{"id":"c14142f4-d17c-446b-9ef7-084aa3403ff7","arxiv_id":"2611.16342","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"HARMONY uses equivariant neural networks to generate harmonic spectra and a Transformer to map them to experimental IR and Raman spectra, achieving mean cosine similarities of 0.912 and 0.907.","lead":"HARMONY is a two-stage deep learning framework that predicts experimental infrared and Raman spectra from molecular structures using harmonic spectra as physical anchors. It achieves high accuracy with second-scale inference, approximately 600 times faster than standard DFT calculations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"HarmoNet's generalization to unseen measurement conditions is untested because metadata fields are sparse and the test set may not sample condition combinations independently of molecular structure.","rationale":"The reader's weakest_assumption correctly identifies that generalization to unseen molecules and conditions is a key assumption, and notes that metadata sufficiency is uncertain. However, the reader's verdict is ACCEPT with HIGH confidence, suggesting that the evidence is considered sufficient. My concern is more specific: the paper does not demonstrate that the model generalizes to unseen condition combinations, as opposed to memorizing condition-dependent corrections for molecules whose condition combinations were seen during training. The scaffold split prevents molecular overlap but does not prevent condition leakage. The adipic acid example is illustrative but not statistically significant. The paper's own limitation statement (b213) acknowledges the need for broader coverage of measurement conditions, which supports the concern. Nevertheless, the paper is well-executed and the concern does not invalidate the central claim; it identifies a specific gap in the evidence. The verdict should remain UNCHANGED because the paper is still a solid contribution, but the confidence might be slightly tempered. The agreement_with_reader is 'partial' because the reader identified the general assumption but did not pinpoint the specific risk of condition-molecule correlation or the lack of cross-condition evaluation.","tokens_in":23184,"tokens_out":746,"duration_ms":11163,"concrete_test":"Identify all molecules in the HarmoSpec test set that appear under at least two distinct metadata condition combinations (e.g., gas cell vs. solid split mull). For each such molecule, compute the cosine similarity between the predicted and experimental spectra for each condition. Then compare the mean similarity for molecules whose condition combinations were both present in the training set versus those where at least one condition combination was absent from training. If the latter group shows significantly lower similarity, the model is memorizing rather than generalizing condition-dependent effects. This test requires no new data, only reanalysis of existing predictions stratified by condition novelty.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that HARMONY accurately predicts experimental vibrational spectra across diverse molecules and conditions rests on the assumption that the learned mapping from harmonic spectra to experimental spectra generalizes to measurement conditions not seen during training. The paper acknowledges this as a limitation (b213), but the evidence provided is insufficient to assess the risk. The HarmoSpec dataset contains only 2,522 Raman spectra versus 33,423 IR spectra (Table S1), and the metadata fields (phase, preparation, matrix) are categorical with many combinations appearing only a few times (Fig. S2). The scaffold-based split (Section 4.1) prevents molecular overlap but does not guarantee that condition combinations are independent of molecular structure. For example, gas-phase measurements are more common for volatile small molecules, while solid-state measurements dominate for larger or ionic compounds. If condition correlates with molecular topology, the model could learn to associate certain molecular features with spectral shifts rather than learning a true condition-dependent mapping. The adipic acid example (Fig. 3D) is a single molecule; it demonstrates that the model can produce different outputs for different metadata inputs, but it does not test whether the mapping generalizes to molecules whose condition combinations were absent from training. The paper does not report a systematic evaluation where the same molecule appears under multiple conditions in the test set, nor does it analyze whether prediction accuracy degrades for condition combinations that are rare or absent in training. Without such analysis, the risk remains that HarmoNet's condition-dependent corrections are memorized rather than generalized.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"This paper introduces HARMONY, a two-stage deep learning framework for predicting experimental vibrational spectra (IR and Raman) from molecular structures. The first stage uses equivariant neural networks (EquiformerV3) to directly predict atomic forces, Hessians, dipole derivatives, and polarizability derivatives, enabling rapid construction of harmonic spectra without DFT calculations. The second stage, HarmoNet, is a Transformer-based model that learns a theory-to-experiment mapping by combining harmonic spectra with molecular topology (via graph neural networks) and measurement metadata (phase, preparation, matrix) through adaptive layer normalization and cross-attention. HARMONY achieves mean cosine similarities of 0.912 for IR and 0.907 for Raman spectra on held-out test sets, with inference times of approximately 2.12 seconds per molecule, roughly 600-fold faster than the evaluated DFT workflow. Beyond forward prediction, the paper demonstrates applications in molecular structure elucidation by using HARMONY to generate pseudo-experimental spectra for pretraining the Vib2Mol inverse model, for spectral reranking of candidate structures, and as a scientific tool within a Codex-based agent (Vib2Mol-HARMONY Agent) that achieves Top-1 recall of 70.53% on NIST and 83.33% on SDBS.","tokens_in":30394,"tokens_out":1694,"duration_ms":175400,"significance":"If the reported results hold, HARMONY represents a significant advance in computational spectroscopy. The two-stage framework with harmonic anchoring is physically motivated and novel, providing a principled way to combine quantum-chemical information with data-driven learning. The paper provides extensive validation, including comparisons with multiple baselines (DFT, MLMD, end-to-end neural predictors), ablation studies demonstrating the complementary roles of harmonic spectra, molecular topology, and measurement metadata, cross-route transfer tests showing robustness to the source of harmonic spectra, and applications to molecular structure elucidation and peptide spectroscopy. The integration with general-purpose agents for adaptive structure elucidation is a creative and potentially impactful extension. The paper is well-written, the methodology is clearly described, and the code and data are presumably available. The work is likely to be of interest to the computational chemistry, spectroscopy, and machine learning communities.","major_comments":[{"comment":"The paper demonstrates that metadata conditioning improves predictions and can distinguish conditions for the same molecule (adipic acid). However, the test set is split by molecular scaffold (Section 4.1), not by condition combinations. The paper does not systematically evaluate whether the model can generalize to condition combinations not seen during training. The authors acknowledge this as a future direction (b213), but given the central claim of predicting spectra 'across diverse molecules and conditions,' this limitation should be discussed more prominently. I recommend adding an analysis of condition combination overlap between train and test sets, and a discussion of how potential correlation between molecular structure and measurement conditions could affect generalization.","section":"Section 2.3, Fig. 3D","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The primary metric is cosine similarity, which measures spectral shape but not absolute intensity. While standard in the field, the paper should also report RMSE or MAE for spectral prediction to provide a more complete picture. Additionally, the paper uses Savitzky-Golay smoothing with a 14-point window and cubic polynomial as postprocessing for the final predictions (Section 4.6). The paper should report raw prediction performance (without smoothing) to allow comparison with methods that do not use such postprocessing, or at least justify why smoothing is necessary and quantify its effect on the reported cosine similarities.","section":"Section 4.6, evaluation metrics","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The comparison with baselines is performed on a fixed panel of 45 molecules with gas-phase IR spectra. While this is a reasonable benchmark, the paper should also report full test set performance for all methods, or explain why this is not feasible. The current presentation gives the impression that HARMONY outperforms all baselines, but the comparison is limited to a specific subset. If full test set comparison is not possible, the paper should discuss potential selection bias and how the 45-molecule panel relates to the full test set.","section":"Section 2.2, Fig. 2A","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The peptide adaptation results show a large improvement (from 0.669 to 0.896) on 15 held-out peptides. However, the training set is very small (27 spectra), and no confidence intervals or statistical significance tests are reported. Given the small sample size, the paper should report bootstrap confidence intervals or similar measures to assess the reliability of the improvement. Additionally, the paper should discuss the potential for overfitting given the small training set and the large number of parameters in HarmoNet.","section":"Section 2.6, peptide adaptation","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The claim 'approximately 600-fold faster than the evaluated DFT workflow' is based on GPU (HARMONY) vs CPU (Gaussian 16) timing. The paper should clarify this in the abstract or main text to avoid misleading readers.","section":"Abstract","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The x-axis is on a log scale, which makes the comparison visually clear. However, the timing for end-to-end predictors (3.1-3.7 ms) is not visible on this scale. Consider adding a zoomed inset or a table with exact values.","section":"Figure 2A","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The regional ablation study is informative. However, the cosine loss values are very small (0.00-0.08). The paper should clarify whether these differences are statistically significant, perhaps by reporting confidence intervals or effect sizes.","section":"Section 2.3, Fig. 3C","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The frequency error plot shows good agreement. The paper should report the number of molecules with imaginary frequencies and how they affect the harmonic spectrum quality, beyond what is already in Supplementary Table S11.","section":"Section 2.4, Fig. 4C","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The masked MSE loss uses 'max(1, sum(m_ip))' as normalization. This is unusual; typically one would divide by the number of valid points. The paper should explain why max(1, ...) is used instead of just sum(m_ip), and whether this choice affects training.","section":"Section 4.3, Eq. (4)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The Raman graph-only HarmoNet achieves 0.730, which is much lower than SpecGNN's 0.874. The paper should comment on this discrepancy in the main text or supplement.","section":"Supplementary Table S3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper uses 'cosine similarity' throughout, but in some places it also reports Pearson correlation coefficient 'r'. The relationship between these metrics should be clarified, as they measure different aspects of similarity.","section":"Fig. 2C,D","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper should ensure all references are complete and correctly formatted. For example, reference [14] (EquiformerV3) is an arXiv preprint; the paper should note if it has been accepted for publication or is under review.","section":"References","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"minor_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Colleague,\n\nThis paper is worth your time. HARMONY is a two-stage framework that first uses equivariant neural networks to predict harmonic spectra from molecular structure, then learns a mapping from those harmonic spectra to experimental IR and Raman spectra, conditioned on molecular topology and measurement metadata. The core idea—using harmonic spectra as a physical anchor rather than predicting experimental spectra directly from structure—is sensible and the execution is thorough.\n\nWhat is new: the combination of direct equivariant tensor prediction (forces, Hessians, dipole/polarizability derivatives) with a Transformer-based theory-to-experiment mapping that incorporates metadata. The direct prediction of spectroscopic tensors without differentiation is a practical advance that gives a 600× speedup over DFT while maintaining high spectral fidelity. The ablation studies are well done and show that all three information sources (harmonic spectra, molecular graph, metadata) contribute. The cross-route transfer experiment (DFT-trained HarmoNet works on DL harmonic inputs and vice versa) is a nice sanity check that the mapping is learning something physically meaningful rather than dataset-specific corrections.\n\nThe paper does what it claims. The numbers are credible: mean cosine similarities of 0.912 (IR) and 0.907 (Raman) on held-out test sets, with second-scale inference. The structure elucidation results (Vib2Mol–HARMONY Agent achieving 70.53% Top-1 recall on NIST) are impressive, though the comparison to prior work is fair and the gains come from multiple components (pseudo-experimental pretraining, reranking, agentic search).\n\nSoft spots: The stress-test concern about metadata generalization is real but overstated. The paper acknowledges this limitation explicitly (b213: \"broader coverage of chemical and measurement conditions\"). The test set uses a scaffold-based split that prevents molecular overlap, but as the stress-test notes, condition combinations may correlate with molecular structure. The paper does not report a systematic evaluation of prediction accuracy for rare or unseen condition combinations. That said, the ablation results show that metadata contributes most in the high-wavenumber region (2500–4000 cm⁻¹) and remains beneficial even in the fingerprint region, which suggests the model is learning genuine condition-dependent corrections rather than memorizing. The adipic acid example (Fig. 3D) is a single molecule but it does demonstrate that the model can produce distinct outputs for different metadata inputs from the same harmonic spectrum. A more rigorous test would be to hold out entire condition combinations and measure degradation, but this is a minor omission in an otherwise thorough evaluation.\n\nAnother soft spot: the Raman dataset is small (2,522 spectra vs 33,423 IR). The peptide adaptation results are promising but based on only 27 training spectra. The authors acknowledge this (b214) and frame it as preliminary. The Raman results are weaker than IR across the board, which is expected given the data imbalance.\n\nThe paper is well-written, the methods are clearly described, and the code and data appear to be available. The authors do not oversell; they state limitations plainly.\n\nWho this is for: anyone working in computational spectroscopy, machine learning for chemistry, or structure elucidation. The framework is practical and the ablation studies provide useful design guidance for similar problems.\n\nRecommendation: This paper deserves a serious referee. I would accept it with minor revisions. The metadata generalization concern should be addressed in revision—either with additional analysis or a clearer statement of the limitation—but it does not undermine the central claims.\n\nBest,\n[You]","headline":"HARMONY is a solid, well-engineered framework that delivers on its claims; the metadata generalization concern is real but the paper acknowledges it and the evidence is stronger than the stress-test suggests.","tokens_in":24097,"tokens_out":826,"duration_ms":9235},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"HARMONY uses harmonic spectra as physical anchors to predict experimental IR and Raman spectra with high accuracy and second-scale inference, achieving mean cosine similarities of 0.912 and 0.907 while running about 600-fold faster than…","keywords":["vibrational spectroscopy","infrared spectroscopy","Raman spectroscopy","equivariant neural networks","harmonic spectra","experimental spectrum prediction","molecular structure elucidation","deep learning"],"falsifier":"Evaluate HARMONY on a diverse set of molecules whose chemical scaffolds and experimental conditions, such as different solvents, temperatures, or instrument configurations, are not represented in the HarmoSpec training set, and measure whether the cosine similarity to experimental spectra drops significantly below the reported means of 0.912 for infrared and 0.907 for Raman.","tokens_in":27655,"feed_emoji":"🔬","tokens_out":1593,"duration_ms":25780,"temperature":0.7,"pith_summary":"HARMONY is a two-stage deep-learning framework that predicts experimental infrared and Raman spectra by using harmonic spectra as physical anchors, derived from quantum-chemical potential-energy surfaces and molecular response derivatives. In the first stage, equivariant neural networks directly predict atomic forces and spectroscopic tensors to construct harmonic spectra without expensive coordinate differentiation. In the second stage, HarmoNet learns a mapping from these harmonic spectra to experimental spectra, conditioned on molecular topology and measurement metadata such as sample phase and preparation method. The paper demonstrates that this harmonic-anchoring approach bridges the persistent theory-to-experiment gap in vibrational spectroscopy while reducing per-molecule inference time to 2.12 seconds, about 600-fold faster than a DFT workflow. If correct, HARMONY enables rapid spectral assignment, molecular identification, and molecular structure elucidation across diverse chemical systems.","feed_headline":"Harmonic anchors predict IR and Raman spectra with 600× speedup","feed_subtitle":"Two-stage deep learning framework achieves 0.91 cosine similarity in seconds per molecule, enabling rapid structure elucidation","key_machinery":"The central mechanism is the two-stage harmonic-anchored framework. Stage one uses EquiformerV3 equivariant neural networks to directly predict atomic forces, Hessians, dipole derivatives, and polarizability derivatives from molecular graphs and three-dimensional geometries, enabling rapid harmonic-spectrum construction without coordinate differentiation. Stage two is HarmoNet, a Transformer architecture that partitions harmonic spectra into patches and processes them through six blocks combining spectral self-attention, molecular cross-attention using a two-dimensional graph neural network encoder, and adaptive layer normalization conditioned on measurement metadata (phase, preparation, matrix). This design integrates three complementary information sources, frequency-resolved vibrational information from harmonic spectra, chemical context from molecular topology, and experimental conditions from metadata, to predict experimental infrared and Raman spectra jointly.","core_discovery":"The paper establishes that computational harmonic spectra, despite deviating from experiment, serve as effective physical anchors for learning a transferable theory-to-experiment mapping. HARMONY achieves mean cosine similarities of 0.912 for infrared and 0.907 for Raman spectra across diverse molecules and measurement conditions, with a mean inference time of 2.12 seconds per molecule, approximately 600-fold faster than the evaluated DFT workflow. Crucially, the learned mapping transfers across different sources of harmonic spectra, DFT versus deep-learning predictions, without retraining, indicating that the model captures vibrational information shared across computational approximations rather than source-specific artifacts. The paper also shows that accurate forward prediction improves inverse molecular inference: pseudo-experimental spectra improve Vib2Mol pretraining, spectral verification reranks candidate structures, and integration with a general-purpose agent achieves Top-1 recall of 70.53 percent on NIST and 83.33 percent on SDBS for structure elucidation.","pith_inferences":["The success of cross-route transfer, DFT versus deep-learning harmonic inputs, implies that the learned mapping captures a robust relationship between local molecular vibrations and experimental spectra, so improvements in harmonic-spectrum quality from more accurate quantum methods could directly improve prediction accuracy without retraining the mapping.","Metadata conditioning could be extended to incorporate additional experimental variables such as temperature, concentration, and instrument parameters, potentially improving accuracy and broadening applicability to more diverse measurement conditions.","The agent-assisted structure elucidation paradigm could generalize to other spectroscopic techniques such as nuclear magnetic resonance, ultraviolet-visible, or mass spectrometry by substituting the forward prediction model while retaining the adaptive search and verification loop.","Using harmonic spectra as explicit intermediate representations rather than end-to-end structure-to-spectrum prediction may improve interpretability, because the harmonic spectrum provides a physically meaningful latent space that can be analyzed independently."],"forward_implications":["Rapid and accurate experimental spectrum prediction enables high-throughput spectral assignment and molecular identification without quantum-chemical calculations, reducing per-molecule cost from minutes to seconds.","Pseudo-experimental spectra generated by HARMONY can serve as training data for inverse structure-elucidation models, alleviating the scarcity of experimental vibrational spectra.","Spectral verification via HARMONY reranking improves molecular structure elucidation accuracy, and when integrated with adaptive agents, can identify structures beyond fixed candidate pools.","The framework can be adapted to new chemical domains, such as short peptides, with limited experimental data through fine-tuning, suggesting potential applicability to broader classes of molecules.","Harmonic anchoring provides a principled way to combine physical priors with data-driven learning for spectroscopic prediction, potentially transferable to other spectroscopic modalities."],"supporting_citations":[{"why":"EquiformerV3 architecture provides the equivariant neural network backbone for direct prediction of forces and spectroscopic tensors.","marker":"[14]"},{"why":"QMe14S dataset supplies the quantum-chemical reference data for training the first-stage models.","marker":"[15]"},{"why":"NIST Chemistry WebBook is the primary source of experimental infrared spectra for the HarmoSpec dataset.","marker":"[16]"},{"why":"SDBS database supplies experimental infrared and Raman spectra for the HarmoSpec dataset.","marker":"[17]"},{"why":"DetaNet establishes the harmonic-spectrum construction methodology that HARMONY adapts.","marker":"[3]"},{"why":"Chemprop-IR serves as an end-to-end neural baseline for infrared spectrum prediction.","marker":"[8]"},{"why":"SpecGNN provides both a baseline for spectral prediction and the molecular graph encoder architecture adapted by HarmoNet.","marker":"[9]"},{"why":"Vib2Mol is the inverse structure-generation model that HARMONY integrates with for molecular structure elucidation.","marker":"[19]"}],"fun_headline_variants":[],"cache_read_input_tokens":0,"weakest_assumption_plain":"The learned mapping from harmonic spectra to experimental spectra, conditioned on molecular topology and measurement metadata, generalizes to molecules and experimental conditions not represented in the training data (HarmoSpec), and the measurement metadata fields (phase, preparation, matrix) capture the relevant experimental effects sufficiently for accurate prediction.","fun_headline_variants_meta":{"raw":{"variants":[]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":9.9e-05,"raw_usage":{"cost":9.8917e-05,"is_byok":false,"cost_details":{"upstream_inference_cost":9.8917e-05,"upstream_inference_prompt_cost":9.8553e-05,"upstream_inference_completions_cost":3.64e-07},"total_tokens":1085,"prompt_tokens":1083,"completion_tokens":2,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":0,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":0}},"tokens_in":1083,"tokens_out":2,"cost_source":"openrouter_usage","duration_ms":2668,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":0,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-13T06:35:11.053407+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Evaluate HARMONY on a diverse set of molecules whose chemical scaffolds and experimental conditions, such as different solvents, temperatures, or instrument configurations, are not represented in the HarmoSpec training set, and measure whether the cosine similarity to experimental spectra drops significantly below the reported means of 0.912 for infrared and 0.907 for Raman.","supporting_citations":[{"cited_title":"EquiformerV3: Scaling Efficient, Expressive, and General SE(3)-Equivariant Graph Attention Transformers","cited_arxiv_id":"2604.09130","evidence_quote":"EquiformerV3 architecture provides the equivariant neural network backbone for direct prediction of forces and spectroscopic tensors."},{"cited_title":"doi:10.1021/acs.jpclett.5c00839 , copyright =","cited_arxiv_id":null,"evidence_quote":"QMe14S dataset supplies the quantum-chemical reference data for training the first-stage models."},{"cited_title":"NIST Chemistry WebBook, NIST Standard Reference Database 69","cited_arxiv_id":null,"evidence_quote":"NIST Chemistry WebBook is the primary source of experimental infrared spectra for the HarmoSpec dataset."},{"cited_title":"https://sdbs.db.aist.go.jp/ Accessed 8 September 2026","cited_arxiv_id":null,"evidence_quote":"SDBS database supplies experimental infrared and Raman spectra for the HarmoSpec dataset."},{"cited_title":"A deep learning model for predicting selected organic molecular spectra","cited_arxiv_id":null,"evidence_quote":"DetaNet establishes the harmonic-spectrum construction methodology that HARMONY adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Chemprop-IR serves as an end-to-end neural baseline for infrared spectrum prediction."},{"cited_title":"Crystal structure prediction via deep learning","cited_arxiv_id":null,"evidence_quote":"SpecGNN provides both a baseline for spectral prediction and the molecular graph encoder architecture adapted by HarmoNet."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Vib2Mol is the inverse structure-generation model that HARMONY integrates with for molecular structure elucidation."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}