{"id":"d73ab48f-1e7d-4396-b2b1-1c56202a783e","arxiv_id":"2412.11793","paper_version":4,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MHCS-BO, a Bayesian optimizer with an extra exploration rule, is applied to polarization gradient cooling with an optical lattice, reportedly producing 10^8 rubidium-87 atoms at 0.4 microkelvin, but the abstract overstates what the experiment actually shows.","lead":"Researchers used an upgraded Bayesian optimizer (MHCS-BO) to tune 16 laser-cooling parameters in a rubidium atom experiment, claiming to reach 10^8 atoms at 0.4 microkelvin in 15 minutes. The headline temperature and speed claims, however, are not fully backed by the experimental details in the paper.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The twofold speedup claim is internally contradicted: MHCS-BO reaches its reported optimum at sample 61 versus BO's 66 while running two experiments per iteration, so the claimed factor-of-two experimental-cost reduction is unsupported.","rationale":"I agree with the reader's overall REJECT verdict and with the concern that the TOF-amplitude-to-atom-number proportionality is uncalibrated, which undermines the absolute 10^8 atom-number claim. However, I identify a more directly load-bearing issue for the paper's central algorithmic claim: the twofold efficiency statement is internally contradicted by the reported convergence samples. The paper's Section IV sets BO at 100 iterations and MHCS-BO at 50 iterations, with MHCS-BO producing two experimental samples per iteration, so both methods total 100 samples. The reported convergence at the 61st versus 66th sample implies only about an 8% reduction in experimental cycles, not a factor of two. The test-function comparison uses 50 MHCS-BO EI evaluations versus 100 BO evaluations, which can justify a computational speedup of the EI acquisition loop but does not transfer to the experimental setting where MHCS samples are real experiments. This is an internal inconsistency, not a disagreement with community consensus, and it targets the abstract's most prominent claim. I recommend maintaining the reader's REJECT verdict because the abstract's headline claims—twofold efficiency and 10^8 atoms—are both unsupported by the evidence presented in the paper.","tokens_in":13704,"tokens_out":5765,"duration_ms":56436,"concrete_test":"Re-extract the raw per-sample objective values behind Fig. 6(a) and compute the first sample index at which each method reaches the same objective threshold, counting every MHCS-EI and MHCS-MHCS implementation as one experimental cycle. If the ratio of BO samples to MHCS-BO samples to reach the same threshold is not approximately 2, the abstract's 'twofold increase in optimization efficiency' is contradicted by the paper's own measurements.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim is that MHCS-BO gives a twofold increase in optimization efficiency. Section IV states that BO and MHCS-BO were run for 100 and 50 iterations, respectively, 'yielding 100 sampling points,' because MHCS-BO tests both the EI and MHCS predictions each iteration. It then reports that MHCS-BO and BO reach the historical global optimum at the 61st and 66th sample. Under the physically relevant cost measure—number of experimental cycles—MHCS-BO used 61 experiments versus BO's 66, an improvement of about 8%, not 100%. The twofold claim appears to compare per-iteration progress while doubling the number of experiments per iteration, which is an apples-to-oranges comparison. The test-function benchmarks in Section III and Appendix C compare 50 MHCS-BO EI evaluations against 100 BO evaluations; treating MHCS evaluations as free is appropriate for computational cost but not for the experimental setting, where every MHCS prediction is an additional apparatus run. Thus the headline 'accelerated' claim is contradicted by the paper's own convergence data. Separately, the atom-number values feeding the 10^8 claim rest on an uncalibrated TOF amplitude proxy (A_s ∝ N_a without a measured proportionality constant), which is a real but secondary concern: even if the proxy were calibrated, the twofold speedup claim would still fail.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Maximum Hypersphere Compensation Sampling Bayesian Optimization (MHCS-BO), a BO variant that augments expected-improvement acquisition with a point selected from the largest empty hypersphere in the observed parameter set. The authors apply the method to a 16-parameter polarization-gradient-cooling sequence enhanced by a one-dimensional optical lattice for 87Rb atoms, reporting a twofold increase in optimization efficiency and superior prediction accuracy over conventional BO, and claiming preparation of approximately 10^8 atoms at 0.4±0.2 μK within 15 minutes. The experimental work includes a 10-segment PGC timing sequence, lattice-assisted cooling, robustness scans, and comparison with manual optimization.","tokens_in":14033,"tokens_out":6640,"duration_ms":55632,"significance":"The experimental apparatus work is genuine and the paper contains useful detail: a 16-parameter optimization of a real cooling sequence, a long-term stability test (Appendix E), robustness scans (Fig. 7), and a ballistic-expansion temperature measurement (Fig. 8). If the claims were supported, MHCS-BO would be a valuable addition to the toolbox of automated cold-atom optimization. However, the central claims of twofold efficiency and the 0.4 μK/15-minute result are not supported by the paper's own data, and the atom-number estimates rest on an uncalibrated proxy. These issues concern the main contribution of the paper rather than presentation, so they are decisive.","major_comments":[{"comment":"The claim in the abstract and Section III that MHCS-BO provides a 'twofold increase in optimization efficiency' is contradicted by the convergence data in Fig. 6(a). The text states that BO and MHCS-BO were run for 100 and 50 iterations, respectively, 'yielding 100 sampling points,' because MHCS-BO tests two points per iteration. MHCS-BO reaches the historical global optimum at the 61st sample and BO at the 66th. Measured in experimental cycles, which is the relevant cost for a cold-atom apparatus, the improvement is about 8%, not 100%. The test-function benchmarks in Fig. 4 and Appendix C compare 50 MHCS-BO EI evaluations with 100 BO evaluations, counting MHCS evaluations as free; this is valid for computational cost but not for experimental cost, where every MHCS prediction is an additional apparatus run. The central efficiency claim therefore needs to be re-evaluated or reframed.","section":"Section IV, Fig. 6(a)"},{"comment":"The abstract's statement that 'approximate 10^8 cold atoms at a temperature of 0.4±0.2 μK can be achieved given the optimal parameters within 15 minutes' is not supported by the optimization results. The MHCS-BO optimized lattice parameters yield T_a=0.91 μK (Section IV, Fig. 6(d)). The 0.4 μK result is obtained only after manually increasing t_h to 20 ms and MOT loading time to 3 s (Fig. 8), which are not the parameters found by the 15-minute MHCS-BO run. The abstract should distinguish the automated optimization outcome from the subsequently hand-tuned result.","section":"Abstract; Section IV, Fig. 8"},{"comment":"The objective function O(X)=T_a/A_s^b and the atom-number estimates throughout rely on the assumption A_s ∝ N_a, where A_s is the TOF signal amplitude. No calibration of the proportionality constant is presented, and the paper does not verify that the proportionality holds across the optimized parameter range (for example, as t_h and lattice power change the cloud size and position at the detection region, the geometric collection efficiency may vary). The 'approximately 10^8 atoms' claim is therefore unsupported. An independent atom-number measurement (e.g., fluorescence imaging calibrated by absorption imaging) is needed.","section":"Section II; Appendix D"},{"comment":"The claim of 'superior prediction accuracy' is not demonstrated quantitatively. The Gaussian-process prediction is never evaluated against held-out experimental data; the comparisons in Fig. 4 and Fig. 6 are historical optima, not prediction error. The statement in Section IV that MHCS 'bring[s] a more precise mapping from the model to apparatus' is an interpretation, not a measurement. A prediction-error comparison on independent test points, or at least a cross-validation of the GP fit, is needed to support the claim.","section":"Appendix B; Section IV"},{"comment":"The definition of the MHCS candidate hypersphere contains an unclear and likely erroneous condition. Step 4 states that a hypersphere is selected as a candidate if any observed parameter satisfies |X_m − X_i| > R_c, but R_c is never defined, and the inequality direction is opposite to the usual 'empty sphere' test (one would check whether all observed points lie outside the sphere of radius R_m). Without a precise definition of R_c and the selection criterion, the algorithm is not reproducible.","section":"Section III, Step 4"}],"minor_comments":[{"comment":"The test-function benchmarks are performed only for two-dimensional functions (Fig. 12), while the experimental problem has 12-16 parameters; the claimed advantage of MHCS in high-dimensional spaces is therefore not tested by these benchmarks.","section":"Appendix C"},{"comment":"The x-axis of Fig. 6(a) is not explicitly defined; because MHCS-BO runs two experiments per iteration, the reader cannot immediately tell whether 'sample' means an iteration or an experimental cycle. Please label the axis as 'experimental cycle' and define it in the caption.","section":"Section IV"},{"comment":"In Appendix B, σ is used both for the GP standard deviation in Eq. (B3) and for the observation noise in Eq. (B1); this notational overlap is confusing and should be disambiguated.","section":"Appendix B"},{"comment":"The conclusion states that MHCS-BO cannot find the complete Pareto frontier, but the paper never defines a Pareto frontier for the two objectives (T_a and N_a) or explains how the scalarized objective with b=0.5 relates to it.","section":"Section V"}],"recommendation":"reject","confidential_remarks":"The manuscript reports real experimental measurements, but the abstract overstates the results. The convergence data in Fig. 6(a) contradict the twofold-speedup claim, and the 0.4 μK result comes from manual parameter changes rather than the optimized parameters. I recommend rejection; if the authors wish to resubmit, they should either provide a fair experimental comparison (equal number of runs) or substantially weaken the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper is a genuine experimental study with a new (if simple) BO variant, but the abstract's two headline numbers—twofold efficiency and 0.4 μK within 15 minutes—don't survive contact with the paper's own data. Worth a referee's eyes, but only if the authors are willing to rewrite claims and redo the comparison.\n\nWhat's new: MHCS-BO adds a diversity-promoting step to BO: after choosing the EI maximizer, you also run an experiment at the center of the largest hypersphere empty of observed points. That is a sensible way to combat the non-sampling blind spots of EI in high dimension, and it's new in this application. The experimental apparatus is real: 16-parameter optimization of PGC plus a 1D optical lattice, with decoupled stages, stability tests (CV=0.035), and robustness scans near the optimum. Those parts are done carefully.\n\nThe soft spots are load-bearing. Section IV says BO ran 100 iterations and MHCS-BO ran 50, both yielding 100 sampling points because MHCS-BO tests two candidates per iteration. In experiment cost, MHCS-BO reaches its historical optimum at sample 61 vs BO's 66: an 8% saving, not a doubling. The 'twofold' claim comes from comparing per-iteration progress while doubling the experiments per iteration. The test-function benchmarks in Section III and Appendix C make the same apples-to-oranges comparison, and treating MHCS evaluations as free is fine for CPU time but not for apparatus runs.\n\nThe 0.4±0.2 μK result is also not what the optimizer found. The optimizer's best with 780 nm lattice was 0.91 μK; the 0.4 μK came after manually extending t_h to 20 ms and MOT loading to 3 s. So the abstract phrase 'given the optimal parameters within 15 minutes' overstates both the role of the optimizer and the time. The atom number rests on an uncalibrated TOF-amplitude proxy (A_s ∝ N_a) with no independent number measurement; that proportionality is asserted, not demonstrated.\n\nThe algorithmic idea is worth exploring, and the experimental data could be useful after a careful revision that states the comparison correctly and tempers the claims. As it stands, the central quantitative claims are unsupported. I'd send it to a serious referee, but with a clear expectation that the authors either provide a correct comparison or withdraw the headline claims.","headline":"The experimental effort and the MHCS heuristic are real, but the paper's headline claims count iterations rather than experiments and credit the optimizer with a result that came from manual tweaks.","tokens_in":14570,"tokens_out":2811,"would_cite":false,"duration_ms":26620,"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":"A modified Bayesian optimizer that also samples the largest empty hypersphere in parameter space doubles optimization efficiency on benchmark functions and, in a rubidium-87 cooling experiment, reaches about 10^8 atoms at 0.4 μK in 15…","keywords":["Bayesian optimization","maximum hypersphere compensation sampling","polarization gradient cooling","optical lattice","cold atoms","rubidium-87","sub-microkelvin cooling","experimental parameter optimization"],"falsifier":"Repeat the optimized sequence while measuring atom number independently (for example, by absorption imaging calibrated with the same cloud) and compare it with the time-of-flight amplitude A_s across the scanned parameter ranges; if the ratio A_s/N_a varies appreciably, the objective ranking and the reported approximately $10^{8}$ atom count are not supported.","tokens_in":13514,"feed_emoji":"⚛️","tokens_out":9454,"duration_ms":78351,"temperature":0.7,"pith_summary":"This paper proposes a modification to Bayesian optimization called Maximum Hypersphere Compensation Sampling (MHCS-BO), in which the optimizer, alongside the usual expected-improvement prediction, also proposes the center of the largest hypersphere in parameter space that contains no previously observed data. The authors try to establish that this extra proposal roughly doubles optimization efficiency and improves prediction accuracy relative to standard Bayesian optimization, on five benchmark functions and in a live cold-atom experiment. The experimental target is a two-stage cooling sequence that first runs polarization gradient cooling (PGC) on rubidium-87 and then adds a one-dimensional optical lattice, with a total of 16 timing, power, and detuning parameters optimized in a decoupled way. They report that within about 15 minutes of optimization the sequence produces approximately $10^{8}$ atoms at 0.4±0.2 μK, near the recoil temperature, and that a 1064 nm lattice variant gives 2×$10^{8}$ atoms at 0.12±0.07 μK. If these results hold, the method offers a fast route to the large, sub-microkelvin samples that atom interferometry and other quantum experiments need.","feed_headline":"Sparse sampling doubles optimizer speed, yields 10^8 cold atoms","feed_subtitle":"A Bayesian optimizer that probes sparse regions tunes cold atoms to near-recoil temperatures in minutes, not hours.","key_machinery":"The engine is the Maximum Hypersphere Compensation Sampling (MHCS) step. Given a set of observed points scaled into [0,1], the algorithm forms every pair of points, treats the segment between them as a hypersphere diameter, and computes the midpoint center and radius. A hypersphere is a candidate if no other observed point lies at distance greater than R_c from its center; the algorithm picks the candidate with the largest radius and proposes its center as the next experimental setting. This proposal runs in parallel with the standard expected-improvement (EI) proposal, so each iteration adds two samples, and the EI prediction is improved because the Gaussian process surrogate sees the previously empty region. The other load-bearing machinery is the decoupled objective: PGC parameters X_P={Att,δ_c,t_i,t_m} are optimized first with the lattice off, then lattice parameters X_L={t_h,t_m,t_u,V_L} with the lattice on, using O(X)=T_a/A_s^b with b=0.5 to balance atom number and temperature.","core_discovery":"The central discovery claimed is that deliberately sampling the largest empty region of parameter space repairs a known weakness of acquisition-function-based Bayesian optimization, namely that expected-improvement predictions cluster in already-sampled regions and leave large parts of the space unobserved. In MHCS-BO, every observed point is scaled to [0,1], all pairwise hyperspheres (spheres whose diameters connect two observed points) are constructed, and the center of the largest hypersphere containing no observed point is selected as a second recommendation, X_MHCS, executed alongside the EI recommendation X_EI in each cycle. On five standard test functions the authors find that MHCS-BO reaches lower historical optima in roughly half the iterations of plain BO. In the atom experiment, the objective function O(X)=T_a/A_s^b (temperature divided by the b-th power of the time-of-flight signal amplitude, with b=0.5) is optimized first over the PGC parameters and then over the optical-lattice parameters; the final sequence yields approximately $10^{8}$ atoms at 0.4±0.2 μK measured by ballistic expansion. The paper frames these results as the first demonstration of this optimizer in preparing sub-μK cold atoms and as a protocol transferable to other high-dimensional experimental optimizations.","pith_inferences":["Because the MHCS step is agnostic to the underlying objective, the same largest-empty-hypersphere rule could be dropped into other acquisition functions or multi-objective Bayesian optimization, potentially addressing the Pareto-front limitation the authors note.","An independent check of the claimed atom numbers would be to calibrate the time-of-flight amplitude A_s against absorption imaging over the full optimized parameter range; if A_s is not strictly proportional to atom number, the reported 10^8 figures would shift.","The observed temperature-floor behavior with holding time suggests the optimizer is trading atom number against temperature at the edge of the lattice well; scanning the balance coefficient b in O(X)=T_a/A_s^b would map the full trade-off curve and tell users where the reported operating point sits on it.","The same protocol should transfer to other expensive high-dimensional alignment tasks in cold-atom and quantum-device laboratories, where standard EI commonly clusters samples and leaves sparse regions unexplored."],"forward_implications":["On the five benchmark functions, MHCS-BO converges to better optima with 50 iterations where standard BO needs 100, so for experiments with expensive objective evaluations the method halves the number of runs needed.","With the optimized 780 nm sequence, roughly 10^8 rubidium-87 atoms at 0.4±0.2 μK are produced within 15 minutes of optimization, a combination of large atom number and near-recoil temperature.","Using a 1064 nm lattice and a 200 ms holding time, the same protocol produces about 2×10^8 atoms at a vertical temperature of 0.12±0.07 μK, about three times below the recoil limit.","Simulated atom interferometry at 0.12 μK versus 4.4 μK shows fringe contrast rising from 71.2% to 99.3% and coherence time from 93 μs to 289 μs, so lower temperature directly translates into better interferometer performance.","Optimizing the 16 parameters in two decoupled groups performs better than optimizing them all at once, avoiding the poor time-of-flight signals and misleading surrogate models seen in coupled optimization."],"supporting_citations":[{"why":"Supplies the 3D-MOT, laser, vacuum, and controller setup inherited from the previous PGC optimization work.","marker":"[34]"},{"why":"Establishes Bayesian optimization as an efficient global optimizer for expensive quantum-gas experiments, the baseline this paper improves on.","marker":"[32]"},{"why":"Demonstrates Bayesian optimization applied to a cold-atom trap, providing the standard-BO comparison context.","marker":"[33]"},{"why":"Provides the benchmarking of machine-learning optimizers against which the two-fold efficiency claim is framed.","marker":"[35]"},{"why":"Source of the five standard test functions used to measure MHCS-BO's convergence and accuracy.","marker":"[36]"},{"why":"Supplies the ballistic-expansion relation used to extract the final 0.4 μK temperature.","marker":"[37]"},{"why":"Shows optical-lattice-assisted laser cooling of rubidium, the physical mechanism this experiment extends with PGC.","marker":"[20]"},{"why":"Gives the two-photon Rabi formula used in the simulation connecting atom temperature to fringe contrast and coherence time.","marker":"[38]"}],"fun_headline_variants":["MHCS-BO doubles Bayesian optimization efficiency for cold atoms","Largest-gap sampling cuts Bayesian opt iterations in half","10^8 atoms at 0.4 μK via smarter Bayesian optimization","New sampler boosts Bayesian optimization for ultracold atoms","Hypersphere-aware Bayesian optimization cools atoms to 0.4 μK"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's atom-number claims rest on the assumption that the strength of the falling-atom fluorescence signal is directly proportional to the number of atoms for every parameter setting tried, and that raising that signal to the 0.5 power in the optimization score faithfully balances temperature against atom number; no independent atom-number calibration is reported.","fun_headline_variants_meta":{"raw":{"variants":["MHCS-BO doubles Bayesian optimization efficiency for cold atoms","Largest-gap sampling cuts Bayesian opt iterations in half","10^8 atoms at 0.4 μK via smarter Bayesian optimization","New sampler boosts Bayesian optimization for ultracold atoms","Hypersphere-aware Bayesian optimization cools atoms to 0.4 μK"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000387,"raw_usage":{"total_tokens":2046,"prompt_tokens":951,"completion_tokens":1095,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":1006}},"tokens_in":567,"tokens_out":1095,"duration_ms":8746,"temperature":1.0,"reasoning_tokens":1006,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:34:41.103721+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the optimized sequence while measuring atom number independently (for example, by absorption imaging calibrated with the same cloud) and compare it with the time-of-flight amplitude A_s across the scanned parameter ranges; if the ratio A_s/N_a varies appreciably, the objective ranking and the reported approximately $10^{8}$ atom count are not supported.","supporting_citations":[{"cited_title":"Liang, S","cited_arxiv_id":null,"evidence_quote":"Supplies the 3D-MOT, laser, vacuum, and controller setup inherited from the previous PGC optimization work."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes Bayesian optimization as an efficient global optimizer for expensive quantum-gas experiments, the baseline this paper improves on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates Bayesian optimization applied to a cold-atom trap, providing the standard-BO comparison context."},{"cited_title":"Anton, V","cited_arxiv_id":null,"evidence_quote":"Provides the benchmarking of machine-learning optimizers against which the two-fold efficiency claim is framed."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the five standard test functions used to measure MHCS-BO's convergence and accuracy."},{"cited_title":"Xin and S","cited_arxiv_id":null,"evidence_quote":"Supplies the ballistic-expansion relation used to extract the final 0.4 μK temperature."},{"cited_title":"Wei and C","cited_arxiv_id":null,"evidence_quote":"Shows optical-lattice-assisted laser cooling of rubidium, the physical mechanism this experiment extends with PGC."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the two-photon Rabi formula used in the simulation connecting atom temperature to fringe contrast and coherence time."}],"review_version":1}