{"id":"e80e1cfc-1d67-44b1-8b3f-37c8d21043e5","arxiv_id":"2610.31183","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A single unconstrained neural network backbone learns several equivariant molecular properties from rotated data, matching or beating equivariant reference models at lower cost.","lead":"This paper shows that an unconstrained graph neural network, trained on rotated copies of molecules, can learn molecular dipoles, polarizabilities, NMR shielding tensors, and electron density coefficients with accuracy comparable to or better than equivariant reference models. It also shows the model is faster and uses less memory, so learned symmetry can replace built-in symmetry for many atomistic property prediction tasks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Load-bearing risk: for SCFBench λ=3,4 the equivariance RMSE stays at half or more of the symmetrized RMSE at every backbone size (Fig. S3), and the 'impact is small' dipole bound rests on a 52-orientation sample that Sec. S6 admits may underestimate the orbit variance.","rationale":"The paper is a systematic and unusually candid empirical study: one PET backbone learns dipoles, polarizabilities, NMR shielding tensors (including the (1,−1) pseudovector block), and auxiliary-basis density coefficients up to λ=4–5, with held-out validation against external DFT references and released reference models, and downstream checks running from IR/Raman spectra and isotropic chemical shifts to one-shot SCF properties and fine-tuning to CCSD densities. The comparison-protocol caveats the reader flags — MACE-MDP timed in float64 without cuEquivariance, NequIP-L compared only against S-fitted targets while PET is also trained on J-fitted ones, untuned SCFBench hyperparameters, post hoc test-set exclusions (Sec. S5, b193), and unconverged FCC-Au PET models — are all disclosed by the authors and mostly bounded, and the unconverged PET runs actually cut in PET's favor. I therefore do not see the accuracy/cost leg as the weakest point. The weakest point is the high-order learnability leg: for the hardest density blocks the equivariance error does not shrink relative to the achievable (symmetrized) error with backbone size, and the observable impact is quantified with a method the authors admit may understate it (Sec. S6, b673). This is load-bearing because the abstract's final sentence ('the impact of approximate equivariance is small for most observables') and the 'high angular order' component of the claim both rest on it. The authors already name the density as the clearest limit (Sec. V, b314) and note the growing relative share (Sec. IV.C), so the concern is about what remains unestablished rather than about a concealed flaw: whether the equivariance/symmetrized ratio for λ=3,4 declines with further scaling, and whether the dipole spread is below model error under an exact quadrature. The proposed check settles the second point directly; a secondary probe would be whether a larger backbone (or extended fine-tuning) moves the λ=3,4 ratio below 0.5. Neither check overturns the demonstrated results, so the reader's CONDITIONAL verdict remains appropriate; the condition should include the quadrature-level evaluation (or release of checkpoints enabling reviewers to run it), alongside the artifacts and tightened protocol the reader already requested.","tokens_in":53415,"tokens_out":18535,"duration_ms":132134,"concrete_test":"Evaluate the one-shot dipole (and total-energy) orientation spread of the best density model (PET-L, J,J, etb-2.0, Sec. IV.C) on the same 50 in-distribution and 50 out-of-distribution molecules as Sec. S6, but over a converged O(3) quadrature of degree 13 — Lebedev axes with all equispaced in-plane rotations and inversion, as used for the spectra in Sec. S3 C — instead of the 52-element unweighted sample with a single in-plane angle. If the fully orbit-averaged dipole spread exceeds the per-molecule one-shot dipole model error on the in-distribution set, where the current estimate already reaches about half of the model error, the abstract's claim that approximate equivariance has a small impact on density-derived observables requires qualification. If the spread stays below the model error, the concern is bounded and the CONDITIONAL verdict stands without revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two load-bearing legs: that learned (approximate) equivariance is good enough at high angular order, and that residual symmetry breaking is benign for observables. The paper's own numbers leave the first leg unproven in its strongest form. For the SCFBench density, the equivariance RMSE of the λ=3 and 4 blocks is half or more of the O(3)-symmetrized RMSE at every backbone size, and from PET-XS to PET-L it falls only about as fast as the symmetrized error (Sec. III.A, Fig. S3; Sec. V), so the relative share does not decrease with scale. No saturation or downturn of that ratio is shown, leaving open a plateau at which augmentation-driven training alone cannot close the gap. The second leg is defended by downstream validation: the one-shot energy spread sits at the integration-grid noise floor, but the dipole — linear in the density, so directly inheriting coefficient equivariance error — shows a median spread reaching about half of the one-shot dipole model error in distribution, and the paper states the relative share grows as models improve (Sec. IV.C, Fig. S8). That quantitative bound itself rests on a 52-element orientation set with a single in-plane angle and no quadrature weights, which Sec. S6 explicitly says 'may underestimate' the true orbit variance (b673). The FCC-Au experiment (Fig. 2) also reports no equivariance errors for λ=5,6, so 'generalizes up to λ=6' demonstrates representability, not learned equivariance at high order. The paper acknowledges the density as 'the clearest limit' (Sec. V, b314) and the growing relative share (Sec. IV.C), so this is not a hidden flaw; what remains unresolved is whether the limit shrinks relative to accuracy with capacity, and whether the impact estimate survives a proper quadrature.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"This paper argues that an unconstrained, non-O(3)-equivariant graph neural network (the Point-Edge Transformer, PET) can learn equivariant atomistic properties — molecular dipoles, polarizabilities, NMR shielding tensors, and auxiliary-basis electron-density coefficients — by learning symmetry from rotation/inversion augmentation rather than from Clebsch-Gordan tensor products. The authors first show that truncated equivariant bases fail on near-octahedral environments while PET fits and generalizes all angular blocks of a model Hamiltonian up to λ=6. They then report systematic improvements in both symmetrized accuracy and equivariance error with backbone size across three datasets, and describe head modifications, linear purification, and ensembling to reduce symmetry breaking. Downstream validation covers IR/Raman spectra, isotropic NMR shieldings, one-shot DFT properties from predicted densities, SCF acceleration, and transfer fine-tuning to different functionals/bases. The paper concludes that PET matches or improves on released equivariant reference models at lower inference cost and memory footprint, while acknowledging that the electron density is the clearest limit: for λ=3 and 4 blocks the equivariance error remains at least half the symmetrized error at all backbone sizes, and the one-shot dipole inherits a non-negligible orientation spread.","tokens_in":53862,"tokens_out":5578,"duration_ms":44788,"significance":"If the central claims hold, the paper provides a substantial demonstration that explicit O(3) symmetry is not a prerequisite for practically useful property prediction, decoupling inference cost from angular order and side-stepping basis-incompleteness problems that affect exactly equivariant models. The study is unusually thorough on the empirical side: it separates symmetrized accuracy from equivariance error, uses held-out external test sets and released third-party models, validates on physically meaningful downstream tasks, and reports inference cost and memory footprint under a consistent protocol. The careful treatment of target scaling, J- versus S-fitted density coefficients, and transfer to CCSD densities are particular strengths. The residual weakness is that the load-bearing claim for high angular order and pseudotensorial blocks is empirical and is only partially supported: the high-order density equivariance error does not shrink relative to model error with scale, and the one-shot dipole bound rests on an orientation sample that the authors themselves state may underestimate the true orbit variance.","major_comments":[{"comment":"For the SCFBench density, the equivariance RMSE of the λ=3 and 4 blocks remains half or more of the O(3)-symmetrized RMSE at every backbone size, and from PET-XS to PET-L it falls only about as fast as the symmetrized error. The paper acknowledges this in Sec. V, but it is load-bearing for the central claim that unconstrained models can learn high-angular-order equivariant targets: the present data show no saturation or downturn of the ratio, leaving open the possibility that augmentation-driven training alone cannot close the gap at any feasible scale. Please either provide evidence that the relative share decreases with further scaling/training, or explicitly restrict the claim to representability plus practically useful but approximate equivariance.","section":"Sec. III.A, Fig. S3, Sec. V","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The bound on the one-shot dipole spread uses 52 orientations, a single in-plane angle, and no quadrature weights; Sec. S6 states that this set \"may underestimate\" the true orbit variance. Since the dipole is linear in the density and directly inherits coefficient equivariance error, the statement in Sec. IV.C that the spread reaches about half the model error \"in distribution\" is not a rigorous quantitative bound. This matters because the abstract and Discussion assert that the impact of approximate equivariance is small for most observables. The authors already have the machinery for weighted O(3) quadrature used elsewhere in the paper; please report a converged weighted orbit variance for the one-shot energy and dipole, or qualify the claim explicitly as a lower-bound estimate.","section":"Sec. IV.C, Sec. S6 (\"Orientation dependence\")","authors_acknowledge":"With a single in-plane angle per direction and no quadrature weights, these are spreads over a sample of orientations, not the weighted orbit variance of Sec. S3 C, and may underestimate it.","authors_acknowledge_p":0.85},{"comment":"The FCC-Au experiment reports relative RMSE per angular block but not equivariance errors, so the statement that \"a small PET model generalizes up to λ=6\" from a single training structure supports representability, not learned equivariance at λ=5,6. Since no equivariance error is reported anywhere for λ=5 and 6, the abstract's phrase \"the same backbone can in principle learn any equivariant property\" should not be read as evidence of learned equivariance at the highest orders. Please either add equivariance-error measurements for the FCC-Au benchmark or rephrase the conclusion to separate expressivity from learned symmetry.","section":"Sec. II.B, Fig. 2, Sec. V","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The caption and text state that MACE-MDP is timed in float64 as released while PET-S is in float32, and the text notes this; the phrase \"Pareto-optimal front\" in Sec. IV.D is still stronger than the controlled evidence, which compares PET with at most one or two external baselines per target. Suggest \"a favorable accuracy/cost front\" or an explicit statement of the precision caveat in the caption.","section":"Fig. 7, Sec. IV.D","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The comparison with NequIP-L mixes the effect of architecture with the effect of J- versus S-fitted targets; the text acknowledges this, but the reader would benefit from a sentence in the SCFBench panel caption stating that no J-fitted NequIP-L model exists, so the gain reflects the target as much as the model.","section":"Sec. IV.D, Fig. 7","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The data availability statement ends with \"(DOI to be confirmed)\"; for a reproducibility-oriented journal, please provide a permanent repository link or a specific expected release date.","section":"Data and code availability","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The abbreviation \"LLPR\" is used heavily in Sec. IV.A and Fig. 4; it is defined in the text, but a one-sentence reminder of what the last-layer prediction rigidity measures would make the figure caption self-contained.","section":"General","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Short version: this is the most complete systematic test yet of PET as a general target model. The same backbone is trained on dipoles and polarizabilities, NMR shielding tensors, and density coefficients. The genuine additions are the per-block heads, density-loss fine-tuning with overlap and Coulomb metrics, the O(3)-character analysis of the last-layer features, the equivariance scaling of ensembles, and the transfer experiments to BFDb and CCSD densities. The MACE-versus-PET single-frame FCC-Au experiment cleanly isolates basis incompleteness from model capacity. The downstream validation is real work: IR/Raman spectra on 7184 molecules, isotropic shieldings against a released ShiftML4 ensemble, and one-shot SCF properties. The accuracy/cost and memory claims are supported, within the caveats below.\n\nWhere I part company with the strongest reading of the abstract: the claim that learned equivariance is good enough at high angular order is not established. In the paper's own numbers, the SCFBench lambda=3 and 4 equivariance RMSE stays at half or more of the symmetrized RMSE at every backbone size, and the relative share does not shrink with scale. The authors call this 'the clearest limit', so it is not hidden; they also do not show a plateau or a path past it. The dipole impact estimate over 52 orientations is explicitly a sample rather than a quadrature, and the SI says it may underestimate the true orbit variance. For the most accurate density models the dipole spread reaches about half the one-shot dipole model error in distribution. That is a load-bearing caveat, not a footnote.\n\nThe benchmark comparisons have loose ends. MACE-MDP runs in float64 against PET in float32; ShiftML4's single global scaling convention de-weights 1H; and there is no J-fitted NequIP-L, so the density comparison reflects the better-posed target as much as the architecture. The FCC-Au PET models are not converged after 50k steps, so Figure 2 demonstrates representability more than final error. The exclusions of 32 ShiftML test structures are post hoc but documented. Code, data, and checkpoints are promised, not yet delivered. None of this sinks the paper; it conditions how the comparisons should be read.\n\nWho gets value: people choosing backbones for property models, and equivariant-GNN researchers who believe truncation and parity costs are unavoidable. The dataset-agnostic backbone heuristic alone is useful. I would send it to peer review, and I would require the artifacts, a quadrature-based dipole estimate, and at least one float64 PET comparison point before accepting. My own verdict is conditional rather than accept.","headline":"A thorough, honest case that unconstrained GNNs can learn practical equivariant targets, with one genuine unresolved problem at high angular order.","tokens_in":54500,"tokens_out":2360,"duration_ms":22066},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"An unconstrained transformer GNN with no explicit rotation algebra learns molecular dipoles, polarizabilities, NMR shielding tensors, and electron-density expansions, matching or improving on exactly equivariant models at lower inference…","keywords":["equivariant atomistic properties","Point-Edge Transformer","learned O(3) equivariance","graph neural networks","electron density prediction","NMR chemical shielding","molecular dipoles and polarizabilities","rotational data augmentation"],"falsifier":"Take a target of angular order $\\lambda=3$ or $\\lambda=4$ (for instance electron-density coefficients on an atom-centered basis) and evaluate a PET model on structures whose local environments are highly symmetric — an octahedral or tetrahedral coordination site, or a symmetric molecule such as $\\mathrm{SF}_6$ — that were absent from training, measuring the equivariance RMSE over a dense $O(3)$ quadrature that includes inversions. If that residual error stays comparable to, rather than far below, the model's physical error on the same structures while the backbone and the augmentation are scaled up, the claim that data augmentation suffices to learn the symmetry breaks.","tokens_in":54856,"feed_emoji":"⚛️","tokens_out":17359,"duration_ms":149752,"temperature":0.7,"pith_summary":"This paper sets out to show that a graph neural network with no rotation symmetry written into its architecture can learn molecular properties that must transform correctly under rotations and inversions, and that it can do so as accurately as, and more cheaply than, networks that enforce the symmetry by construction. The test cases span molecular dipoles and polarizabilities, NMR chemical-shielding tensors (whose antisymmetric block is a pseudovector), and atom-centered expansions of the electron density up to angular order $\\lambda=5$. Against exactly equivariant models, the paper argues, such targets carry a hidden cost: Clebsch–Gordan product bases truncated at low angular order become incomplete or ill-conditioned for near-symmetric environments, and their cost climbs steeply with angular order. The authors report that the unconstrained Point-Edge Transformer learns all four target classes from rotationally augmented data, that larger backbones improve accuracy and equivariance error together, and that for the downstream observables they test the residual symmetry breaking is not what limits accuracy.","feed_headline":"Learned symmetry matches built-in symmetry at lower cost","feed_subtitle":"No rotation algebra: one GNN learns dipoles, NMR tensors and densities as accurately as equivariant rivals.","key_machinery":"The load-bearing mechanism is the replacement of explicit $O(3)$ algebra with learned symmetry in the Point-Edge Transformer (PET): the backbone computes edge-to-edge attention among the edges of each atomic neighborhood, with no Clebsch–Gordan couplings and no angular-order cutoff, so the cost and expressivity do not depend on the target's irreps. Around that backbone sit three further pieces of machinery. First, per-block heads: a separate nonlinear head and linear readout for each $O(3)$ irrep block $(\\lambda,\\sigma)$ of the target, letting each last-layer feature specialize to one angular character (pseudotensorial characters develop latest and benefit most). Second, random rotation-and-inversion augmentation of every structure in each batch, which is what teaches the symmetry; the paper analyzes the $O(3)$-character content of the last-layer features to show that the linear readout filters the character it needs from a mixed spectrum. Third, for the electron density the learning problem is re-posed in the metric of the target field: losses on auxiliary-basis coefficients weighted by the overlap or Coulomb metric, and reference coefficients produced by S- or J-fitting, which together account for most of the density-model gains. The motivation is pinned down in the rattled FCC-Au benchmark: a truncated product basis fails to fit or generalize Hamiltonian blocks with $\\lambda$ above its reach, while a matched-size PET fits all blocks up to $\\lambda=6$ from a single training frame.","core_discovery":"The paper's central claim is that equivariance for atomistic properties can be learned rather than engineered. The Point-Edge Transformer (PET) replaces the Clebsch–Gordan tensor products of equivariant GNNs with edge-to-edge attention, so its cost is independent of the target's angular order and no irrep block is excluded by truncation; the symmetry is acquired from random rotation-and-inversion augmentation during training. On that backbone the authors demonstrate dipoles and polarizabilities from SPICE-α, NMR shielding tensors including the $(1,-1)$ pseudovector block from ShiftML-molecules, and electron-density coefficients on auxiliary atom-centered bases up to $\\lambda=4$ ($\\lambda=5$ in transfer) from SCFBench. They show that scaling the backbone under simple heuristics improves symmetrized accuracy and equivariance error together, that per-block heads and deep ensembling (equivariance error dropping as $N^{-1/2}$) reduce symmetry breaking further, and that PET matches or improves on the released MACE-MDP, ShiftML4, and NequIP-L references at lower inference cost and larger batch capacity on a GPU. They also record the clearest limit of the approach: for the $\\lambda=3$ and $\\lambda=4$ density blocks the equivariance error falls with backbone size only about as fast as the physical error and remains at half or more of it, propagating into one-shot observables such as the dipole.","pith_inferences":["If learned equivariance keeps improving with size and data, the angular-order ceiling that shapes equivariant GNN design stops being a fundamental constraint, and capacity can be spent on width, coverage, and training protocol instead of higher-$\\lambda$ tensor-product channels — a shift the paper's roadmap implies but does not fully state.","For the $\\lambda=3/4$ density blocks, where equivariance error persists at half the symmetrized error, exact $O(3)$-quadrature averaging of predictions (cited by the paper but not used in production) would remove the residual breaking at a known multiplicative cost, and whether that is cheaper than ensembling is a testable question the paper leaves open.","The same per-block-head and metric-loss recipe transfers, by construction, to any atom-centered scalar field — spin densities, electrostatic potentials, local densities of states — and to electric-field-gradient tensors, which are pure $(2,+1)$ targets; the paper names these successors but does not run them.","Because equivariance error behaves like an out-of-distribution detector, it could be calibrated against the paper's LLPR (last-layer prediction rigidity) uncertainty flag on the spectral benchmark to give one cheap reliability score for unconstrained atomistic models — a diagnostic the paper gestures at without formalizing."],"forward_implications":["Angular-order truncation is not a safe shortcut for equivariant architectures: near-octahedral environments defeat any product basis cut below the target's $\\lambda$, while a matched-size unconstrained model fits and generalizes every block up to $\\lambda=6$ from a single training frame.","A single backbone, sized by width and read out per irrep block, spans scalars, vectors, rank-2 tensors with a pseudovector part, and $\\lambda\\le5$ density expansions, so the $L_{\\max}$ and correlation-order hyperparameters of equivariant design drop out of model selection.","Downstream observables inherit little harm from approximate symmetry: IR and Raman spectra reach mean match scores of 0.994–0.997 at DFT normal modes, a single PET-L model matches the 8-member ShiftML4 ensemble on isotropic shieldings, and predicted densities save about 40% of SCF cycles.","For the electron density, the reference data and loss metric matter more than the architecture: J-fitted coefficients, a Coulomb-metric loss, and a larger auxiliary basis cut median one-shot energy errors by more than an order of magnitude with the backbone untouched, and ten CCSD geometries suffice to fine-tune a PBE-pretrained model below B3LYP accuracy in the Coulomb metric.","Residual equivariance error correlates with model error, so the orientation spread of a prediction flags under-trained environments, and ensembling reduces it systematically as $N^{-1/2}$."],"supporting_citations":[{"why":"Supplies the O(3)-character analysis of last-layer features, the equivariance-error measure, the finding that pseudotensorial characters develop later and remain weaker, and the linear-purification procedure reused throughout.","marker":"[35]"},{"why":"Introduces the PET architecture and the idea of learning symmetries from data through rotational augmentation, which this paper extends to general equivariant targets.","marker":"[46]"},{"why":"The equivariant MACE architecture used as the baseline in the FCC-Au benchmark that exposes product-basis incompleteness.","marker":"[4]"},{"why":"Provides the rattled FCC-Au on-site Hamiltonian dataset whose angular blocks up to $\\lambda=6$ form the single-frame generalization benchmark.","marker":"[39]"},{"why":"Supplies the SPICE-α dataset, the MACE-MDP reference model, and the IR-R-7193 spectral benchmark used for dipoles and polarizabilities.","marker":"[20]"},{"why":"Provides the ShiftML-molecules dataset and the ShiftML4 architecture used as the external reference for NMR chemical shieldings.","marker":"[59]"},{"why":"Provides the SCFBench density dataset, the NequIP-L reference model, and the SCF-acceleration evaluation pipeline.","marker":"[30]"},{"why":"Defines the BFDb-SSI benchmark and SA-GPR baselines against which transfer of density models to a new functional and basis is tested.","marker":"[105]"},{"why":"Reports that ensembling suppresses the rotational fluctuations of a single unconstrained shielding model, the basis of the deep- and partial-ensemble strategy.","marker":"[24]"}],"fun_headline_variants":["Equivariance learned from data, not built in — with lower cost","PET learns equivariance from data, matches O(3) rivals","Skip Clebsch-Gordan: attention learns all equivariant targets","One GNN, no O(3) algebra: as accurate, cheaper"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The approach stands or falls on one empirical premise: that random rotation-and-inversion augmentation during training teaches the network near-exact target equivariance on every structure it will face; the paper itself reports that the $\\lambda=3$ and $\\lambda=4$ electron-density blocks still carry equivariance error at half or more of their physical error even at the largest backbone, so the premise is not fully satisfied there.","fun_headline_variants_meta":{"raw":{"variants":["Equivariance learned from data, not built in — with lower cost","PET learns equivariance from data, matches O(3) rivals","Skip Clebsch-Gordan: attention learns all equivariant targets","One GNN, no O(3) algebra: as accurate, cheaper"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000608,"raw_usage":{"cost":0.00060816,"is_byok":false,"cost_details":{"upstream_inference_cost":0.00060816,"upstream_inference_prompt_cost":0.00016016,"upstream_inference_completions_cost":0.000448},"total_tokens":2744,"prompt_tokens":1144,"completion_tokens":1600,"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":1521}},"tokens_in":1144,"tokens_out":1600,"cost_source":"openrouter_usage","duration_ms":8823,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":1521,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-23T10:31:40.764927+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":"Take a target of angular order $\\lambda=3$ or $\\lambda=4$ (for instance electron-density coefficients on an atom-centered basis) and evaluate a PET model on structures whose local environments are highly symmetric — an octahedral or tetrahedral coordination site, or a symmetric molecule such as $\\mathrm{SF}_6$ — that were absent from training, measuring the equivariance RMSE over a dense $O(3)$ quadrature that includes inversions. If that residual error stays comparable to, rather than far below, the model's physical error on the same structures while the backbone and the augmentation are scaled up, the claim that data augmentation suffices to learn the symmetry breaks.","supporting_citations":[{"cited_title":"Impact of quantum-chemical metrics on the machine learning prediction of electron density, The Journal of Chemical Physics155, 024107","cited_arxiv_id":null,"evidence_quote":"Defines the BFDb-SSI benchmark and SA-GPR baselines against which transfer of density models to a new functional and basis is tested."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}