{"id":"c1d229b2-0186-4a4d-ac98-5b29febcfca7","arxiv_id":"2506.02325","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Bayesian-optimized fillet geometry lifts the quality factor of centimeter-scale Si3N4 torsion nanoribbons above 1e8 at room temperature, with Q-frequency products above 1e13 Hz.","lead":"The paper uses Bayesian optimization to shape the clamp fillets of strained silicon nitride nanoribbons, achieving the best torsion mode quality factors yet reported, above 100 million at room temperature. A smart generalist would read this because ultra-low-loss torsion oscillators at the centimeter scale could improve sensors for gravity, dark matter, and quantum optomechanics.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Eq. 5 loss model is the load-bearing weak point: the paper's own flexural-mode data show the same simulation misses the dominant loss by ~100x, and only two of seven fabricated ribbons are reported, leaving open that the two Q>1e8 torsion devices are selected outliers.","rationale":"The reader's weakest_assumption correctly identifies the FEM loss model as the key risk. I agree that the ~100x flexural-mode discrepancy is the best in-paper evidence that Eq. (5) omits relevant dissipation. However, the more decisive framing for the central claim is the combination of that model failure with the selective reporting of device outcomes: the paper fabricates seven ribbons but reports Q values for only two, and even the count is inconsistent ('seven' vs 'six'). If the unreported devices have substantially lower torsion Q, the claim that Bayesian optimization robustly realizes Q>1e8 is not established, regardless of whether the flexural loss mechanism transfers to torsion. This is why I mark agreement as partial rather than full. The proposed test directly addresses both issues: full device-level reporting settles the selection concern, and the flexural-mode comparison provides a quantitative bound on the missing loss that can be propagated into the torsion prediction. The reader's CONDITIONAL verdict already captures this uncertainty, so no verdict adjustment is needed.","tokens_in":9338,"tokens_out":10594,"duration_ms":114658,"concrete_test":"Provide a complete data table for all fabricated ribbons, including the four unreported devices, with measured Q, frequency, and corresponding COMSOL prediction for each device. Then perform the same Eq. (5) simulation for the fundamental flexural mode of the same optimized geometry and compare to the measured ~100x-lower flexural Q; if the flexural discrepancy persists, add the implied unmodeled loss rate (1/Q_extra = 1/Q_flex_meas - 1/Q_flex_sim) in quadrature to the simulated torsion loss and recompute the predicted torsion Q. If the corrected torsion prediction drops below 1e8 for any highlighted device, the optimization objective is not a valid proxy for total torsional loss.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The claim that Bayesian-optimized fillets 'realize' Q>1e8 torsion modes rests on Eq. (5), which is both the optimization objective and the basis for attributing the measured Q to the designed geometry. Eq. (5) reduces dissipation to a uniform material loss Q0 plus clamp-induced bending loss computed from the two-step COMSOL model; it contains no other dissipation channel. The manuscript itself provides a direct in-device test of this model: the fundamental and second-order flexural modes of the first optimized ribbon are reported to be ~100 times lower than simulated. For a mode of the identical chip, the model misses the dominant loss mechanism by two orders of magnitude, and the paper offers no argument or measurement showing that the torsion mode is immune to that missing mechanism. The only torsion evidence is that two highlighted devices fall within a factor of two of simulation, while the text first says five 400-um and two 600-um ribbons were fabricated and then says 'six ribbons fabricated and characterized', without reporting the Q values of the remaining devices. If the unmodeled loss also affects torsion, the optimized geometry's advantage could shrink below the claimed threshold, and the headline result could be a selected outlier rather than a reproducible consequence of Bayesian optimization. The central existence claim (a single measured Q=1.5e8 device) is not directly threatened, but the broader claim that optimization produced the ultrahigh Q is not supported without addressing the model's demonstrated incompleteness and the unreported device outcomes.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This Letter reports the use of Bayesian optimization over fillet geometry to maximize the simulated dissipation-dilution quality factor of the fundamental torsion mode of strained Si3N4 nanoribbons, using Eq. (5) as the FEM-based objective. The authors fabricated centimeter-scale, 90-nm-thick ribbons with optimized diagonal fillets and report stroboscopic ringdown Q factors of 1.5e8 and 1.2e8 at room temperature for two 400-um-wide devices, with Q*f > 1e13 Hz, plus a 4-K measurement showing Q = 1.7e8. They frame these results as the first solid-state torsion oscillator with Q > 1e8 and as evidence that Bayesian-optimized clamp geometry extends torsional dissipation dilution beyond the hard-clamping limit.","tokens_in":9628,"tokens_out":5941,"duration_ms":53895,"significance":"If the reported values are reproducible and attributable to the optimized geometry, the result is significant for torque sensing and room-temperature quantum optomechanics: the Q*f product exceeds the 300-K thermal decoherence threshold, and the inferred torque sensitivity (1e-20 N m / sqrt(Hz)) and zero-point angular displacement spectral density are competitive. The paper has strengths that should be credited: the optimization objective is a forward FEM calculation with material parameters from prior literature, so the geometry optimization is not circularly fit to the measured torsion Q; the ringdown method is standard; and the authors explicitly flag the large discrepancy between simulated and measured flexural-mode Q. The main weaknesses are under-reported device statistics and the absence of uncertainty quantification, both of which are load-bearing for the headline attribution claim.","major_comments":[{"comment":"The manuscript reports measured Q for only two of the fabricated ribbons and does not state the Q values of the remaining devices; the fabrication count is also internally inconsistent, as the text first says 'five 400-um-wide and two 600-um-wide ribbons' (seven total) and then says 'six ribbons fabricated and characterized'. Without complete reporting of all devices or an explicit pre-registered selection rule, the claim that Bayesian optimization reproducibly realizes Q > 1e8 is not supported, because the two highlighted devices could be selected outliers.","section":"Experimental results; Fig. 3"},{"comment":"Equation (5) is both the optimization objective and the basis for attributing the measured Q to the designed geometry, yet the same experimental section reports that the fundamental and second-order flexural modes of the first optimized device are roughly 100 times lower than simulated, with the paper attributing this to 'sensitivity to other forms of loss'. No argument or measurement is given that the torsion mode is immune to this missing loss channel; the factor-of-two agreement for two torsion devices is therefore not, by itself, sufficient evidence that the unmodeled loss is negligible. The authors should provide a quantitative test, for example a torsion-Q measurement across a geometry variation predicted by Eq. (5), an estimate of the missing loss contribution from the flexural discrepancy, or a direct comparison with non-optimized control devices of identical material.","section":"Fillet optimization via finite element simulation; Experimental results"},{"comment":"No uncertainty is reported for any Q value and no repeated ringdown measurements are shown. This matters for several comparisons in the paper: the 'within a factor of two' agreement with simulation, the 40% increase from 1.2e8 to 1.7e8 at 4 K, and the claimed distinction from the earlier Q = 1.0e8 device in Ref. [5]. Please provide standard errors or repeated-measurement statistics, and state the systematic uncertainties in the ringdown extraction method.","section":"Experimental results"}],"minor_comments":[{"comment":"The figure label gives Q = 147 million while the text and summary quote Q = 150 million; these numbers should be made consistent.","section":"Fig. 1(d) and main text"},{"comment":"The phrase 'coated with on both sides' should read 'coated on both sides'.","section":"Fabrication paragraph"},{"comment":"The caption says 'V on Mises stress profile'; it should be 'von Mises stress profile'.","section":"Fig. 1 caption"},{"comment":"The subscripts in k_shear,ext_E and k_bend,cl_E are not defined in the text; a one-line definition would improve readability.","section":"Eq. (2)"},{"comment":"The denominator 'Q0/h / 60 nm^-1' is awkwardly parenthesized; rewriting as (Q0/h)/(60 nm^-1) would clarify the scaling.","section":"Eq. (3)"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the central existence claim of a single Q > 1e8 torsion device is likely sound, but the broader claim that Bayesian optimization produced this result requires complete device statistics and at least one concrete test that the torsion mode is not affected by the missing loss mechanism that creates the 100x flexural discrepancy. The Bayesian optimization methodology is appropriate and the forward nature of the simulation is a positive feature; I see no novelty or scope concern. A major revision that reports all measured devices, adds uncertainties, and addresses the torsion-versus-flexural model discrepancy would make the paper suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline is a real measured thing: a torsion mode at Q=1.5e8 with Qf>1e13 Hz at room temperature, a record for solid-state torsion oscillators. The catch is that the Bayesian optimization story is only partially supported by the data, and the paper's own numbers raise a selection question.\n\nWhat's new: applying Bayesian optimization to the fillet geometry to push torsional dissipation dilution beyond the hard-clamping limit. The underlying physics is from their earlier PRX paper, and the optimization method is imported from prior nanomechanics work, so novelty is moderate. That said, the result is significant if reproducible. The ringdown measurements look standard, and the two highlighted devices land within a factor of two of the FEM prediction. The cryogenic measurement is a nice extra.\n\nThe soft spots are real. The text says five 400-um and two 600-um ribbons were fabricated, then later says 'six ribbons fabricated and characterized' and only reports the Qs of two. That's an inconsistency and a red flag for cherry-picking. A referee needs to see the full distribution of Qs across devices. Second, the same simulation used as the optimization objective predicts flexural modes about 100 times higher than measured. The paper acknowledges this and attributes it to 'other forms of loss,' but offers no argument that torsion modes are immune to those losses. The two torsion data points matching within a factor of two could be luck on a small sample. The model's incompleteness doesn't kill the existence claim, but it weakens the causal claim that the optimized geometry is what produces the Q.\n\nMinor: no stated uncertainty on the Q values, and the device count discrepancy should be fixed.\n\nOverall, this is a serious experimental paper with a genuine record Q. It deserves peer review. The referee should push for complete device statistics, error bars, and some discussion of why the loss model is trustworthy for torsion when it fails so badly for flexure.","headline":"A genuine room-temperature Q>1e8 torsion oscillator, but the Bayesian-optimization attribution leans on two devices and a loss model that misses flexural losses by ~100x.","tokens_in":10189,"tokens_out":2787,"would_cite":true,"duration_ms":25233,"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":"The paper reports the first solid-state torsion oscillator with a Q factor above 100 million, reached at room temperature in silicon nitride nanoribbons whose clamp fillets are shaped by Bayesian optimization.","keywords":["Bayesian optimization","torsion oscillator","dissipation dilution","silicon nitride nanoribbon","quality factor","soft clamping","torque sensing","nanomechanics"],"falsifier":"Measure the ringdown Q of torsion modes on a set of ribbons whose fillet shapes are chosen to span the optimizer's predicted response surface, including shapes the optimizer judged poor. If the measured Q does not follow the simulated K_tot/U_s curve—for instance, if all shapes give nearly the same Q, or if the best measured geometry is not the simulated optimum—then clamp-region bending is not the dominant loss and the central claim fails. A simpler check: if the torsion Q is found to depend strongly on surface preparation or on ambient conditions, an unmodeled surface loss channel is at play.","tokens_in":9168,"feed_emoji":"🔬","tokens_out":6944,"duration_ms":58536,"temperature":0.7,"pith_summary":"Torsion modes of tensioned nanoribbons can have their mechanical loss dramatically reduced by a 'dissipation dilution' effect, but the bending of the mode at the clamps eventually limits this gain. This paper argues that the shape of the fillet where the ribbon meets its support controls that clamp loss, and that a Bayesian optimization loop over fillet geometries in a finite-element simulator can find shapes with far lower loss. Applied to centimeter-scale silicon nitride nanoribbons, the search yields torsion oscillators with quality factors above 100 million at room temperature—the first solid-state torsion oscillator to reach that mark—and with quality-factor–frequency products above $10^{13}$ Hz, the threshold for quantum coherence at 300 K. The resulting thermal torque noise is at the level of $10^{-20}$ N·m/√Hz and the zero-point angular displacement noise near $10^{-10}$ rad/√Hz, which the authors argue makes these devices attractive for weak-torque sensing and room-temperature quantum optomechanics.","feed_headline":"Torsion nanoribbons pass 100-million Q at room temperature","feed_subtitle":"Bayesian-optimized clamp shapes put torque sensors in the quantum-coherent regime.","key_machinery":"The central object is the clamp fillet, whose geometry sets the mode curvature near the support and therefore the bending-loss term in the dissipation-dilution formula. The load-bearing identity is Q/Q0 ≈ K_tot/U_s (Eq. 5), which turns the finite-element mode solution into a predicted quality factor by comparing total kinetic energy with stored strain energy; minimizing U_s relative to K_tot maximizes Q. The optimization machinery is Bayesian optimization with Gaussian-process regression and an expected-improvement acquisition function, used to search the fillet parameters (e.g., elliptical fillet radii rx and ry, or a circular fillet with diagonal boundary) at a computational cost low enough to be practical.","core_discovery":"The paper's central claim is that the fundamental torsion mode of a strained silicon nitride nanoribbon can be 'soft-clamped' far beyond the conventional hard-clamping limit by optimizing the fillet geometry at the clamps. The authors identify the relevant loss as clamp-region bending, proportional to the integral of the squared mode curvature, and compute the dissipation-diluted Q in a two-step finite-element simulation using the ratio of total kinetic energy to stored strain energy. They then run a Bayesian optimizer with an expected-improvement acquisition function over a one- or two-parameter fillet description, obtaining predicted Q factors near 2×$10^{8}$. Fabricated 400-µm-wide, 7-mm-long, 90-nm-thick ribbons reach Q = 1.5×$10^{8}$ at room temperature and 1.7×$10^{8}$ at 4 K, with Q·f > $10^{13}$ Hz in both the fundamental and second torsion modes; the authors report this as the first solid-state torsion oscillator with Q exceeding 100 million.","pith_inferences":["The paper's own data show flexural modes of the same ribbons are roughly a hundred times lower than the finite-element prediction; if a similar unmodeled loss affects the torsion mode, the absolute Q ceiling would be lower than the simulated optimum, even though the optimizer's ranking of geometries could still be correct.","A direct test of the mechanism would be to measure torsion Q while sweeping the fillet shape through the optimizer's predicted landscape; if the measured Q fails to track the simulated K_tot/U_s curve, then clamp-bending loss is not the only important loss channel.","The single-parameter fillet parametrization leaves unexplored geometries such as non-elliptical tapers, multiple fillets, or three-dimensional clamp structures; a higher-dimensional Bayesian search could push Q further if the model is trusted.","Because the objective is a scalar Q at a fixed width, an obvious extension is to maximize Q at a target frequency or moment of inertia, which would directly address sensing applications where the resonance frequency matters."],"forward_implications":["The reported Q·f > 10^13 Hz means the fundamental and second torsion modes of the optimized ribbons meet the room-temperature quantum-coherence condition Q·f > k_BT/h, opening a route to feedback cooling and quantum optomechanics without dilution refrigeration.","Thermal torque sensitivity near 10^-20 N·m/√Hz and zero-point angular displacement spectral density near 10^-10 rad/√Hz put weak-force sensing tasks—magnetometry, gravimetry, tests of short-range gravity and dark matter—within reach of a simple, photolithographically defined device.","Because torsional soft-clamping is preserved under heavy mass loading, a central pad can be added to the optimized ribbon without degrading Q, enabling the design of micro- to milligram-scale torsion pendula.","Cooling the device to 4 K increases Q from 1.2×10^8 to 1.7×10^8 and raises the quantum-coherence number about a hundredfold, suggesting further gains at millikelvin temperatures."],"supporting_citations":[{"why":"Provides the prior discovery of torsional dissipation dilution in strained nanoribbons and the baseline devices this work improves upon.","marker":"[5]"},{"why":"Introduces soft clamping and dissipation dilution as a modeshape-engineering method that this work extends to torsion modes.","marker":"[13]"},{"why":"Reviews dissipation dilution and supplies the general framework for Q enhancement in strained resonators.","marker":"[14]"},{"why":"Provides the modern continuum viscoelastic theory of dissipation dilution used to derive the hard-clamping limit and the role of clamp curvature.","marker":"[19]"},{"why":"Shows that clamp-tapering raises Q of stressed nanobeams, the flexural analogue of the fillet optimization here.","marker":"[21]"},{"why":"Supplies the strain-energy/total-kinetic-energy expression (Eq. 5) used as the objective function in the optimizer.","marker":"[22]"},{"why":"Demonstrates Bayesian optimization of nanomechanical resonators (spiderweb geometry), the methodological precedent for the design loop.","marker":"[33]"},{"why":"Reports centimeter-scale low-dissipation resonators obtained by Bayesian optimization, a direct precursor to this work's scale and method.","marker":"[34]"},{"why":"Provides the quantum-limited optical lever measurement technique used to read the ringdowns and estimate imprecision.","marker":"[42]"}],"fun_headline_variants":["Bayesian design yields torsion oscillators with Q over 100M","Optimized clamps push torsion Q past 100 million","Torsion resonators reach 100M quality factor at 300K","Bayesian-optimized nanoribbons hit record torsion Q"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the finite-element model, which computes Q from the total-kinetic-to-strain-energy ratio accounting only for material loss and clamp-bending loss, faithfully captures every significant damping channel of the torsion mode; the paper's own observation that the same model overestimates flexural-mode Q by about a factor of one hundred shows this premise is not guaranteed.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian design yields torsion oscillators with Q over 100M","Optimized clamps push torsion Q past 100 million","Torsion resonators reach 100M quality factor at 300K","Bayesian-optimized nanoribbons hit record torsion Q"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000392,"raw_usage":{"total_tokens":2062,"prompt_tokens":946,"completion_tokens":1116,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":1041}},"tokens_in":562,"tokens_out":1116,"duration_ms":7486,"temperature":1.0,"reasoning_tokens":1041,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:25:49.173846+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the ringdown Q of torsion modes on a set of ribbons whose fillet shapes are chosen to span the optimizer's predicted response surface, including shapes the optimizer judged poor. If the measured Q does not follow the simulated K_tot/U_s curve—for instance, if all shapes give nearly the same Q, or if the best measured geometry is not the simulated optimum—then clamp-region bending is not the dominant loss and the central claim fails. A simpler check: if the torsion Q is found to depend strongly on surface preparation or on ambient conditions, an unmodeled surface loss channel is at play.","supporting_citations":[{"cited_title":"Nanoscale torsional dis- sipation dilution for quantum experiments and precision mea- surement,","cited_arxiv_id":null,"evidence_quote":"Provides the prior discovery of torsional dissipation dilution in strained nanoribbons and the baseline devices this work improves upon."},{"cited_title":"Ultra- coherent nanomechanical resonators via soft clamping and dis- sipation dilution,","cited_arxiv_id":null,"evidence_quote":"Introduces soft clamping and dissipation dilution as a modeshape-engineering method that this work extends to torsion modes."},{"cited_title":"Ultrahigh- quality-factor micro-and nanomechanical resonators using dis- sipation dilution,","cited_arxiv_id":null,"evidence_quote":"Reviews dissipation dilution and supplies the general framework for Q enhancement in strained resonators."},{"cited_title":"Generalized dissipation dilution in strained mechanical resonators,","cited_arxiv_id":null,"evidence_quote":"Provides the modern continuum viscoelastic theory of dissipation dilution used to derive the hard-clamping limit and the role of clamp curvature."},{"cited_title":"Clamp-tapering increases the quality factor of stressed nanobeams,","cited_arxiv_id":null,"evidence_quote":"Shows that clamp-tapering raises Q of stressed nanobeams, the flexural analogue of the fillet optimization here."},{"cited_title":"Influ- ence of clamp-widening on the quality factor of nanomechani- cal silicon nitride resonators,","cited_arxiv_id":null,"evidence_quote":"Supplies the strain-energy/total-kinetic-energy expression (Eq. 5) used as the objective function in the optimizer."},{"cited_title":"Spiderweb nanomechanical resonators via bayesian optimization: inspired by nature and guided by machine learning,","cited_arxiv_id":null,"evidence_quote":"Demonstrates Bayesian optimization of nanomechanical resonators (spiderweb geometry), the methodological precedent for the design loop."},{"cited_title":"Centimeter-scale nanomechanical resonators with low dissipation,","cited_arxiv_id":null,"evidence_quote":"Reports centimeter-scale low-dissipation resonators obtained by Bayesian optimization, a direct precursor to this work's scale and method."},{"cited_title":"Quantum- 6 limited optical lever measurement of a torsion oscillator,","cited_arxiv_id":null,"evidence_quote":"Provides the quantum-limited optical lever measurement technique used to read the ringdowns and estimate imprecision."}],"review_version":1}