acceptodds
Under review as a conference paper at ICLR 2027

Solver-Oriented Electron Density Learning with Equivariant Gaussian Shells

Abstract

Learning electron density from molecular geometry supports electronic-property recovery and self-consistent field (SCF) initialization. We introduce GEM-FIELD, an equivariant Gaussian-shell representation with learnable radial parameters, nonlinear angular distributions, and exact electron-number conservation. Its Gaussian expansion provides analytical Coulomb integrals for target atomic-orbital bases, connecting predicted densities to property evaluation and SCF solvers. On QM9-PBE, GEM-FIELD achieves 0.226% dense-grid density NMAE; its compact variant achieves 0.237% while reducing evaluation time by 96.3% relative to ChargE3Net. On larger OE62 molecules, density-derived frontier-orbital errors are 86.7%–92.8% lower than task-specific predictions. On SCFBench OOD molecules with up to 60 atoms, GEM-FIELD achieves 100% convergence, reduces SCF iterations by 40.5%, and lowers end-to-end runtime by 8.98% relative to MINAO. Further evaluations demonstrate transferable initialization across target AO bases and exchange–correlation functionals.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Electron density learning, Equivariant graph neural networks, Gaussian shell representations, Gaussian-type orbitals, Density functional theory, Self-consistent field acceleration, Electronic-structure prediction, Basis-set transferability

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_DThFOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces GEM-FIELD, which predicts electron densities from molecular geometries and uses them for property evaluation and SCF initialization. It places multiple Gaussian shells around atoms and learns their radii, widths, and nonlinear angular distributions to represent the residual relative to the SAD density. The Gaussian expansion of the predicted density is used to construct the Coulomb matrix in the target AO basis. DFT evaluation and a generalized eigensolution then provide orbitals and approximate properties, or the resulting density matrix is used as an initial guess for SCF.

Critical strengths

The solver-oriented motivation is supported by actual calculations. All 1,050 OOD molecules in SCFBench converge, and the number of SCF iterations is reduced by approximately 40.5% relative to MINAO initialization. When model inference and initial density-matrix construction are included, the total runtime reduction with the analytical Coulomb implementation is smaller, at 8.98%. I appreciate that the paper explicitly reports results that include these overheads.

Critical weaknesses

The additional contribution of the model and training approach over existing methods is not sufficiently established.

GEM-FIELD constructs a Fock matrix from the predicted density and diagonalizes it to obtain a density matrix and electronic properties, so its downstream computations are closely related to those of Hamiltonian prediction models. However, Tables 2–3 mainly compare against models that directly predict properties or energies and forces, while comparisons with Hamiltonian models are limited to the SCF initialization evaluation in Table 4. Methods such as QHFlow2, which derive orbitals, energies, and forces from predicted Hamiltonians, should be compared using the same data and DFT settings. The SCDP and BOA models already trained for Table 1 should also be evaluated through the same solver pipeline in Tables 2–3, and trained under the SCFBench setting to compare performance and cost in Table 4. If this is technically difficult, the necessary transformations, approximations, integrals, and additional costs should be explained explicitly. The absence of trained checkpoints alone does not make a convincing case for the necessity of the GEM-FIELD model or methodology over existing approaches. (QHFlow2, Kim et al., 2026)

Also, it needs to be compared with the recent benchmark of the SCF acceleration. SAIL was posted on arXiv on April 23, 2026, and reports that solver-aligned training can improve extrapolation to larger molecules for Hamiltonian and density-matrix models. However, this paper neither cites nor discusses this work, and provides no comparison with it. It is therefore difficult to assess the limitations of matrix-based approaches or the relative transfer advantage of density representations from the OOD results of existing QHNet and QHFlow models alone. Baselines that apply solver-aligned training to existing representations are important for assessing the benefits for SCF initialization. Beyond the density-error ablations in Table 6, it would be helpful to distinguish the effect of changing the representation within the same solver pipeline from the effect of changing the training objective using a common model where feasible. (SAIL, Eberhard et al., 2026) Without this comparison, the message is easily misled into thinking that the matrix-based or coefficient-based methods are not a good approach.

Training and initialization settings and total computational cost are unclear.

The exact PySCF routine used to construct SAD, and its relationship to the MINAO evaluation baseline, should be specified. A SAD-only baseline, with the learned residual set to zero, should also be evaluated using the same Fock construction and diagonalization. Reporting errors in both the predicted density and the resulting density matrix would help clarify the respective contributions of the learned correction and the subsequent electronic-structure calculations.

The end-to-end SCF timing in Appendix C is useful, but the energy and force results in Tables 2–3 do not report the time required for Coulomb/XC matrix construction, diagonalization, and property evaluation after model inference. Comparisons with MLIPs and Hamiltonian predictors should include total evaluation time on the same hardware, including the derivatives and post-processing required by each method.

Questions for authors

Comparison with density and Hamiltonian models. Could Tables 2–3 include QHFlow2 or other Hamiltonian prediction models evaluated using the same data, functional, and AO basis? Please also train ELECTRA, SCDP, and BOA for the respective benchmarks and evaluate them through the same solver pipeline to compare accuracy, SCF convergence, and total runtime in Tables 2–4. If this is difficult, please explain the specific technical constraints and additional costs, beyond the absence of trained checkpoints.

Total runtime and model scaling. Could you break down the time spent on inference, Coulomb/XC matrix construction, diagonalization, and energy/force evaluation in Tables 2–3, and compare total runtime against MLIPs and Hamiltonian predictors, including all required derivatives and post-processing? I would also like to understand the accuracy–runtime trade-off when increasing backbone size while keeping the number of shells and quadrature fixed. In particular, can a larger, more accurate model reduce total SCF runtime?

Level-3 training under the same budget. How do density NMAE and the property results in Tables 2–3 change when the same model is trained on level-3 grids with the same number of sampled points per molecule and optimization updates? Please also report actual training time and explain why level-0 was selected. If this comparison is difficult, please specify the computational constraints.

Definition of SAD and the effect of the learned correction. Could you specify the exact PySCF routine and basis settings used to construct SAD, its relationship to the MINAO initial guess, and how SAD/MINAO is used by each baseline in Table 4? Please also evaluate a SAD-only baseline with the predicted residual set to zero using the same Fock construction and diagonalization, and report errors in both the original predicted density and the resulting density matrix.

Effect of the parameter scaling on downstream performance In the OOD SCF calculations in Appendix C, inference takes 0.025 seconds out of a total of 9.910 seconds, or approximately 0.25%. A larger, more accurate model could therefore save more time by reducing SCF iterations than it adds through inference. The current angular-cutoff ablation reports only density accuracy, so the relationship between model size, property accuracy, and total runtime needs further evaluation.

6 comments
AuthorsOpenReview ↗

Response to Reviewer DThF

Thank you for your rigorous review and constructive suggestions. We have conducted the requested experiments—including evaluating QHFlow2 across Tables 2–4, benchmarking SCDP/BOA/ELECTRA and a SAD-only baseline through the one-shot solver pipeline, training level-3 and 50M-parameter GEM-FIELD models, and adding stage-wise runtime breakdowns—and updated the manuscript accordingly.

W1&Q1. Evaluate QHFlow2 and density baselines (SCDP, BOA, ELECTRA) under the same solver pipeline across Tables 2–4, and clarify technical constraints and runtime overheads.

Response: We added QHFlow2 to Tables 2–4 and Appendix Tables C1–C2, and evaluated SCDP (trainable exponents), BOA, and ELECTRA through the same one-shot solver pipeline in Tables 2–3:

  1. Comparison with QHFlow2 (Tables 2–4): Despite strong ID RIC (50.84%) and energy components (1.54–24.03 meV/atom), QHFlow2 trails GEM-FIELD (level3) on ID total energy (44.92 vs. 0.22 meV/molecule) and force MAE (42.33 vs. 3.21 meV/Å), and incurs an ID force RMSE of 6,793.06 meV/Å (vs. 11.77 meV/Å). On molecules such as gdb_033390 (C5N4\mathrm{C_5N_4}, 92.93% of QHFlow2's total squared ID force error), Hamiltonian errors along near-singular overlap directions create a spurious occupied orbital (ε≈−17.38 Ha\varepsilon \approx -17.38\text{ Ha} vs. +1.21 Ha+1.21\text{ Ha}), corrupting Pulay forces. On OE62 OOD (Tables 2–3), QHFlow2 degrades to 33,750.53 meV/molecule energy MAE and 4,470.92 meV/Å force MAE (vs. 13.32–16.38 and 24.65–29.72 for GEM-FIELD), and on SCFBench OOD (Table 4), all GEM-FIELD variants (59.53%/57.50%/56.79%) beat QHFlow2 (61.22%).
  2. Comparison with SCDP, BOA, and ELECTRA (Tables 2–3) and omission from Table 4: Under the same level-0 training grid and one-shot pipeline, GEM-FIELD (level-0) outperforms SCDP, BOA, and ELECTRA on ID total energy (0.61 vs. 1.57–201.35 meV/molecule), forces (4.81 vs. 12.18–178.48 meV/Å), gap (4.23 vs. 9.90–71.93 meV), and dipole (0.0083 vs. 0.0351–0.3674 D), and on every OE62 OOD metric (where baseline energy, force, and dipole MAEs degrade to 175.98–18,162.53 meV/molecule, 164.44–1,194.53 meV/Å, and 26.36–93.96 D, vs. 13.32, 24.65, and 0.1830 for level-0 GEM-FIELD).

We did not retrain them on SCFBench (Table 4) because initial-guess Coulomb construction makes them slower end-to-end than MINAO, even when they reduce SCF cycles. Because they lack a standard atom-centered 3-center Coulomb interface in GPU4PySCF (SCDP uses atom/bond-centered uncontracted functions with scaled exponents; ELECTRA uses floating anisotropic Gaussians; BOA applies nonlinearities to GTO products), our pipeline constructs ΔJμν\Delta J_{\mu\nu} by contracting grid densities with analytical AO-pair integrals Aμν(rg)A_{\mu\nu}(\mathbf r_g):

=\int \widehat{\Delta\rho}(\mathbf r') A_{\mu\nu}(\mathbf r') d\mathbf r' \approx\sum_g w_g \widehat{\Delta\rho}(\mathbf r_g) A_{\mu\nu}(\mathbf r_g), \quad A_{\mu\nu}(\mathbf r') =\int\frac{\phi_\mu(\mathbf r)\phi_\nu(\mathbf r)} {\lVert\mathbf r-\mathbf r'\rVert} d\mathbf r.$$ On 51-atom folic acid (RTX 4070 Laptop GPU, QM9 checkpoints), initial-guess construction with this grid-based $J$ takes **33.75–59.53 s** (residual $J$ alone ~30 s)—already matching or exceeding the **entire 36.26 s MINAO SCF run**. Thus, even though SCDP and BOA cut SCF cycles from 19 to 14 and 12, $J$-construction overhead overwhelms SCF savings, causing a **2.08×–2.38× net slowdown (0.42×–0.48×)** vs. MINAO (while ELECTRA rises to 22 cycles). Combined with their OOD degradation (Tables 2–3), retraining them on SCFBench is unnecessary. Conversely, GEM-FIELD's analytical and FMM-numerical $J$ paths take only **4.18–5.82 s** for initial-guess construction, achieving **1.34×–1.44×** speedups on folic acid and up to **1.27×** (21.15% savings) on all 1,050 SCFBench OOD molecules (Appendix C, Table C1). | Initialization scheme | Inference (s) | Initial-guess construction (s) | SCF (s) | SCF iterations | Total wall-clock (s) | Speedup vs. MINAO | |:---|---:|---:|---:|---:|---:|---:| | MINAO (default) | – | – | 36.26 | 19 | 36.26 | 1.00× | | GEM-FIELD (numerical $J$) | 0.065 | 4.18 | 21.01 | 11 | 25.25 | **1.44×** | | GEM-FIELD (analytical $J$) | 0.061 | 5.82 | 21.12 | 11 | 27.00 | 1.34× | | SCDP (density-grid $J$) | 0.096 | 59.53 | 26.59 | 14 | 86.21 | 0.42× | | ELECTRA (density-grid $J$) | 0.069 | 33.75 | 41.69 | 22 | 75.51 | 0.48× | | BOA (density-grid $J$) | 0.136 | 57.05 | 22.77 | 12 | 79.95 | 0.45× |

W2. Discuss SAIL (Eberhard et al., 2026) and distinguish the effect of density representation from solver-aligned training objectives.

Response: Thank you for pointing out SAIL. We have added a citation and discussion of SAIL in Section 5.3. We fully agree that the OOD degradation of directly supervised Hamiltonian and density-matrix models reflects a mismatch between ground-state supervision and SCF solver dynamics, and that SAIL's solver-aligned training effectively restores OOD SCF transfer for matrix-based models. At the same time, our revised experiments demonstrate the advantage of the representation under a shared direct-supervision objective and identical solver pipeline: in Tables 1–3, GEM-FIELD, SCDP, BOA, and ELECTRA are all trained with the same direct L1L_1 density loss without solver backpropagation, yet GEM-FIELD transfers substantially better to OOD properties and forces; similarly, in Tables 4–5, GEM-FIELD outperforms the directly supervised auxiliary-basis model NequIP-L, especially under AO basis expansion. Because differentiating through SCF iterations (including DIIS) requires a differentiable solver and increases training memory and runtime, we view representation design and solver-aligned training as complementary and note their combination as a promising future direction.

W4&Q2. Break down stage-wise runtimes for force evaluation (Tables 2–3) on identical hardware against MLIPs, QHFlow2, and DFT.

Response: We added a complete stage-by-stage runtime and accuracy breakdown on the 51-atom folic-acid geometry in Appendix C.4 (Table C2), evaluated on the same NVIDIA RTX 4070 Laptop GPU (mean ±\pm std over 3 runs after 3 warm-ups) using QM9-trained checkpoints. Timings cover preparation, network forward pass, Coulomb/XC construction, diagonalization, and force evaluation (including automatic energy differentiation for NequIP/MACE, MINAO input-Fock construction for QHFlow2, and full grid/auxiliary-basis derivative responses for density and Hamiltonian models):

  • Self-consistent DFT takes 29.243 ± 0.327 s.
  • Direct MLIPs run in 0.0303–0.0720 s, but incur large OOD force MAEs of 300.47–333.07 meV/Å.
  • QHFlow2 takes 10.024 ± 0.051 s and yields a force MAE of 193.78 meV/Å.
  • GEM-FIELD (level-0, numerical JJ) takes 11.354 ± 0.079 s (0.287 s prep, 0.073 s network forward, 3.879 s Coulomb/XC, 0.023 s diagonalization, and 7.092 s force evaluation)—running 61.2% faster than DFT while achieving a force MAE of 65.04 meV/Å, which is 78.4% lower than the best MLIP and 66.4% lower than QHFlow2 at only a 13.3% runtime increase. Analytical JJ takes 12.843 s with the same 65.04 meV/Å MAE, and the level-3 numerical model takes 11.116 s with 69.20 meV/Å MAE.

The effect of scaling the backbone size on SCF runtime is detailed in Q5 below.

Q3. Report density, property, and SCF results when training on subsampled level-3 grids, along with training costs and the grid-generation bottleneck.

Response: We initially trained on level-0 grids because on-the-fly CPU generation of complete level-3 grids was a severe data-loading bottleneck, largely due to spatial grid sorting. During the rebuttal, we switched to GPU4PySCF grid generation with spatial sorting disabled, accelerating level-3 grid construction by 11.4× on water and 24.5× on benzene:

MoleculeGrid levelCPU, default (ms)CPU, no sorting (ms)GPU4PySCF, default (ms)GPU4PySCF, no sorting (ms)
Water01.7210.3771.7941.259
Water322.6831.7872.5791.997
Benzene09.0222.5212.5292.015
Benzene3135.42133.0036.1985.527

This enabled training GEM-FIELD on level-3 grids (sampling 1,000 points per molecule per visit) for 500,000 steps on QM9-PBE (~120 RTX 5090 GPU hours) and 300,000 steps on SCFBench (~60 RTX 5090 GPU hours). As updated in Tables 1–4, level-3 training substantially improves performance across all tasks despite fewer optimization steps:

  • Density NMAE (Table 1): Level-0 and unpruned level-3 NMAEs drop to 0.035% and 0.037% (down from 0.037% and 0.226% for level-0 training), achieving the best accuracy on both grids.
  • Properties & forces (Tables 2–3): Outperforms the level-0 model on 10 of 11 ID metrics and 6 of 11 OOD metrics, reducing ID energy-component MAEs by 77.5%–83.8%, ID total-energy MAE from 0.61 to 0.22 meV/molecule, ID force MAE from 4.81 to 3.21 meV/Å, and OOD energy-component MAEs by 36%–58%.
  • SCF acceleration (Table 4): Lowers SCFBench ID and OOD RICs by 5.90 and 2.03 percentage points (from 58.40%/59.53% to 52.50%/57.50%), and reduces mean OOD end-to-end runtime from 8.959 s to 8.566 s (Table C1).

W3&Q4. Specify the exact PySCF SAD definition, its relation to MINAO in Table 4, and evaluate a zero-residual SAD-only baseline through the same one-shot pipeline.

Response: We have clarified these definitions and added the SAD-only baseline to Tables 1–3:

  1. SAD vs. MINAO and Table 4 baseline protocols: SAD is constructed via pyscf.scf.hf.init_guess_by_atom(mol) in the spherical-harmonic def2-SVP AO basis (cart=False), which block-diagonally superposes spherically averaged neutral-atom HF density matrices DSAD\mathbf D_{\mathrm{SAD}}. In contrast, PySCF's default MINAO initial guess projects minimal atomic natural orbitals onto the target AO basis. In Table 4, all RIC values are measured relative to MINAO: GEM-FIELD constructs F[ρSAD+Δρ^]\mathbf F[\rho_{\mathrm{SAD}} + \widehat{\Delta\rho}] and diagonalizes it once to obtain D(0)\mathbf D^{(0)}; NequIP-L constructs F\mathbf F directly from its full predicted def2-universal-jfit density without SAD; QHFlow2 takes F[DMINAO]\mathbf F[\mathbf D_{\mathrm{MINAO}}] as conditioning input to predict F\mathbf F and diagonalizes it; QHNet and QHFlow follow the official SCFBench protocol.
  2. SAD-only baseline (Δρ^=0\widehat{\Delta\rho}=0, Tables 1–3): We evaluated raw SAD densities and passed DSAD\mathbf D_{\mathrm{SAD}} through the exact same DF-PBE/def2-SVP Fock construction and single diagonalization DSAD(0)\mathbf D^{(0)}_{\mathrm{SAD}}. On QM9-PBE, SAD-only has level-0 and unpruned level-3 density NMAEs of 12.805% and 12.766% (and 12.768% / 12.734% on OE62-OOD). After one Fock diagonalization, DSAD(0)\mathbf D^{(0)}_{\mathrm{SAD}} still incurs huge errors on QM9-PBE (energy MAE of 15,306.77 meV/molecule, HOMO–LUMO gap MAE of 860.41 meV, and force MAE of 2,276.34 meV/Å) and OE62-OOD (37,831.62 meV/molecule, 482.45 meV, and 2,269.55 meV/Å). Adding GEM-FIELD's learned residual reduces these errors by 2 to 5 orders of magnitude (to 0.22 meV/molecule, 3.23 meV, and 3.21 meV/Å on ID), confirming that the accuracy stems from the learned density correction rather than the single diagonalization step.

Q5. Evaluate backbone parameter scaling (50M model) on end-to-end SCF runtime and report downstream property errors in the Table 6 ablations.

Response: We trained a 50M-parameter backbone variant, GEM-FIELD-L (level3), keeping the shell counts and degree-15 Lebedev quadrature fixed. As shown in Table 4, scaling from 3.6M to 50M parameters further reduces ID and OOD RICs from 52.50% and 57.50% to 51.58% and 56.79% with 100% convergence, achieving the lowest OOD RIC among all compared methods. In the end-to-end timing across all 1,050 OOD molecules (Appendix C, Table C1), enlarging the model under the numerical JJ branch increases mean inference time by only 0.006 s (0.025 s →\rightarrow 0.031 s) while reducing subsequent SCF time by 0.063 s (6.546 s →\rightarrow 6.483 s, the shortest SCF time of any method), lowering total wall-clock time from 8.718 s to 8.642 s (increasing aggregate speedup from 1.25× to 1.26×). For the analytical JJ branch, SCF time similarly drops from 6.539 s to 6.505 s (total time 8.672 s vs. 8.566 s for the 3.6M model).

Finally, we added one-shot HOMO–LUMO gap MAE and RMSE (evaluated on a 1% random test subset) to the angular-cutoff and backbone ablations in Table 6. Increasing the EquiformerV2 cutoff from Lmax⁡=3L_{\max}=3 to Lmax⁡=6L_{\max}=6 steadily lowers level-0 density NMAE from 0.066% to 0.037% and reduces the downstream gap MAE from 8.91 meV to 4.60 meV (and RMSE from 13.02 to 6.76 meV) overall, linking higher-order angular expressivity to downstream electronic-property accuracy.

Reviewer_efJqOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a novel method for predicting the ground state electron density of molecules. The proposed GEM-FIELD model mainly introduces a new representation for the electron density. Instead of predicting the density directly on the grid or using a predefined basis set, a predicted set of basis functions is used. These basis functions are separated into an angular part and a radial part. While the angular part is predicted using spherical harmonics in an equivariant way, the radial part is build from Gaussian functions placed on shells around the atoms. This representation fits well into standard Gaussian DFT codes, enabling the analytic evaluation of integrals necessary for property evaluation or DFT initialization. The models are trained and evaluated on QM9-PBE and SCFBench, and additionally generalization to larger molecules is analyzed using structures from OE62-OOD. Not only the accuracy of the density directly is evaluated but also how well it can be used for the calculation of downstream properties and the initialization of DFT calculations.

The introduced density representation is novel and interesting for the community. However the given results do not seem convincing enough as they stand.

Critical strengths

The introduced electron density representation is a novel approach and interesting in itself. It seems well suited for the ground state electron density prediction context.

Critical weaknesses

The improvements on SCF acceleration seem marginal. The fact that the density errors on QM9-PBE look great on level 0 grids but much worse on level 3 seem to signal that the model is overfitting to the level 0 grids. The property results, which look good in itself, are made much less attractive by the fact that basically one SCF iteration has to be performed to calculate these properties from the electron density, inheriting the scaling of DFT. The energy and force results in Table 3 suffer from the same problem.

A more direct comparison with ground state density prediction benchmarks on already existing datasets would be preferable.

Questions for authors
  • Are all properties calculated with the density matrix after the one step diagonalization? If yes, could also properties, e.g. the dipole moment, of the original predicted density be evaluated?
  • Could density prediction results on more established benchmarks be provided? For example the original QM9 dataset which was used by BOA, ELECTRA, SCDP and others?
  • The models seem to overfit on the level 0 grid used in training. Since this grid is anyways subsampled in training, why not just train using grid level 3?

Minor questions:

  • To my understanding the Section 4.1 has paragraphs describing EquiformerV2 and ELECTRA (“Molecular graph and equivariant features” and the first part of “Continuous Gaussian shells”). Maybe one could mention that more explicitly?
  • It is mentioned that numerical Coulomb construction is faster that analytical construction and gives nearly the same reduction in SCF iteration count. If that is the case why is the fact that integrals can be calculated analytically with the introduced density representation important?
7 comments
AuthorsOpenReview ↗

Response to Reviewer efJq

Thank you for your thorough review and constructive feedback. We have performed the suggested experiments—including training on level-3 grids and expanding our baseline and runtime comparisons—and updated the revised PDF accordingly.

W1&W3. SCF gains appear marginal, and property evaluation inherits DFT-like computational costs.

Response: We agree with both observations and clarify the significance and positioning of our results:

  1. SCF acceleration. On SCFBench OOD, GEM-FIELD achieves the best performance with 100% convergence, and the newly added 50M level-3 model further lowers OOD RIC to 56.79%—reducing RIC by 9.89 percentage points over NequIP-L (66.68%) and 4.43 percentage points over the state-of-the-art Hamiltonian baseline QHFlow2 (61.22%), with up to 21.15% (1.27×) end-to-end wall-clock savings over MINAO (Table C1). Because strong learned initial guesses already eliminate the rapid initial relaxation steps, further reducing the remaining near-convergence iterations represents a substantial gain.
  2. Accuracy–speed trade-off in property evaluation. While one-shot property evaluation requires single-step DFT-level cost, GEM-FIELD complements existing methods by bridging fast MLIPs and full DFT: it extrapolates substantially better than MLIPs to larger molecules while running much faster than self-consistent DFT. As shown in the updated 51-atom folic-acid force benchmark (Appendix C.4, Table C2), full DFT takes 29.243 s, and MLIPs take 0.0303–0.0720 s but suffer large OOD errors (300.47–333.07 meV/Å); GEM-FIELD (level-0, numerical JJ) takes 11.354 s (61.2% faster than DFT) and reaches 65.04 meV/Å, reducing force error by 78.4% over the best MLIP and by 66.4% over QHFlow2 (193.78 meV/Å in 10.024 s).

Q1. Which properties use the first-diagonalization density matrix, and can properties such as dipoles be evaluated directly from the predicted density?

Response: Spatial multipole moments—specifically the dipole moment reported in Table 2—are already evaluated directly from the predicted continuous density ρ^=ρSAD+Δρ^\hat\rho = \rho_{\mathrm{SAD}} + \widehat{\Delta\rho} prior to any diagonalization. Because each Gaussian shell component is normalized and centered at Cikj=Ri−dikuj\mathbf C_{ikj}=\mathbf R_i-d_{ik}\mathbf u_j with charge qikjq_{ikj}, its first moment is analytically qikjCikjq_{ikj}\mathbf C_{ikj} (Section 4.2 and Appendix G.4, Equation G11). In contrast, orbital energies (HOMO, LUMO, gap), kinetic and total Kohn–Sham energy components, and analytic nuclear forces require molecular orbitals or an idempotent density matrix, and are therefore evaluated from the one-shot diagonalization density matrix D(0)\mathbf D^{(0)} and orbitals (Appendix G.1–G.2).

W4&Q2. Evaluate density prediction on established benchmarks such as the QM9 dataset used by BOA, ELECTRA, and SCDP.

Response: Models trained on prior public density benchmarks such as QM9-VASP are unsuitable for Gaussian-AO solver workflows and fall outside the scope of our study for two reasons: (i) Pseudopotential labels: QM9-VASP uses plane-wave PAW pseudopotentials that omit core electrons, whereas Gaussian-AO Kohn–Sham solvers require all-electron densities to construct the Fock matrix; and (ii) Uniform FFT grids: supervising on uniform periodic FFT grids is mismatched with the nonuniform atom-centered Becke–Lebedev quadratures required to resolve steep near-nuclear radial structure in molecular GTO codes. To align with our solver-oriented theme, we built QM9-PBE on the standard 133,885 QM9 molecular geometries and retrained ChargE3Net, SCDP (including Adam-based exponent fine-tuning), ELECTRA, and BOA under a unified protocol (identical splits, pretraining optimizer, and integrated L1L_1 objective), ensuring that the density supervision labels strictly correspond to the exact same all-electron PBE/def2-SVP electronic-structure problem used for downstream one-shot property recovery, analytic nuclear forces, and SCF initialization (Tables 1–3).

W2&Q3. Does the level-0/level-3 error gap indicate grid overfitting? Why not train directly on level-3 grids with subsampling?

Response: Yes, training on sparse level-0 grids caused mild grid-specific overfitting. Following your suggestion, we trained GEM-FIELD directly on subsampled level-3 grids (1,000 points sampled per molecule per visit, matching the level-0 protocol). Note that PySCF level-0 and level-3 grids are constructed independently from distinct radial/angular quadrature rules rather than one being a subset of the other. Initially, we used level-0 because on-the-fly CPU construction of full level-3 grids was a severe bottleneck due to spatial grid sorting. During the rebuttal, we switched to GPU4PySCF grid generation with spatial sorting disabled, reducing level-3 grid construction time by 11.4× (water) and 24.5× (benzene):

MoleculeGrid levelCPU, default (ms)CPU, no sorting (ms)GPU4PySCF, default (ms)GPU4PySCF, no sorting (ms)
Water01.7210.3771.7941.259
Water322.6831.7872.5791.997
Benzene09.0222.5212.5292.015
Benzene3135.42133.0036.1985.527

This optimization made level-3 training practical (~120 and ~60 RTX 5090 GPU hours for 500k and 300k steps on QM9-PBE and SCFBench). As updated in Tables 1–4, level-3 training virtually eliminates the cross-grid gap:

  • Density accuracy (Table 1): Level-0 and unpruned level-3 NMAEs drop to 0.035% and 0.037% (an 83.6% reduction on the unpruned level-3 grid over the level-0-trained GEM-FIELD model's 0.226%).
  • Downstream properties & SCF (Tables 2–4): ID energy-component MAEs drop by 77.5%–83.8%, ID total-energy MAE drops from 0.61 to 0.22 meV/molecule, OOD energy-component MAEs drop by 36%–58%, and SCFBench ID/OOD RICs improve from 58.40%/59.53% to 52.50%/57.50% (and 51.58%/56.79% for the 50M model).

Response to Minor questions:

MQ1. Clarify which parts of Section 4.1 follow EquiformerV2 and ELECTRA.

Response: We have revised Section 4.1 to state explicitly at the start of "Molecular graph and equivariant features" that we adopt EquiformerV2 as our equivariant backbone, and to cite ELECTRA at the start of "Continuous Gaussian shells" when discussing how local stabilizer symmetries constrain discrete floating Gaussian displacements before introducing our continuous spherical-shell formulation.

MQ2. Why are analytical Coulomb integrals important if numerical construction is faster and achieves similar SCF acceleration?

Response: The analytical Gaussian-shell structure is essential for three reasons:

  1. Exactness and closed-form properties: Analytical 3-center 2-electron integrals (μν∣g)(\mu\nu\mid g) are free of spatial-grid quadrature error (equivalent to an infinite grid) and allow exact closed-form evaluation of multipole moments (Equation G11). Moreover, after replacing legacy auxiliary-molecule construction with direct GPU Gaussian-charge contraction (Appendix C.2), analytical JJ is faster than numerical JJ for the 3.6M level-3 model on SCFBench OOD (8.566 s vs. 8.718 s end-to-end in Table C1, achieving a 1.27× speedup over MINAO).
  2. Our "numerical JJ" itself relies on analytical Gaussian integration: Evaluating ΔJμν\Delta J_{\mu\nu} directly from a real-space density involves a 6D two-electron Coulomb integral. Our numerical branch exploits the analytical Hartree potential vH(r)=q erf⁡(r/σ)/rv_H(\mathbf r) = q\,\operatorname{erf}(r/\sigma)/r of each isotropic Gaussian to reduce this 6D integral to a single 3D quadrature sum (accelerated via FMM).
  3. Efficiency over general grid-contracted JJ construction: For density models that do not expose a standard atom-centered 3-center integral or analytical Hartree-potential interface in GPU4PySCF, in our unified pipeline we construct ΔJμν\Delta J_{\mu\nu} by contracting grid densities with analytical AO-pair integrals at every grid point, ΔJμν≈∑gwgΔρ^(rg)∫ϕμ(r)ϕν(r)/∣r−rg∣dr\Delta J_{\mu\nu} \approx \sum_g w_g \widehat{\Delta\rho}(\mathbf r_g)\int \phi_\mu(\mathbf r)\phi_\nu(\mathbf r)/|\mathbf r-\mathbf r_g|d\mathbf r (Equation 13). As shown below on an RTX 4070 Laptop GPU, our analytical JJ is 11×–13× faster than this general grid-to-AO approach:
MoleculeAtomsAOsGEM-FIELD analytical JJ (s)GEM-FIELD analytical vHv_H + numerical JJ (s)Grid density + analytical AO integrals (s)
Benzene (C6H6\mathrm{C_6H_6})121140.0809 ± 0.00780.5644 ± 0.00091.0911 ± 0.0030
Naphthalene (C10H8\mathrm{C_{10}H_8})181800.2360 ± 0.00600.8204 ± 0.00253.0826 ± 0.0059
Coronene (C24H12\mathrm{C_{24}H_{12}})363961.8799 ± 0.05722.3431 ± 0.058220.9073 ± 0.0627

Dear Reviewer efJq,

Thank you sincerely for the time and care you have devoted to reviewing our manuscript. Your thoughtful feedback has helped us strengthen the work, and we have invested substantial time and effort in additional experiments, analyses, and careful revisions to address your concerns.

As the discussion period draws to a close, we would be very grateful to hear whether our responses and new results have satisfactorily addressed your concerns. We sincerely value the opportunity to exchange views with you during the remaining discussion period.

Thank you again for your thoughtful consideration of our revised work.

Best regards,
The Authors

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

This paper studies how electron densities predicted by machine learning can be used in existing electronic-structure solvers. Given a molecular geometry, GEM-FIELD predicts the density residual relative to a baseline formed by a superposition of atomic densities. It represents the continuous density using Gaussian shells with learnable radii, widths, and angular weights. Normalization and a charge correction enforce the electron count. Nonlinear spherical decoding increases angular expressivity, while the Gaussian expansion permits analytical evaluation of Coulomb integrals. The predicted density can be used to approximate molecular properties or initialize self-consistent field (SCF) iterations. Experiments cover density prediction on QM9-PBE, property extrapolation on OE62, initialization on SCFBench, and transfer across basis sets and functionals. The results indicate a favorable trade-off between density error and computational cost, and fewer SCF iterations for the larger molecules tested.

Critical strengths

S1 The shell representation integrates well with the solver, and ablations support the role of nonlinear decoding Gaussian shells can learn radial and angular density structure while allowing analytical evaluation of Coulomb integrals. Appendices D and E use theoretical analysis and angular spectra to explain how nonlinear decoding generates angular components beyond the backbone cutoff. In Table 6, using the same NequIP backbone with Lmax = 3, nonlinear decoding achieves a dense-grid NMAE of 0.266%, compared with 0.660% for linear decoding. This comparison provides relatively direct evidence for the contribution of nonlinear decoding. (pp. 4–5; Table 6, p. 9; Appendices D and E) S2 The SCF experiments show practical benefits after accounting for initialization costs Table 4 reports convergence for all 1,050 out-of-distribution (OOD) molecules. The relative iteration count (RIC) is 59.49%, compared with 66.68% for the listed auxiliary-density baseline. Appendix C includes inference, initial density-matrix construction, and subsequent SCF iterations in its timing. For the analytical path, the mean OOD end-to-end runtime decreases from 10.888 seconds with MINAO to 9.910 seconds with GEM-FIELD. The numerical J path achieves an aggregate speedup of 1.1667×. The speedup after including initialization costs supports the method’s practical utility. The experiments across basis sets also demonstrate that the interface can be reused. (Table 4, p. 8; Table 5, p. 9; Tables C1 and C2, pp. 17–18)

Critical weaknesses

W1 The distinction from prior work and the contribution of the new shell design need clarification G2Sphere already applies a shared pointwise multilayer perceptron (MLP) to spherical-harmonic fields. It shows that this nonlinearity can generate frequencies beyond the input cutoff while preserving SO(3) equivariance (G2Sphere, §4.2, p. 6). Equations 4–6 and Appendix D of this paper use a related mechanism, but I did not find a discussion of G2Sphere in the references. The relationship should be clarified, with the new contributions identified in the Gaussian shell design, constraints, and solver interface. In addition, the SCDP baseline disables learning of Gaussian exponents, which the original method supports (Related Work, p. 2; Appendix B.1, p. 14). The property experiments also lack results from other density representations passed through the same solver. Tables 1, 2, and 3 therefore do not yet isolate how much of the improvement comes from the new shell design. An SCDP comparison with trainable Gaussian exponents and at least one density model using the same solution procedure would help resolve this question. This mainly affects the assessment of the extent of the contribution. W2 It is unclear whether the SCF baselines are evaluated on the same test set Table 4 cites ID baseline results from SCFBench. The original SCFBench paper (§5.1, p. 7) states that its ID evaluation uses a random 1% subset of the test set. This paper instead reports results for all 4,386 ID molecules (p. 8). Please clarify whether the baselines were rerun on a common set. If not, differences in the test sets could affect the ID ranking. This concern does not directly affect the results on the full OOD set. Appendix C demonstrates an end-to-end speedup over MINAO, but does not report end-to-end timings for the main learning-based baselines on the same hardware. Determining whether a lower RIC translates into a shorter runtime requires timing that includes initial-guess construction.

Questions for authors

Q1 Related to W1 What does the proposed angular mechanism add beyond G2Sphere’s shared pointwise spherical MLP? Why was learning of Gaussian exponents disabled for SCDP, and what are the results when it is enabled? Q2 Related to W2 Were the ID baselines in Table 4 rerun on the same set of molecules? If the results come from the original paper’s 1% subset, please report GEM-FIELD’s results on that subset or reevaluate the methods on a common set. Q3 Related to W2 What are the end-to-end runtimes of NequIP-L and GEM-FIELD on the same hardware, with identical convergence thresholds and solver settings? Please include inference, initial-guess construction, and subsequent SCF iterations, and report variability across repeated timings. This would clarify the method’s practical efficiency relative to existing learning-based baselines. Q4 Related to W1 and numerical reliability Could an existing density model or a linear-decoder variant be evaluated through the same one-shot pipeline to compare downstream property errors?

7 comments
AuthorsOpenReview ↗

Response to Reviewer mHei

Thank you for your time and constructive feedback. We have conducted the requested experiments and updated the manuscript and Appendix accordingly. Below we address each point in detail.

W1&Q1. What distinguishes the angular decoder from G2Sphere, and why were SCDP exponents fixed? Please report results with trainable exponents.

Response: Thank you for pointing out G2Sphere. While our angular decoder was originally motivated by the pointwise spherical nonlinearities in SCN, eSCN, and EquiformerV2, G2Sphere shares a closely related formulation; we have added an explicit citation and discussion in Section 4.1. Crucially, GEM-FIELD goes beyond a standalone spherical MLP by embedding nonlinear angular decoding into continuous Gaussian shells with learnable radii, widths, and signed charges to model 3D all-electron densities. Combined with spherical-mean subtraction and molecule-wide charge correction, our formulation guarantees exact electron-number conservation after Lebedev discretization, while decomposing each shell into displaced isotropic Gaussians that admit analytical three-center Coulomb integrals against target GTO bases. Thus, our core contribution lies in unifying high-frequency angular expressivity with physical charge conservation and an analytical solver interface.

Regarding SCDP, our initial experiment followed the default single-stage training configuration in the official SCDP repository, which keeps Gaussian exponents fixed. Following your suggestion, we performed the official 50,000-step exponent-scaling fine-tuning on the trained SCDP checkpoint and re-evaluated it on the same QM9-PBE test split:

SCDP configurationAdditional fine-tuning stepsLevel-0 NMAE (%) ↓Unpruned level-3 NMAE (%) ↓
Fixed Gaussian exponents00.1070.359
Trainable Gaussian exponents50,0000.1030.319

Trainable exponents moderately reduce SCDP's NMAE on both grids, and we have updated Table 1 (as well as the downstream property evaluations in Tables 2–3) with this fine-tuned checkpoint. Even against fine-tuned SCDP (2.31M parameters), GEM-FIELD at a smaller parameter count (Lmax⁡=4L_{\max}=4, 1.53M) achieves 46.6% lower level-0 NMAE (0.055%) and 25.7% lower unpruned level-3 NMAE (0.237%) while evaluating faster (1.00× vs. 1.73×). We also added the SAD-only baseline and a level-3-trained GEM-FIELD (Lmax⁡=6L_{\max}=6) model to Table 1, which further reduces level-0 and unpruned level-3 NMAEs to 0.035% and 0.037%, respectively.

W2&Q2. Are Table 4 ID results evaluated on the same molecules? Please align comparisons on a common set, including the reported 1% subset.

Response: Thank you for catching this discrepancy. Our original manuscript reported GEM-FIELD on the full 4,386-molecule ID test set, whereas the SCFBench baselines were evaluated on a 1% ID subset. Following the exact sampling protocol of the public SCFBench evaluation code, we re-evaluated GEM-FIELD (and the newly added QHFlow2 baseline) on the same fixed 1% ID subset across PBE/def2-SVP (Table 4) and all cross-functional and cross-basis settings (Table 5). On this aligned 1% ID subset at PBE/def2-SVP, GEM-FIELD (level-0), GEM-FIELD (level3), and GEM-FIELD-L (level3) all maintain 100% convergence with RICs of 58.40%, 52.50%, and 51.58%, consistently outperforming NequIP-L (63.78%) and QHNet (63.20%). All rankings and conclusions in Tables 4 and 5 remain unchanged.

W2&Q3. Compare NequIP-L and GEM-FIELD inference, initial-guess construction, SCF, and total runtimes under identical hardware, convergence thresholds, and solver settings.

Response: To provide a fair end-to-end runtime comparison, we first optimized GEM-FIELD's initial-guess construction by reusing shared shell parameters, fusing density and spatial-gradient evaluation into a single GPU kernel, and applying distance-based Gaussian screening. For the numerical JJ path, we further introduced a range-separated Hartree potential combined with the Fast Multipole Method (FMM):

=\sum_i q_i\frac{\operatorname{erf}(\lVert\mathbf r-\mathbf R_i\rVert/\sigma_i)} {\lVert\mathbf r-\mathbf R_i\rVert} =\sum_i q_i\left[ \frac{1}{\lVert\mathbf r-\mathbf R_i\rVert} -\frac{\operatorname{erfc}(\lVert\mathbf r-\mathbf R_i\rVert/\sigma_i)} {\lVert\mathbf r-\mathbf R_i\rVert} \right],$$ where the long-range $1/r$ sum is evaluated via FMM and the short-range $\operatorname{erfc}$ correction is truncated locally on the GPU. We then benchmarked NequIP-L, QHFlow2, and GEM-FIELD across all **1,050 SCFBench OOD molecules** on identical hardware (**one NVIDIA RTX 5090 GPU with 8 CPU threads**) and identical GPU4PySCF solver settings (**restricted Kohn–Sham, PBE/def2-SVP, level-3 grids, energy convergence tolerance $10^{-9}\text{ Ha}$, and at most 50 SCF iterations**). The results are updated in Appendix C (Table C1): | Initialization scheme | Method | Inference (s) | Initial-guess construction (s) | SCF (s) | Total wall-clock (s) | MINAO time (s) | Aggregate speedup | |:---|:---|---:|---:|---:|---:|---:|---:| | NequIP-L (50M) | – | 0.045 | 0.622 | 7.455 | 8.123 | 10.862 | **1.34×** | | QHFlow2 (20.5M) | – | 0.137 | 1.299 | 7.117 | 8.552 | 10.875 | 1.27× | | GEM-FIELD (3.6M) | analytical | 0.024 | 2.225 | 6.710 | 8.959 | 10.861 | 1.21× | | GEM-FIELD (3.6M) | numerical | 0.025 | 2.132 | 6.706 | 8.863 | 10.862 | 1.23× | | GEM-FIELD (3.6M, level3) | analytical | 0.023 | 2.004 | 6.539 | 8.566 | 10.864 | 1.27× | | GEM-FIELD (3.6M, level3) | numerical | 0.025 | 2.147 | 6.546 | 8.718 | 10.875 | 1.25× | | GEM-FIELD-L (50M, level3) | analytical | 0.030 | 2.136 | 6.505 | 8.672 | 10.903 | 1.26× | | GEM-FIELD-L (50M, level3) | numerical | 0.031 | 2.128 | **6.483** | 8.642 | 10.865 | 1.26× | Because NequIP-L predicts coefficients in a fixed pre-tabulated auxiliary basis (def2-universal-jfit), its initial-guess construction is faster (0.622 s vs. 2.004 s), giving it a slightly higher end-to-end speedup (1.34× vs. 1.27×, or a 4% difference in wall-clock reduction: 25.2% vs. 21.2%). However, with $14\times$ fewer parameters (3.6M vs. 50M), GEM-FIELD (level3) achieves **shorter subsequent SCF runtimes** (6.539–6.546 s vs. 7.455 s for NequIP-L, and **6.483–6.505 s** for the 50M GEM-FIELD-L) due to its closer-to-ground-state density, and transfers substantially better to larger AO bases (def2-TZVP/TZVPPD/QZVP in Table 5, lowering RIC by **3.69–4.80 percentage points on ID** and **6.25–7.09 percentage points on OOD** over NequIP-L) while also supporting accurate one-shot property recovery.

W1&Q4. Compare downstream property errors of an existing density model or linear-decoder variant through the same one-shot pipeline.

Response: We evaluated the fine-tuned SCDP, BOA, and ELECTRA models through the exact same one-shot Fock-diagonalization pipeline on both QM9-PBE (ID) and OE62 (OOD) and added all results to Tables 2 and 3. Because these baselines do not expose a standard atom-centered three-center Coulomb integral interface in GPU4PySCF (ELECTRA uses floating anisotropic Gaussians, BOA applies nonlinearities to GTO products, and SCDP introduces bond-centered functions with scaled exponents), in our unified pipeline we construct their residual Coulomb matrices by contracting predicted grid densities with analytical AO-pair electrostatic integrals Aμν(rg)A_{\mu\nu}(\mathbf r_g):

=\int \widehat{\Delta\rho}(\mathbf r') A_{\mu\nu}(\mathbf r') d\mathbf r' \approx\sum_g w_g \widehat{\Delta\rho}(\mathbf r_g) A_{\mu\nu}(\mathbf r_g)$$ where $$A_{\mu\nu}(\mathbf r') =\int\frac{\phi_\mu(\mathbf r)\phi_\nu(\mathbf r)} {\lVert\mathbf r-\mathbf r'\rVert} d\mathbf r$$ This general quadrature-based $J$ construction is substantially slower than GEM-FIELD's analytical Gaussian-shell integrals and FMM-accelerated numerical branch, as shown on an RTX 4070 Laptop GPU (mean $\pm$ std over 3 runs after warm-up): | Molecule | Atoms | AOs | GEM-FIELD analytical $J$ (s) | GEM-FIELD analytical $v_H$ + numerical $J$ (s) | Grid density + analytical AO integrals (s) | |:---|---:|---:|---:|---:|---:| | Benzene ($\mathrm{C_6H_6}$) | 12 | 114 | **0.0809 ± 0.0078** | 0.5644 ± 0.0009 | 1.0911 ± 0.0030 | | Naphthalene ($\mathrm{C_{10}H_8}$) | 18 | 180 | **0.2360 ± 0.0060** | 0.8204 ± 0.0025 | 3.0826 ± 0.0059 | | Coronene ($\mathrm{C_{24}H_{12}}$) | 36 | 396 | **1.8799 ± 0.0572** | 2.3431 ± 0.0582 | 20.9073 ± 0.0627 | More importantly, under the same level-0 training grid and one-shot solver pipeline (Tables 2–3), GEM-FIELD (level-0) substantially outperforms SCDP, BOA, and ELECTRA on downstream properties and OOD transfer: - **In distribution (QM9-PBE)**: GEM-FIELD (level-0) achieves 0.61 meV/molecule total-energy MAE, 4.81 meV/Å force MAE, 4.23 meV HOMO–LUMO gap MAE, and 0.0083 D dipole MAE, outperforming the strongest density baseline SCDP (1.57 meV/molecule, 12.18 meV/Å, 9.90 meV, and 0.0351 D) as well as BOA and ELECTRA on total energy, forces, all frontier-orbital energies, and dipoles (while SCDP is lower only on the four individual uncancelled ID energy components, and level-3 training further lowers GEM-FIELD's ID errors to 0.22 meV/molecule, 3.21 meV/Å, and 3.23 meV). - **Out of distribution (OE62)**: While SCDP, BOA, and ELECTRA degrade sharply on larger molecules (OOD energy MAEs of 175.98–18,162.53 meV/molecule, force MAEs of 164.44–1,194.53 meV/Å, and dipole MAEs of 26.36–93.96 D), GEM-FIELD (level-0) outperforms all three baselines across every OOD metric (13.32 meV/molecule energy MAE, 24.65 meV/Å force MAE, and 0.1830 D dipole MAE), demonstrating the advantage of our charge-conserving Gaussian-shell representation under a shared supervision and solver pipeline. Finally, we added one-shot HOMO–LUMO gap errors (evaluated on a 1% random test subset) to the ablation study in Table 6. Because orbital energies have a first-order dependence on the Fock matrix, they directly reflect how density accuracy propagates to downstream observables: across backbones and angular cutoffs ($L_{\max}=3\rightarrow 6$), lowering density NMAE (0.117% → 0.037%) shows a clear overall reduction in the gap MAE (22.87 → 4.60 meV) and RMSE (29.23 → 6.76 meV).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.