{"id":"047fd580-361d-4ef1-be88-e9ddd9b3dbc6","arxiv_id":"2412.15063","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Equivariant graph neural networks accurately predict full magnetic shielding and electric field gradient tensors for SiO2 structures, enabling fast simulation of static solid-state NMR spectra.","lead":"Researchers trained graph neural networks to predict solid-state NMR tensors from atomic structure, then used them to simulate NMR spectra for silica glasses and for the cristobalite phase transition. The framework could speed up NMR-based materials analysis by replacing expensive first-principles calculations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Zeolite generalization rests on hand-picked spectra; no quantitative spectral metric or full 50-zeolite error distribution is reported, so 'good agreement' is unverified out-of-domain.","rationale":"The reader's weakest assumption—training-set representativeness for zeolites and the cristobalite transition—is indeed the most load-bearing gap. The in-domain results are credible: the ISD and tensor-product models track GIPAW on the a-SiO2 test set, learning curves show remaining error is data-limited, and the physical decomposition into spherical ranks is well motivated. However, the manuscript's own Section III C and Fig. 6 contain explicit caveats: ring-size distributions in the training set differ from those in low-density a-SiO2 and zeolites, and the four zeolite spectra shown are selected as best and worst cases based on the ISD model's error. These caveats are not flaws by themselves—showing best/worst is a reasonable stress test—but they mean the main-text evidence does not establish that the models produce 'good agreement' for typical zeolites. The cristobalite demonstration in Section IV relies on the same models applied to crystalline environments not represented in the training set, with validation against experiment rather than against GIPAW on the same snapshots, so it inherits the same extrapolation risk. This concern is directly testable: computing aggregate spectral errors for all 50 zeolites, or an ablation that removes large rings from training, would settle whether the out-of-domain claim lands. Because the verdict is already CONDITIONAL and the proposed check would strengthen or qualify the claim, no change to the verdict is needed—only an explicit condition on the evidence.","tokens_in":17085,"tokens_out":10441,"duration_ms":73600,"concrete_test":"Compute the normalized RMSE between the ML-predicted and GIPAW static 1D 29Si and 17O spectra for all 50 hypothetical zeolites in Ref. 75, using the same Gaussian broadening (0.5 ppm) and frequency grid as Fig. 6. Report the median and worst-of-50 spectral RMSE, and compare with the median spectral RMSE on the 150 a-SiO2 test structures. If the zeolite median is more than 2x the a-SiO2 median, or if the worst case exceeds a pre-defined agreement threshold (e.g., 20% spectral RMSE), the 'good agreement' claim does not generalize to the zeolite class.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that both ML models produce static 1D NMR spectra in good agreement with GIPAW is established mainly for in-distribution a-SiO2. Section III C states that the training set covers ring sizes 6-16 (max 18), while a low-density test structure has rings of size 20/22 and zeolites have different ring distributions. The zeolite demonstration in Fig. 6 explicitly selects four representatives based on the ISD model's accuracy (\"We choose these four representative structures based on the accuracy of the ISD models\"), i.e. best and worst cases, so it does not provide an unbiased estimate of typical out-of-domain performance. The paper also notes a \"possible correlation between the ring distribution in the zeolites and ML prediction errors\" and defers further investigation. If the typical zeolite environment lies outside the learned local representation, the predicted tensor orientations and asymmetries, and hence the static spectra, could be unreliable. No quantitative spectral-similarity metric is reported for the zeolite set, and the aggregate errors on all 50 zeolites are only relegated to the Supplementary Materials. The extrapolation from a-SiO2 to zeolites and the cristobalite transition is therefore the most load-bearing unproven step.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents two graph-neural-network strategies for predicting solid-state NMR tensors (magnetic shielding and electric field gradient) from spherical tensor decompositions: an irreducible spherical decomposition (ISD) and a tensor-product (TP) combination with learnable weights. The models are trained on a diverse dataset of 1000 amorphous SiO2 structures and benchmarked against GIPAW reference data in terms of irreducible tensor components, anisotropy parameters, Euler angles, and simulated static 1D NMR spectra. The authors then apply the models to larger a-SiO2 configurations, 50 hypothetical zeolites, and the alpha-to-beta cristobalite transition, reporting good agreement with quantum-mechanical and experimental spectra. The central claim is that machine-learned tensor predictions can substitute for DFT-based NMR calculations in silica-based materials, including out-of-domain applications.","tokens_in":17309,"tokens_out":3955,"duration_ms":27523,"significance":"If the central claim holds, this work represents a substantial step toward ML-accelerated solid-state NMR prediction, with potential high-throughput impact on structural characterization of framework and amorphous silicates. The paper is methodologically strong: it uses a clear train-validation-test split, normalized error metrics, learning curves, direct spectral comparison to GIPAW, and a systematic comparison of two tensor-decomposition strategies. The transparent discussion of the tensor-product representation and its parameter count is a useful contribution. The main scientific value lies in demonstrating that equivariant GNNs can predict full tensorial NMR parameters, not only isotropic shifts, and that static spectra can be reconstructed from these predictions. However, the out-of-domain claims, especially for zeolites, are supported by only a biased subset of examples, and the quantitative evidence for the full zeolite set is deferred to the Supplementary Material.","major_comments":[{"comment":"The claim that the ML models generalize to zeolites is not quantitatively supported in the main text. The authors state that they \"choose these four representative structures based on the accuracy of the ISD models,\" which explicitly selects best- and worst-case examples rather than a representative or random sample. No spectral similarity metric or aggregate error distribution for the full 50-zeolite set is reported in the main text; the paper instead notes a \"possible correlation between the ring distribution in the zeolites and ML prediction errors\" and defers further investigation. Because the training set contains ring sizes mostly between 6 and 16 while zeolites have different ring distributions, the typical out-of-domain performance remains unverified. The authors should either report the full 50-zeolite error distribution (or a quantitative spectral similarity metric) in the main text, or substantially soften the extrapolation claim in the conclusions.","section":"Section III C (Fig. 6) and Section V"},{"comment":"The conclusion states that \"the irreducible spherical decomposition yielded the lowest prediction errors,\" but Figure 3 shows no significant difference between the ISD and tensor-product models across the tested magnetic-shielding parameters for either silicon or oxygen. This overstatement should be corrected to say that the two decompositions perform comparably, with only minor differences that are within the spread of the reported errors.","section":"Section V (Conclusions) vs. Figure 3"},{"comment":"The central claim that \"both models produce simulated static 1D NMR spectra in good agreement with quantum-mechanical computations\" is directly demonstrated only for the a-SiO2 structures in Figure 5. For zeolites, only a hand-picked subset is shown and no aggregate metric is provided; for cristobalite, the comparison in Figure 7 is to an experimental spectrum rather than to GIPAW calculations. The abstract and conclusions should be rephrased to distinguish the in-distribution demonstration from the exploratory out-of-domain applications.","section":"Abstract and Conclusions"}],"minor_comments":[{"comment":"The text contains a typographical error: \"We use the the root mean square error (RMSE)\" should read \"We use the root mean square error (RMSE).\"","section":"Section III (error metrics)"},{"comment":"The phrase \"based on test errorsor 50 independent configurations\" should be corrected to \"based on test errors on 50 independent configurations.\"","section":"Section II E"},{"comment":"There is a duplicate word in \"for all tensor product models models when the training set size is 10 structures,\" which should be \"for all tensor product models when the training set size is 10 structures.\"","section":"Section III B"},{"comment":"The hyperparameter optimization is performed only for the ISD model and a TP model with a l=1 target; it would be helpful to state explicitly whether the same hyperparameters were used for all other TP variants compared in Figures 2 and 3, since differences in hyperparameters could confound the decomposition comparison.","section":"Section II E"},{"comment":"The comparison of models with different numbers of tensor products is made at a fixed maximum rank lmax=4 and a fixed number of features per rank; the authors should note explicitly in the text or a footnote that the 4096-tensor-product model has a different parameter count than the single-tensor-product model, since this affects the interpretation of the performance gains.","section":"Section III A"}],"recommendation":"major_revision","confidential_remarks":"The paper is a strong, well-executed contribution to ML-driven solid-state NMR. The main issue is the overstatement of zeolite extrapolation: the main text presents a biased selection of best/worst cases and defers the full 50-zeolite aggregate metrics to the Supplementary Material. This is fixable within the scope of the manuscript by adding the aggregate error distribution or a spectral similarity metric to the main text, and by tempering the conclusions. I would be willing to review a revised version that addresses these points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis is a solid, well-executed study of equivariant GNNs for predicting tensorial NMR parameters in SiO2, and it earns its place as a reference for this kind of work. The main new contribution is the systematic comparison of two spherical-tensor decompositions—irreducible spherical decomposition (ISD) and a learnable tensor-product combination—within the NequIP architecture. The paper does a careful job with train/validation/test splits, normalized error metrics, learning curves, and comparing predicted tensors directly to GIPAW-computed spectra. The in-distribution results for amorphous silica are convincing: the static 1D spectra generated from ML tensors overlap well with DFT. The cristobalite dynamics section is a nice demonstration of coupling an MLIP with ML tensors, and it is honest about limitations, such as the over-stabilization of the β phase and observing only a single transition event.\n\nThe soft spots are real but not fatal. The conclusion states that ISD yielded the lowest prediction errors, but Figure 3 shows essentially no difference between ISD and the tensor-product model; the text should reflect that more carefully. More substantively, the out-of-domain zeolite claim rests on hand-picked best/worst examples in the main text. The full 50-zeolite errors are in the SI, and the authors themselves note a possible correlation with ring sizes, so the evidence for extrapolation is suggestive rather than conclusive. They do not overclaim—they explicitly defer further investigation—but the abstract's 'structurally diverse dataset' framing slightly oversells the transferability. Code and data are only promised upon acceptance, which limits reproducibility right now.\n\nThe citation practice is honest; they build directly on Venetos and Harper, which is the right prior work. For a reader working on ML for spectroscopic properties, this is a useful benchmark and a clear demonstration of what works (and what doesn't) with tensor-valued targets. It deserves a serious referee. My recommendation: send it to peer review, with a request to release the code/data or provide a clear timeline, and to soften the ISD/TP comparison claim.\n\nBest,\n\n[Your name]","headline":"Solid methods paper for in-distribution a-SiO2, with thinner out-of-domain evidence on zeolites and a minor overstatement of the ISD advantage.","tokens_in":17848,"tokens_out":2649,"would_cite":true,"duration_ms":24148,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Graph-neural-network models trained on amorphous silica predict the full tensor quantities that determine solid-state NMR spectra—magnetic shielding and electric field gradients—well enough to reproduce static 1D spectra from…","keywords":["graph neural networks","solid-state NMR","magnetic shielding tensor","electric field gradient","spherical tensor decomposition","amorphous silica","zeolites","cristobalite phase transition"],"falsifier":"If a zeolite or low-density silica structure with rings larger than those in the training set (say 20- and 22-membered rings) systematically yields ML spectra whose peak positions or line shapes deviate from the DFT reference by much more than the training error, the claimed out-of-domain accuracy would be refuted; a direct test is to compare ML and GIPAW spectra for a library of zeolites with ring sizes 18–24 and check whether error grows with ring-size mismatch.","tokens_in":16874,"feed_emoji":"🧲","tokens_out":8648,"duration_ms":48506,"temperature":0.7,"pith_summary":"This paper tries to establish that an equivariant graph neural network can learn the full tensor quantities that determine solid-state NMR spectra—the anisotropic magnetic shielding tensor and the electric field gradient—from atomic coordinates alone, not just the isotropic shift. If that works, computational NMR for materials becomes much cheaper: once a model is trained, spectra for large, disordered, or dynamically evolving structures can be generated without repeated quantum-mechanical calculations. The authors test this on a diverse set of 1,000 amorphous SiO2 configurations, show that simulated static 1D NMR spectra match DFT calculations, and then apply the models out-of-domain to hypothetical zeolites and to the α–β inversion in cristobalite, where thermally averaged tensors agree with experimental peak separations and asymmetry parameters. The upshot is a route to ML-driven NMR for both static and dynamic behavior of complex oxides.","feed_headline":"Graph networks learn full NMR tensors, matching quantum spectra","feed_subtitle":"Machine-learned tensors reproduce NMR spectra for silica, zeolites, and a phase transition.","key_machinery":"The central machinery is the spherical tensor decomposition of a rank-2 tensor, $T = T^{(0)} \\oplus T^{(1)} \\oplus T^{(2)}$, in which each piece transforms independently under spatial rotation and all pieces are parity-even. The models are equivariant graph neural networks that map atom-centred local environments to these irreducible components; the tensor-product variant replaces a single decomposition with a learnable linear combination of tensor products of spherical-harmonic features, using Clebsch–Gordan coefficients to extract ranks 0–2. This construction makes the tensor orientation (Euler angles) and anisotropy descriptors (span, skew, asymmetry) available as outputs, not just the isotropic value.","core_discovery":"The central claim is that decomposing rank-2 NMR tensors into irreducible spherical components—rank 0 (isotropic), rank 1 (antisymmetric), and rank 2 (anisotropic symmetric)—makes them natural targets for equivariant graph neural networks, and that models trained this way are accurate enough to reproduce static 1D NMR spectra from quantum-mechanical computations. Magnetic shielding uses all three irreducible components; the electric field gradient, being symmetric and traceless, is fully captured by the rank-2 part alone. An alternative tensor-product decomposition, built from learnable linear combinations of Clebsch–Gordan-coupled spherical-harmonic tensors, reaches comparable accuracy with roughly three times fewer parameters. Beyond the training distribution, the models predict spectra for hypothetical zeolites and track the α–β cristobalite transition through simulated magic-angle-spinning spectra, where the ML-derived oxygen electric-field-gradient asymmetry ($\\eta_Q = 0.13$) matches an experimental value.","pith_inferences":["A natural stress test would be to train on structures with ring sizes larger than 18 and see whether zeolite and low-density predictions improve, since the paper itself reports that ring-size distribution correlates with error.","The oxygen $\\sigma^{(1)}$ component is the least accurate target; a weighted loss function or separate loss balance could recover it, potentially improving the oxygen spectra further.","The same spherical-tensor scheme could be applied to other tensor-valued NMR quantities such as spin–spin coupling tensors or to non-NMR properties like the Born effective charge tensor, where orientation information matters.","Combining the tensor model with a fast machine-learned interatomic potential could enable closed-loop screening: propose a hypothetical structure, generate its spectrum, and refine, all without first-principles NMR."],"forward_implications":["Static 1D NMR spectra for amorphous silica at densities from 2.2 to 2.6 g/cm³ can be generated from ML tensor predictions in good agreement with DFT reference spectra.","The same ML models transfer, without retraining, to hypothetical zeolites, with the best spectral agreement occurring when ring sizes resemble the training distribution.","For the α–β cristobalite transition, ML-driven molecular dynamics combined with tensor prediction yields thermally averaged 29Si MAS spectra whose peak separation and oxygen EFG asymmetry match experiment, demonstrating a practical route to studying phase-transition dynamics without per-snapshot DFT NMR.","Because NMR tensors are learned per atom in an equivariant representation, the same framework applies to other chemical species and to other tensorial properties such as dielectric response or polarization, without retraining from scratch.","Predicting anisotropy parameters (span, skew, asymmetry) and tensor orientations from learned tensors is more reliable than specialized scalar models, so spectral shape, not just peak position, becomes accessible."],"supporting_citations":[{"why":"Establishes that predicting the full magnetic-shielding tensor, then extracting anisotropy, outperforms dedicated scalar models; this motivates the tensor-target approach.","marker":"51"},{"why":"Defines performance metrics and finds learning the full EFG tensor beats scalar models for Vzz, guiding how EFG accuracy is evaluated here.","marker":"53"},{"why":"Supplies the melt-quench-anneal protocol and the machine-learned potential used to generate the 1,000 amorphous SiO2 training structures.","marker":"64"},{"why":"Provides the machine-learned interatomic potential used to build larger amorphous models and to drive the 100 ps cristobalite transition trajectory.","marker":"72"},{"why":"Contributes the experimental 29Si MAS powder spectrum of a mixed α/β cristobalite sample at ~500 K used as the benchmark for thermally averaged ML spectra.","marker":"5"},{"why":"Defines the GIPAW method for periodic solid-state NMR tensor computations, the quantum-mechanical reference against which ML predictions are measured.","marker":"13"},{"why":"Introduces the equivariant graph-neural-network architecture on which the present models are built.","marker":"54"},{"why":"Supplies the database of hypothetical SiO2 zeolites used as the out-of-domain extrapolation test.","marker":"75"}],"fun_headline_variants":["Spherical tensor GNNs predict NMR spectra for complex materials","Graph nets learn NMR tensors, reproduce quantum spectra","ML matches NMR spectra from spherical tensor decomposition","Equivariant GNNs capture NMR tensors and spectra","Graph neural nets reproduce NMR spectra via tensor decomposition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the training set of 1,000 melt-quench-anneal amorphous SiO2 structures covering local environments well enough that hypothetical zeolites and the α–β cristobalite transition, with their different ring-size distributions, stay within the learned representation.","fun_headline_variants_meta":{"raw":{"variants":["Spherical tensor GNNs predict NMR spectra for complex materials","Graph nets learn NMR tensors, reproduce quantum spectra","ML matches NMR spectra from spherical tensor decomposition","Equivariant GNNs capture NMR tensors and spectra","Graph neural nets reproduce NMR spectra via tensor decomposition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001064,"raw_usage":{"total_tokens":4458,"prompt_tokens":937,"completion_tokens":3521,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":3444}},"tokens_in":553,"tokens_out":3521,"duration_ms":20331,"temperature":1.0,"reasoning_tokens":3444,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:39:16.125834+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If a zeolite or low-density silica structure with rings larger than those in the training set (say 20- and 22-membered rings) systematically yields ML spectra whose peak positions or line shapes deviate from the DFT reference by much more than the training error, the claimed out-of-domain accuracy would be refuted; a direct test is to compare ML and GIPAW spectra for a library of zeolites with ring sizes 18–24 and check whether error grows with ring-size mismatch.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines performance metrics and finds learning the full EFG tensor beats scalar models for Vzz, guiding how EFG accuracy is evaluated here."},{"cited_title":"Spearing , author I","cited_arxiv_id":null,"evidence_quote":"Contributes the experimental 29Si MAS powder spectrum of a mixed α/β cristobalite sample at ~500 K used as the benchmark for thermally averaged ML spectra."}],"review_version":1}