{"id":"76d6fc77-69e1-4e8c-a14a-bdb485be86ae","arxiv_id":"2505.09174","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Crystal structures encoded as quotient complexes with triangle features give improved perovskite bandgap prediction when processed by a simplex-based transformer.","lead":"A new machine learning model, QCformer, represents crystal structures as quotient complexes, with edges for atomic bonds and triangles for three-atom interactions, and uses a transformer to predict material properties. It reports better bandgap predictions for perovskite solar-cell materials than existing graph networks, and the authors apply it to screen new 2D perovskites.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Triangle features improve MAE mainly on the small-data perovskite tasks; the five-compound DFT check does not resolve whether the QC triangles carry physical three-body information, because the new-compound MAE (0.385 eV) is driven by one outlier and is comparable to DFT functional scatter.","rationale":"The reader's weakest_assumption is exactly the load-bearing issue: degenerate self-loop triangles in the quotient complex are treated as valid higher-order features without physical validation, and no sensitivity analysis separates them from genuine three-body interactions. I agree with the reader's diagnosis and with the CONDITIONAL verdict. The paper is not internally inconsistent: the mathematics of the homotopy-equivalent construction is coherent, the reported results are plausible, and the method is potentially reproducible from the stated architecture and linked code. The concern is evidential: the central novelty claim rests on triangle features, but (a) the ablation is missing on the perovskite datasets, (b) the self-loop triangles are known to exist and are unvalidated, and (c) the external check is too small and too noisy to rescue the interpretation. The proposed concrete test would settle whether the improvement is due to genuine three-body features or to quotient-complex construction artifacts. If the test passes, ACCEPT would be justified; until then CONDITIONAL is the right verdict. As a secondary note, Theorem S1.4 cites the authors' prior work (ref 74), so the mathematical support is not fully independent, though this does not change the verdict.","tokens_in":22545,"tokens_out":6968,"duration_ms":60905,"concrete_test":"Run a controlled experiment on HOIP and HOIP2D: (1) repair the quotient complex by constructing the homotopy-equivalent simplicial complex eK of Lemma 4.3 as the actual model input, replacing each self-loop triangle with the two new edges; (2) alternatively, remove all 2-simplices that contain a self-loop edge, retaining only genuine triangles with three distinct atoms; (3) also run 1D-QCformer on HOIP and HOIP2D. Use identical hyperparameters and at least five random seeds, reporting mean and standard deviation of MAE for Tables 3 and 4. If the QCformer gains over 1D-QCformer and MEGNet persist without self-loop triangles, the physical-interpretation claim survives; if the gains vanish or change sign, the headline claim is falsified.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that quotient-complex triangles capture chemically meaningful three-body interactions and drive QCformer's state-of-the-art HOIP bandgap prediction. The paper's own construction undercuts this attribution. Supplementary S2.1 and Figure S4 state that after gluing periodic copies, the two 2-simplices of ZnS consist of two normal edges and one self-loop edge. A 2-simplex with a self-loop is not a simplex; after the quotient operation the clique complex is no longer a simplicial complex. The paper repairs this topologically by constructing a homotopy-equivalent simplicial complex eK (Lemma 4.3 and Theorem 4.4), but the model is not described as operating on eK; it operates on the glued object, and the triangle features include products and squares of the self-loop edge length. Second, the triangle ablation (Tables 6-7) is only run on Materials Project and JARVIS, not on HOIP and HOIP2D, so the 30.59% HOIP improvement over MEGNet is attributed to triangles without direct support. Third, the five-compound external validation (Table 5) has MAE 0.3849 eV against GGA-PBE, comparable to the 0.4297 eV GGA-PBE/r2SCAN spread; compound 452 alone contributes 0.6818 eV error. No error bars, repeated seeds, or significance tests are reported, so the 2-11% ablation deltas and the HOIP gains could be training noise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Quotient Complex Transformer (QCformer), a topological deep learning model for predicting crystal and perovskite properties. A crystal is represented by taking the 12-nearest-neighbor graph of atoms in a unit cell, forming its clique complex, and gluing periodically equivalent atoms to obtain a quotient complex; vertex, edge, and triangle features are updated by a simplex-level Transformer with attention-based message passing. The model is pretrained on JARVIS and evaluated on Materials Project and JARVIS benchmarks, then fine-tuned on HOIP and HOIP2D bandgap datasets, where it reports state-of-the-art MAE values (0.0969 eV on HOIP, 0.0754 eV on HOIP2D), with ablations attributing part of the gain to triangle features. A theoretical section states that the quotient complex is homotopy equivalent to a true simplicial complex and describes its homology relative to the original crystal complex, but the proof is delegated to a prior paper by the same authors.","tokens_in":22853,"tokens_out":5306,"duration_ms":49366,"significance":"If the method's representational assumptions hold, this is a meaningful contribution: it is among the first applications of topological deep learning to materials property prediction, it combines periodic quotient structure with higher-order simplex features, and it reports consistent improvements over strong baselines on HOIP and HOIP2D benchmarks, with code and data availability stated. The reported benchmark numbers do support the central empirical claim that QCformer is competitive or better than ComFormer and Matformer on the tested inorganic tasks and substantially better on HOIP bandgap prediction. However, the significance is conditional on resolving whether the higher-order 'triangles' used by the model are well-defined geometric objects and whether the performance gain is reproducible beyond single runs; these points are load-bearing for the core novelty claim. No machine-checked proofs are supplied, and the central homology theorem is cited from the authors' prior work rather than proved in this manuscript.","major_comments":[{"comment":"Supplementary Figure S4's caption states that after gluing periodic copies, 'the two 2-simplices consist of two normal edges and one self-loop edge.' A 2-simplex whose boundary contains a self-loop is not a simplex, so the quotient object used as the model input is not a simplicial complex; the paper's own repair (Lemma 4.3 and Theorem 4.4) constructs a homotopy-equivalent simplicial complex eK, but the model is not described as operating on eK. Since the triangle features defined in Supplementary S2.2 include products and squares of edge lengths, and self-loop edge lengths appear in those combinations, the claimed chemical meaning of the three-body features is not established. Please specify whether triangles with self-loop edges are kept, replace the model input by eK (or define the quotient complex as a cell complex), and add a sensitivity or physical validation that separates genuine three-body information from artifacts of degenerate triangles.","section":"§2.1.1 and Supplementary S2.1/S2.2 (Figs. S3–S4)"},{"comment":"The triangle ablation (1D-QCformer vs QCformer) is only run on Materials Project and JARVIS, not on HOIP or HOIP2D. Consequently, the headline result of a 30.59% MAE improvement over MEGNet on the HOIP dataset is attributed in the text to higher-order triangle interactions without direct experimental support. Please run the ablation on the perovskite datasets, or explicitly state and justify that the gain transfers.","section":"§4.3.2 Tables 6–7; §2.3.1 Table 3"},{"comment":"The five-compound external validation does not resolve whether the model's predictions carry physical three-body information. The reported MAE of 0.3849 eV against GGA-PBE is comparable to the 0.4297 eV spread between GGA-PBE and r2SCAN, compound 452 alone contributes an absolute error of 0.6818 eV, and the text even notes that QCformer's prediction for compound 110 deviates strongly from both DFT functionals. With n=5 and no error bars, this experiment cannot discriminate the quotient-complex representation from a nontopological baseline.","section":"§2.3.2 and Supplementary Table S2"},{"comment":"Theorem 4.5 is stated in the main text as a theorem justifying that the quotient complex captures periodic patterns, but its proof is not included in this manuscript; Supplementary Section S1.4 presents it as 'Theorem S1.4 (74)', where reference 74 is the authors' own prior paper. Please include a self-contained proof or a precise statement of the provenance, because this theorem is the theoretical basis for the claim that the quotient operation preserves the relevant topological information.","section":"§4.2, Theorem 4.5; Supplementary S1.4"}],"minor_comments":[{"comment":"The abstract says QCformer is pretrained on the Materials Project and JARVIS, but §2.3.1 states that pretraining for the perovskite tasks uses only the JARVIS Bandgap(OPT) task; please harmonize the wording.","section":"Abstract; §2.3.1"},{"comment":"In the attention coefficient formula, the query vector appears as [q_sigma, q_sigma], which looks like a typo for a concatenation involving q_sigma and a key- or message-related vector; please check.","section":"§2.1.2, Eq. (3)"},{"comment":"The text refers to 'Figure S3.3' while the figure is captioned 'Figure S5'; please fix the cross-reference.","section":"Supplementary S3.3"},{"comment":"All results are reported without run-to-run variance; adding standard deviations over multiple seeds or folds would help the reader assess whether small deltas, such as 0.191 vs 0.193 eV on the MP bandgap task, are meaningful.","section":"Tables 1–2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central theoretical result, Theorem 4.5 / Theorem S1.4, is cited from reference 74, which is the authors' own prior work, and is not proved here; this should be addressed editorially. In addition, the core novelty claim depends on the validity of the degenerate quotient triangles, which is currently unresolved; the empirical benchmark numbers in Tables 1–4 are otherwise consistent with the reported central claim, but the missing ablations on perovskite datasets and the lack of error bars make the attribution to higher-order interactions weaker than the text suggests."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: QCformer is a genuine architectural contribution—a simplex Transformer running on quotient complexes of crystals, with triangle features standing in for three-body interactions—and it produces strong bandgap numbers on perovskite benchmarks. But the story about why the triangles help is weaker than the paper suggests, and the statistics are thin enough that several headline claims could be noise.\n\nNew and good: the quotient complex lifts a familiar idea (quotient graphs, used in MEGNet/ALIGNN/Matformer) to include 2-simplices, and the model updates node, edge, and triangle features with attention over neighbors and cofaces, using more layers at the node level then letting triangle updates flow down. JARVIS pretraining with transfer to HOIP is sensible and appears to pay off. Code is public. The authors are also straight about some limitations—they admit the HOIP2D MAE trails GDA-GBT and that the five-compound DFT check has functional-scatter problems.\n\nSoft spots in order of size. First, no variance: every benchmark number is a single run, and on MP bandgap the edge over ComFormer is 0.191 vs 0.193 eV. The 2–11% ablation gains are exactly the kind of deltas that need repeated seeds. Second, the triangle ablation is only on MP and JARVIS, not on HOIP/HOIP2D, so the headline 30.59% HOIP improvement over MEGNet is attributed to triangles without a direct test. Third, the construction itself: Supplementary Figure S4 says the two 2-simplices in the ZnS quotient complex 'consist of two normal edges and one self-loop edge.' A 2-simplex with a self-loop is not a simplex; the homotopy-equivalent repair (eK, Lemma 4.3–Theorem 4.4) is proven, but the model consumes the glued object, including triangle features built from the self-loop length. That does not necessarily invalidate the empirical gains—it may still be a useful inductive bias—but the paper's 'higher-order interaction' language overclaims. Fourth, the five-compound validation is one outlier (452, error 0.68 eV) away from being a different story, and its 0.385 eV MAE is comparable to the 0.430 eV spread between GGA-PBE and r2SCAN; it doesn't settle the physical-meaningfulness question. Fifth, HOIP baselines are SchNet and MEGNet only—no Matformer or ComFormer—and pretraining is never ablated, so the source of the HOIP gain is underdetermined.\n\nWorth noting: Theorem 4.5 is cited as Theorem S1.4 of the authors' own JCIM paper. That is an honest citation, but it means part of the 'mathematics' here is prior work, and the main text could say so more clearly.\n\nBottom line: the empirical claims are plausible, the representation is reusable, and the paper deserves a real referee. My recommendation: send it to review with requests for repeated-seed runs and significance tests, the triangle ablation on the perovskite datasets, a clarification of what exactly the model sees from the glued complex versus eK, and stronger HOIP baselines.","headline":"A novel and promising quotient-complex transformer for crystal bandgap prediction, with the triangle-feature story undermined by degenerate self-loop simplices and a lack of error bars.","tokens_in":23416,"tokens_out":5666,"would_cite":true,"duration_ms":54378,"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":"A crystal becomes a quotient complex—edges for bonds, triangles for three-atom interactions—and a simplex Transformer over it predicts perovskite bandgaps more accurately than graph networks.","keywords":["quotient complex","simplicial complex","higher-order interactions","perovskite bandgap prediction","topological deep learning","geometric deep learning","material property prediction","transfer learning"],"falsifier":"Retrain QCformer on the HOIP benchmark twice: once with all 2-simplices that contain a self-loop edge removed, and once with triangle features replaced by features of randomly selected triples of edges of the same size and dimension. If removing self-loop triangles leaves MAE essentially unchanged, or if random triples reproduce the 0.0969 eV MAE, then the claimed three-body encoding is not what drives the improvement.","tokens_in":1793,"feed_emoji":"🔺","tokens_out":2204,"duration_ms":96594,"temperature":0.7,"pith_summary":"QCformer models a crystal as a quotient complex: periodically equivalent atoms are glued together into one finite cell, pairwise bonds become edges, and three-atom interactions become triangles. The paper's central claim is that this representation, processed by a Transformer that passes messages from triangles to edges to vertices, captures periodic structure and higher-order interactions that ordinary graph neural networks miss. On the HOIP perovskite bandgap benchmark it reports a 30.59% reduction in MAE over MEGNet (0.0969 eV versus 0.1396 eV), with a coefficient of determination of 0.9829 and a Pearson correlation coefficient of 0.9916, and on the HOIP2D benchmark it lowers MAE by about 9.2% relative to the best previous GNN while improving RMSE. An ablation that removes triangle features, leaving only the quotient graph, degrades performance on every task tested, which the paper presents as evidence that the three-body features are what carry the improvement. If correct, the quotient complex gives a general way to embed periodicity and many-body physics into deep-learning models for materials.","feed_headline":"Triangles in crystal graphs cut perovskite bandgap error 30.6%","feed_subtitle":"Encoding three-atom interactions as triangle features lets a transformer beat graph networks on HOIP perovskite bandgaps.","key_machinery":"The load-bearing object is the quotient complex. For a given crystal, the paper builds a directed 12-nearest-neighbor graph on the unit cell's atoms, forms its clique complex so that mutually connected triples become 2-simplices (triangles), and then identifies all atoms related by lattice translations to obtain a finite quotient complex with multi-edges, self-loops, and triangles. Triangle features are built from the three edge lengths, their pairwise products, and their squares, expanded through radial basis functions. The Simplex Transformer (Sformer) block updates an n-simplex by attending from its neighboring n-simplices through shared (n+1)-simplex cofaces; QCformer runs five edge-to-vertex layers followed by two triangle-to-edge-to-vertex layers. The mathematical analysis proves that the quotient complex is homotopy equivalent to an auxiliary simplicial complex K~ obtained by adding star-shaped sets, so the homology groups are preserved; this is how the paper argues the periodic topology of the crystal survives the quotient.","core_discovery":"The paper's core claim is that representing a periodic crystal as a quotient complex rather than a plain graph improves material property prediction. Starting from a 12-nearest-neighbor graph around each atom in the unit cell, the construction takes its clique complex and glues together atoms related by lattice translations, yielding vertices, multi-edges, self-loops, and triangles; triangles are assigned features from the three edge lengths and their products and squares. A Simplex Transformer then updates simplex embeddings through attention over cofaces, with the hierarchy flowing from triangles to edges to vertices. On the HOIP dataset the model reaches MAE 0.0969 eV, a 30.59% improvement over MEGNet's 0.1396 eV, with COD 0.9829 and PCC 0.9916; on HOIP2D it reaches MAE 0.0754 eV and RMSE 0.1038 eV, lower than the strongest prior GNN SIGNNA_c and competitive with descriptor-based gradient-boosting models. On the Materials Project and JARVIS benchmarks it posts the best or tied-best MAE on most tasks, including a 13.16% improvement over the second-best model on Bulk Moduli and 7.38% and 7.69% improvements on the two JARVIS bandgap tasks, and removing triangle features in the 1D-QCformer ablation hurts every task.","pith_inferences":["The paper does not test whether triangles containing a self-loop edge after the quotient gluing are physically meaningful; deleting exactly those degenerate 2-simplices and re-running the HOIP experiment would isolate whether the 30.59% gain comes from three-body chemistry or from extra message-passing capacity.","If the gain survives randomly permuted triangle features, then the advantage may be architectural rather than physical; the same parameter budget spent on extra edge-to-vertex attention layers could be compared against QCformer to separate the two explanations.","The construction's 12-nearest-neighbor cutoff is a free parameter, and its sensitivity was not reported; a scan over k would test whether the triangle benefits are robust or tied to this specific graph density."],"forward_implications":["Pretrained on large inorganic datasets (Materials Project and JARVIS) and fine-tuned on small perovskite datasets, the model transfers well enough to give HOIP bandgap MAE under 0.1 eV from only 1346 data points.","Because the graph-only ablation is worse on all nine benchmark tasks, the triangle-to-edge message flow, not just the Transformer backbone, is the source of the gain; adding tetrahedral 3-simplices is the natural next extension.","On five unlabeled 2D perovskites, QCformer's predicted bandgaps fall between GGA-PBE and r2SCAN DFT values in most cases, suggesting it can serve as a cheap screening estimator when experimental labels do not exist.","The same quotient-complex representation should apply to any periodic solid, not just perovskites, wherever many-body interactions matter."],"supporting_citations":[{"why":"Supplies the periodic-graph Transformer architecture, attention scheme, and train/validation/test splits on the Materials Project and JARVIS that QCformer builds on and compares against.","marker":"(24)"},{"why":"ComFormer is the strongest prior GNN baseline on the Materials Project and JARVIS benchmarks; QCformer must beat it.","marker":"(25)"},{"why":"MEGNet is the quotient-graph GNN baseline whose HOIP MAE of 0.1396 eV defines the 30.59% improvement claimed for QCformer.","marker":"(15)"},{"why":"Provides the topological-descriptor gradient-boosted-tree baselines (TPD, persistent Ricci curvature, persistent homology) on HOIP that the paper compares against.","marker":"(67)"},{"why":"Supplies the HOIP dataset with HSE06 bandgaps used in the main downstream evaluation.","marker":"(65)"},{"why":"Supplies the HOIP2D database (NMSE) of two-dimensional perovskite structures and DFT bandgaps.","marker":"(66)"},{"why":"Provides the SIGNNA and SIGNNA_c GNN baselines on HOIP2D that incorporate organic/inorganic substructures and are the strongest prior GNNs there.","marker":"(26)"},{"why":"Provides the 92-dimensional vertex features adopted for 0-simplices and serves as a CGCNN baseline in the benchmarks.","marker":"(18)"}],"fun_headline_variants":["QCformer's triangles cut perovskite bandgap error 30.6%","Triangle-laden quotient complex beats GNNs on perovskites","Perovskite bandgap error drops 30.6% with QCformer","From crystal triangles to better perovskite predictions: QCformer","Quotient complex transformer outperforms graph nets on HOIP properties"],"cache_read_input_tokens":25472,"weakest_assumption_plain":"The method's advantage rests on the assumption that the triangles formed by gluing periodic copies of the 12-nearest-neighbor graph encode real three-atom chemical interactions, even when a triangle contains a self-loop edge rather than three ordinary bonds.","fun_headline_variants_meta":{"raw":{"variants":["QCformer's triangles cut perovskite bandgap error 30.6%","Triangle-laden quotient complex beats GNNs on perovskites","Perovskite bandgap error drops 30.6% with QCformer","From crystal triangles to better perovskite predictions: QCformer","Quotient complex transformer outperforms graph nets on HOIP properties"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000289,"raw_usage":{"total_tokens":1752,"prompt_tokens":1060,"completion_tokens":692,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":602}},"tokens_in":676,"tokens_out":692,"duration_ms":6960,"temperature":1.0,"reasoning_tokens":602,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:37:55.601471+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain QCformer on the HOIP benchmark twice: once with all 2-simplices that contain a self-loop edge removed, and once with triangle features replaced by features of randomly selected triples of edges of the same size and dimension. If removing self-loop triangles leaves MAE essentially unchanged, or if random triples reproduce the 0.0969 eV MAE, then the claimed three-body encoding is not what drives the improvement.","supporting_citations":[],"review_version":1}