{"id":"a695f715-e3ef-4b51-9027-2d933b0fa2a0","arxiv_id":"2607.08617","paper_version":1,"verdict":"CONDITIONAL","confidence":"UNKNOWN","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"Large-scale MLIP simulations of amorphous In₂O₃ reveal percolating edge-sharing polyhedra chains as the structural basis for high mobility and O–O peroxide bonds that introduce in-gap states detectable by Raman at 850 cm⁻¹.","lead":"Researchers used a machine-learned interatomic potential to simulate amorphous indium oxide at unprecedented scales (5120 atoms), finding that chains of edge-sharing polyhedra explain its high electron mobility and that O–O peroxide bonds create in-gap states detectable by Raman spectroscopy. This matters because amorphous indium oxide is a key material in display transistors and could become a better photoanode for solar water splitting.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"MLIP O–O energetics are not independently validated despite energy differences of only ~0.015 eV/f.u. separating the central claim from its negation.","rationale":"The reader correctly identified the MLIP's accuracy for O–O bond energetics as the weakest assumption. I concur fully: the ~0.015 eV/f.u. energy difference is smaller than typical MLIP force/energy errors, and no independent DFT benchmarking on the large cells is provided. This is the single most load-bearing concern because the entire electronic-structure narrative—in-gap states, n-type self-doping, enhanced sub-gap absorption, the Raman prediction—depends on O–O bonds being genuinely present in the amorphous network, not artifacts of MLIP extrapolation beyond its training distribution. The paper does provide strong independent support for its structural claims (X-ray structure factors in excellent agreement with experiment, Figure 3), and the Raman prediction at 850 cm⁻¹ is a genuinely falsifiable experimental signature. The electronic-structure calculations at both SCAN and DDH levels (Figures 9–11) are internally consistent in showing that O–O bonds introduce in-gap states. The concern is specifically about whether the MLIP correctly predicts the thermodynamic prevalence of O–O bonds in the large cells, not about the electronic-structure methodology itself. The single 5120-atom sample is a secondary limitation—statistical confidence in the 2% O–O fraction would require multiple independent samples, but the 640-atom statistics (40/45 samples) are more robust if the energetics are trustworthy. The broken figure cross-references ('??') are a presentation issue, not a substantive concern. I recommend keeping the verdict at CONDITIONAL: the work is potentially important and the Raman prediction is testable, but the central claim about O–O bond prevalence needs the DFT validation check described above before it can be fully accepted. If that check confirms MLIP–DFT agreement on O–O energetics within ~0.01 eV/f.u., the verdict should move to ACCEPT.","tokens_in":15198,"tokens_out":994,"duration_ms":283329,"concrete_test":"Select 5–10 representative 640-atom structures from the MLIP trajectories (half with O–O bonds, half without, spanning the energy distribution in Figure 5). Recompute total energies with single-point SCAN DFT calculations. Check: (1) Does DFT agree with the MLIP on the sign of the energy difference between O–O and non-O–O structures? (2) Is the DFT energy difference still within ~0.015 eV/f.u.? If the MLIP systematically over-stabilizes O–O bonds by even 0.02 eV/f.u. relative to DFT, the claimed ~90% prevalence in 640-atom cells would be unreliable.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim—that O–O peroxide-like bonds form in a-In₂O₃ and introduce in-gap states near the CBM—rests on the MLIP correctly capturing O–O bond energetics in 640- and 5120-atom cells. The training set is drawn from 80-atom FPMD trajectories (27 trajectories), where O–O bonds appear in ~50% of samples. The MLIP is then extrapolated to 640-atom cells (where O–O bonds appear in ~90% of samples) and one 5120-atom cell (32 O–O bonds). This dramatic increase in O–O prevalence with cell size could reflect genuine physics, but it could also reflect a systematic bias in the MLIP's extrapolation regime. The critical problem is that the energy difference between structures with and without O–O bonds is only ~0.015 eV/f.u. (Figure 5, §III.B). The paper does not report any independent DFT validation of the MLIP's O–O energetics on the larger cells—no single-point DFT calculations on 640-atom MLIP-generated structures with and without O–O bonds to confirm that DFT agrees with the MLIP on which is lower in energy, or on the magnitude of the difference. Given that the entire electronic-structure argument (in-gap states, n-type self-doping, enhanced sub-gap absorption) depends on O–O bonds actually being thermodynamically accessible (not just kinetically trapped), this is the weakest link. The 80-atom DDH electronic structure calculations (Figure 10) validate the electronic consequences of O–O bonds but not their thermodynamic prevalence in the large cells where they are claimed to be abundant.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"This manuscript presents large-scale simulations of amorphous In₂O₃ (a-In₂O₃) using a MACE-based machine-learned interatomic potential (MLIP) trained on first-principles molecular dynamics (FPMD) data. The authors generate amorphous structures of up to 5120 atoms—two orders of magnitude larger than typical ab initio studies—and report X-ray structure factors in quantitative agreement with experiment. The paper makes three principal claims: (i) extended chains of edge-sharing InOₖ polyhedra provide a structural basis for high electron mobility; (ii) O–O peroxide-like bonds (mean length ~1.5 Å) form in the amorphous network, introducing localized in-gap states near the conduction band minimum that act as intrinsic n-type self-doping; and (iii) these bonds produce a distinct Raman feature near 850 cm⁻¹ absent in the IR spectrum, offering an experimentally testable prediction. The combination of structural, vibrational, and electronic-structure analysis, together with the scale of the simulations, represents a substantial contribution to the understanding of a-In₂O₃.","tokens_in":16021,"tokens_out":1511,"duration_ms":363373,"significance":"The paper addresses a long-standing debate about the structure and transport properties of a-In₂O₃, a technologically important semiconductor. The key strengths include: (1) the achievement of simulation scales (5120 atoms) that reveal percolating polyhedral chains inaccessible to prior ab initio work; (2) quantitative agreement with experimental X-ray scattering data (Figure 3C), providing external validation of the structural model; (3) a falsifiable experimental prediction—the Raman-active O–O stretch near 850 cm⁻¹—that is directly testable; and (4) electronic-structure calculations at both SCAN and dielectric-dependent hybrid (DDH) levels connecting the O–O bonds to in-gap states and sub-gap optical absorption. The structure–property picture is coherent and the photoelectrochemical implications are well-motivated.","major_comments":[{"comment":"§III.B, Figure 5: The central claim that O–O peroxide-like bonds form in a-In₂O₃ rests on the MLIP correctly capturing O–O bond energetics in 640- and 5120-atom cells. The reported energy difference between structures with and without O–O bonds is only ~0.015 eV/f.u., which is comparable to the expected accuracy of the MLIP relative to DFT. The paper does not report any independent DFT validation (e.g., single-point DFT calculations) on MLIP-generated 640-atom structures with and without O–O bonds to confirm that DFT agrees with the MLIP on the sign and magnitude of this energy difference. Given that the entire electronic-structure argument (in-gap states, n-type self-doping, enhanced sub-gap absorption) depends on O–O bonds being thermodynamically accessible and not merely kinetically trapped, at least one direct DFT cross-check on a representative 640-atom structure should be performed","section":null},{"comment":"§III.C, Figure 8: The identification of extended chains of edge-sharing InOₖ polyhedra is a headline result, but the statistics are based on a single 5120-atom sample. The reported average chain length (12.18 ± 2.41 In atoms) carries an uncertainty derived from one configuration. Since this result is presented as a structural basis for high electron mobility, generating at least one additional independent 5120-atom sample (or several 640-atom samples) to confirm the persistence and typical length of these chains would substantially strengthen the claim. The authors should also clarify what fraction of In atoms belongs to percolating chains versus isolated clusters.","section":null}],"minor_comments":[{"comment":"The title reads 'first-principle simulations' but should read 'first-principles simulations.'","section":null},{"comment":"Several cross-references in the manuscript are broken (e.g., 'Figure ??' appears multiple times in §III.B and §III.C, and 'section ??' in §III.C). These should be fixed before publication.","section":null},{"comment":"§II.A: The DDH exact exchange fraction α is stated to be determined from ε∞ = 4.2 evaluated at the hybrid SE-RSH level for crystalline In₂O₃. It would be helpful to state the resulting value of α explicitly for clarity.","section":null},{"comment":"§III.A: The comparison of total g(r) with previous works (Figure 2) notes quantitative discrepancies in In–O bond distances attributed to the use of SCAN versus PBE. A brief comment on whether these differences are expected from known SCAN versus PBE trends in oxide bond lengths would strengthen this argument.","section":null},{"comment":"§IV: The 80-atom DDH calculations (Figure 10) yield a band gap of 2.27 eV, while the 640-atom SCAN calculations (Figure 9) are used for the EDOS analysis. The authors note finite-size effects on the conduction density of states. A brief discussion of whether the gap reduction and in-gap state positions are expected to change at the hybrid level for the 640-atom cells would aid comparison.","section":null},{"comment":"Figure 6B: The Raman spectrum is computed at the LDA level for an 80-atom cell. The authors note that VDOS calculations with SCAN and LDA yield consistent results, but the Raman intensities depend on the Raman tensor. A brief justification for using LDA for Raman intensities, or a note on the expected level of agreement, would be helpful.","section":null},{"comment":"§III.B: The defect formation energy of 0.04 eV/f.u. for the peroxide interstitial in the recovered crystalline structure is reported as an average over 10 snapshots. The spread of these values and the structural definition of the defect should be briefly described for direct comparison.","section":null},{"comment":"References [10] and [38] are cited with 2026 dates. If these are accepted/in-press manuscripts, the DOI and journal information should be complete and not placeholder dates.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The reader's stress-test concern about MLIP O–O energetics validation is well-placed and is the basis for my first major comment. The energy difference of 0.015 eV/f.u. is small enough that MLIP–DFT disagreement could plausibly flip the thermodynamic conclusion. However, I note that the paper's language is appropriately hedged ('strongly support,' 'likely presence'), and the O–O bond is also observed in 50% of 80-atom FPMD samples (i.e., directly from DFT, not just MLIP). The MLIP is used to extend statistics and access larger scales, not as the sole evidence for O–O bond existence. Thus the concern is about quantitative prevalence and thermodynamic accessibility in large cells, not about whether the bonds exist at all. A single DFT cross-check on a 640-atom structure would likely resolve this. I do not view the single 5120-atom sample as a fatal limitation given the 640-atom ensemble, but additional statistics on chain lengths would strengthen the mobility claim."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The headline: this paper does two genuinely new things. It scales amorphous In2O3 simulations to 5120 atoms using a MACE-based MLIP trained on SCAN-level FPMD, and it uses that scale to identify percolating edge-sharing InOk polyhedra chains as a structural basis for high electron mobility. The X-ray structure factor agreement with experiment (Figure 3C) is strong and provides real external validation. The Raman prediction at 850 cm-1 for O-O peroxide bonds, absent in IR, is a clean falsifiable claim that elevates the work above pure computation. These are real contributions and the paper deserves credit for them. The structural analysis is thorough—RDFs, angular distributions, polyhedral connectivity—and the comparison with prior computational work is fair and specific about where results differ (e.g., edge-sharing angle at 97 degrees vs. 104 degrees from Buchholz et al.). The electronic structure story (in-gap states from O-O bonds near CBM, n-type self-doping, enhanced sub-gap absorption) is internally consistent across SCAN and DDH levels. Now the soft spot, and it is a real one. The central claim about O-O bond prevalence rests on the MLIP correctly capturing O-O energetics in 640- and 5120-atom cells, but the energy difference between structures with and without O-O bonds is only ~0.015 eV/f.u. (Figure 5). The training set is 27 FPMD trajectories on 80-atom cells. There is no independent DFT validation on the larger cells—no single-point SCAN calculations on MLIP-generated 640-atom structures with and without O-O bonds to confirm DFT agrees with the MLIP on which is lower in energy. The increase in O-O prevalence from ~50% of 80-atom samples to ~90% of 640-atom samples could be genuine physics (better sampling of rare motifs in larger cells) or could reflect MLIP extrapolation bias. At 0.015 eV/f.u., small systematic errors in the potential flip the conclusion. This is the load-bearing gap. Secondary issues: only one 5120-atom sample was generated, so chain-length statistics and the 2% O-O fraction are single-realization numbers. The DDH electronic structure is limited to 80-atom cells with acknowledged finite-size effects on conduction band DOS. Several figure cross-references are broken ('??'). None of these secondary issues are dealbreakers on their own. The stress-test concern about O-O energetics is the one that matters. It does not invalidate the paper—the X-ray validation and structural analysis stand independently—but it leaves the O-O bond prevalence claim under-supported. A few single-point DFT calculations on selected 640-atom MLIP structures would close the gap. This paper is for researchers working on amorphous oxide semiconductors, MLIP development for oxides, and photoelectrochemistry. It deserves a serious referee who can assess the MLIP methodology and push for the missing DFT cross-check on O-O energetics.","headline":"Solid large-scale MLIP study of a-In2O3 with one load-bearing gap: O-O bond energetics are not independently validated at the scale where they matter most.","tokens_in":16207,"tokens_out":722,"would_cite":true,"duration_ms":162458,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Peroxide bonds in amorphous indium oxide act as built-in dopants","keywords":["amorphous indium oxide","peroxide bonds","machine-learned interatomic potential","intrinsic n-type doping","electron mobility","Raman spectroscopy","photoanode","water oxidation"],"falsifier":"Measure Raman spectra of amorphous In₂O₃ films deposited under controlled conditions. If no peak appears near 850 cm⁻¹, or if the peak's intensity does not correlate with carrier density, the claim that peroxide bonds are the source of intrinsic n-type doping is undermined.","tokens_in":15345,"feed_emoji":"⛓️","tokens_out":1273,"duration_ms":199216,"temperature":0.7,"pith_summary":"Amorphous indium oxide is a widely used semiconductor that retains high electron mobility even when its crystalline order is lost — a behavior that is unusual and poorly understood. This paper uses a machine-learned interatomic potential, trained on first-principles molecular dynamics data, to simulate amorphous structures of up to 5120 atoms, far larger than previous studies. At this scale, the authors identify two structural features they argue are central to the material's properties. First, they find extended chains of edge-sharing indium–oxygen polyhedra that span the simulation cell, providing a structural pathway for the high electron mobility that has been observed experimentally but never connected to a specific atomic-scale motif. Second, they find that oxygen–oxygen peroxide-like bonds, about 1.5 Å long, form readily in the amorphous network. These bonds introduce filled electronic states just below the conduction band edge, functioning as an intrinsic source of n-type doping — meaning the material can generate charge carriers without any external impurity. The bonds also enhance absorption of below-gap light, which matters for using the material as a photoanode to split water. The authors predict that these bonds produce a sharp Raman signal near 850 cm⁻¹ that is absent from infrared spectra, offering a way to detect and quantify them experimentally. They also find that the amorphous phase is only about 0.7 eV per formula unit above the crystal in energy, consistent with the experimental observation that indium oxide is a poor glass former prone to quasi-crystalline regions.","feed_headline":"Peroxide bonds in amorphous indium oxide act as built-in dopants","feed_subtitle":"Large-scale simulations reveal the structural motifs behind high mobility and self-doping, with a testable Raman signature at 850 cm⁻¹.","key_machinery":"The argument works through a chain of scale and observation: (1) a MACE machine-learned potential, trained on first-principles molecular dynamics trajectories, enables simulations at 5120 atoms — large enough to see percolating chains of polyhedra that 80-atom cells cannot resolve; (2) at these scales, O–O peroxide bonds appear in the majority of independently generated samples, with a mean length of 1.5 Å; (3) electronic structure calculations at both SCAN and hybrid DDH levels show these bonds introduce filled states just below the conduction band minimum; (4) inverse participation ratio analysis confirms the in-gap states are localized on the peroxide bond while the associated donor state","core_discovery":"The central claim is that amorphous indium oxide contains two previously unresolved structural motifs — extended chains of edge-sharing InO_k polyhedra and short O–O peroxide-like bonds — and that these motifs respectively explain the material's anomalously high electron mobility and its intrinsic n-type conductivity. The peroxide bonds introduce localized in-gap states near the conduction band minimum, acting as self-dopants, and produce a detectable Raman signature at 850 cm⁻¹ that is absent from IR spectra.","pith_inferences":["The energy difference between structures with and without O–O bonds is only ~0.015 eV per formula unit, meaning that whether these bonds form could be highly sensitive to deposition conditions, annealing history, and oxygen partial pressure — suggesting a route to controllable doping.","If peroxide bonds are a general feature of amorphous wide-bandgap oxides (as hinted by the authors' reference to other systems), the self-doping mechanism identified here may extend beyond indium oxide to materials like amorphous zinc oxide or gallium oxide.","The quasi-crystalline regions found in the amorphous samples raise the question of whether the high mobility actually requires a fully amorphous phase or whether it is sustained by these crystalline remnants — a distinction that matters for device engineering."],"forward_implications":["If the Raman feature at 850 cm⁻¹ is confirmed experimentally, it would provide a direct, quantitative diagnostic for O–O bond concentration in deposited amorphous films, linking deposition conditions to carrier density.","The identification of percolating edge-sharing polyhedra as the mobility pathway suggests that deposition protocols favoring longer chains could further boost electron transport.","The self-doping mechanism via peroxide bonds offers a way to tune n-type carrier density without introducing external impurities, which could reduce scattering and improve mobility.","If amorphization reliably activates optically dark transitions present in the crystal, a-In₂O₃ becomes a more viable photoanode candidate for solar water splitting."],"fun_headline_variants":["Amorphous indium oxide structure decoded in 5120-atom simulations","Edge-sharing polyhedra explain high mobility in amorphous In₂O₃","O-O peroxide bonds cause intrinsic n-type doping in In₂O₃","Raman signature at 850 cm⁻¹ tracks peroxide bonds in amorphous In₂O₃","Quasi-crystalline regions confirmed in amorphous indium oxide"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The machine-learned potential, trained primarily on 80-atom first-principles trajectories, is assumed to accurately reproduce the energetics of O–O peroxide bonds when extrapolated to cells 64–640 times larger. The energy difference between structures with and without these bonds is only about 0.015 eV per formula unit, so even small errors in the potential's description of the O–O interaction could change the conclusion about whether these bonds are genuinely prevalent or an","fun_headline_variants_meta":{"raw":{"variants":["Amorphous indium oxide structure decoded in 5120-atom simulations","Edge-sharing polyhedra explain high mobility in amorphous In₂O₃","O-O peroxide bonds cause intrinsic n-type doping in In₂O₃","Raman signature at 850 cm⁻¹ tracks peroxide bonds in amorphous In₂O₃","Quasi-crystalline regions confirmed in amorphous indium oxide"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":795,"prompt_tokens":684,"completion_tokens":111,"prompt_tokens_details":null},"tokens_in":684,"tokens_out":111,"duration_ms":83592,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T04:10:38.503411+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Measure Raman spectra of amorphous In₂O₃ films deposited under controlled conditions. If no peak appears near 850 cm⁻¹, or if the peak's intensity does not correlate with carrier density, the claim that peroxide bonds are the source of intrinsic n-type doping is undermined.","supporting_citations":[],"review_version":1}