acceptodds
Under review as a conference paper at ICLR 2027

Orbital Specific Exchange: advancing hybrid density functional theory with geometric deep learning

Abstract

One of the main challenges of hybrid Density Functional Theory (DFT) is selection of the exact-exchange fraction (). While state-of-the-art methods adapt per molecule, they remain limited to a global scalar, ignoring that distinct electronic environments require distinct physical treatments, i.e., the optimal would be different for various parts of the same molecule. We propose a generalized formalism that treats exchange mixing not as a scalar hyperparameter, but as a learnable orbital-specific interaction map. We illustrate it with OSE26 — an equivariant neural network predicting a dense mixing matrix that modulates the Hartree–Fock exchange operator. This exchange-functional implementation provides self-consistency and roto-translational energy invariance. Trained and validated on MSR-ACC-TAE and RDB7, OSE26 achieves near-chemical accuracy for reaction-barrier prediction, attaining a 1 kcal/mol MAE on RDB7 and reducing MAE by 11–56% relative to the considered baselines. This move from system-wide scalars to learnable orbital-level matrices promises a more broadly applicable hybrid-functional paradigm.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
density functional theory, hybrid density functionals, machine learning for quantum chemistry, geometric deep learning, E(3)-equivariant neural networks, exchange--correlation functionals, reaction-energy prediction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_z2SqOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

OSE26 uses MACE and Graph2Mat to predict a symmetric, geometry-dependent AO-space matrix AA. This matrix enters the exchange energy through −12∑σTr⁡[PσAK(Pσ)]-\frac12\sum_\sigma\operatorname{Tr}[P^\sigma A K(P^\sigma)], recovering the usual scaled HF exchange when A=αIA=\alpha I. Training uses reaction and atomization energies with precomputed B3LYP/def2-QZVP density and exchange matrices. The reported test MAEs are 0.85 kcal/mol on MSR-ACC and 1.23 kcal/mol on RDB7. Appendices include scalar, delta-learning, and diagonal-matrix baselines, a 30-reaction self-consistent check, equilibrium bond-length results, and water-cluster evaluations.

Critical strengths

The learned quantity has a useful physical structure. Predicting an exchange-modulation matrix provides a concrete connection between geometric learning and electronic-structure calculations. The scalar limit is explicit, and Appendix A derives the density derivative of the exchange energy for use in SCF. The potential contribution to machine learning is this structured parameterization and its physical readout.

Critical weaknesses
  1. The evaluation protocol is insufficiently specified. Table 1 gives dataset counts without explaining subset selection, split construction, or the validation set used for model selection. Section 4.3 alternates between calling the evaluated data test and validation sets. For RDB7, it is important to report how forward/reverse barriers, shared transition states, and related reactions are distributed across splits. For MSR, structure identifiers and duplicate checks are needed. The settings and provenance of the conventional-functional baselines are also unclear. Table 2 reports the spread of test errors, not uncertainty across training runs; no repeated-run or paired uncertainty is given for the 0.15 kcal/mol RDB7 advantage over MACE delta-learning.

  2. The ablations do not fully isolate the claimed machine-learning advantage. MACE and Graph2Mat are established components [1]. Both OSE and MACE delta-learning correct B3LYP energies. The OSE correction explicitly contracts with full precomputed density and exchange matrices, whereas the delta head predicts a scalar correction from geometry. Sharing an encoder does not match the information available at the energy readout, and the statement that the heads have similar sizes is not supported by parameter counts. The scalar baseline also varies a conventional hybrid mixing fraction, while OSE keeps the semilocal contribution fixed. A scalar restriction A=a(R)IA=a(R)I within the same energy expression would isolate the effect of matrix mixing more directly. The diagonal-only ablation needs a precise symmetry definition: elementwise diagonal AO matrices are generally not closed under rotations of p or higher-angular-momentum shells. The discussion should also position the contribution relative to equivariant learned functionals such as EG-XC [2].

  3. The main benchmark does not establish the reported accuracy of the self-consistent functional. Training and the principal test results use frozen B3LYP orbitals. In this setting, the learned term is a structured energy correction contracted with fixed electronic quantities. The 30-reaction SCF check is useful, but it primarily reports agreement with post-SCF energies. It does not provide a clear comparison of post-SCF and self-consistent MAEs against the reference, or sufficient convergence statistics across chemical classes. The broader functional claims need to be distinguished from the demonstrated frozen-orbital prediction results.

  4. The energy definition contains an omission that must be resolved. Equation (5) contains nuclear repulsion, kinetic, Coulomb, semilocal XC, learned exchange, and dispersion terms, but no explicit electron–nuclear attraction energy Tr⁡(PVne)\operatorname{Tr}(PV_{\mathrm{ne}}). The accompanying definitions identify EnucE_{\mathrm{nuc}} as nuclear repulsion and TsT_s as kinetic energy; the precomputation list in Section 4.1 omits the attraction term as well. The authors should give the complete energy decomposition used in the calculations and clarify whether this contribution is included under another name. The interpretation of AA also needs precision: Equations (2)–(3) use ordinary matrix multiplication to filter the density, which differs from assigning independent elementwise weights to individual exchange contractions.

  5. The quantitative computational claims are unsupported. The paper claims an approximately two-order-of-magnitude cost advantage over the double hybrid, but provides no timings under common settings. The reported neural-network latency of less than one second excludes the baseline hybrid calculation and exchange construction. Appendix A explicitly requires two exchange builds per spin channel at each OSE SCF step. A hybrid-like formal cost is plausible; the claimed numerical speedup requires measurements. The O(N2Naux)O(N^2N_{\mathrm{aux}}) RI-K statement also needs the contraction algorithm and its assumptions, including the treatment of occupied orbitals and screening.

  6. The implementation and derivative claims cannot be checked from the supplied materials. The manuscript gives training and network settings but identifies no OSE26 code, weights, or split files. Appendix A involves K(PσA)K(P^\sigma A), whose matrix argument is generally nonsymmetric; the PySCF handling of this intermediate is unspecified. Appendix B gives a force expression and geometry results without an energy–force finite-difference consistency check. It should state whether the analytic force applies to self-consistent OSE, post-SCF evaluation, or both, and specify the orbital-response and Pulay treatment. These details are central to reproducing the proposed functional.

[1] Febrer et al. Graph2Mat: universal graph to matrix conversion for electron density prediction. Machine Learning: Science and Technology 6, 025013 (2025).

[2] Gao, Eberhard, and Günnemann. Learning Equivariant Non-Local Electron Density Functionals. ICLR 2025.

Questions for authors
  1. Please provide the split identifiers, subset-selection rules, overlap checks, and validation/checkpoint-selection procedure. State the source and computational settings of each conventional-functional baseline, the number of training runs, and paired uncertainty for the main ML comparisons.

  2. Please specify parameter counts and the exact scalar-baseline energy expression. Clarify whether diagonal-only OSE means AO-element diagonality, atomic-block diagonality, or a rotation-invariant shell restriction. A focused scalar control A=a(R)IA=a(R)I using the same fixed semilocal term would be the most useful additional training comparison.

  3. For the existing 30-reaction SCF subset, please provide identifiers, post-SCF and self-consistent MAEs against the reference, convergence thresholds, iteration statistics, and the largest energy difference. Add energy–force consistency checks on a few representative molecules.

  4. Please resolve the energy decomposition in Equation (5), explain the exchange routine used for the nonsymmetric intermediate, and clarify code and model availability. Report representative end-to-end timings for B3LYP, post-SCF OSE, self-consistent OSE, and the double-hybrid comparator, or remove the unsupported quantitative speedup claim.

4 comments
AuthorsOpenReview ↗

Official Comment by Authors. Part 1

We thank the reviewer for the detailed critique. We address each point below and will update the manuscript accordingly.

W1, Q1.

Dataset splits.
RDB7: 3,902 train / 6,883 test reaction entries with no shared reaction IDs. Forward and reverse barriers of the same reaction, and their reaction-energy labels, remain in one partition. A geometry-overlap check (0.01 Å) finds no shared transition-state geometries; ~86% of test reactions reuse at least one reactant/product geometry from training. The split is therefore reaction-disjoint, not molecule-disjoint; this is not label leakage. Forward and reverse reactions are assigned to the same split: training or test. For MSR-ACC, we randomly selected 1,200 test structures using a reaction-ID-based split, keeping all entries with the same reaction ID in a single partition to prevent reaction-level overlap between training and test; we used this subset because the original databases provides reference energies but not the density and KK matrices required by our model, making calculations for the full dataset substantially more expensive, while the selected subset is sufficient to support the main claims of the paper.

OSE26 and B3LYP-D3(BJ) on the same geometries/basis are computed by us (B3LYP/def2-QZVP + D3(BJ) [2]). MSR values for other methods were taken directly from original MSR database with similar basis set. RDB7 values were computed using PySCF UKS RIJK with energy tolerance of 10−6Eh10^{-6} E_{h}, there are no double hybrids in official PySCF, we used ORCA 6.1.1 for ω\omegaB97M(2) with ω\omegaB97M-V orbitals. The standard deviations in Table 2 are the spread of per-example test errors, not uncertainty across training runs; we clarify this in the caption. OSE26 and MACE Δ-learning were each trained once. We have launched additional training runs with 3 different seeds for the ablation study and will include the results in the revised manuscript once they are completed.

W2, Q2

We respectfully disagree that the ablations fail to isolate the claimed advantage. The claim of the paper is not that MACE or Graph2Mat are new. The claim is that, under a shared encoder and matched capacity, parameterizing the correction as an AO-pair exchange modulation (OSE) improves over (i) adaptive global hybrid mixing and (ii) unconstrained scalar Δ-learning, and that off-diagonal AO-pair terms contribute beyond a diagonal restriction. Appendix C is constructed exactly for that comparison and nothing else.

MACE and Graph2Mat are deliberately used as established components. They are the shared implementation substrate, not the scientific claim. Keeping the encoder fixed and changing only the energy readout is the standard way to attribute differences to the readout.

Shared MACE (r_max=6.0, num_bessel=6, polynomial_cutoff=6, max_ell=2, num_interactions=2, hidden_irreps=8×0e+8×1o, MLP_irreps=8×0e, avg_num_neighbors=8.0, correlation=1).

OSE26: MACE + Graph2Mat (edge_hidden_irreps=4×0e+4×1o, basis def2-QZVP) total: 2.168.388.

Adaptive-α\alpha, Δ\Delta-learning: Shared MACE + MLP: Linear(16x0e, 652), 5×Linear(652,652), Linear(652,1) total: 2,166,933

We agree that elementwise diagonal AO matrices are generally not closed under rotations, we will state this limitation when interpreting the diag vs full comparison. To address the requested concern directly, we additionally trained and evaluated the scalar restriction inside the same OSE energy expression, with the B3LYP semilocal XC held fixed:

E=Eunxc+ExcDFA+a(Z,R)ITr(PK)E = E_{\mathrm{unxc}} + E_{xc}^{\mathrm{DFA}} + a(Z,R) I \mathrm{Tr}(P K)

Test MAEs (kcal/mol):

MethodRDB7MSR
Adaptive hybrid α\alpha1.651.57
MACE Δ\Delta-learning1.381.23
OSE scalar A=aIA=aI1.883.67
OSE26 (full AA)1.230.85
B3LYP-D3(BJ)2.795.26

The adaptive-α\alpha baseline predicts a system-dependent scalar α(Z,R)\alpha(Z,R) and evaluates the energy as

Eadaptive=Eunxc+α(Z,R)ExHF+[1−α(Z,R)]ExDFA+EcB3LYP+ED3(BJ).E_{\mathrm{adaptive}} = E_{\mathrm{unxc}}+ \alpha(Z,R) E_x^{\mathrm{HF}}+ [1-\alpha(Z,R)] E_x^{\mathrm{DFA}}+ E_c^{\mathrm{B3LYP}}+ E_{\mathrm{D3(BJ)}}.

In adaptive hybrid α\alpha baseline, α\alpha changes the weights of both exact and semilocal exchange. In OSE26, the semilocal contribution is held fixed. The Δ-learning head predicts a scalar correction to the B3LYP energy from atomic species and coordinates. OSE26 predicts AA from the same inputs and contracts it with the frozen B3LYP density and exchange matrices to evaluate the exchange energy. The comparison therefore evaluates the full OSE26 energy readout, including its use of these electronic quantities

Relative to EG-XC [Gao et al., ICLR 2025], OSE learns an AO-pair modulation of hybrid exact exchange with an explicit SCF potential and analytical forces, rather than a learned nonlocal density-to-energy XC map.

AuthorsOpenReview ↗

Official Comment by Authors. Part 2

W3, Q3.

Main MSR-ACC / RDB7 MAEs (0.85 and 1.23 kcal/mol) are post-SCF on frozen B3LYP/def2-QZVP orbitals and KK matrices. We will label them as post-SCF in Sec. 4.3, Table 2, and the conclusions.

For the existing 30-reaction SCF subset (official RDB7 database indices; FBH: 17, 25, 44, 47, 56, 57, 61, 72, 94; RBH: 14, 15, 19, 22, 38, 41, 42, 47, 56, 64, 84; RE: 12, 22, 30, 35, 47, 48, 62, 68, 71, 98):

SCF timing and iteration statistics versus system size are reported for water clusters in the revised manuscript (Appendix D, Figure 8). Mean SCF iterations per system: B3LYP 7.4, OSE 7.4, ωB97X-V 7.0, DM21 8.9. We used 10−6Eh10^{-6} E_{h} convergence treshold.

Post-SCF MAE vs reference: 1.37 kcal/mol.
SCF MAE vs reference: 1.35 kcal/mol.
Largest ∣ΔESCF−ΔEpost∣|\Delta E_{\mathrm{SCF}}-\Delta E_{\mathrm{post}}|: 0.143 kcal/mol (index 64, RBH).

Energy-force consistency is added in Appendix B, Figure 5 (analytical forces overlay finite-difference −dE/dR and potential energy curve).

Q4, W4

We have corrected Eq. (5).

Q4, W5

We remove the unsupported numerical speedup vs double hybrids. Remaining claim: hybrid cost class. NN latency <1 s excludes hybrid SCF and exchange builds. Self-consistent OSE needs two exchange builds per spin channel per iteration. RI-K follows the standard density-fitting exchange path in PySCF. We will not claim a numerical advantage over ω\omegaB97M(2) without matched timings. Regarding availability, we will publicly release the code and trained model upon publication. Their release is not mandatory at the time of submission.

W6, Q4

The nonsymmetric PAPA intermediate is handled with PySCF get_k(..., hermi=0). Analytic forces apply to self-consistent OSE (Appendix B); energy-force checks are under W3/Q3.

[2] https://github.com/dftd3/simple-dftd3
[3] https://github.com/Eipgen/BHpredict

AuthorsOpenReview ↗

Official Comment by Authors

If our responses have resolved the issues, would you consider raising your score?

Thank you for the responses, whcih address some of my concerns. I have raised my ratings.

Reviewer_1yJCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

OSE26 replaces the single exact-exchange mixing coefficient in hybrid DFT with a geometry-dependent atomic-orbital matrix predicted by MACE and Graph2Mat. The formulation supports self-consistent calculations and analytical nuclear forces, with ablations showing benefits over scalar mixing and direct energy corrections. Using fixed B3LYP densities, it achieves MAEs of 0.85 kcal/mol for atomization energies and 1.23 kcal/mol for reaction barriers. Additional tests cover self-consistency on 30 reactions, equilibrium bond lengths, and neutral and charged water clusters.

Critical strengths
  1. The orbital-level exchange formulation is well motivated. Ablations show improvements over global mixing, direct energy corrections, and diagonal-only modulation, supporting the benefit of the proposed matrix structure.

  2. The model performs well on both atomization energies and reaction barriers. Additional tests on self-consistent calculations, bond lengths, and charged water clusters further support its practical potential.

Critical weaknesses
  1. Limited out-of-distribution validation. Beyond the supplementary water-cluster and geometry tests, evaluation largely relies on held-out samples from the training databases. Broader testing on independent datasets is needed to establish generalization to unfamiliar chemical environments.

  2. Although the paper includes useful ablations, comparisons with other current models under matched training data, model sizes, and evaluation settings remain limited.

Questions for authors

Beyond the supplementary water-cluster and geometry tests, can you provide evidence that OSE26 generalizes to systems with chemical environments or electronic states substantially different from those represented in training?

4 comments
AuthorsOpenReview ↗

Official Comment by Authors

We thank the reviewer for highlighting the need for independent out-of-distribution validation and evaluation settings.

W1, Q1. To address this concern, we performed an additional evaluation across different chemical reactions using reference data from database [1]. The results are summarized below, all MAEs are in kcal/mol and calculated in def2-QZVP basis set.

Reaction classB3LYP MAEωB97X-D MAEMN15 MAEOSE26 MAE
Acid–base reactions1.010.701.950.79
Diels–Alder reactions8.022.432.541.46
Electrocyclic ring opening and closing3.171.570.560.99
Electrophilic addition1.932.830.932.54
Electrophilic aromatic substitution2.361.971.211.41
Overall3.171.761.491.33

All errors are in kcal/mol. Across these reactions, OSE26 reduces the overall MAE from 3.17 to 1.33 kcal/mol and the RMSE from 4.16 to 1.75 kcal/mol. We will add this additional test results in main text. Calculations with ωB97M(2) are more time-consuming and are still in progress, we will add the corresponding results once they are available.

W2. We agree that the ablation study evaluations could have been clearer in the original manuscript. We clarify the experimental configuration below and will make this more prominent in the revised main text.

Evaluation setting: All models use the same post-SCF protocol as the main benchmarks: energies on precomputed unrestricted B3LYP/def2-QZVP densities, energies and exchange matrices K(P)K(P) (calculated with standard PySCF def2-universal-jkfit RIJK), with D3(BJ) dispersion using B3LYP parameters [2].

Model hyperparameters:

  1. Shared MACE (r_max=6.0, num_bessel=6, polynomial_cutoff=6, max_ell=2, num_interactions=2, hidden_irreps=8×0e+8×1o, MLP_irreps=8×0e, avg_num_neighbors=8.0, correlation=1).

  2. OSE26: MACE + Graph2Mat (edge_hidden_irreps=4×0e+4×1o, basis def2-QZVP) total: 2.168.388.

  3. Adaptive-α\alpha, Δ\Delta-learning: Shared MACE + MLP: Linear(16x0e, 652), 5×Linear(652,652), Linear(652,1) total: 2,166,933

We also include OSE26-diag (same OSE architecture, off-diagonals of AA forced to zero) and DeePHF with published hyperparameters [3] on the same B3LYP-D3(BJ) baseline.

[1] Karton, A. (2026). Benchmarking electronic structure methods using generative AI to construct diverse reaction databases. Chemical Physics Letters, 891, 142765. https://doi.org/10.1016/j.cplett.2026.142765

[2] https://github.com/dftd3/simple-dftd3

[3] https://github.com/Eipgen/BHpredict

AuthorsOpenReview ↗

Official Comment by Authors

If our responses have resolved the issues for broader testing on independent datasets, would you consider raising your score?

Thank you for the detailed response and the additional experiments. I appreciate the authors’ efforts to address my concerns. The new reaction-energy evaluation provides useful evidence of generalization and addresses my first concern. The clarification of evaluation settings and parameter counts also helps substantiate the ablation comparisons. It would also be helpful to discuss the variation in performance across reaction classes. OSE26 achieves the lowest overall MAE among the reported methods, but its advantage is not uniform: it performs particularly well on Diels–Alder reactions, while some conventional functionals perform better on other classes. What characteristics of these reactions might explain this pattern, and how might they relate to the proposed orbital-specific exchange formulation? A brief discussion would help clarify both the strengths and the limitations of the method.

My remaining concern is the comparison with contemporary learned functionals. I acknowledge the existing comparisons with strong conventional functionals, including ωB97M(2), as well as the DeePHF baseline and the additional DM21 results. However, Skala appears to be a particularly relevant baseline given its reported accuracy and computational efficiency. A direct comparison on a common independent benchmark would strengthen the evidence for the proposed method’s competitiveness. Since Skala was trained on MSR-ACC data, potential training-data overlap should be considered when selecting this benchmark.

I also look forward to the pending ωB97M(2) results on the additional reaction benchmark.

AuthorsOpenReview ↗

Official Comment by Authors

We thank the reviewer for the follow-up comments.

We have added Skala 1.0/1.1 and preliminary ωB97M(2) on the reaction benchmark of Ref. [1] (def2-QZVP). All MAEs are in kcal/mol:

Reaction classB3LYPωB97X-DMN15OSE26Skala 1.0Skala 1.1ωB97M(2)
Acid–base reactions1.010.701.950.791.951.421.03 (99/99)
Diels–Alder reactions8.022.432.541.461.781.342.01 (41/70)
Electrocyclic ring opening and closing3.171.570.560.991.351.071.28 (11/73)
Electrophilic addition1.932.830.932.541.710.751.05 (18/52)
Electrophilic aromatic substitution2.361.971.211.411.410.830.62 (25/77)
Overall3.171.761.491.331.651.121.20 (194/371)

The class-wise pattern is chemically expected. Acid–base, Diels–Alder, and electrocyclic reactions are linked to strong orbital interactions—Diels–Alder and electrocyclic cases via Woodward–Hoffmann orbital-symmetry control, acid–base via orbital hardness/softness—where an orbital-specific exchange correction is most relevant; electrophilic addition and electrophilic aromatic substitution are more charge-controlled, so density/charge-driven errors dominate and Skala 1.1 is stronger there. We report Skala on Ref. [1] rather than MSR-ACC, to the best of our knowledge Skala’s full training dataset is not publicly released, so an independent test evaluation against its training data is not possible. DM21 is too expensive for a full pass at this stage (runtime comparable to CCSD(T); see plot); we will include full ωB97M(2), DM21, DeepHF results in the camera-ready version.

If our responses have resolved the issues for broader comparison with this methods, would you consider raising your score?

[1] Karton, A. (2026). Benchmarking electronic structure methods using generative AI to construct diverse reaction databases. Chemical Physics Letters, 891, 142765. https://doi.org/10.1016/j.cplett.2026.142765

Reviewer_mzMVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes OSE26, a geometry-dependent orbital-specific exchange formalism that replaces the scalar exact-exchange fraction in conventional hybrid density functionals with a learned AO-by-AO mixing matrix. An equivariant neural network predicts the matrix, which modulates the Hartree–Fock exchange operator. The paper evaluates atomization energies, reaction barriers, and noncovalent interactions, compares against adaptive global-alpha and delta-learning baselines, and reports additional experiments on self-consistent convergence, analytical forces, and water-cluster transfer.

Critical strengths
  1. The AO-pair exchange formulation is a clear and potentially useful conceptual extension of global hybrid functionals. The comparisons with adaptive scalar alpha, delta-learning, and diagonal-only OSE are useful.\n\n2. The paper goes beyond predictive accuracy by deriving the self-consistent exchange potential and analytical forces, and by reporting SCF stability and computational-scaling considerations.
Critical weaknesses
  1. The motivation that a global exchange fraction is insufficient is well established in local hybrids, range-separated hybrids, and adaptive functionals. The paper should more clearly explain why AO-pair modulation is a distinct and necessary contribution.\n\n2. The train/test split is insufficiently described, and reaction-level random splitting could permit overlap in molecules, scaffolds, or reaction families. Further, the main accuracy results appear to rely on precomputed B3LYP quantities, whereas fully self-consistent OSE is evaluated only on a limited subset. The basis dependence and transferability of the learned AO-space matrix are also not adequately evaluated.
Questions for authors
  1. What are the exact molecule-level, scaffold-level, and reaction-family split protocols for MSR-ACC and RDB7?\n\n2. Which reported test numbers are obtained post-SCF, and which are obtained with fully self-consistent OSE calculations?\n\n3. Does the method preserve performance under alternative basis sets, open-shell systems, charged systems, or basis transformations within the same Hilbert space?\n\n4. Can the authors provide systematic timing and SCF-convergence statistics versus standard hybrid functionals across increasing system sizes?
3 comments
AuthorsOpenReview ↗

Responce. Part 1

W1 We respectfully disagree with this assessment. The Introduction and Section 5.1 explain how AO-pair modulation complements local, range-separated, and adaptive global hybrids. In particular, Section 5.1 explains how OSE can modulate distinct exchange channels associated with orbitals that overlap in the same region of space.

  1. Local hybrids already provide real-space flexibility in α, but their practical SCF cost is often prohibitive: evaluating local Hartree–Fock features on the grid can make them comparable in wall time to high-level wavefunction methods even for moderate systems (https://doi.org/10.1063/5.0266500) and Appendix D (Fig. 8). OSE is deliberately formulated to remain in the hybrid exact-exchange cost class while still allowing orbital-resolved and therefore position-dependent mixing.

  2. Range-separated hybrids improve over a single global α, but they still use a fixed functional-level parametrization. That can transfer poorly to particular chemical regimes, for example, on large π-conjugated systems, on the benchmark of fullerene isomerization (https://pubs.acs.org/doi/10.1021/acs.jpclett.6c01636) ωB97X-V gives MAE 16.27 kcal/mol, whereas B3LYP and OSE26 give 2.22 and 1.71 kcal/mol, even though OSE26 was not trained on large π-conjugated systems of that type.

  3. Adaptive hybrids learn one system-wide mixing coefficient. A single scalar cannot assign different exchange treatments to chemically distinct regions within the same system.

The contribution of this work is a self-consistent hybrid functional whose exact-exchange mixing is an AO-pair matrix A: a complementary axis of flexibility to local hybrids (real-space a(r)), RSHs (fixed range separation), and adaptive global-α models, together with the corresponding exchange potential and practical SCF implementation.

We disagree that basis dependence and transferability is missing from the manuscript. The Limitations section already states that the present OSE26 is trained for def2-QZVP and that transferring to another AO basis requires recomputing the basis-dependent DFT quantities and retraining (or using basis-independent reformulation).

Q1 For RDB7, the split is reaction-ID disjoint: 3,902 training reactions and 6,883 held-out test reactions, with 0% shared reaction IDs. Forward and reverse barriers of the same reaction, together with their reaction-energy labels, remain entirely on one side of the split.

The split is not molecule-disjoint: approximately 86% of test reactions reuse at least one reactant/product geometry present in training, using a 0.01 Å matching tolerance, whereas the transition-state geometries are distinct. Thus, this protocol evaluates generalization to unseen reactions, not necessarily to entirely unseen molecules or scaffolds. Reaction IDs and their associated targets remain disjoint, but we acknowledge that molecular overlap makes this evaluation less stringent than molecule-disjoint testing.

For MSR-ACC, the targets are molecular atomization energies, with 16,502 training examples and 1,200 held-out test examples. Each labeled example corresponds to a molecular structure and its atomization energy. The split was likewise not constructed to enforce scaffold- or chemical-family-disjointness. Given the limited availability of CCSD(T)-quality reference data for both reaction barriers and atomization energies, we did not impose additional scaffold- or chemical-family-disjoint constraints on either benchmark.

AuthorsOpenReview ↗

Responce. Part 2

Q2 The primary aim of this study was to demonstrate the OSE formalism in its simplest practical setting: post-SCF evaluation on precomputed B3LYP/def2-QZVP densities and exchange matrices. This is precisely the setting in which OSE26 was trained. Accordingly, the main MSR-ACC and RDB7 accuracy results and the matched ablation comparisons are post-SCF results. They establish the accuracy of the proposed implementation in its training and evaluation setting, rather than constituting full-dataset benchmarks of self-consistent OSE26.

Fully self-consistent OSE26 calculations are reported for water clusters (Appendix D; Fig. 7 and 9) and for a randomly sampled RDB7 subset (Appendix A; Fig. 3). Energy–force consistency of the analytical gradients is shown in Appendix B (Fig. 5).

The SCF and geometry-optimization experiments in Appendices A and B are additional results illustrating the practical capabilities of the formalism. OSE admits a straightforward self-consistent implementation and analytical nuclear gradients, enabling efficient geometry optimization at hybrid-like computational cost. These capabilities are nontrivial: analytical gradients are not broadly available for local hybrids, while fully self-consistent treatments and geometry optimization are more involved for double hybrids, although analytical gradients exist for selected implementations. These experiments therefore illustrate a potentially useful combination of orbital-resolved exchange flexibility, practical SCF evaluation, and analytical forces. They complement, rather than replace, the main post-SCF accuracy evaluation.

Q3 The scope and limitations of transferability are discussed in Section 5.3 and Appendix D. OSE26 was trained in def2-QZVP; preservation of its accuracy under alternative AO basis sets has not been established. Transfer to another basis requires recomputing the basis-dependent quantities and retraining, or developing a basis-independent reformulation. The present implementation is tied to a fixed AO basis and basis transformations have not been investigated.

No open-shell or charged systems were included in training. Charged-system transfer is nevertheless evaluated on water clusters, as described in Appendix D (Fig. 7). This provides a specific out-of-training-domain test, not a general validation across charge states or open-shell chemistry. As acknowledged in Section 5.3, the current mixing matrix depends only on nuclear geometry (simplest form under the study).

Q4 Self-consistent OSE requires two exchange builds per spin channel per iteration, corresponding to K[P]K[P] and K[PA]K[PA], with the latter evaluated using PySCF’s exchange routine for a nonsymmetric input matrix. Representative SCF wall times versus system size for water clusters are shown in Appendix D (Fig. 8) with statistics of SCF iterations. SCF convergence for the RDB7 subset is shown in Appendix A (Fig. 3).

We have removed the unsupported numerical claim of an approximately 100-fold speedup over double hybrids. The remaining claim is that OSE retains hybrid-class computational cost without the additional virtual-orbital correlation step of a double hybrid.

AuthorsOpenReview ↗

Responce. Part 3

If our responses have resolved the issues you raised regarding the reference comparisons and the methodological advances presented, would you consider raising your score?

Comments on the paper

1 comment

We want to sincerely thank all reviewers for their productive engagement. Their insight contributed greatly to clarification of OSE26's contribution and to the development of its validation. We also thank the Area Chair for their time and consideration.

Reviewer mzMV motivation, transferability, and cost. In our responses, we clarified how AO-space exchange modulation complements existing hybrid approaches, referring to the Introduction and Section 5.1. Our responses explain the reaction-level splits, highlight the difference between post-SCF and SCF results, and outline that Section 5.3 presents basis-dependent scope. Appendix D now includes SCF timings and iteration statistics for water clusters (Figure 8). We have dropped the unsubstantiated quantitative speedup claim.

Reviewer 1yJC independent test and learned-functional comparisons. We reported an additional external benchmark spanning five reaction classes. We have shown direct comparisons with Skala 1.0/1.1 models, preliminary results from double hybrids with explicit coverage counts, and an analysis of how performance varies across different reaction classes.

Reviewer z2Sq - ablations and self-consistent implementation. Our responses specify parameter counts, baseline energy expressions, and the symmetry limitation of the diagonal ablation. We reported an additional scalar-OSE control on MSR-ACC and RDB7 to assess the benefit of matrix parameterization. Analytical-force versus finite differences comparisons have been put into Appendix B with a figure 5. Equation (5) now explicitly includes the electron–nuclear attraction term.

We consider the revised discussion more persuasive in favor of a physically motivated learnable parameterization of AO-space exchange modulation. We hope that the additional results and clarifications will support a favorable assessment and acceptance of our work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.