{"id":"70deaec1-8f59-4a80-ba78-d5a10119ea57","arxiv_id":"2611.11019","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":6,"one_line_summary":"A local excitation Hamiltonian readout achieves exact size consistency and improves transferability for excited-state machine-learned potentials.","lead":"This paper shows that using a local excitation Hamiltonian readout in machine-learned interatomic potentials gives exact size consistency for excited-state energies, meaning the predicted spectrum of two separated molecules is exactly the union of their individual spectra. This property improves transferability to larger molecules and dimers, bringing universal excited-state simulations closer.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"No significant objection identified","rationale":"The Reader's weakest assumption concerns the local representation's ability to capture the relevant excited-state physics. The paper acknowledges this limitation and provides evidence that the model works well for the tested classes of molecules (closed-shell organics with C, H, N, O, F). No additional critical flaw emerged. The central claim (exact size consistency) is mathematically clean and empirically supported. Therefore my assessment does not change the verdict.","tokens_in":40955,"tokens_out":2181,"duration_ms":22485,"concrete_test":"Train LEH and a direct-softmin baseline on a dataset that includes a few donor–acceptor complexes where low-lying charge-transfer states exist within the cutoff. Test both models on a held-out set of such complexes. If LEH fails to reproduce the CT states (expected, given the model's scope), this confirms the acknowledged limitation; if it unexpectedly succeeds, the scope of size consistency might be broader than claimed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that LEH achieves exact size consistency by construction—is mathematically proven (Eq. 4) and empirically validated with dimer spectra matching monomer spectra to 10⁻¹⁴ eV. The paper acknowledges the crucial scope condition: this holds only for local single excitations within the cutoff radius; charge-transfer states, double excitations, and long-range couplings are not represented. Extensive comparisons against multiple baselines on varied test sets systematically support that LEH's advantage grows with distance from the training distribution. No internal inconsistency or hidden assumption undermines the core argument. The study is thorough, the code and data are released, and limitations are honestly stated. I find no load-bearing concern beyond those already noted by the authors and the Reader.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"This paper introduces LEH-MACE, a machine-learned interatomic potential for excited states that replaces the conventional multistate readout with a local excitation Hamiltonian (LEH) assembled from atom-centred onsite and pair blocks. The eigenvalues of this matrix yield the singlet excitation energies, and the block-diagonal structure ensures exact size consistency: for noninteracting fragments the spectrum is the union of the spectra of the fragments. The authors generate a large TD-ωB97X-D/6-31G* dataset of ~300,000 geometries of ~100,000 molecules with energies and forces, covering equilibrium, thermal, and excited-state-sampled geometries. They compare LEH against multiple baselines (direct readouts, X-MACE, OMNI-P2x, MS-ANI) on varied test sets including larger molecules, conjugated dyes, ladder molecules, and dimers. LEH shows accuracy comparable to the best baselines on in-distribution molecules, and its advantage grows with distance from the training distribution, especially for states above S1 and for molecule pairs. The paper also includes an active-learning loop that improves accuracy on excited-state dynamics. Limitations (charge-transfer states, double excitations, long-range couplings, nonadiabatic couplings, S1/S0 conical intersections) are explicitly acknowledged.","tokens_in":41122,"tokens_out":1840,"duration_ms":19332,"significance":"The paper addresses a fundamental challenge in excited-state machine-learned potentials: size consistency. The LEH architecture is a principled solution that guarantees the union-of-spectra property exactly, which no previous readout achieves. The empirical results convincingly show that this property translates to improved transferability to larger molecules, conjugated systems, and dimers, especially for higher excited states. The dataset and open-source code are valuable resources. The work is a clear step toward universal excited-state MLIPs and will likely influence subsequent research.","major_comments":[{"comment":"The benchmark models (LEH, Direct-softmin, X-MACE) include three validation molecules (styrene, benzaldehyde, cinnamaldehyde) in their training set, as acknowledged in Supplementary Note 14. This is a mild contamination of the validation set, though the main test sets are unaffected. The authors should discuss the potential impact on the reported validation errors and the active-learning stopping criteria, which rely on these validation trajectories. The paper already notes this, but a brief quantitative statement of the effect would strengthen the report.","section":"Table 1 and Fig. 3","authors_acknowledge":"The benchmark models contain three of the seven validation molecules (styrene, benzaldehyde, and cinnamaldehyde) in their training set, and the sampling models of rounds 3 and 4, fine-tuned from the dye-augmented LEH model, additionally azobenzene (Methods), so the criterion values of these models are slightly optimistic.","authors_acknowledge_p":0.88},{"comment":"The choice of the weighted loss uses validation data (mean signed S1 error below 2 eV must decrease by at least 30%, validation S1–S3 error and trajectory error must not increase by more than 5%). The authors apply this rule uniformly across architectures. While the procedure is transparent, using the same validation data both to design the loss and to monitor progress is a mild degree of optimism. The paper acknowledges this, but a note about potential overfitting to the validation set would be appropriate. The results on the held-out test trajectories (not used for any decision) largely confirm the trends, so the concern is limited.","section":"Methods: Weighted loss decision","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The phrase 'with about half their S2 and S3 errors for long conjugated oligomers' could be clarified: 'about half the errors of the direct readouts for S2 and S3' would avoid ambiguity.","section":"Abstract","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The x-axis labels include 'Dyes', 'Ladders', 'Dimers'. Consider adding a note that the 'Dimers' error is for both different and identical molecules, since the text discusses both cases.","section":"Fig. 4a","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Equation (2) shows Hij with a symmetrization. It may be worth noting explicitly that this ensures the matrix is symmetric, which is already stated but could be highlighted for clarity.","section":"Methods: Local excitation Hamiltonian","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Table S2 is very comprehensive. Consider adding a column for the number of molecules in each test set to help readers gauge statistical significance.","section":"Supplementary Note 2","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"accept"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"This paper is a genuine step forward. The core idea — assembling a local excitation Hamiltonian from atom-centred blocks whose eigenvalues are the excitation energies — is mathematically clean and directly solves the size-consistency problem that all multistate pooled readouts have. The proof is trivial once you see it, but nobody had built it into a working transferable MLIP before. The empirical validation is thorough: large dataset (300k geometries, 100k+ molecules), multiple baselines (X-MACE, OMNI-P2x, MS-ANI, and their own direct readouts on the same encoder), and systematic tests on random molecules, unseen scaffolds, larger molecules, conjugated dyes, ladder molecules, and dimers. The advantage of LEH grows exactly where a universal model needs it most — far from the training distribution. The authors also ran an active-learning loop to improve coverage of twisted geometries, and they release code, data, and trained models. The limitations are stated with unusual candour: no charge-transfer states, no double excitations, no long-range couplings, and the direct softmin readout is still more accurate for S1 dynamics of single large chromophores. That last point is the main practical trade-off: for photodynamics of individual molecules, LEH is not yet the best choice. But for spectra, for states above S1, and for multiple molecules, it is. The paper's one soft spot is that the model's systematic overestimation of S1 for large conjugated molecules is not fully explained — the authors say they don't know why the eigenvalue readout learns this lowering less well. That's an honest admission, not a flaw in the evaluation. The results are reproducible: the math is exact, the data are public, and the trained models are released. This paper deserves a serious referee. I would bring it to reading group and would cite it for the architecture and dataset.","headline":"LEH-MACE delivers exact size consistency for excited-state MLIPs via a local Hamiltonian readout, with thorough evaluation and honest limitations; the main trade-off is that direct softmin readouts still beat it for S1 dynamics of individual large chromophores.","tokens_in":41676,"tokens_out":1291,"duration_ms":12688},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"A local excitation Hamiltonian readout makes machine-learned excited-state potentials exactly size consistent, improving transfer to larger and paired molecules.","keywords":["machine-learned potentials","excited states","size consistency","transferability","local excitation Hamiltonian","TD-DFT","active learning","photodynamics"],"falsifier":"If a model with a non-size-consistent readout (e.g., the softmin readout) were to match or exceed LEH's accuracy for S2 and S3 of noninteracting molecule pairs, the claim that exact size consistency is necessary for transferability to higher-lying states would be disproven.","tokens_in":42332,"feed_emoji":"","tokens_out":2281,"duration_ms":20550,"temperature":0.7,"pith_summary":"The paper introduces LEH-MACE, a machine-learned potential for excited states whose excitation energies are the eigenvalues of a local excitation Hamiltonian assembled from atom-centered blocks. This construction guarantees exact size consistency: for noninteracting molecules the predicted spectrum is exactly the union of the molecules' spectra. The approach is tested on a new dataset of 300,000 TD-DFT geometries and shown to transfer better than standard multistate readouts to larger molecules, conjugated dyes, ladder molecules, and molecule pairs, especially for states above the first excited state. The authors argue that exact size consistency is a necessary step toward universal excited-state MLIPs.","feed_headline":"Exact size consistency boosts excited-state ML potentials","feed_subtitle":"A local Hamiltonian readout makes machine-learned excitation energies transfer to larger, conjugated, and dimer systems.","key_machinery":"The local excitation Hamiltonian (LEH) is a symmetric matrix built from atom-centred blocks: a diagonal onsite block from each atom's MACE features and off-diagonal coupling blocks for atom pairs within a 5 Å cutoff. Its sorted eigenvalues are the singlet excitation energies. For two fragments beyond the cutoff the matrix is block diagonal, so its spectrum is the union of the fragments' spectra by construction. This gives exact size consistency for local single excitations.","core_discovery":"Only a local excitation Hamiltonian readout is exactly size consistent for excited-state MLIPs: the spectrum of noninteracting fragments is the multiset union of their spectra. This property, which readouts based on summing, averaging, or soft-min pooling do not share, improves transferability to larger molecules, conjugated systems, and dimers, with the largest advantage for higher-lying excited states. LEH-MACE achieves half the errors of the softmin readout for S2 and S3 of long conjugated oligomers and retains single-molecule accuracy for dimers, whereas all other readouts degrade by factors of three to four.","pith_inferences":["LEH could be combined with ground-state foundation models (e.g., MACE-MP-0) to create a universal excited-state potential that inherits ground-state transferability.","The failure to learn S1 lowering for large chromophores suggests that multiple latent channels per atom or longer-range couplings are needed; these are testable extensions that the paper's framework supports.","The exact size consistency property could be exploited in fragment-based methods to predict spectra of molecular clusters from monomer models without retraining.","The paper's dataset, covering many geometries per molecule, enables studying whether excited-state MLIPs benefit more from more molecules or more geometries per molecule; the authors find they are equally valuable."],"forward_implications":["LEH can be used to predict absorption spectra and photophysical processes involving higher-lying excited states with transferable accuracy.","The readout is well suited for simulations of chromophores in explicit solvent, where large solvation shells contain many noninteracting molecules.","Active learning on excited-state dynamics selects geometries that standard thermal sampling misses; combining it with a size-consistent readout may extend the reach of excited-state MLIPs to nanosecond photodynamics.","The open dataset and code provide a benchmark for developing size-consistent excited-state models and for testing extensions to charge-transfer and long-range couplings."],"supporting_citations":[{"why":"The MACE encoder that provides atom features for the local excitation Hamiltonian.","marker":"[29]"},{"why":"The QCDGE dataset from which most training molecules and reference excitation energies are derived.","marker":"[26]"},{"why":"X-MACE, a baseline model using a DeepSets autoencoder readout, compared against LEH.","marker":"[18]"},{"why":"OMNI-P2x, a baseline model using per-state atomic sums, compared against LEH.","marker":"[19]"},{"why":"Extremal pooling (softmin readout) that defines the Direct-softmin baseline.","marker":"[20]"},{"why":"Latent Hamiltonian approach for individual-molecule excited states, which LEH extends to transferable atom-centred blocks.","marker":"[24]"},{"why":"Active-learning protocol for excited-state dynamics that the paper adapts.","marker":"[8]"}],"fun_headline_variants":["Local Hamiltonian gives exact size consistency for excited-state ML","Only local Hamiltonian readout keeps accuracy for dimers and large molecules","Exact size consistency in ML: local Hamiltonian beats softmin for higher states","LEH-MACE halves errors for S2 and S3 of conjugated oligomers","Local excitation Hamiltonian keeps dimer accuracy while others degrade threefold"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The representation is limited to local single excitations within a 5 Å cutoff; charge-transfer states, double excitations, and long-range excitonic couplings are not captured by the model.","fun_headline_variants_meta":{"raw":{"variants":["Local Hamiltonian gives exact size consistency for excited-state ML","Only local Hamiltonian readout keeps accuracy for dimers and large molecules","Exact size consistency in ML: local Hamiltonian beats softmin for higher states","LEH-MACE halves errors for S2 and S3 of conjugated oligomers","Local excitation Hamiltonian keeps dimer accuracy while others degrade threefold"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":9.4e-05,"raw_usage":{"cost":9.440256e-05,"is_byok":false,"cost_details":{"upstream_inference_cost":9.440256e-05,"upstream_inference_prompt_cost":2.953216e-05,"upstream_inference_completions_cost":6.48704e-05},"total_tokens":1588,"prompt_tokens":864,"completion_tokens":724,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":632}},"tokens_in":864,"tokens_out":724,"cost_source":"openrouter_usage","duration_ms":7351,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":632,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-10T20:04:19.370464+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":"If a model with a non-size-consistent readout (e.g., the softmin readout) were to match or exceed LEH's accuracy for S2 and S3 of noninteracting molecule pairs, the claim that exact size consistency is necessary for transferability to higher-lying states would be disproven.","supporting_citations":[{"cited_title":"MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields","cited_arxiv_id":"2206.07697","evidence_quote":"The MACE encoder that provides atom features for the local excitation Hamiltonian."},{"cited_title":"Quantum chemistry dataset with ground- and excited-state properties of 450 kilo molecules.Scientific Data2024,11, 948, DOI: 10.1038/s41597-024-03788-x","cited_arxiv_id":null,"evidence_quote":"The QCDGE dataset from which most training molecules and reference excitation energies are derived."},{"cited_title":"Transferable Machine Learning Potential X-MACE for Excited States using Integrated DeepSets","cited_arxiv_id":"2502.12870","evidence_quote":"X-MACE, a baseline model using a DeepSets autoencoder readout, compared against LEH."},{"cited_title":"OMNI-P2x universal neural network potential for excited-state simulations","cited_arxiv_id":null,"evidence_quote":"OMNI-P2x, a baseline model using per-state atomic sums, compared against LEH."},{"cited_title":"Transferable machine learning of excited-state dynamics with extremal pooling","cited_arxiv_id":"2606.16859","evidence_quote":"Extremal pooling (softmin readout) that defines the Direct-softmin baseline."},{"cited_title":"Latent unified smooth Hamiltonians for excited state chemistry","cited_arxiv_id":"2609.01871","evidence_quote":"Latent Hamiltonian approach for individual-molecule excited states, which LEH extends to transferable atom-centred blocks."},{"cited_title":"Machine learning enables long time scale molecular photodynamics simulations","cited_arxiv_id":null,"evidence_quote":"Active-learning protocol for excited-state dynamics that the paper adapts."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}