{"id":"9a4b7908-6ffb-42c9-b8b3-a35536777a57","arxiv_id":"2412.02916","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"CuXASNet predicts Cu L3/L2 XAS spectra from atomic structure with accuracy comparable to FEFF9 simulations (MAE 0.125 vs 0.131 against experiments).","lead":"CuXASNet is a neural network that predicts copper L-edge X-ray absorption spectra from atomic structure, trained on thousands of simulated spectra. It matches the accuracy of the FEFF9 simulation code while running orders of magnitude faster, which could accelerate screening of candidate materials.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 14-spectrum experimental validation has no stated energy-alignment protocol, so the reported MAE/Spearman equivalence between CuXASNet and FEFF9 is not yet interpretable.","rationale":"The strongest claim has two parts: (1) CuXASNet reproduces FEFF9 spectra, and (2) this reproduction is accurate enough for experimental use because FEFF9 itself matches 14 experimental spectra. Part 1 is supported by the held-out test metrics (median MAE 0.039, edge errors mostly within 0.2 eV). Part 2 is the load-bearing external validation. The reader's weakest assumption, that the fixed M3GNet embedding retains enough information, is a legitimate generalization concern, but it is not the bottleneck for the stated claim: the test set already demonstrates that the embedding carries sufficient information for the training distribution, and a missing ablation does not invalidate the reported accuracy. The experimentally-facing claim, by contrast, is threatened by an omitted protocol that the paper itself shows is needed in the organometallic example. This is a concrete, checkable omission rather than a speculative representational limit. It does not prove the model is wrong, but it prevents the reported experimental metrics from being verified, which is exactly what the CONDITIONAL verdict should require. The reader's rationale also noted the absence of explicit energy alignment, so the concerns partially overlap.","tokens_in":16295,"tokens_out":6081,"duration_ms":64006,"concrete_test":"Re-derive the Section III.D comparison for all 14 materials with an explicit alignment step: for each material, shift the FEFF9 and CuXASNet spectra over a range of ±3 eV (for example, maximize cross-correlation with the digitized experimental spectrum over the measured window), then recompute per-spectrum and average MAE and Spearman correlation. Report the optimal shifts for FEFF9 and CuXASNet separately. If the optimal shifts differ between the two methods by more than about 0.2 eV, or if the average MAE or Spearman correlation changes by more than about 10% relative to the unaligned numbers, the abstract's equivalence claim is not established without alignment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central external claim, that CuXASNet matches FEFF9's agreement with experiments (abstract and Section III.D), rests on comparing simulated spectra to 14 webplotdigitizer-extracted experimental spectra. Section II.D describes only cropping to the measured energy range and computing MAE; it never states that the simulated energy axes were aligned to the experimental axes. In XAS analysis, FEFF9 spectra are not automatically on an absolute energy scale and are normally shifted to an experimental reference; indeed, the paper itself manually aligns the L3 peak for the molecular case in Section III.G. If such a shift was not applied to the 14 spectra, the computed average MAE of 0.125 and Spearman correlation of 0.891 can change substantially with an energy offset, and the claimed 'comparable to FEFF9' may reflect a shared systematic offset rather than true spectral accuracy. No per-spectrum metrics or alignment parameters are reported, so the result cannot be independently checked.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CuXASNet, a dense neural network that predicts Cu L3- and L2-edge X-ray absorption spectra from a 64-dimensional M3GNet embedding of the absorbing Cu site. The model is trained on 5497 FEFF9 multiple-scattering spectra of Cu-containing materials, and the authors report median test-set MAE of 0.0391, accurate edge-onset prediction to within 0.2 eV for roughly half of the test set, and agreement with FEFF9 across oxidation states, coordination numbers, and neighbor types. The external validation uses 14 experimental spectra digitized from the literature, for which the authors report an average MAE of 0.125 and Spearman correlation of 0.891, comparable to the FEFF9 values of 0.131 and 0.898. The authors also demonstrate a molecular example outside the crystalline training distribution and analyze how model accuracy scales with training-set size.","tokens_in":16423,"tokens_out":6492,"duration_ms":69026,"significance":"If the claims hold, CuXASNet would be a useful, fast surrogate for FEFF9 L-edge simulations, enabling structure screening and rapid generation of reference spectra for Cu systems. The internal validation against FEFF9 is strong and clearly presented: the test-set MAE distributions, edge-location histograms, and error breakdown by chemical environment give a convincing picture of surrogate accuracy. The experimental comparison is potentially valuable evidence that the model is not systematically worse than FEFF9, but its current lack of methodological detail limits the strength of the external claim. The paper also provides useful scaling data on training-set size and an honest discussion of outlier classes such as antiperovskites.","major_comments":[{"comment":"The manuscript does not state whether the simulated energy axes were aligned to the experimental axes before computing MAE and Spearman correlation. Section III.G explicitly mentions manually aligning the L3 peak for the molecular comparison, which is standard practice for FEFF9 L-edge calculations because their absolute energy scale is not directly transferable to experiment. If no such alignment was applied to the 14 literature spectra, a uniform energy offset could substantially change both metrics, and the claimed equivalence between CuXASNet and FEFF9 might reflect a shared systematic offset rather than true spectral accuracy. Please specify the alignment procedure used for all 14 spectra, report any alignment shifts, and provide metrics both with and without alignment.","section":"Sections II.D and III.D"},{"comment":"The external validation is summarized only by averages over 14 spectra, with no per-spectrum table and no uncertainty estimates. Figure 5(d) shows the distribution of MAE values, but the reader cannot tell which systems drive the average, whether any single spectrum is an outlier, or whether the FEFF9 and CuXASNet values are statistically distinguishable. Provide a table with per-spectrum MAE and Spearman correlation for both methods, together with the experimental energy range used for each system and the corresponding digitization source.","section":"Section III.D"},{"comment":"The train/test split is not described. The paper reports that the test set contains 1375 spectra but does not state how the split was constructed (e.g., random by spectrum, random by material, stratified by oxidation state) or whether test structures can share near-duplicate local environments with training structures. This detail is load-bearing for the central claim that CuXASNet generalizes to new atomic structures, and it should be stated explicitly.","section":"Sections II.C and III.B"}],"minor_comments":[{"comment":"Several typos should be cleaned up: 'igure S4' is missing the capital F, 'epocs' should be 'epochs', 'principle component' should be 'principal component', and 'For for biological' has a duplicated word.","section":"Throughout"},{"comment":"The text refers to bar plots in Figure 4(a-b) and two-dimensional histograms in Figure 4(c-d), but the figure caption and the actual panel arrangement appear to place the two-dimensional histograms in (a-b) and the bar plots in (c-d). Please correct the subplot references so the text matches the figure.","section":"Section III.C and Figure 4"},{"comment":"The data and code repository is described as private and 'will be shared upon request.' For a reproducibility claim, the authors should either make the repository accessible to reviewers before publication or provide a review copy, since the paper's central results depend on the exact training and evaluation pipeline.","section":"Section V"},{"comment":"The description of experimental data extraction is thin: the authors cite webplotdigitizer and the source references, but do not state which curve from each reference was digitized, how the smoothing broadening function was chosen, or how the intensity normalization was performed before computing MAE. A short supplementary table or paragraph would make the experimental comparison reproducible.","section":"Section II.D"}],"recommendation":"major_revision","confidential_remarks":"The internal FEFF9-surrogate validation appears sound and the paper is a good fit for the journal. The main obstacle is the under-specified experimental comparison, which is fixable by adding an alignment protocol and per-spectrum reporting. I do not see evidence of misleading intent; the missing details are presentation and methodology gaps rather than fundamental flaws. The private code repository should be made available to reviewers before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CuXASNet is a workmanlike ML surrogate for FEFF9 Cu L-edge XAS, and the internal test-set results are convincing. The experimental comparison is the weak link: the paper never explains how the simulated energy axes were aligned to the digitized spectra, and that makes the headline equivalence claim uninterpretable as written.\n\nWhat is actually new: the authors train two dense networks (L3 and L2) on 5497 site-specific FEFF9 spectra, using M3GNet node features as input, and they validate the models against 14 experimental spectra extracted from the literature. The Cu-specific dataset, the two-edge models, and the experimental screening are real additions beyond the same group's earlier universal XAS framework (Ref. 10). The paper also does several things well: it filters unconverged FEFF9 calculations, removes the experimental comparison structures from training, shows PCA and chemical-trend analyses, reports edge-location error histograms, and includes a training-data-volume curve. The internal accuracy is demonstrated convincingly: median MAE around 0.039, edge onsets mostly predicted within 0.2 eV, and the worst outliers are discussed honestly.\n\nThe soft spots are real but separable. First, the experimental validation protocol is thin. Section II.D says only that spectra were cropped to the measured energy range and MAE was computed; no energy calibration or alignment is mentioned. FEFF9 spectra are not automatically on an absolute experimental energy scale, and the paper itself manually aligns the L3 peak for the molecular complex in Section III.G, which proves alignment matters. Without knowing whether the 14 spectra were aligned, the reported MAE=0.125 and Spearman=0.891 (vs FEFF9's 0.131/0.898) could reflect a shared systematic offset rather than genuine spectral accuracy. There are also no per-spectrum metrics or error bars. This is the load-bearing weakness in the external claim. Second, the code and dataset are private until publication, so the results cannot currently be checked independently. Third, the M3GNet embedding is not ablated; the test-set performance suggests it is adequate, but the paper does not test alternatives. That is minor. The novelty is incremental over Ref. 10, but that is fine.\n\nThis paper is for practitioners who want a fast FEFF9-quality surrogate for Cu L-edge screening in high-throughput studies. It deserves a serious referee. I would send it to review with a request that the authors state the energy-alignment procedure explicitly, provide per-spectrum MAE and Spearman values, and release the code and data. With those revisions, the experimental claim would be checkable and the paper would be a solid contribution.","headline":"Useful FEFF9 surrogate for Cu L-edge XAS with solid internal validation, but the 14-spectrum experimental equivalence claim is undercut by an unstated energy-alignment protocol.","tokens_in":17005,"tokens_out":2208,"would_cite":true,"duration_ms":22211,"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":"A neural network predicts Cu L-edge X-ray absorption spectra from atomic structure at FEFF9 accuracy.","keywords":["Cu L-edge X-ray absorption spectroscopy","machine learning spectral prediction","M3GNet featurization","FEFF9 multiple-scattering simulation","dense neural network","structure-spectrum relationship","XANES","transition metal L-edge"],"falsifier":"A decisive test: find two distinct Cu coordination environments that receive exactly the same 64-dimensional M3GNet embedding but produce measurably different FEFF9 L3 spectra; because CuXASNet is a deterministic function of that vector, it would be unable to distinguish them, falsifying the claim that the embedding contains the spectral information. A practical approximation is to train the same network from an alternative local-environment descriptor on the same split and show a significantly lower test MAE, which would disprove the sufficiency of the M3GNet featurization.","tokens_in":16059,"feed_emoji":"🔬","tokens_out":7254,"duration_ms":85441,"temperature":0.7,"pith_summary":"CuXASNet is a machine-learning model that claims to predict the Cu L3 and L2 X-ray absorption near-edge spectra of a material directly from its atomic structure, replacing the computationally expensive multiple-scattering simulations of FEFF9. The paper argues that a 64-dimensional site embedding from M3GNet, fed through a small dense neural network, captures the local Cu environment well enough to reproduce FEFF9 spectra with a median test-set MAE of 0.0391. The decisive evidence is transfer to experiment: on fourteen spectra extracted from the literature, CuXASNet achieves an average MAE of 0.125 and Spearman correlation 0.891, essentially matching FEFF9's own averages of 0.131 and 0.898. If this holds, the model can generate large Cu L-edge spectral databases in minutes, which matters because most Cu-containing structures in the Materials Project currently lack simulated L-edge spectra.","feed_headline":"Neural net matches FEFF9 on copper X-ray spectra","feed_subtitle":"Trained on simulated spectra, it reproduces 14 experimental Cu L-edge measurements as faithfully as the reference code does.","key_machinery":"The load-bearing object is the M3GNet-derived 64-dimensional site vector: for each symmetrically unique Cu absorber, M3GNet's graph representation of the unit cell is evaluated before its readout layer, and the resulting node features are the sole input to CuXASNet. The spectral branch is a pair of fully connected networks with hidden layers of 120, 240 and 480 units and ReLU activations, a softmax output layer spanning 312 energy points, and separate weights for the L3 and L2 edges. FEFF9 provides the training signal at the multiple-scattering level of theory; the interpolated 0.2 eV grid and fixed energy windows standardize the output. What the machinery does is compress the structure-spectrum relationship into a learned map from a local-environment embedding to a normalized intensity profile.","core_discovery":"The central claim of the paper is that the Cu L-edge X-ray absorption spectrum is, for practical purposes, a function of a short vector encoding the absorbing site's local atomic environment: specifically, the 64-dimensional node embedding produced by M3GNet. Trained on 4,123 of 5,497 site-specific FEFF9 multiple-scattering spectra, two dense networks (one for L3, one for L2) predict the full 312-point spectrum on a 0.2 eV grid. The paper reports median test-set MAE of 0.0391, edge-onset errors that are exact or within a grid step for most spectra, and, in the experimental validation, aggregate errors indistinguishable from FEFF9's own. The authors present this as establishing CuXASNet as a drop-in replacement for FEFF9 in screening candidate structures against measured spectra, including structures outside its crystalline training distribution.","pith_inferences":["Since CuXASNet only learns to imitate FEFF9, its agreement with experiment is capped by FEFF9's own systematic errors: any level-of-theory deficiency in the multiple-scattering method will be inherited by the network, so the model cannot be expected to improve on FEFF9 for elements such as Ti or Fe where FEFF9 is known to be weaker.","The absence of a featurizer ablation leaves open that a cheaper or more physically interpretable descriptor could match or beat the M3GNet embedding; a head-to-head comparison against SOAP or atom-centered symmetry functions on the same 5,497 labels would settle where the accuracy resides.","The model's speed invites an inverse-design loop not implemented in the paper: search over structures or dopant configurations to match a target experimental spectrum, then validate the top hits with full simulation.","Extension to other 3d transition-metal L-edges will require new training labels but, if the data-scaling trend holds, would need only a few thousand spectra per element, a modest simulation budget relative to database-scale studies."],"forward_implications":["CuXASNet can generate thousands of Cu L3/L2 spectra per minute, so most Cu-containing Materials Project structures without L-edge spectra could be populated at FEFF9-equivalent quality without running FEFF9.","Experimental analysis can use CuXASNet as a fast filter: candidate structural models of an unknown Cu material are screened against a measured spectrum before expensive FEFF9 refinement, since the ML-versus-experiment error is statistically the same as FEFF9-versus-experiment.","The data-scaling result, accuracy plateauing after roughly 500 training spectra, indicates that the same architecture can be retrained on more expensive levels of theory with a computationally feasible dataset.","The successful prediction for a peroxo-Cu(II) organometallic complex implies the embedding generalizes beyond crystalline periodic structures, extending the model's use to molecular Cu sites.","Because the two worst outliers are antiperovskites with unusual FEFF9 spectra, the model's reliability is chemistry-dependent and its limits cluster around rare coordination motifs rather than being uniform."],"supporting_citations":[{"why":"Supplies the M3GNet-based site featurization and the prior finding that this representation is effective for X-ray spectral prediction; forms the input to CuXASNet.","marker":"[10]"},{"why":"FEFF9 multiple-scattering code that generated the simulated spectra used as training labels and as the accuracy benchmark.","marker":"[34]"},{"why":"Materials Project L-edge XAS database that supplies initial Cu spectra and motivates the need for faster spectral generation.","marker":"[29]"},{"why":"Lightshow workflow used to generate FEFF9 input files for the additional simulated spectra in the dataset.","marker":"[33]"},{"why":"Prior work whose dataset construction and spectral processing procedure CuXASNet adopts, and source of the additional 3,387 Cu site spectra.","marker":"[45]"},{"why":"Materials Project repository of structures used to select stable and experimentally synthesized Cu-containing materials and the source of most training structures.","marker":"[35]"},{"why":"Original M3GNet graph neural network whose pretrained node embedding provides the 64-dimensional featurization used as model input.","marker":"[60]"},{"why":"Literature experimental Cu L-edge spectra used to validate that CuXASNet matches experiments as well as FEFF9 does.","marker":"[63–70]"}],"fun_headline_variants":["CuXASNet: neural net predicts Cu L-edge spectra fast","ML matches FEFF9 accuracy for Cu X-ray absorption spectra","Neural net matches FEFF9 for Cu L-edge X-ray spectra","CuXASNet: structure in, copper X-ray spectrum out","Deep learning predicts Cu X-ray spectra as well as FEFF9"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim collapses if M3GNet's 64-dimensional embedding of a Cu site does not contain enough information about the local atomic environment to determine the L-edge spectrum; the paper takes this representation as fixed and does not ablate or compare it against alternative featurizations.","fun_headline_variants_meta":{"raw":{"variants":["CuXASNet: neural net predicts Cu L-edge spectra fast","ML matches FEFF9 accuracy for Cu X-ray absorption spectra","Neural net matches FEFF9 for Cu L-edge X-ray spectra","CuXASNet: structure in, copper X-ray spectrum out","Deep learning predicts Cu X-ray spectra as well as FEFF9"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000854,"raw_usage":{"total_tokens":3729,"prompt_tokens":980,"completion_tokens":2749,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":2658}},"tokens_in":596,"tokens_out":2749,"duration_ms":19420,"temperature":1.0,"reasoning_tokens":2658,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:57:18.256394+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test: find two distinct Cu coordination environments that receive exactly the same 64-dimensional M3GNet embedding but produce measurably different FEFF9 L3 spectra; because CuXASNet is a deterministic function of that vector, it would be unable to distinguish them, falsifying the claim that the embedding contains the spectral information. A practical approximation is to train the same network from an alternative local-environment descriptor on the same split and show a significantly lower test MAE, which would disprove the sufficiency of the M3GNet featurization.","supporting_citations":[{"cited_title":"OmniXAS: A Universal Deep-Learning Framework for Materials X-ray Absorption Spectra","cited_arxiv_id":"2409.19552","evidence_quote":"Supplies the M3GNet-based site featurization and the prior finding that this representation is effective for X-ray spectral prediction; forms the input to CuXASNet."},{"cited_title":"Rehr, Joshua J","cited_arxiv_id":null,"evidence_quote":"FEFF9 multiple-scattering code that generated the simulated spectra used as training labels and as the accuracy benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Materials Project L-edge XAS database that supplies initial Cu spectra and motivates the need for faster spectral generation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Lightshow workflow used to generate FEFF9 input files for the additional simulated spectra in the dataset."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior work whose dataset construction and spectral processing procedure CuXASNet adopts, and source of the additional 3,387 Cu site spectra."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Materials Project repository of structures used to select stable and experimentally synthesized Cu-containing materials and the source of most training structures."}],"review_version":1}