{"id":"246adf3d-7c59-4e74-8718-18fa5d75ffa5","arxiv_id":"2607.06470","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":10,"one_line_summary":"Machine-learning interatomic potentials (GAP, MACE) combined with lattice dynamics and NEMD predict thermal boundary resistance at silicon grain boundaries that is sensitive to interfacial roughness and differs substantially from classical potential predictions at high roughness.","lead":"This paper uses machine-learning interatomic potentials to simulate how heat moves through silicon with grain boundaries, finding that ML potentials predict different thermal boundary resistance than classical potentials, especially at rough interfaces. The work matters for designing nanoscale devices where heat dissipation across interfaces limits performance.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The retrained GAP (Si_Phonon.xml) used for headline NEMD TBR results is trained exclusively on bulk crystalline Si configurations; its accuracy for grain-boundary environments is never validated against DFT, making the claim that GAP gives 'more accurate' TBR than classical potentials rest on an unt","rationale":"The reader correctly identified a real circularity concern (MACE trained on GAP-generated GB data for Figs. 5–9) and noted the ambiguous provenance of the retrained GAP. However, I believe the more load-bearing concern is slightly different: it is not just that MACE inherits GAP's potential GB inaccuracy for the lattice-dynamical figures, but that the retrained GAP itself — which produces the headline TBR numbers in Table I — is trained exclusively on bulk crystalline configurations and is never validated for grain-boundary environments against any DFT reference. This strikes the central claim more directly because the TBR comparison is the paper's strongest quantitative result. The chain of training is: DFT → original GAP [54] (crystalline + amorphous) → retrained GAP (bulk crystal only) → GB MACE (GB configs from retrained GAP). Each distillation step narrows the training domain, and the retrained GAP's application to disordered GB structures is an unvalidated extrapolation. That said, the CONDITIONAL verdict with MODERATE confidence is appropriate: the bulk validation is sound, the NEMD methodology is standard, and the experimental TBR ranges cited are broad enough that the GAP values are plausible. The concern does not invalidate the work but identifies a critical gap in validation that should be addressed before the TBR comparison can be considered reliable. The verdict remains CONDITIONAL; the confidence in correctness_risk should perhaps be lowered from 'unknown' to 'moderate concern' given the unvalidated extrapolation.","tokens_in":20544,"tokens_out":3861,"duration_ms":296844,"concrete_test":"Compute DFT (PBE) energies and forces for ~50–100 grain-boundary configurations sampled from the NEMD equilibration trajectories (including the A=3 Å roughened bicrystal). Compare the retrained GAP (Si_Phonon.xml) force RMSE on these GB configurations against its force RMSE on bulk test configurations. If the GB force RMSE exceeds the bulk RMSE by more than a factor of 2–3, the GAP's transferability to GB environments is unreliable and the TBR comparison in Table I cannot be attributed to improved physical accuracy. Additionally, compare GAP-predicted GB formation energies against available DFT literature values for symmetric tilt boundaries in Si.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that GAP predicts more accurate thermal boundary resistance (TBR = 1.66 m²K/GW at A=3 Å) than SW (1.08) and Tersoff (1.17), bringing predictions closer to experiment. The NEMD TBR values in Table I are computed using the retrained GAP (Si_Phonon.xml), which was trained only on configurations from 300 K MD of bulk crystalline silicon (Sec. II: 'Atomic configurations extracted from the equilibrated trajectories... were stored... and used as the reference dataset for subsequent model training'). The original GAP of Bartók et al. [54] was trained on DFT data spanning crystalline and amorphous configurations, giving it transferability to defected environments. But the retrained GAP only sees bulk crystalline environments filtered through the original GAP — it is a distilled model with a narrower training domain. When applied to grain-boundary structures with significant structural disorder (roughness amplitudes up to 3 Å, misorientation angles up to 40°), this model is extrapolating beyond its training distribution. No validation against DFT energies, forces, or stresses for grain-boundary configurations is presented anywhere in the paper. The claim that GAP's higher TBR reflects 'more accurate' phonon scattering at the GB (rather than an artifact of poor transferability, e.g., incorrectly stiff or soft bonds at the interface) is therefore unverified. This concern is distinct from but related to the reader's identified issue of MACE being trained on GAP-generated GB data (which affects Figs. 5–9). The concern here directly undermines the headline TBR comparison in Table I, which is the paper's strongest quantitative claim. If the retrained GAP misrepresents the potential energy surface at the GB, the entire TBR comparison between GAP and classical potentials could be systematically biased.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This manuscript combines Gaussian Approximation Potential (GAP) and MACE machine-learning interatomic potentials with lattice-dynamical calculations (Phonopy/Phono3py) and non-equilibrium molecular dynamics (NEMD) to study phonon-mediated thermal transport in bulk and nanocrystalline silicon. The bulk validation (Figs. 1–3) demonstrates that the retrained GAP and MACE models reproduce experimental phonon dispersions, lifetimes, and thermal conductivity more faithfully than Stillinger–Weber and Tersoff potentials. The paper then proceeds to grain-boundary (GB) phonon properties (Figs. 5–9) computed via a MACE model trained on GAP-generated GB data, and to NEMD thermal boundary resistance (TBR) calculations (Table I, Fig. 11) using the retrained GAP directly. The headline finding is that GAP predicts substantially higher TBR (1.66 m²K/GW at A=3 Å) than SW (1.08) and Tersoff (1.17), bringing predictions closer to experimental values for rough interfaces.","tokens_in":20806,"tokens_out":4170,"duration_ms":197189,"significance":"The integrated framework combining MLIP-derived force constants with both lattice dynamics and NEMD for interfacial thermal transport is timely and addresses a genuine gap in the literature. The bulk silicon benchmarking (Figs. 1–3) is well executed, with both GAP and MACE showing good agreement with experimental phonon dispersions and thermal conductivity. The systematic study of TBR as a function of roughness amplitude across multiple potentials (Table I) provides useful comparative data. The provision of the trained potential file (Si_Phonon.xml) supports reproducibility. However, the significance of the interfacial results is diminished by the training-data lineage issues identified below.","major_comments":[{"comment":"Section II (paragraph beginning 'Atomic configurations extracted from the equilibrated trajectories') and Section III (paragraph beginning 'Atomistic simulations of both near-GB and GB regions'): The retrained GAP (Si_Phonon.xml) used for all grain-boundary simulations and NEMD TBR calculations (Table I, Fig. 11) is trained exclusively on configurations from 300 K MD of bulk crystalline silicon. The original GAP of Bartók et al. [54] was trained on DFT data spanning crystalline and amorphous configurations, but the retrained model sees only bulk crystalline environments filtered through that original GAP. When applied to grain-boundary structures with significant structural disorder (roughness amplitudes up to 3 Å, misorientation angles up to 40°), this model is extrapolating beyond its training distribution. No validation against DFT energies, forces, or stresses for any grain-boundary,","section":null},{"comment":"Section III (paragraph beginning 'Atomic positions, forces, and per-atom energies were recorded'): The MACE model used for all grain-boundary lattice-dynamical results (phonon DOS, lifetimes, mean free paths, thermal conductivity in Figs. 5–9) is trained on energies and forces generated by the retrained GAP (Si_Phonon.xml), not on DFT reference data. This creates a dependency chain: the GB phonon properties reflect the retrained GAP's extrapolation to GB environments, not independent first-principles-level predictions. If the retrained GAP's description of GB bonding is inaccurate (see comment 1), MACE inherits that inaccuracy. The manuscript does not acknowledge this limitation. At minimum, the authors should (a) explicitly state that the GB MACE training labels come from GAP, not DFT, (b) provide DFT validation for a subset of GB configurations to assess the fidelity of the GAP-derived","section":null}],"minor_comments":[{"comment":"The relationship between the original GAP [54] and the retrained GAP (Si_Phonon.xml) should be stated more explicitly. A reader could easily miss that the retrained model has a narrower training domain than the original.","section":null},{"comment":"Table I: the experimental TBR ranges are listed without specific references in the table footnote. The text cites Ref. [30] (Isotta et al.) in the paragraph following Table I, but the table itself should cite its sources.","section":null},{"comment":"Figure 9(d): the label 'GB core 0°' is somewhat confusing since the text describes this as a reference system without a true grain boundary. A clearer label such as '0° (no misorientation)' would help.","section":null},{"comment":"Section IV, Eq. (1): the notation T_R(x_GB) and T_L(x_GB) is introduced but the subscripts R and L are not defined at that point. Adding 'right' and 'left' would improve readability.","section":null},{"comment":"The SOAP descriptor parameters (nmax=7, lmax=6, ζ=2, 800 sparse points) are given in Section II, but the two-body descriptor parameters and regularization values are not specified. These should be included or referenced in the Supplementary Material.","section":null},{"comment":"Supplementary Material S2: the MACE training convergence is shown only for energy RMSE (Fig. S5). Force RMSE curves would provide a more complete picture of model quality, especially for the GB-core configurations.","section":null},{"comment":"The NEMD production run of 2 ns (Section IV) is relatively short for convergence of thermal conductivity. The κ(t) curves in Fig. 11(a) and Figs. S9–S10 show ongoing fluctuations. A brief discussion of statistical uncertainty (error bars on the TBR values in Table I) would strengthen the quantitative claims.","section":null}],"recommendation":"major_revision","confidential_remarks":"The core issue is that the paper's headline interfacial results rest on a retrained GAP whose training domain (bulk crystalline Si only) does not include the disordered GB environments to which it is applied. This is compounded by the MACE-for-GB being trained on GAP labels rather than DFT. The bulk validation is solid and the framework is interesting, but without any DFT validation for GB configurations, the claim that MLIPs give 'more accurate' interfacial TBR than classical potentials is not substantiated—it could be that the retrained GAP is simply wrong in a different way than SW/Tersoff. The authors should be asked to either (a) validate the retrained GAP against DFT for representative GB structures, or (b) use the original GAP [54] (which was trained on DFT data including amorphous configurations) for the GB/NEMD calculations, or (c) substantially temper their accuracy claims for the interfacial results. Option (b) seems most straightforward if the original GAP is available. If the authors can address this, the paper could become a solid contribution."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful reading and for identifying two important limitations of the manuscript regarding the training-data lineage of the retrained GAP and the GB MACE model. Both comments are substantively correct: the retrained GAP (Si_Phonon.xml) was trained exclusively on bulk crystalline configurations, and the GB MACE model was trained on GAP-derived labels rather than DFT data. We will revise the manuscript to explicitly acknowledge these limitations and will add DFT validation for a subset of grain-boundary configurations to quantify the fidelity of the GAP-derived labels. We also note that the original Bartók et al. GAP, from which our training labels were generated, was itself trained on DFT data spanning crystalline and amorphous environments, which provides partial coverage of disordered bonding motifs.","responses":[{"response":"The referee is correct that the retrained GAP (Si_Phonon.xml) was trained on configurations sampled from 300 K MD of bulk crystalline silicon, and that direct DFT validation for grain-boundary configurations is not provided in the current manuscript. We will address this in two ways. First, we will add an explicit statement in the revised manuscript acknowledging that the retrained GAP's training distribution is limited to bulk crystalline environments and that application to grain-boundary structures constitutes extrapolation beyond this distribution. Second, we will perform DFT calculations (PBEsol or similar) on a representative subset of the grain-boundary configurations used in this work—including the 20° and 40° misorientation angles and the A = 3 Å roughness case—to validate the retrained GAP's energies and forces against first-principles reference data. We note that the training labels for Si_Phonon.xml were themselves generated using the original Bartók et al. GAP [Ref. 54], which was trained on DFT data spanning crystalline and amorphous silicon configurations including liquid and defected structures. The amorphous and liquid training environments in the original GAP's DFT dataset include under-coordinated atoms and distorted bonding geometries that partially overlap with grain-boundary local environments. This provides a degree of indirect coverage of disordered bonding motifs through the teacher model, though we agree it does not substitute for direct DFT validation of GB structures. We will report the DFT validation results in the revised manuscript and Supplementary Material.","revision_made":"yes","referee_comment":"The retrained GAP (Si_Phonon.xml) used for all grain-boundary simulations and NEMD TBR calculations is trained exclusively on configurations from 300 K MD of bulk crystalline silicon. The original GAP of Bartók et al. was trained on DFT data spanning crystalline and amorphous configurations, but the retrained model sees only bulk crystalline environments filtered through that original GAP. When applied to grain-boundary structures with significant structural disorder, this model is extrapolating beyond its training distribution. No validation against DFT energies, forces, or stresses for any grain-boundary configuration is provided."},{"response":"The referee is correct on both points. The GB MACE model was trained on energies and forces generated by the retrained GAP, not on DFT data, and the manuscript does not currently acknowledge this dependency chain. We will revise the manuscript to: (a) explicitly state in Section III that the MACE model used for grain-boundary lattice-dynamical calculations was trained on labels from the retrained GAP (Si_Phonon.xml), not from DFT, and that this creates a teacher–student dependency in which any inaccuracy in the GAP's description of GB bonding is inherited by MACE; (b) provide DFT validation for a subset of GB configurations, comparing DFT energies and forces against both the retrained GAP and the GB MACE model. This will allow us to quantify the fidelity of the GAP-derived labels and the extent to which the MACE model reproduces them. We will add these validation results to the revised manuscript and Supplementary Material. We note that the GB lattice-dynamical results (Figs. 5–9) and the NEMD TBR results (Table I, Fig. 11) are based on different models: the former use the GB MACE model trained on GAP labels, while the latter use the retrained GAP directly. The DFT validation we add will cover configurations relevant to both sets of results.","revision_made":"yes","referee_comment":"The MACE model used for all grain-boundary lattice-dynamical results is trained on energies and forces generated by the retrained GAP (Si_Phonon.xml), not on DFT reference data. This creates a dependency chain: the GB phonon properties reflect the retrained GAP's extrapolation to GB environments, not independent first-principles-level predictions. If the retrained GAP's description of GB bonding is inaccurate, MACE inherits that inaccuracy. The manuscript does not acknowledge this limitation. At minimum, the authors should (a) explicitly state that the GB MACE training labels come from GAP, not DFT, (b) provide DFT validation for a subset of GB configurations to assess the fidelity of the GAP-derived labels."}],"tokens_in":20401,"tokens_out":1056,"duration_ms":189968,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The headline: this paper applies GAP and MACE interatomic potentials to phonon transport in nanocrystalline silicon, including NEMD-based thermal boundary resistance (TBR) at roughened grain boundaries. The bulk validation (Figs. 1–3) is genuinely well done — GAP and MACE reproduce phonon dispersions, lifetimes, and thermal conductivity in good agreement with experiment and first-principles data, while SW and Tersoff show the expected systematic overestimation of lifetimes and conductivity. That part is clean and useful as a benchmarking exercise. The NEMD framework for TBR with controlled roughness amplitudes is a legitimate methodological application, and the systematic comparison across potentials at varying roughness has not been reported before. Credit is earned for the bulk work and for the NEMD setup itself. The paper also ships the trained GAP as an XML file, which is a real reproducibility plus. The TBR trends — increasing resistance with roughness, GAP predicting larger TBR than classical potentials at high roughness — are qualitatively plausible and land within experimental ranges cited from Ref. 30. So the framework is reasonable and the qualitative story is sound. The problem is that the central quantitative claim — GAP gives more accurate TBR than SW/Tersoff — is not actually validated. The retrained GAP (Si_Phonon.xml) used for the NEMD TBR results in Table I is trained exclusively on configurations from 300 K MD of bulk crystalline silicon (Sec. II). The original Bartók GAP (Ref. 54) was trained on DFT data spanning crystalline and amorphous configurations, giving it transferability to defected environments. The retrained version is a distilled model with a narrower training domain. When applied to grain boundaries with misorientation angles up to 40° and roughness amplitudes up to 3 Å, it is extrapolating beyond its training distribution. No validation against DFT energies, forces, or stresses for grain-boundary configurations is presented anywhere. So the claim that GAP's higher TBR reflects more accurate phonon scattering — rather than an artifact of poor transferability at the interface — is unverified. This is the load-bearing concern. The reader also flagged that the grain-boundary lattice-dynamical results (Figs. 5–9) use a MACE model trained on GAP-generated data, not DFT data. That concern is real but secondary: those figures are descriptive (phonon DOS, lifetimes, MFPs in the GB region) and not the headline quantitative claim. The circularity means those results should be read as GAP-consistent, not independently validated. The NEMD TBR values in Table I also lack error bars and use short production times (2 ns), which is a minor but real statistical concern. The stress-test concern about the retrained GAP's training domain is the most important issue and it does land on reading the paper. The paper does not address it. Who is this for? Researchers working on MLIP-based thermal transport simulations who want a methodological template and a benchmarking study. The bulk validation and NEMD setup are useful references. But the claim that MLIPs give more accurate TBR than classical potentials needs DFT validation at grain-boundary configurations before it can be taken as established. The paper deserves a serious referee who can push the authors on this point. If they can show that the retrained GAP reproduces DFT forces for even a few grain-boundary structures, the claim becomes much stronger. Without that, the TBR comparison is suggestive but not conclusive.","headline":"MLIPs applied to grain-boundary thermal transport in nanocrystalline Si — bulk validation is solid, but the headline TBR comparison rests on a GAP model whose transferability to grain-boundary environments is never validated against DFT.","tokens_in":21697,"tokens_out":838,"would_cite":false,"duration_ms":189759,"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":"Machine-learned potentials predict 50% higher thermal resistance at rough grain boundaries","keywords":["machine-learning interatomic potentials","thermal boundary resistance","grain boundary","phonon transport","nanocrystalline silicon","non-equilibrium molecular dynamics","lattice dynamics"],"falsifier":"If direct DFT calculations of phonon properties at a silicon grain boundary disagree substantially with the MACE-predicted lifetimes or mean free paths reported in Figures 6–7, the grain-boundary lattice-dynamical results would be shown to reflect GAP's extrapolation rather than first-principles physics. Similarly, if experimental TBR measurements for roughness-controlled silicon grain boundaries fall closer to the SW/Tersoff predictions than to the GAP value of 1.66 m²K/GW, the central claim that MLIPs are more predictive at disordered interfaces would be undermined.","tokens_in":20773,"feed_emoji":"🔥","tokens_out":1456,"duration_ms":149856,"temperature":0.7,"pith_summary":"This paper argues that machine-learning interatomic potentials (MLIPs) — specifically a Gaussian Approximation Potential (GAP) and a MACE neural-network model — provide a more accurate and internally consistent description of phonon-mediated heat transport in both bulk and nanocrystalline silicon than the classical Stillinger-Weber and Tersoff potentials that have dominated the field for decades. The authors train GAP and MACE on the same reference dataset and benchmark them against experimental phonon dispersions, lifetimes, and thermal conductivity for bulk silicon, showing that the MLIPs reproduce the full vibrational picture — from second- and third-order force constants to mean free paths — while the classical potentials systematically overestimate phonon lifetimes and thermal conductivity. The paper then applies these potentials to symmetric tilt grain boundaries in silicon. Using lattice-dynamical calculations, the authors show that the grain-boundary core suppresses phonon lifetimes to a few picoseconds across all frequencies and reduces mean free paths to tens of nanometers or less, with the effect governed by local atomic disorder rather than the specific crystallographic misorientation angle. Finally, using non-equilibrium molecular dynamics (NEMD) with controlled interfacial roughness, the authors quantify thermal boundary resistance (TBR). At the highest roughness studied (amplitude 3 Å), GAP predicts a TBR of 1.66 m²K/GW — substantially higher than SW (1.08) or Tersoff (1.17) — and closer to experimental values reported for complex interfaces such as nanotwins. The central claim is that the choice of interatomic potential matters most precisely where it matters most: at highly disordered interfaces, where classical potentials underpredict phonon scattering and thus underpredict thermal resistance.","feed_headline":"Machine-learned potentials predict 50% higher thermal resistance at rough grain boundaries","feed_subtitle":"MLIPs reveal that classical potentials underpredict phonon scattering at disordered silicon interfaces, with direct consequences for nanosca","key_machinery":"The load-bearing machinery is the combination of (1) MLIP-derived harmonic and anharmonic force constants fed into a Phonopy/Phono3py lattice-dynamical workflow for phonon dispersions, lifetimes, and thermal conductivity, and (2) MLIP-driven non-equilibrium molecular dynamics (NEMD) with controlled sinusoidal roughness amplitudes at grain boundaries to extract thermal boundary resistance from the steady-state temperature discontinuity. The MACE model trained on GAP-generated grain-boundary configurations serves as the bridge for lattice-dynamical calculations at interfaces.","core_discovery":"The paper's central discovery is that machine-learning interatomic potentials, trained on first-principles reference data, predict substantially stronger phonon scattering at disordered silicon grain boundaries than classical empirical potentials do, and that this difference becomes dramatic as interfacial roughness increases. At a roughness amplitude of 3 Å, the GAP model yields a thermal boundary resistance of 1.66 m²K/GW versus 1.08–1.17 for the classical potentials — a gap of roughly 50 percent that places the MLIP prediction closer to experimental measurements for structurally complex interfaces. The paper also establishes that once a grain-boundary core becomes structurally disordered,","pith_inferences":["The MACE model used for grain-boundary lattice-dynamical calculations (phonon DOS, lifetimes, MFP, thermal conductivity in Figs. 5–9) is trained on energies and forces generated by the GAP model itself, not on independent DFT reference data for grain-boundary configurations. This means the reported grain-boundary phonon properties reflect GAP's extrapolation to these environments and are not indep","The NEMD thermal boundary resistance results (Table I, Fig. 11) use GAP directly, not the MACE surrogate, so the TBR comparison with classical potentials stands on more independent ground — though it still rests on GAP's accuracy for grain-boundary bonding, which is validated here only against bulk silicon benchmarks and experimental TBR ranges, not against DFT-level grain-boundary reference data.","The weak dependence of grain-boundary phonon lifetimes on misorientation angle (above 10°) suggests a saturation of disorder-induced scattering, but the limited set of four angles (10°–40°) and finite system size (~550 atoms in the GB core) leave open whether this saturation is a physical effect or a consequence of the simulation geometry."],"forward_implications":["Thermal management simulations for nanocrystalline silicon devices that still rely on SW or Tersoff potentials may systematically underestimate grain-boundary thermal resistance, leading to overly optimistic heat-dissipation predictions.","The framework can be extended to other polycrystalline semiconductors (e.g., germanium, SiGe alloys) where grain-boundary scattering limits thermal conductivity and where classical potentials are known to be unreliable.","The demonstrated sensitivity of TBR to interfacial roughness at the atomic scale suggests that grain-boundary engineering — controlling roughness and misorientation — could be used as a tuning knob for thermal transport in nanocrystalline materials.","The unified MLIP + NEMD + lattice-dynamical workflow could be applied to other interfaces (heterojunctions, phase boundaries) where the interplay of harmonic and anharmonic phonon physics controls heat flow."],"fun_headline_variants":["Machine-learned interatomic potentials predict stronger phonon scattering at rough silicon","ML potentials reveal 50% higher thermal boundary resistance at disordered silicon interfac","Gaussian approximation potentials expose 50% larger thermal resistance at rough grain boun","Machine-learning potentials reconcile predicted and measured thermal boundary resistance a","Machine-learned potentials show classical models underpredict phonon scattering at disorde"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The grain-boundary phonon properties computed via the MACE model (Figures 5–9) are derived from a MACE potential trained on GAP-generated data for grain-boundary configurations, not on independent first-principles reference data. If GAP's description of grain-boundary bonding is inaccurate, MACE inherits that inaccuracy, and the reported phonon lifetimes and thermal conductivities for the grain-boundary core cannot be considered independently validated.","fun_headline_variants_meta":{"raw":{"variants":["Machine-learned interatomic potentials predict stronger phonon scattering at rough silicon grain boundaries than classical potentials","ML potentials reveal 50% higher thermal boundary resistance at disordered silicon interfaces than classical models","Gaussian approximation potentials expose 50% larger thermal resistance at rough grain boundaries than Tersoff or Stillinger-Weber models","Machine-learning potentials reconcile predicted and measured thermal boundary resistance at disordered silicon grain boundaries","Machine-learned potentials show classical models underpredict phonon scattering at disordered silicon grain boundaries"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":751,"prompt_tokens":627,"completion_tokens":124,"prompt_tokens_details":null},"tokens_in":627,"tokens_out":124,"duration_ms":34253,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T04:38:03.481863+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If direct DFT calculations of phonon properties at a silicon grain boundary disagree substantially with the MACE-predicted lifetimes or mean free paths reported in Figures 6–7, the grain-boundary lattice-dynamical results would be shown to reflect GAP's extrapolation rather than first-principles physics. Similarly, if experimental TBR measurements for roughness-controlled silicon grain boundaries fall closer to the SW/Tersoff predictions than to the GAP value of 1.66 m²K/GW, the central claim that MLIPs are more predictive at disordered interfaces would be undermined.","supporting_citations":[],"review_version":1}