{"id":"d721ede4-0f60-4463-894b-b608e1fd0279","arxiv_id":"2607.24327","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A GPTFF graph network maps PBE formation energies to r2SCAN with 14.3 meV/atom MAE on 380k MatPES pairs, improving phase, voltage, and reaction predictions.","lead":"A graph neural network learns the energy gap between two common quantum-chemistry methods so older, cheaper calculations can be upgraded to a more accurate standard. That could let materials researchers finally merge huge mismatched databases and train better AI models for batteries and new compounds.","discovery_kind":"new_application","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The headline 14.3 meV/atom MAE rests on a random 80:10:10 split of MatPES, a dataset dominated by near-duplicate structures along relaxation/MD trajectories — so train/test leakage likely inflates the central accuracy claim, even in its narrowly scoped form.","rationale":"The reader flagged the untested generality claim (transfer beyond MatPES-style pairs), which is real but concerns the Discussion/Conclusion framing rather than the empirical core. I agree with that concern, but the more load-bearing issue sits one level deeper: the internal validity of the 14.3 meV/atom figure itself. MatPES's trajectory-sampled structure distribution makes random splits leaky in a way that is well documented for this dataset family, and the paper's own bonding-category analysis (Fig. 1c) shows much of the residual is composition-level signal — precisely the kind that a random split lets the model memorize rather than learn. The fix is cheap and standard (composition-grouped splitting), so the verdict remains CONDITIONAL rather than REJECT: the empirical core is plausibly sound, and the application tests (hull recovery, voltages, reactions) provide some independent evidence of real signal, but the central quantitative claim cannot be taken at face value until a composition-disjoint evaluation is reported. I also concur with the reader's secondary notes (no classical baselines, no code/weights), but these are subordinate to the split-validity question.","tokens_in":9328,"tokens_out":1381,"duration_ms":49681,"concrete_test":"Retrain GPTFF with a composition-disjoint split: group all 380,190 entries by reduced formula, assign entire compositions to train/val/test (80:10:10), and report the test MAE. Also compute a trivial baseline on the same split: the per-element (or per-anion-class) mean residual learned on train. If the composition-disjoint MAE rises well beyond ~2× the reported 14.3 meV/atom, or approaches the element-wise baseline, the headline accuracy is substantially a memorization artifact of the random split and must be re-reported; if it stays near 14.3, the central claim stands.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim is that adding predicted residuals to PBE yields r2SCAN-level energies at 14.3 meV/atom test MAE. That number comes from §2.2: \"The paired dataset was randomly divided into training, validation, and test sets in an 80:10:10 ratio.\" MatPES is explicitly constructed from sampled configurations along relaxation trajectories and perturbed/near-equilibrium structures of a much smaller set of distinct compositions (~380k structures but on the order of tens of thousands of compositions). Under a purely random split, the test set will contain structures that are slight geometric perturbations of training structures of the same composition — the model then effectively interpolates a per-composition constant offset plus a small structure-dependent term it has already seen. The chemical-dependence analysis in §3.1 (Fig. 1c) actually strengthens this worry: the residual is strongly composition/bonding-dependent (ionic 181 meV/atom mean vs. metallic 62), i.e., much of the learnable signal is a composition-level correction. If test compositions overlap training compositions, a 14.3 meV/atom MAE tells us the model memorized per-composition residuals, not that it generalizes the structure-dependent part to unseen materials. Yet every downstream use the paper advertises — correcting legacy PBE databases, new hull constructions, voltages for new orderings — is prediction on unseen compositions/structures. The 75% hull-recovery result (§3.3) and the missed LiMn2O4 plateau (§3.4) may already be hints of weaker extrapolation than the test MAE suggests. This concern is distinct from, and prior to, the reader's transferability-to-other-DFT-setups point: it challenges whether the reported MAE is representative even for MatPES-style PBE→r2SCAN correction on new materials. No composition-disjoint or structure-deduped split is reported anywhere in the paper or (as far as the text indicates) the SI.","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript presents a residual-learning framework for aligning DFT formation energies across computational settings. Using 380,190 structurally paired PBE–r2SCAN entries from the MatPES dataset, the authors train graph neural networks (GPTFF and CHGNet backbones) to predict the per-atom r2SCAN-minus-PBE formation-energy residual. Adding predicted residuals to PBE energies yields a test-set MAE of 14.3 meV/atom (GPTFF) versus 18.2 meV/atom (CHGNet). The corrected energies are then validated on three downstream tasks: binary convex-hull stability for six systems (recovering 6 of 8 phases whose PBE and r2SCAN classifications differ), Li deintercalation voltages for LiCoO2 and LiMn2O4, and reaction enthalpies along four solid-state reaction pathways, in each case showing substantial improvement over raw PBE relative to r2SCAN and experiment. The authors claim the pipeline generalizes to arbitrary heterogeneous DFT dataset pairs (different codes, pseudopotentials, cutoffs) and to computation–experiment alignment.","tokens_in":9693,"tokens_out":2346,"duration_ms":79616,"significance":"If the results hold, the work addresses a real and widely felt problem: the systematic energy offsets that prevent merging large legacy PBE databases with higher-fidelity datasets. The strengths of the manuscript are concrete: a clean, non-tautological learning target (the residual is measured from independent paired DFT labels, not defined by the model); a large paired dataset (380,190 entries); two independent backbone architectures giving consistent results; and, notably, honest multi-task downstream validation — the authors report the 75% hull-recovery rate with explicit residual errors (e.g., MoCl4 at 115.9 meV/atom, Mn5O8 at 64.8 meV/atom), the missed LiMn2O4 plateau at x = 0.25, and the spurious plateau at x = 0.083, rather than cherry-picking successes. The reaction-pathway test, which shows correct product recovery against experiment (BaTiO3, LiFeO2), is a meaningful functional check beyond pointwise MAE. However, the central accuracy figure currently rests on a random split of a trajectory-dominated dataset, and the broad generality claims are untested outside the single MatPES pair, which limits the demonstrated scope relative to the stated conclusions.","major_comments":[{"comment":"The headline result — 14.3 meV/atom test MAE — is computed on a random 80:10:10 split of MatPES. MatPES is constructed from configurations sampled along relaxation/MD trajectories and perturbed near-equilibrium structures, so a random split places near-duplicate geometries of the same composition in both training and test sets. This is a standard leakage mode for PES-derived datasets and plausibly inflates the reported MAE. The concern is sharpened by the paper's own Fig. 1(c): the residual is strongly composition/bonding-dependent (ionic 181 vs. metallic 62 meV/atom mean), i.e., a large fraction of the learnable signal is a per-composition correction. Under a random split, the model can interpolate a composition-level offset it has effectively already seen, and 14.3 meV/atom would then measure memorization of composition-level residuals rather than generalization of the structure-depend","section":"§2.2, §3.2 (Fig. 2)"},{"comment":"The Discussion and Conclusion assert that the pipeline 'is applicable to any pair of DFT datasets with differing computational configurations' and can 'standardize energies across these databases' (Materials Project, OMat24, Alexandria), and the Conclusion claims it 'readily standardizes misaligned energy values across all global public materials databases.' None of this is tested. The model is trained exclusively on MatPES pairs sharing the same code, pseudopotentials, cutoffs, k-point protocols, and magnetic settings — only the functional differs. A residual learned for (PBE, r2SCAN) under MatPES settings does not, without evidence, transfer to pairs that also differ in pseudopotential, cutoff, or magnetism protocol, where the residual landscape is different and possibly larger. The Abstract's 'effectively upgrades massive legacy PBE datasets to high-precision r2SCAN standards' oversta","section":"§4 (Discussion), §5 (Conclusion), Abstract"},{"comment":"The elemental reference states are chosen independently per functional ('the lowest-energy unary phase available in the corresponding dataset'). The learned residual therefore absorbs both the functional dependence of the compound energy and the functional dependence of the elemental references. This is fine within the MatPES pair, but it means the model's target is defined relative to a specific reference-state convention; applying the correction to legacy PBE data referenced to different unary phases (e.g., MP's reference scheme, or +U-mixed references) would introduce a systematic offset of exactly the type the method is meant to remove. The manuscript should state this explicitly and quantify the reference-state contribution to the residual (e.g., the per-element PBE–r2SCAN unary energy differences), so users can re-reference correctly. This is load-bearing for the advertised use cas","section":"§2.1"}],"minor_comments":[{"comment":"The learning target is rendered as 'ΔE!' (twice), apparently a typesetting artifact of a subscript; please define the residual symbol cleanly and use it consistently.","section":"§3.1, §3.2"},{"comment":"The experimental comparison (1,234 compounds; PBE MAE 164, r2SCAN MAE 87 meV/atom) is used to motivate r2SCAN as the target scale. It would help to note that 87 meV/atom is itself a substantial residual versus experiment, so the corrected energies inherit this ceiling; readers should not interpret 'r2SCAN-level' as 'experiment-level.' A sentence to this effect would calibrate the framing in the Abstract.","section":"§3.1 (Fig. 1a)"},{"comment":"The phase-diagram analysis is commendably honest. Please clarify whether the eight phases whose classifications differ are drawn only from the six binary systems shown, and give the denominator context (total number of hull vertices evaluated), so the 75% recovery rate can be interpreted.","section":"§3.3 (Fig. 3)"},{"comment":"The LiMn2O4 panel is labeled 'LixMnO2, normalized to MnO2 formula units' — presumably LixMn2O4; please check. Also state how many Li/vacancy orderings were sampled per composition, since the missed x = 0.25 plateau may reflect sampling rather than model error.","section":"§3.4 (Fig. 4)"},{"comment":"Training details are thin: hyperparameters, whether backbones were trained from scratch or fine-tuned from pretrained weights (relevant given both are pretrained MLIPs), and compute cost. A short table or SI pointer would aid reproducibility.","section":"§2.2"},{"comment":"Refs. [13] and [41] are dated 2026 / arXiv 2512 — please verify the OMat24 citation year and update preprint status at revision. Related cross-functional learning work [28,29] is cited; a brief quantitative comparison to multifidelity results in [29] (e.g., their correction accuracy) would help situate the 14.3 meV/atom figure.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The residual-learning idea is not novel per se (the authors appropriately cite cross-functional transfer and multifidelity prior work), but the scale of the paired dataset and the honest downstream validation give the paper practical value for the journal's audience. The two senior authors are developers of GPTFF and the Atomly database, which the method uses — not a conflict requiring action, but the editor may wish to ensure the requested composition-split benchmark is run with a held-out protocol independent of the authors' own tooling choices. My recommendation hinges on the leakage test in Major Comment 1; if the composition-held-out MAE comes back near the random-split value, I would support acceptance after textual revisions."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful core is simple: take 380k MatPES structures that already have both PBE and r2SCAN formation energies, train GPTFF (and CHGNet) to predict the per-atom residual, and show that PBE + residual gets you to 14.3 meV/atom MAE on a held-out random split, with clearer gains than raw PBE on a handful of binary hulls, two Li-cathode voltage curves, and a few solid-state reactions.\n\nWhat is actually new is the scale of the paired supervision and the decision to reuse mature MLIP backbones as residual heads rather than invent another architecture. The chemical breakdown in Fig. 1c is honest—ionic systems carry most of the bias—and the downstream checks are the right ones for formation energies. Circularity is low; the labels are external DFT pairs.\n\nThe soft spots are real but scoped. First, MatPES is built from relaxation/MD trajectories and near-equilibrium perturbations of far fewer distinct compositions. A pure random 80/10/10 split almost certainly puts near-duplicates of the same composition in train and test, so the 14.3 number is partly memorizing composition-level offsets plus small geometry wiggles the model has already seen. That is prior to the transferability problem the authors themselves over-sell in the Discussion: nothing here tests pseudopotential, cutoff, code, or magnetism mismatches, nor true composition-disjoint generalization to “massive legacy PBE datasets.” Hull recovery is only 6/8 of the disagreeing phases, with residual ordering errors of tens of meV/atom; the LiMn2O4 plateau glitch is consistent with that. No elemental/FERE/linear baseline is reported, and no code or weights ship.\n\nFor people already building multifidelity DFT corpora or battery/phase-diagram pipelines inside the MatPES/PBE–r2SCAN lane, this is worth reading and probably citing for the empirical residual numbers. It is not yet a universal alignment tool. I would send it to referees with a clear request to add composition-disjoint splits, classical baselines, and a tightened claim set. Engage if that is your problem; do not treat the abstract’s generality as demonstrated.","headline":"Solid residual-learning demo on MatPES pairs, but the 14.3 meV/atom headline is soft under random splits and the “upgrade any legacy PBE set” claim is untested.","tokens_in":10939,"tokens_out":560,"would_cite":false,"duration_ms":21630,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A graph network learns PBE-to-r2SCAN formation-energy residuals at 14.3 meV/atom, so large cheap DFT sets can be upgraded toward higher-fidelity thermodynamics.","keywords":["DFT formation energies","PBE","r2SCAN","graph neural networks","cross-functional alignment","phase stability","materials datasets","transfer learning"],"falsifier":"Take a chemically diverse hold-out set computed with a different code or pseudopotential family, apply the frozen MatPES-trained residual model to its PBE energies, and check whether the corrected energies still lie within ~15–20 meV/atom of the corresponding r2SCAN (or experimental) values and preserve convex-hull rankings.","tokens_in":10651,"feed_emoji":"⚛️","tokens_out":905,"duration_ms":16705,"temperature":0.7,"pith_summary":"Different density-functional choices leave systematic offsets of tens to hundreds of meV per atom in formation energies, so multi-source DFT tables cannot be merged without corrupting phase stability, voltages, and reaction energies. This paper shows that those offsets are largely structure-dependent and therefore learnable. Trained on 380,190 paired PBE–r2SCAN structures, a graph neural network predicts the per-atom residual; adding the residual to raw PBE recovers r2SCAN-scale energies with 14.3 meV/atom test error. The corrected energies restore several experimentally or r2SCAN-stable phases that PBE misplaces, improve Li-ion voltage plateaus, and fix product rankings in solid-state reactions. The practical payoff is that the vast existing PBE corpora can be lifted toward higher-fidelity thermodynamics without recomputing every entry at meta-GGA cost, opening a route to larger, more consistent training sets for materials models.","feed_headline":"Graph net lifts PBE energies to r2SCAN at 14.3 meV/atom","feed_subtitle":"Structure-aware residuals let cheap legacy DFT match higher-fidelity thermodynamics for phases, voltages, and reactions","key_machinery":"The cross-functional residual ΔE_f = E_f(r2SCAN) − E_f(PBE) predicted by a graph neural network (GPTFF backbone) from crystal structure alone; adding that residual to the original PBE formation energy produces the aligned energy.","core_discovery":"A structure-aware graph neural network trained solely on per-atom formation-energy differences between paired PBE and r2SCAN calculations can convert ordinary PBE energies into energies that match the r2SCAN scale to 14.3 meV/atom MAE, recovering the principal thermodynamic orderings needed for phase diagrams, battery voltages, and reaction enthalpies.","pith_inferences":["If residual models prove portable across codes, the community could maintain a small library of ‘alignment heads’ rather than re-relaxing entire databases whenever a new functional appears.","The largest remaining errors on ionic compounds suggest that explicit charge or oxidation-state features may further tighten the residual, especially near convex-hull boundaries.","Extending the same supervised residual idea to DFT-versus-experiment pairs would directly attack the theory–measurement gap the discussion flags as future work."],"forward_implications":["Large public PBE databases can be bulk-corrected toward r2SCAN thermodynamics without new high-level DFT for every structure.","Phase-stability screens and Li-voltage estimates that previously used raw PBE become closer to meta-GGA and experiment.","Multi-source DFT tables that once could not be merged because of functional offsets become usable as single training corpora for foundation models.","The same residual-learning pattern can be retrained for any other paired functional or parameter combination that supplies matched structures."],"fun_headline_variants":["GNN aligns PBE to r2SCAN formation energies at 14.3 meV/atom","Structure-aware graph net maps PBE energies onto r2SCAN scale","Paired MatPES residuals train GNN to close 107 meV/atom DFT gap","Graph transfer learning upgrades legacy PBE to r2SCAN accuracy","Cross-functional GNN residual model hits 14.3 meV/atom MAE"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That a residual model trained only on one paired PBE–r2SCAN corpus will transfer, without retraining, to legacy PBE sets and other DFT pairs that differ in codes, cutoffs, pseudopotentials, and magnetism protocols.","fun_headline_variants_meta":{"raw":{"variants":["GNN aligns PBE to r2SCAN formation energies at 14.3 meV/atom","Structure-aware graph net maps PBE energies onto r2SCAN scale","Paired MatPES residuals train GNN to close 107 meV/atom DFT gap","Graph transfer learning upgrades legacy PBE to r2SCAN accuracy","Cross-functional GNN residual model hits 14.3 meV/atom MAE"]},"model":"grok-4.5","effort":"low","cost_usd":0.00396,"raw_usage":{"total_tokens":1243,"prompt_tokens":820,"num_sources_used":0,"completion_tokens":90,"cost_in_usd_ticks":39604000,"prompt_tokens_details":{"text_tokens":820,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":333,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":820,"tokens_out":90,"duration_ms":6947,"temperature":1.0,"reasoning_tokens":333,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T18:01:01.008326+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Take a chemically diverse hold-out set computed with a different code or pseudopotential family, apply the frozen MatPES-trained residual model to its PBE energies, and check whether the corrected energies still lie within ~15–20 meV/atom of the corresponding r2SCAN (or experimental) values and preserve convex-hull rankings.","supporting_citations":[],"review_version":1}