{"id":"652b2b07-94dd-41db-9be7-332eb7c9f9ab","arxiv_id":"2507.08089","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A pad-etch surface treatment removes germanium residue under Josephson junctions and improves median transmon T1 from 157 to 334 microseconds.","lead":"Coherence measurements across 182 devices show that a fluorine-based pad etch plus argon ion milling before junction deposition doubles the median energy relaxation time of aluminum-on-silicon transmon qubits, from 157 to 334 microseconds. The work gives hardware teams a concrete, evidence-backed way to remove a specific germanium contaminant that limits quantum processor performance.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 2x improvement claim rests on qubit-level statistics applied to wafer-level process splits; the paper does not report the number of wafers per split, so wafer-to-wafer variability may be the actual source of the difference.","rationale":"The paper's own disclosure about excluded qubits is honest, and the reader's conditional verdict is reasonable. However, I think the exclusion issue is not the most load-bearing: with only 6 exclusions out of 48 in Process E, even a pessimistic imputation cannot move the median below the POR level unless the included distribution is radically different, and the authors already state the exclusions are dominated by Jc/frequency targeting rather than T1. The more serious threat is the experimental unit. The process variables are wafer-level, and the statistics in Appendix C treat qubits as if they were independent draws from a process distribution. If each process condition is one wafer, the effective replication for the headline comparison is one or two wafers per arm, and the reported p<0.0001 is not a valid estimate of the probability that the process effect is real. The materials analysis is a genuine strength: EELS and SEM directly show Ge residue removal with pad etch, which supports the proposed mechanism and makes a pure measurement artifact unlikely. But it does not substitute for wafer-level replication of the qubit coherence comparison. I therefore recommend keeping the verdict at CONDITIONAL; the condition should be that the authors provide wafer identifiers and either demonstrate multiple wafers per split or perform a wafer-level analysis. If they cannot, the statistical claim of significance should be downgraded.","tokens_in":15972,"tokens_out":10903,"duration_ms":131001,"concrete_test":"Ask the authors for per-qubit data with wafer IDs, chip IDs, qubit frequency, T1, and exclusion reason. First count the number of wafers per process split. If any split has fewer than two wafers, recompute the POR-vs-Process E comparison with wafer as the experimental unit (e.g., a wafer-level permutation test or a mixed-effects model with wafer as a random effect). If the 2x difference remains significant at the wafer level, the concern is resolved. As a secondary sensitivity check, recompute the Process E median after assigning the six excluded qubits the lowest T1 observed in that split; if the POR-vs-E ratio drops materially below 2x, report that sensitivity explicitly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of a statistically significant 2x improvement (Table I, Appendix C) is established by treating each measured qubit as an independent observation (N=42 for Process E, N=59 for POR) in a Tukey-Kramer HSD test. The pad etch, ion mill, and straps decisions are applied at wafer scale before dicing, so all qubits in a split share the same wafer-level processing and deposition run. The manuscript does not state how many wafers underlie each process split. If, as the wafer-level materials study suggests, each condition is represented by one wafer, then the POR-vs-E comparison is effectively a one- or two-wafer comparison; the qubit-level p-values do not capture wafer-to-wafer variability in Jc targeting, oxidation conditions, evaporator state, or handling. Any such batch effect would be fully confounded with the treatment. The consistent ordering of treated splits above POR reduces but does not eliminate this worry, because all treated wafers could share a temporal or tooling batch. The acknowledged exclusion of qubits (6/48 in E) is a secondary, bounded concern; a worst-case reassignment would not obviously erase the median gap, so I would not hang the verdict on it. The absence of wafer-level replication is the more load-bearing issue.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript reports an empirical study of fluorine-based wet surface treatments (dilute HF and a commercial aluminum pad etch) applied to the silicon substrate surface immediately before Josephson-junction deposition in Al-based transmon qubits. Six process splits are compared across a dataset of 182 measured qubits. The central claim is that the combination of pad-etch surface treatment, pre-Junction Ar ion milling, and no strap connections (Process E) improves the median qubit quality factor by roughly 2x over the process of record (median Q = 6.6e6 and T1 = 334 us for E versus Q = 3.3e6 and T1 = 157 us for POR). Supporting materials characterization including STEM, EDS, EELS, and SEM is used to identify germanium-containing residues under the Dolan bridge that arise from the Ge-based shadow-mask dry etch, and to argue that the pad etch removes these residues, thereby reducing a dominant loss channel at the metal-substrate interface underneath the Josephson junction.","tokens_in":16193,"tokens_out":3080,"duration_ms":36952,"significance":"If the central claim holds, the paper identifies a specific, previously underappreciated fabrication-induced loss source (germanium dry-etch byproducts under the Dolan bridge) and demonstrates a simple, manufacturable process modification that more than doubles median transmon energy relaxation time. The study is notable for its scale (182 qubits), for combining device-level coherence statistics with independent materials analysis (STEM/EDS/EELS), and for transparently reporting which qubits were excluded from the analysis. However, the statistical case for the 2x improvement rests on qubit-level p-values from a Tukey-Kramer test applied to wafer-level process splits, and the paper does not report the number of wafers underlying each split. This missing replication information is the main load-bearing weakness; the exclusion pattern of qubits is a second, smaller concern. The materials-science narrative is internally consistent and the attributed mechanism is plausible, but the current statistical framing does not fully rule out wafer-level batch effects as an alternative explanation for part or all of the observed difference.","major_comments":[{"comment":"The central statistical claim (2x improvement, statistically significant) is based on a Tukey-Kramer HSD test that treats each measured qubit as an independent observation, but the process conditions being compared are applied at wafer scale before dicing. The manuscript does not report the number of wafers per process split. If each split is represented by only one or two wafers, then wafer-to-wafer variability in Jc targeting, oxidation conditions, evaporator state, or handling is fully confounded with the treatment, and the qubit-level p-values substantially overstate the significance of the treatment effect. Please report the number of wafers per split and either perform a wafer-level or mixed-effects analysis, or provide evidence (e.g., between-wafer scatter within a split) that wafer-to-wafer variability is negligible relative to the observed 2x effect.","section":"Section IV A, Table I, Appendix C"},{"comment":"The exclusion of qubits from the T1 dataset is acknowledged, but the manuscript states that low-Jc frequency targeting and calibration failures were 'more common on wafers with pre-JJ deposition surface treatments.' Since the excluded fraction is higher in treated splits (e.g., 6/48 in Process E, larger fractions in HF splits), and since low-Jc or miscalibrated qubits could plausibly have different (typically worse) coherence properties, the reported medians may be biased upward. The sentence 'we have no evidence that this is correlated with the coherence improvements' is not a substitute for a quantitative sensitivity analysis. Please provide per-wafer exclusion counts, the frequency distributions of excluded qubits, and a robustness check (e.g., reassigning excluded qubits to low percentiles) to show that the 2x median improvement survives reasonable worst-case assumptions.","section":"Section IV A"},{"comment":"The Tukey-Kramer HSD test compares group means, while Table I reports medians and quartiles. With skewed T1/Q distributions, a statistically significant difference in means does not directly establish the claimed difference in medians, and the p-values do not quantify the 2x median improvement that is the headline claim. Please either report mean-based comparisons alongside the medians or replace the test with a nonparametric comparison (e.g., Wilcoxon/Mann-Whitney) and state unambiguously whether the '2x improvement' statement refers to medians or means.","section":"Table C.1 and Table I"}],"minor_comments":[{"comment":"The abstract states the surface treatments are applied 'with no other changes to the overall process flow,' but the paper actually varies the presence and size of strap connections across splits. Please qualify this statement to avoid overstating the controlled nature of the comparison.","section":"Abstract and Section II"},{"comment":"The footnote for 'Qubits Measured' is defined only in Section IV A; including a brief definition in the table caption would improve readability.","section":"Table I"},{"comment":"The sentence 'P-values of less than 0.05 represent a rejection of the null hypothesis of equal means' is repeated in the caption; please trim redundancy and state the test assumptions explicitly.","section":"Appendix C"},{"comment":"The EELS discussion correctly notes that Ga contamination can produce a false Ge signal, but the statement that the Ge signal in Figure 6b is genuine because 'no corresponding Ga signal is present at that position' would benefit from showing the Ga profile in the same figure or panel for direct comparison.","section":"Section IV B, Figure 6"},{"comment":"The paper does not report the number of cooldowns or refrigerators used for each process split. Since the dataset was acquired over multiple cooldowns and four test setups, a sentence on whether process splits were balanced across measurement setups would help rule out systematic measurement drift as a confound.","section":"Section III"}],"recommendation":"major_revision","confidential_remarks":"The core empirical observation is likely of interest to the superconducting-qubit fabrication community, and the materials evidence is strong. The main gate for acceptance is whether the authors can provide wafer-level replication information or otherwise defend the qubit-level statistical test against the wafer-confounding concern. If the wafer counts are small (e.g., one wafer per split), the manuscript may need to be reframed as a demonstration of a promising process direction rather than a statistically definitive 2x improvement. I would also encourage the editor to ask for the raw per-wafer data to be included as supplementary material."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Steve — quick take. The paper is worth engaging with. The genuinely new piece is that germanium residue from the Ge/MMA Dolan mask, left at the metal-substrate interface under the junction, is a dominant loss source, and that a short pad etch plus Ar ion mill removes it and doubles median T1/Q (334 us vs 157 us, Q 6.6e6 vs 3.3e6). The materials characterization is independent of the device data and makes the mechanism plausible. This is exactly the kind of practical process change the qubit fabrication community needs.\n\nThe soft spots are statistical, not conceptual. The Tukey-Kramer test treats each qubit as independent, but the treatments are wafer-level. The paper doesn't state the number of wafers per split, so we can't rule out a batch effect. The fact that all five treated splits beat POR, and that similar treatments give similar results, argues against a pure batch artifact, but the paper should report wafers and, ideally, do a wafer-level analysis or at least show a wafer-by-wafer breakdown. The second issue is the excluded qubits: 6/48 in E and larger fractions in other treated splits were dropped from the T1 dataset. The authors disclose this and say it's not correlated with coherence, but they don't bound the possible bias. My guess is it wouldn't erase the 2x gap, but it could change the headline number.\n\nThe paper is honest and the data are well presented. I'd send it out. Ask for wafer counts, a wafer-level analysis, and a sensitivity check on the exclusions. The lack of raw data is a minor annoyance; the tables and figures give enough to see the main effect.\n\nThis is for the transmon fabrication crowd, especially anyone using shadow-evaporation masks. I'd take it seriously for its engineering value even if the statistics need tightening.","headline":"Genuinely useful process result with a real statistical caveat: wafer-level replication is not reported.","tokens_in":16850,"tokens_out":3070,"would_cite":true,"duration_ms":36141,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["85.25.Cp","81.65.Cf"],"model":"deepseek-v4-flash","headline":"Fluorine-based surface treatments that remove germanium residue from Josephson-junction fabrication double the median energy relaxation time of aluminum-on-silicon transmon qubits, to 334 microseconds.","keywords":["transmon qubits","Josephson junction fabrication","fluorine surface treatment","germanium contamination","energy relaxation time","quality factor","STEM/EELS materials analysis","shadow evaporation mask"],"falsifier":"Re-measure Process E on wafers after correcting the junction oxidation dose so that all 48 qubits meet frequency targets, and include every qubit in the median; if the all-inclusive median $Q$ drops to the process-of-record level near $3.3 \\times 10^6$, the reported $2\\times$ improvement is an artifact of dataset exclusion.","tokens_in":15777,"feed_emoji":"⚛️","tokens_out":6242,"duration_ms":65328,"temperature":0.7,"pith_summary":"This paper tries to establish that a major source of decoherence in aluminum-on-silicon transmon qubits sits at the metal-substrate interface directly under the Josephson junction, in the form of germanium-containing residue left behind by the shadow-mask fabrication process. The authors show that a fluorine-based wet etch (a commercial pad etch) applied to the silicon before junction deposition removes this residue, and that combining it with an in-vacuo argon ion mill produces a statistically significant twofold improvement in qubit quality. Their best process yields a median energy relaxation time $T_1 = 334\\,\\mu\\mathrm{s}$ and quality factor $Q = 6.6 \\times 10^6$, up from $T_1 = 157\\,\\mu\\mathrm{s}$ and $Q = 3.3 \\times 10^6$ for the process of record. If correct, this gives other qubit foundries a simple, targeted way to lift coherence by cleaning one specific interface rather than changing the base material or junction design.","feed_headline":"Pad etch and ion mill double transmon lifetime to 334 µs","feed_subtitle":"Removing germanium residue left by shadow-mask fabrication gives a 2x quality-factor gain over the process of record.","key_machinery":"The mechanism that carries the argument is a pre-deposition surface-treatment sequence: a buffered fluorine-based pad etch (a commercial etchant containing HF, ammonium fluoride, and acetic acid with aluminum corrosion inhibitors) that strips oxide and residue from the exposed silicon at the junction footprint, followed by in-vacuo argon ion milling that cleans the surrounding silicon surface. The contamination it removes is produced by the CF$_4$ dry etch of a germanium hard mask in the suspended-bridge (Dolan-style) resist stack, whose byproducts are redeposited onto the silicon and remain trapped beneath the bridge where no line-of-sight ion mill can reach them. Materials analysis using cross-sectional scanning transmission electron microscopy with energy-dispersive X-ray and electron-energy-loss spectroscopy is the tool that identifies the germanium layer and verifies its removal.","core_discovery":"The central claim is that the loss limiting these fixed-frequency transmons is dominated by a contamination layer at the silicon-aluminum interface under the Josephson junction, and that this layer originates as germanium dry-etch byproducts from the Ge/MMA resist stack used to form the suspended shadow-evaporation mask. Cross-sectional STEM with EDS and EELS shows a germanium-bearing layer at that interface in untreated devices, absent after pad etch plus argon ion mill; top-down SEM of removed masks shows residue under the bridge that only the wet etch can reach, and residue on the open silicon that the ion mill removes. The authors attribute the measured $2\\times$ improvement in median $Q$ ($6.6 \\times 10^6$ versus $3.3 \\times 10^6$) to removal of this residue, and note that removing the strap connections between junction leads and base metallization gives an additional significant improvement once the interface is clean.","pith_inferences":["If interface cleaning is the true cause, then a direct test would compare quality factors on wafers whose mask stack contains no germanium at all: the pad-etch step should then produce a much smaller improvement, since the presumed contaminant would never be introduced.","The authors report that low-$J_c$ frequency targeting and calibration failures were more common on treated wafers; a testable prediction is that correcting the junction oxidation parameters restores target frequencies without erasing the $Q$ gain, removing the current yield penalty.","Because pad etch attacks exposed aluminum only mildly, similar buffered-fluoride chemistries may be adaptable to niobium- or tantalum-base qubit processes, where sub-junction interfaces are also suspected loss sources.","The observation that $T_2^E$ stays well below the $2T_1$ limit suggests that once relaxation improves to $334\\,\\mu\\mathrm{s}$, dephasing from residual resonator photons or two-level-system dynamics will become the next bottleneck; the paper does not resolve this channel."],"forward_implications":["Other groups fabricating Al/AlOx/Al transmons with germanium-based shadow masks can apply the same pad-etch step without changing junction geometry, base metal, or oxidation parameters, and should expect a similar reduction in sub-junction interface loss.","Because the ion mill regrows a sub-junction oxide yet the pad-etch-plus-mill process still wins, the dominant loss before treatment was the contaminant layer rather than the oxide itself; future work can target the residue directly.","Eliminating strap connections adds a further significant quality-factor improvement once the interface is cleaned, pointing to strap processing as the next interface-loss bottleneck.","Combining device-level $T_1$ statistics with STEM, EDS, and EELS metrology provides a template for attributing qubit decoherence to specific fabrication residues."],"supporting_citations":[{"why":"Establishes the suspended-bridge shadow-evaporation method whose mask stack is the source of the germanium contamination.","marker":"[19]"},{"why":"Documents the germanium-based resist stack that the paper's junction process uses and identifies as the residue source.","marker":"[32]"},{"why":"Shows that sub-micron Josephson-junction fabrication introduces capacitive loss and motivates the strap-processing comparison.","marker":"[28]"},{"why":"Provides the pad-etch chemistry (Silox Vapox III) used as the fluorine-based surface treatment.","marker":"[36]"},{"why":"Demonstrates that microfabrication-induced residues and oxides limit superconducting circuit coherence, the premise the paper builds on.","marker":"[34]"},{"why":"Frames decoherence in terms of two-level-system defects at dielectric interfaces, the loss model the paper invokes.","marker":"[8]"},{"why":"Recent aluminum-on-silicon interfacial dielectric loss mitigation that this work extends with a junction-footprint-specific treatment.","marker":"[17]"}],"fun_headline_variants":["Fluorine etch clears germanium, boosts transmon T1 to 334 µs","Germanium residue removal doubles transmon quality factor","Fluorine surface treatment lifts transmon T1 to 334 µs","Etch removes germanium, transmon T1 doubles to 334 µs","Cleaning germanium from silicon boosts transmon T1 to 334 µs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported $2\\times$ improvement assumes that qubits excluded from the $T_1$ dataset, including those with low-$J_c$ frequency targeting or calibration failures that were more common on treated wafers, are not systematically worse than the qubits that were included.","fun_headline_variants_meta":{"raw":{"variants":["Fluorine etch clears germanium, boosts transmon T1 to 334 µs","Germanium residue removal doubles transmon quality factor","Fluorine surface treatment lifts transmon T1 to 334 µs","Etch removes germanium, transmon T1 doubles to 334 µs","Cleaning germanium from silicon boosts transmon T1 to 334 µs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000562,"raw_usage":{"total_tokens":2653,"prompt_tokens":917,"completion_tokens":1736,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":1636}},"tokens_in":533,"tokens_out":1736,"duration_ms":13452,"temperature":1.0,"reasoning_tokens":1636,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:26:57.368287+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-measure Process E on wafers after correcting the junction oxidation dose so that all 48 qubits meet frequency targets, and include every qubit in the median; if the all-inclusive median $Q$ drops to the process-of-record level near $3.3 \\times 10^6$, the reported $2\\times$ improvement is an artifact of dataset exclusion.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the suspended-bridge shadow-evaporation method whose mask stack is the source of the germanium contamination."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the germanium-based resist stack that the paper's junction process uses and identifies as the residue source."},{"cited_title":"Dunsworth, A","cited_arxiv_id":null,"evidence_quote":"Shows that sub-micron Josephson-junction fabrication introduces capacitive loss and motivates the strap-processing comparison."},{"cited_title":"Dai, A maskless wet etching silicon dioxide post- CMOS process and its application, Microelectronic En- gineering Materials for Advanced Metallization (MAM 2006), 83, 2543 (2006)","cited_arxiv_id":null,"evidence_quote":"Provides the pad-etch chemistry (Silox Vapox III) used as the fluorine-based surface treatment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates that microfabrication-induced residues and oxides limit superconducting circuit coherence, the premise the paper builds on."},{"cited_title":"M¨ uller, J","cited_arxiv_id":null,"evidence_quote":"Frames decoherence in terms of two-level-system defects at dielectric interfaces, the loss model the paper invokes."},{"cited_title":"Bizn´ arov´ a, A","cited_arxiv_id":null,"evidence_quote":"Recent aluminum-on-silicon interfacial dielectric loss mitigation that this work extends with a junction-footprint-specific treatment."}],"review_version":1}