{"id":"0e6b9bf6-c86d-46b4-b1be-d43541037373","arxiv_id":"2602.18477","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"A two-stage optimizer (iterative time reversal plus genetic-algorithm temporal interference) focuses 700-MHz amplitude-modulated fields at chosen deep-brain points in a 77-tissue voxel head model.","lead":"This paper simulates a non-invasive way to reach deep brain targets: an optimized array of microwave antennas whose signals beat together to create a slow, focused electrical envelope inside a realistic head model. If the principle works in biology, it could offer deep-brain stimulation without drilling through the skull.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (17)'s E_AM is polarization-blind; the TI optimizer maximizes an upper bound, so the reported focal envelope may not be the low-frequency envelope neurons experience.","rationale":"The reader correctly identifies the microwave-TI biological premise and field-strength uncertainty as important. However, the most directly load-bearing weakness of the quantitative claim is more specific: Eq. (17) is used as if the envelope of a vector sum were determined by the norms of the two frequency phasors. The optimizer does not enforce any relationship between E_f1 and E_f2 at the target, so the reported focal contrast could be an artifact of an upper-bound metric. This is not a matter of consensus; it is a correctness risk inside the paper's own electromagnetic and signal-processing definitions, and it can be settled from the existing simulation outputs. The missing SAR/perturbation/statistical analyses and the normalized-only figures are additional concerns, but they do not strike as directly at the central claim. The appropriate outcome remains conditional: the paper should recompute the physiologically relevant envelope and report whether the focusing survives. I therefore retain the reader's CONDITIONAL verdict, with a sharper technical condition.","tokens_in":15339,"tokens_out":19960,"duration_ms":225900,"concrete_test":"Using the optimized arrays from the center- and side-focus cases, reconstruct for each voxel the instantaneous vector field E(r,t)=Re[E_f1(r)e^{j2π f1 t}] + Re[E_f2(r)e^{j2π f2 t}] over one 10 ms beat period. For a dense set of unit vectors u spanning possible axon orientations, form the complex envelope at the average carrier, A_u(t)=u·E_f1 e^{jπΔf t} + u·E_f2 e^{-jπΔf t}, and compute its slow-envelope swing max_t|A_u(t)| - min_t|A_u(t)|. Define the effective envelope at each voxel as the maximum over u. Compare the focal peak and focal-to-background ratio with the E_AM maps in Figs. 13 and 14, and report the angle between E_f1(r_f) and E_f2(r_f) at the focal point. If the effective envelope is substantially below E_AM at the focus (>20%) or the focal-to-background ratio degrades, Eq. (17) is an inadequate surrogate for the TI envelope and the central claim requires re-analysis.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central demonstration uses Eq. (17), E_AM = 2 min(||E_f1||, ||E_f2||), which depends only on the per-frequency norms and is insensitive to the vector directions and relative phase of the two fields. For a neuron with axis u, the scalar driving signal is s_u(t) = u·Re(E_f1 e^{jω1t}) + u·Re(E_f2 e^{jω2t}); its slow-envelope swing is 2 min(|u·E_f1|, |u·E_f2|), not 2 min(||E_f1||, ||E_f2||). Since ||E_fi|| ≥ |u·E_fi|, Eq. (17) is an upper bound over neuron orientations. For equal-magnitude fields orthogonal at the focus, the best achievable projection envelope is √2 A, while Eq. (17) reports 2A. The objective JTI (Eq. (19)) and all E_AM maps use this norm-based quantity, so the GA has no incentive to make E_f1 and E_f2 parallel at the target. The paper itself postpones neuronal-orientation effects to future work in Sec. V-A2. Therefore the 'focused TI fields' may be a focusing of a norm-derived upper bound rather than of the actual low-frequency envelope that would drive a neuron. This is an internal, checkable issue, not merely an extrapolation beyond current biological consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a computational proof-of-principle for non-invasive microwave deep brain stimulation using an external array of magnetic point dipoles. A two-stage pipeline is presented: iterative time reversal (iTR) selects antenna positions and orientations, and a genetic algorithm optimizes amplitudes, phases, and frequency assignment to minimize the background-to-focal ratio of the temporal-interference envelope E_AM defined in Eq. (17). The method is evaluated in a 77-tissue voxel model of the Duke head for a center focus and a side focus at 700 MHz and Δf = 100 Hz. The central claim is that amplitude-modulated microwave fields can be effectively focused at deep brain targets, with normalized field plots as the main evidence.","tokens_in":15706,"tokens_out":4413,"duration_ms":53193,"significance":"If the central claim holds, the work would be a useful computational framework combining iTR and TI in a realistic head model, with potential value as a benchmark for non-invasive DBS. Strengths include the use of the IT'IS Duke model, a large number of tissue types, full-wave COMSOL simulations, and a clearly described optimization pipeline. The method description is sufficiently detailed to allow reproduction in principle. However, the quantitative support is currently incomplete, and the TI envelope metric in Eq. (17) is polarization-blind, which materially weakens the link between the reported field plots and the actual neuronal stimulation that the paper claims to demonstrate. These issues significantly reduce the current support for the central conclusion.","major_comments":[{"comment":"The TI envelope is defined as E_AM = 2 min(||E_f1||, ||E_f2||), which depends only on field magnitudes. For a neuron with axis u, the slow envelope of the projection u·(E_f1 e^{jω1t} + E_f2 e^{jω2t}) is 2 min(|u·E_f1|, |u·E_f2|), not the norm-based quantity. Since ||E_fi|| ≥ |u·E_fi|, Eq. (17) is an upper bound over neuron orientations; for orthogonal equal-magnitude fields at the focus it reports 2A while the best achievable projection envelope is √2 A. The objective J_TI in Eq. (19) and all E_AM maps (Figs. 13–14) therefore optimize and report an upper bound, not the low-frequency envelope that would actually drive a neuron. This is an internal consistency problem, not merely a biological extrapolation. The authors acknowledge in Sec. V-A2 that orientation effects are left to future work, but that acknowledgment does not rescue the current central claim. The objective should be revised","section":"§II-B, Eq. (17)"},{"comment":"The abstract states that 'Systematic numerical studies, including perturbation analysis and statistical evaluation, demonstrate consistent spatial localization and robustness across all reported configurations' and that 'safety is quantified using specific absorption rate (SAR), ensuring compliance with exposure limits.' The full text contains no perturbation analysis, no statistical evaluation, no numeric SAR values, and no comparison with IEEE or ICNIRP limits. The results section reports only normalized field plots. These are load-bearing claims in the abstract and must either be added to the manuscript or removed from the abstract. As written, the abstract overstates the content of the paper.","section":"Abstract and §IV, §VI"},{"comment":"The focal maxima shown in Figs. 10–14 are the fitted optima of objective functions that explicitly reward concentration of the optimized quantity in the focal region: J_iTR combines HIR and FIQ in Eq. (11), and J_TI minimizes the background-to-focal envelope ratio in Eq. (19). Normalized plots alone therefore provide no independent evidence of focusing performance. The authors should report the numerical values of HIR, FIQ, and J_TI for both targets, compare them with a non-optimized or single-frequency baseline, and state the achieved background/focal suppression. Without these quantitative results, the claim that the method 'effectively focuses' is largely a restatement of the optimization objective rather than a measured outcome.","section":"§IV, Eqs. (11), (19)"},{"comment":"The medical safety section is entirely qualitative. It asserts that SAR and thermal effects can be kept within limits and that therapeutic activation occurs below heating thresholds, but no actual SAR or temperature calculation is presented anywhere in the paper. Since the abstract explicitly promises safety quantification through SAR, this is a missing load-bearing analysis. The authors should compute and report SAR distributions for the optimized arrays and assess them against the ICNIRP/IEEE limits for head tissues, including off-target regions.","section":"§VI"}],"minor_comments":[{"comment":"Typos and formatting issues: 'V oxelized' appears in several places; 'V .' is used before reference numbers in the bibliography; the caption of Fig. 9 says 'dipole direction' but the displayed quantity is not described in the text.","section":"Throughout"},{"comment":"The statement that 'GA outperforms PSO in both runtime and final objective values' is given without any supporting comparison data. Either report the comparison or remove the claim.","section":"§II-B"},{"comment":"The values of the iTR weights c1 and c2, the attenuation-compensation exponent b, and the GA settings (population size, number of generations, termination criteria) are not specified. These parameters are needed for reproducibility and to understand the sensitivity of the results.","section":"§II-A and §II-B"},{"comment":"The units of the field magnitudes in Figs. 10–14 are labelled '(V/m)' but the fields are normalized and the absolute scaling is not defined in the text. This is confusing; either provide absolute values or remove the unit labelling.","section":"§IV"}],"recommendation":"major_revision","confidential_remarks":"The paper presents an interesting and broadly credible computational framework, but the central demonstration is undercut by the polarization-blind TI metric in Eq. (17) and by the absence of the quantitative analyses promised in the abstract. These issues are substantive but fixable: a vector-aware TI objective and the promised SAR/robustness evaluations are within the scope of a revision. I would encourage the editor to request a major revision rather than reject, provided the authors are willing to recompute the results with a physically appropriate envelope definition."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — quick read of arXiv:2602.18477. The paper is a simulation proof-of-principle combining iterative time reversal (iTR) antenna placement with a genetic-algorithm temporal interference (TI) optimization for microwave DBS, evaluated in the 77-tissue Duke voxel head. That specific combination appears new, and the full-wave treatment with 2 mm voxels is a step beyond earlier spherical or few-tissue models. The pipeline is clearly described, and the iTR stage as a physically informed initialization seems reasonable.\n\nThe soft spots are real. The abstract promises perturbation analysis, statistical evaluation, and SAR-compliant safety quantification. The full text contains neither the analysis nor numeric SAR values; all results are normalized field plots. That mismatch alone would justify a major revision. There are also no absolute field strengths, no FIQ/HIR numbers, and no baseline comparison, so 'demonstrated effective focusing' rests on visual inspection.\n\nThe larger issue is internal to the TI objective. Eq. (17) defines E_AM = 2 min(||E_f1||, ||E_f2||), a norm-based quantity. But the signal a neuron experiences is the projection of the field along its axis, so the slow envelope is 2 min(|u·E_f1|, |u·E_f2|). Because the norm is an upper bound on any projection, the optimizer can produce focal fields that are large in norm yet poorly aligned with any real axon orientation. The paper itself kicks axonal orientation to future work, but as written the claim of 'selective neural stimulation' is a claim about the wrong quantity. The authors say MPD results are a lower bound on array performance; the polarization blindness actually makes it an upper bound on neuromodulatory envelope. So the central conclusion is overstated.\n\nNone of this kills the idea. The iTR+TI pipeline is worth testing with a projection-aware objective, numeric metrics, and realistic field strengths. But as submitted, the evidence doesn't support the abstract's claims. I'd send it to peer review, because the method and model are serious and reproducible in principle; but the referee report should require the missing quantitative analysis, a polarization-aware objective or an explicit frame as an upper-bound focusing study, and honest rewriting of the abstract. With those changes it could be a solid contribution. For my own work I wouldn't cite it yet; I'd bring it to a reading group as a good case study in why the optimization objective has to match the physiology.","headline":"A useful proof-of-principle for microwave TI focusing, but the headline claim runs ahead of what the simulations actually show—especially since the optimized 'envelope' is a norm-based upper bound that ignores field polarization.","tokens_in":16199,"tokens_out":2791,"would_cite":false,"duration_ms":34078,"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":"Microwave temporal interference can focus low-frequency stimulation envelopes at deep brain targets in a realistic head model.","keywords":["deep brain stimulation","temporal interference","microwave focusing","iterative time reversal","antenna array optimization","amplitude-modulated fields","voxel head model","non-invasive neuromodulation"],"falsifier":"An in-vitro or in-vivo neuron preparation exposed to two 700 MHz fields with a 100 Hz offset, at the amplitudes the optimized array would produce at a deep target: if action potentials do not phase-lock to the 100 Hz envelope, or if the required amplitudes exceed head SAR limits, the central claim fails.","tokens_in":15227,"feed_emoji":"🧠","tokens_out":5983,"duration_ms":66712,"temperature":0.7,"pith_summary":"This paper seeks to establish a proof-of-principle for non-invasive deep brain stimulation using microwave fields rather than implanted electrodes. Its central claim is that an external array of about 35 magnetic dipole antennas, arranged around a realistic 77-tissue voxel head model, can concentrate an amplitude-modulated electric-field envelope at a chosen deep target while keeping the envelope low elsewhere. The method combines iterative time-reversal to choose antenna positions and orientations with genetic-algorithm temporal-interference optimization to assign each element a frequency, amplitude, and phase. Reported simulations at 700 MHz carriers with a 100 Hz offset show focused envelopes for both a center and a side-shifted target, and the authors interpret the result as a lower bound on what physical antennas could achieve. The significance is that, if the temporal-interference hypothesis holds at microwave carriers, the approach could deliver DBS-like stimulation without surgery.","feed_headline":"Microwave array focuses deep-brain stimulation without surgery","feed_subtitle":"Simulations show a 100-Hz stimulation envelope can be focused at deep brain targets using external antennas.","key_machinery":"The load-bearing object is the temporal-interference envelope amplitude E_AM(r) = 2 min(||E_f1(r)||, ||E_f2(r)||), which the paper treats as the quantity neurons respond to. The two-stage pipeline is the mechanism: iterative time-reversal (a virtual source at the target, back-propagated fields to rank candidate antenna positions, with Bayesian updates of source parameters and an attenuation-compensation exponent) supplies positions, orientations, and initial excitations; then a genetic algorithm over the Green's-function representation E_fi = sum(G_ni * m_ni * exp(j*alpha_ni)) refines amplitudes, phases, and binary frequency assignment to minimize J_TI = (background mean E_AM)/(focal mean E_","core_discovery":"The paper's core discovery is that adding iterative time reversal (iTR) as an initializer makes temporal-interference (TI) optimization practical and precise in a heterogeneous, lossy head. After iTR fixes the array geometry and initial steering parameters, a genetic algorithm minimizes the ratio of background-to-focal envelope amplitude E_AM = 2 min(|E_f1|,|E_f2|) by assigning each element to one of two carriers, 700 MHz and 700.0001 MHz, and adjusting amplitude and phase. The resulting normalized maps show a confined envelope at the intended location, with noticeable buildup only in superficial non-neuronal tissues, and the side-focus case shows a deliberate asymmetry in frequency assignme","pith_inferences":["A testable extension suggested by the side-focus result: because E_AM uses min(|E_f1|,|E_f2|), asymmetric tissue losses must be compensated by asymmetric element counts, so one could verify that the optimal solution indeed balances the two carrier amplitudes near the focal point.","The paper does not model the neuron itself; feeding the simulated field maps into single-compartment or cable neuron models at 700 MHz carriers would test whether a 100 Hz envelope actually drives spiking at usable amplitudes.","The safety argument treats off-target envelope amplitude as benign because skin and skull lack neurons; an inference beyond the paper is that patient-specific variation in skull thickness and tissue layering will alter the envelope maps, so clinical translation would likely require per-patient optimization.","The 100 Hz difference frequency sits in the classic DBS range, and the same framework could in principle deliver multiple simultaneous targets by adding more carriers or different difference frequencies, though the paper does not explore this."],"forward_implications":["If the central claim is correct, non-invasive DBS becomes a numerical optimization problem over an external array: positions, orientations, frequency split, amplitudes, and phases can all be solved jointly.","The frequency-split asymmetry for off-center targets—putting more f1 elements near the target hemisphere and f2 elements opposite—provides a general steering principle for TI arrays.","Because results are normalized, the immediate next step is to determine the absolute envelope amplitudes required for neuron activation and compare them against SAR and thermal limits; the paper identifies this as an open issue.","The idealized magnetic-point-dipole sources imply that the reported focusing is a lower bound; physical directive antennas should perform at least as well, motivating dedicated antenna design.","The framework extends naturally to multi-objective optimization, such as balancing stimulation envelope against energy absorption along a Pareto front, as the paper suggests."],"fun_headline_variants":["Non-invasive deep brain stimulation via microwave time-reversal focusing","Microwave temporal interference targets deep brain without surgery","Simulation shows focused microwave stimulation for deep brain","Time-reversal boosts microwave focus for non-invasive brain stimulation","External antennas focus microwave fields for non-invasive DBS"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole method rests on the temporal-interference hypothesis at microwave carriers: neurons will demodulate two 700 MHz fields separated by 100 Hz and respond to the envelope E_AM = 2 min(|E_f1|, |E_f2|), at field strengths this array can deliver without exceeding safety limits.","fun_headline_variants_meta":{"raw":{"variants":["Non-invasive deep brain stimulation via microwave time-reversal focusing","Microwave temporal interference targets deep brain without surgery","Simulation shows focused microwave stimulation for deep brain","Time-reversal boosts microwave focus for non-invasive brain stimulation","External antennas focus microwave fields for non-invasive DBS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000162,"raw_usage":{"total_tokens":1076,"prompt_tokens":742,"completion_tokens":334,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":258}},"tokens_in":486,"tokens_out":334,"duration_ms":3873,"temperature":1.0,"reasoning_tokens":258,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T03:09:35.142269+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An in-vitro or in-vivo neuron preparation exposed to two 700 MHz fields with a 100 Hz offset, at the amplitudes the optimized array would produce at a deep target: if action potentials do not phase-lock to the 100 Hz envelope, or if the required amplitudes exceed head SAR limits, the central claim fails.","supporting_citations":[],"review_version":1}