],"review_version":1}