{"id":"c9b8b1b7-ebd4-406d-87fd-162d1702dc8a","arxiv_id":"2506.11908","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"XAStruct learns both structure-to-spectrum and spectrum-to-structure mappings for X-ray absorption spectroscopy across more than 70 elements, including the first reported ML prediction of neighbor atom types from spectra.","lead":"This paper introduces XAStruct, a machine learning system that predicts X-ray absorption spectra from crystal structures and, in reverse, infers local structural details such as neighbor atom types and bond distances from spectra. The authors train it on more than 70 elements, aiming to make X-ray spectroscopy analysis faster and less dependent on expensive simulations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Inverse-task metrics may reflect element-identity shortcuts rather than spectral inference; no element-only or spectrum-ablated baseline is reported.","rationale":"After reading the manuscript, the central claim hinges on the inverse tasks: neighbor-atom-type prediction and generalizable MNND regression. The reported results for these tasks are the most novel, yet they are also the most vulnerable to a simple confound: the model inputs include the absorbing element (Eq. 2), and the energy axis encodes the edge position, which identifies the element. Since bond lengths and coordination chemistry are strongly element-dependent, the model could achieve low MAE and high accuracy by predicting element-conditional statistics without ever using the spectral shape. The paper never rules this out: no element-only baseline, no spectral ablation, and the random-forest baseline uses the same spectra and element information. The reader's concern about dataset simulation is valid but pertains to external validity; the element-identity shortcut, if it holds, would undermine the internal validity of the central claim that structure is inferred 'from XAS spectra.' A decisive test is straightforward: replace the absorption signal with a constant and see if performance collapses. If it does not, the claims must be reframed as element-property prediction rather than spectra-to-structure inference. This keeps the verdict at CONDITIONAL.","tokens_in":14411,"tokens_out":6149,"duration_ms":158940,"concrete_test":"Retrain the MNND and neighbor-atom models exactly as in §5.2 but replace the absorption vector x used in Eq. (8) with a constant (e.g., a per-spectrum mean or zeros), keeping the energy axis and element identity unchanged. If validation MAE stays below ~0.04 Å and neighbor-atom accuracy above ~90%, the model is not using spectral shape, and the 'spectra-to-structure' claim is not supported. Additionally, compare with a model that takes only the absorbing element one-hot and predicts the element-wise mean MNND / majority neighbor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (2) defines the inverse task as gφ(x, z), so the absorbing element is always provided; for MNND the energy axis itself reveals the element through the edge position. Because bond lengths and common neighbor species are strongly element-specific, a model that ignores spectral shape and predicts the conditional mean MNND or majority neighbor per element could plausibly achieve the reported MAE of 0.0350 Å and 92.96% accuracy. Section 5.2 and Table 2 report only a random-forest baseline that also receives the same spectra and element information; no baseline is given for element-only input, nor is any ablation that removes or shuffles the absorption signal. The paper's claim to be the first to predict neighbor atom types 'directly from XAS spectra' and to provide a generalizable MNND model therefore rests on the unstated assumption that the model's accuracy is driven by spectral features rather than by element identity. This is the weakest link in the central claim because it is internal to the evaluation and directly testable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces XAStruct, a two-pipeline machine learning framework for X-ray absorption spectroscopy. The forward pipeline predicts XANES and EXAFS spectra from crystal graphs using a CHGNet-based GNN encoder followed by a gated MLP head, trained per element-edge-spectrum combination. The inverse pipeline predicts local structural descriptors from K-edge XAFS spectra: coordination number via per-element random forests, nearest-neighbor atom type via per-element gated classifiers, and mean nearest-neighbor distance (MNND) via a single shared-weight SGMLP-convolution model. The system is trained on a Materials Project-derived dataset of roughly 43,000 structures and 120,000 spectra spanning 70+ elements. The central claims are that XAStruct is the first ML approach to predict neighbor atom types directly from XAS spectra and the first generalizable MNND regression model across 70+ elements without element-specific tuning. Reported results include XANES K-edge MAE 0.0537, L-edge 0.0031, EXAFS K-edge 0.0302, MNND MAE 0.0350 Å, and neighbor atom accuracy 92.96%.","tokens_in":14587,"tokens_out":4281,"duration_ms":55831,"significance":"If the dataset and simulation provenance are sound, and if the inverse models are genuinely using spectral features rather than element-identity shortcuts, the paper would make a substantial contribution: it provides a large cross-element dataset, a unified forward/inverse framework, and the first reported cross-element MNND model. The paper has clear strengths: systematic ablations in Table 3, comparisons against multiple GNN baselines and random forests, and element-wise radar plots in the appendix. The authors also explicitly acknowledge several limitations in Section 6, including per-element training for CN and neighbor-atom classification and the closed-set nature of the classification tasks. However, the load-bearing generalization claims rest on unverified simulation details and on the absence of a shortcut-control baseline, so the results as currently presented cannot be taken as evidence of interpretation of real XAS data.","major_comments":[{"comment":"The manuscript never specifies how the 120,000 XANES and EXAFS spectra were simulated. There is no mention of the simulation code (e.g., FEFF, FDMNES, or an alternative), the underlying electronic-structure method, the energy grid, the broadening scheme, or the normalization procedure. Because every forward and inverse model is trained and evaluated against these spectra, the reported numbers in Tables 1 and 2 are not interpretable with respect to real XAS measurements, and the work is not reproducible as written. The authors must document the full simulation pipeline, including parameters and version numbers, and ideally release the dataset and simulation scripts.","section":"§5 (Dataset construction)"},{"comment":"The inverse model receives the absorbing element z and the energy axis e as inputs in addition to the absorption vector x. No baseline or ablation has been run with element-only or energy-axis-only input, nor with the absorption signal shuffled or zeroed. Because MNND and neighbor-species distributions are strongly element-dependent and the edge position is encoded in e, the reported MNND MAE of 0.0350 Å and neighbor atom accuracy of 92.96% could, in principle, be achieved by a model that ignores the spectral shape entirely. This shortcut test is essential to the central claim of predicting structural descriptors 'directly from XAS spectra' and should be reported.","section":"§3.3, Eq. (2); §5.2, Table 2"},{"comment":"The CuXASNet comparison is not controlled. The 0.0391 MAE for CuXASNet is taken from the original paper, which used different data, different energy grids, and different normalization and split protocols. Reporting that number alongside XAStruct's 0.0012 in the same table implies a direct comparison that the manuscript does not support. The authors should either retrain CuXASNet under the identical evaluation protocol used for the other baselines or explicitly state that the comparison is cross-paper and not controlled.","section":"§5.1, Table 1"},{"comment":"The periodic-table-wide generalization claim is substantially weaker than the abstract and introduction suggest. Section 6 states that CN and neighbor-atom classifiers are trained separately per element, and only the MNND model uses shared weights across all elements. The abstract claim of a generalizable model that 'requires no element-specific tuning' should be restricted to MNND. In addition, the 8:2 random split of the simulated dataset does not control for structural similarity between training and validation materials, so the validation numbers measure interpolation within one simulation distribution, not generalization to unseen chemistries or to experimental spectra.","section":"§5.2 and §6; §5 (data split)"}],"minor_comments":[{"comment":"The text contains a broken cross-reference: 'see Figure ?? and Figure S12'; the figure number should be inserted.","section":"§5, first paragraph"},{"comment":"Equation (4) divides by |V| even though the mask m selects only the absorber and its immediate neighbors; clarify whether the intended normalization is the number of masked atoms, sum(m_i), and adjust the formula or description accordingly.","section":"§4, Eq. (4)"},{"comment":"The inverse experiments use only K-edge XAFS because L-edge EXAFS is unavailable; this should be stated at the start of Section 5.2 rather than in passing, since the earlier sections discuss both K and L edges.","section":"§5.2, first paragraph"},{"comment":"The row labels 'SwiGLU (use ReLU)' and 'GatedLinear (use nn.Linear)' are ambiguous; make clear that these variants replace the named component with the specified alternative.","section":"§5.3, Table 3"},{"comment":"The introduction contains an unresolved citation placeholder '[?]' in the list of AI breakthroughs; the reference should be completed.","section":"§1 and Reference list"},{"comment":"The periodic table of model availability does not define the threshold for 'successfully trained and evaluated' coverage; specify the minimum data count or performance criterion used for the green tags.","section":"§A.5, Figure S12"}],"recommendation":"major_revision","confidential_remarks":"The main risk to the paper's claims is dataset provenance and the absence of a shortcut-control baseline for the inverse tasks. If the authors can provide the full simulation details, release the code and data, and add the element-only/spectrum-ablated experiments, the paper would be much stronger and potentially acceptable. I would also ask the editor to ensure that the 'first' novelty claims are checked against the broader XAS-ML literature, since the manuscript's related-work section is selective."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's what I took away from XAStruct. The genuinely new piece is the inverse task: predicting neighbor atom types from XAS, and a single MNND regressor across 70+ elements. That's not in the cited literature, and the 120k-spectrum dataset is a useful contribution. The paper does solid engineering: ablations show the SGMLP components matter, the element-wise radar plots are helpful, and the limitations section is honest.\n\nThe soft spots are real, though. First, the paper never describes how the spectra were simulated. \"Collected from Materials Project\" isn't enough; without the simulation code, energy grids, broadening, and normalization, the reported MAEs are not reproducible. Second, the headlines say \"across the periodic table,\" but most tasks (forward spectral prediction, CN, neighbor-atom classification) are trained per element. Only MNND is trained jointly. So the periodic-table-wide claim applies to one task, not the framework. Third, the baseline comparison is uneven: CuXASNet numbers are taken from the original paper rather than retrained, which is not a like-for-like comparison.\n\nThe stress-test concern fits with the paper's framing. Eq. (2) always feeds the absorbing element z to the model, and the energy axis itself strongly encodes the element. Since bond lengths and neighbor species are element-specific, a model that mostly predicts conditional per-element means could get fairly low MAE and high accuracy without \"reading\" the spectrum. The paper reports only a random forest baseline that also gets the same x and z. It does not include an element-only baseline or any ablation that removes or shuffles the absorption signal. That is directly testable and should be the first thing the authors add. If the element-only baseline is much worse, the claim stands; if it's close, then the inverse task is less about spectral inference than the paper argues.\n\nWhom is this for? Anyone working on ML for XAS or materials characterization. The dataset and the new tasks will be useful even if the element-identity concern changes some conclusions. I would not cite it yet, mainly because code and data aren't available and the simulation pipeline is undocumented.\n\nBottom line: this deserves a serious referee, not a desk reject, but it needs major revision. I'd send it out with a request for the element-only baseline, full dataset provenance, and for the authors to either release code/data or show it is available on request.","headline":"Useful engineering with a genuinely new inverse task, but the central generalization claim needs an element-only baseline and full dataset details before it convinces.","tokens_in":15124,"tokens_out":2772,"would_cite":false,"duration_ms":36461,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"One AI framework maps X-ray spectra to structure and back.","keywords":["X-ray absorption spectroscopy","XANES","EXAFS","graph neural networks","structure-to-spectrum prediction","spectra-to-structure inference","mean nearest-neighbor distance","neighbor atom classification"],"falsifier":"Take a set of experimental XAS spectra for a dozen well-characterized compounds spanning light, transition, and heavy elements whose structures are not in the training set, normalize them as the paper does, and compare XAStruct's predicted mean nearest-neighbor distance and neighbor atom type against crystallographic values. If per-element systematic biases appear, or if the reported errors grow sharply outside the training distribution, the periodic-table generalization claim is falsified.","tokens_in":14200,"feed_emoji":"⚛️","tokens_out":7192,"duration_ms":78337,"temperature":0.7,"pith_summary":"XAStruct is a single machine-learning system that treats X-ray absorption spectroscopy (XAS) as a two-way mapping: given a crystal structure it predicts the XANES and EXAFS spectra, and given a spectrum it predicts local structural descriptors—mean nearest-neighbor distance, coordination number, and the types of neighboring atoms. The paper's central claim is that one framework, trained on roughly 120,000 simulated spectra covering more than 70 elements and both K and L absorption edges, can replace element-specific heuristics for both spectral simulation and local-structure inference. It reports the first machine-learning model to predict neighbor atom types directly from XAS spectra, and the first shared-weight regression model for mean nearest-neighbor distance that works across the periodic table without per-element tuning. If the claim holds, experimental XAS data from diverse chemistries could be interpreted automatically, and candidate structures could be checked against spectra without expensive quantum simulations.","feed_headline":"One AI framework maps X-ray spectra to structure and back","feed_subtitle":"Trained on 120k simulated spectra across 70+ elements, it reads bond distances and neighbor atom types directly from spectra.","key_machinery":"The load-bearing component is the SGMLP block—a SwiGLU-gated multi-layer perceptron in which each layer uses a GatedLinear projection (a linear map multiplied by a sigmoid-gated branch) followed by SwiGLU activation and LayerNorm. The forward pipeline is carried by a physics-aware graph neural network encoder (the CHGNet backbone) that produces element-aware node embeddings, combined with a binary mask that isolates the absorbing atom's local environment; the masked mean embedding feeds the SGMLP head that emits the spectrum. The inverse pipeline embeds the energy axis and the absorption axis separately for XANES and EXAFS through parallel SGMLP blocks, concatenates the latent vectors, and passes them through convolutional pooling and further SGMLP layers for MNND regression, while a random forest handles coordination-number classification. The ablation study identifies the SGMLP block as the mechanism responsible for most of the accuracy gain over simpler MLP baselines.","core_discovery":"XAStruct couples a forward pipeline, which maps a crystal graph to a spectrum through a physics-aware graph neural network encoder followed by a gated MLP head, with an inverse pipeline that maps a spectrum to numerical and categorical structural descriptors. On the forward side, the authors report mean absolute errors of 0.0537 for K-edge XANES, 0.0031 for L-edge XANES, and 0.0302 for K-edge EXAFS, with the L-edge figure far below the 0.0391 reported for a prior copper-specific model. On the inverse side, a joint model regresses mean nearest-neighbor distance with a MAE of 0.0350 Å and an R2 of 0.985 across all elements, while per-element classifiers predict coordination number (69.26% accuracy) and neighbor atom type (92.96% accuracy, 88.76% macro F1). The paper frames the neighbor-atom-type prediction as the first machine-learning solution to a long-standing inverse problem in XAS analysis, and the shared-weight MNND model as the first element-agnostic geometric surrogate for spectral interpretation.","pith_inferences":["The paper's generality claim rests most firmly on spectral prediction and MNND regression; coordination number and neighbor-atom models are still trained per element, so a truly unified classifier across the periodic table remains an open problem.","A natural next test is to feed experimental (not simulated) XAS spectra into the MNND and neighbor-atom heads; if the reported errors survive that transfer, the models would be immediately useful to synchrotron users.","The reliance on simulated Materials Project spectra means the models learn that simulation's physics; cross-validation against experimental spectra would reveal how much of the reported accuracy reflects the simulation code's idiosyncrasies rather than real XAS physics."],"forward_implications":["If the simulated spectra are faithful, XAStruct can act as a fast surrogate for XANES and EXAFS simulation, replacing DFT-level calculations for screening candidate structures across the periodic table.","The element-agnostic MNND model gives experimentalists a parameter-free way to read average bond lengths directly from raw spectra, without per-element calibration or iterative EXAFS fitting.","Neighbor atom type prediction from spectra creates a new fingerprint for local chemical environments, useful for operando studies of catalysts and battery electrodes where the bonding shell changes during operation.","The released dataset of over 120,000 structure-spectrum pairs spanning 70+ elements becomes a reusable benchmark for future spectroscopy models.","Because the SGMLP components are shared across tasks, the architecture suggests that a single gated-MLP design can serve both forward and inverse spectroscopy problems."],"supporting_citations":[{"why":"Supplies the 43,000+ crystal structures and roughly 120,000 simulated spectra that both the forward and inverse pipelines are trained on.","marker":"[34]"},{"why":"The pre-trained graph neural network encoder whose node embeddings the forward spectral head builds on.","marker":"[21]"},{"why":"The prior copper-specific L-edge model whose reported MAE of 0.0391 is the benchmark XAStruct's 0.0012 is compared against.","marker":"[17]"},{"why":"The random forest algorithm used as the coordination-number classifier and as the main inverse-pipeline baseline.","marker":"[31]"},{"why":"Prior random-forest XANES property models that are element-specific, the approach XAStruct claims to generalize beyond.","marker":"[20]"},{"why":"Prior random-forest coordination-environment identification from XANES, whose element-specific setup frames the inverse task.","marker":"[19]"},{"why":"A recent graph-neural-network XANES analysis method that still lacks multi-element generalization, positioning XAStruct's scope.","marker":"[23]"}],"fun_headline_variants":["Bidirectional AI links X-ray spectra to crystal structure across 70 elements","AI flips between X-ray spectra and atomic structure for any element","From X-ray spectra to structure and back: one AI for 70+ elements","XAS spectra decoded to structure and predicted from structure by one network","One neural net reads and writes X-ray spectra across the periodic table"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The simulated X-ray absorption spectra taken from the Materials Project are accurate and representative enough to stand in for real experimental spectra across all 70+ elements and both absorption edges, and the random 8:2 split does not leak structurally similar materials into the validation set.","fun_headline_variants_meta":{"raw":{"variants":["Bidirectional AI links X-ray spectra to crystal structure across 70 elements","AI flips between X-ray spectra and atomic structure for any element","From X-ray spectra to structure and back: one AI for 70+ elements","XAS spectra decoded to structure and predicted from structure by one network","One neural net reads and writes X-ray spectra across the periodic table"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001013,"raw_usage":{"total_tokens":4299,"prompt_tokens":983,"completion_tokens":3316,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":3221}},"tokens_in":599,"tokens_out":3316,"duration_ms":23926,"temperature":1.0,"reasoning_tokens":3221,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T01:00:49.239272+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of experimental XAS spectra for a dozen well-characterized compounds spanning light, transition, and heavy elements whose structures are not in the training set, normalize them as the paper does, and compare XAStruct's predicted mean nearest-neighbor distance and neighbor atom type against crystallographic values. If per-element systematic biases appear, or if the reported errors grow sharply outside the training distribution, the periodic-table generalization claim is falsified.","supporting_citations":[{"cited_title":"CuXASNet: Rapid and Accurate Prediction of Copper L-edge X-Ray Absorption Spectra Using Machine Learning","cited_arxiv_id":"2412.02916","evidence_quote":"The prior copper-specific L-edge model whose reported MAE of 0.0391 is the benchmark XAStruct's 0.0012 is compared against."},{"cited_title":"Random forest machine learning models for interpretable x-ray absorption near-edge structure spectrum-property relationships","cited_arxiv_id":null,"evidence_quote":"Prior random-forest XANES property models that are element-specific, the approach XAStruct claims to generalize beyond."},{"cited_title":"Random forest models for accurate identification of coordination environments from x-ray absorption near-edge structure","cited_arxiv_id":null,"evidence_quote":"Prior random-forest coordination-environment identification from XANES, whose element-specific setup frames the inverse task."},{"cited_title":"A graph neural network-based approach to xanes data analysis","cited_arxiv_id":null,"evidence_quote":"A recent graph-neural-network XANES analysis method that still lacks multi-element generalization, positioning XAStruct's scope."}],"review_version":1}