acceptodds
Under review as a conference paper at ICLR 2027

Neural-Network Solutions to Real-Space Charge Density and Generalization

Abstract

The Hohenberg-Kohn theorem establishes that, in principle, the ground state (GS) charge density contains all GS information of a many-electron system, such that all GS observables can be expressed as functionals of the GS charge density. Conventional Kohn-Sham density functional theory requires iterative solution of the self-consistent-field equations at substantial computational cost, motivating the development of deep learning surrogates for electronic structure calculations and, in turn, accelerating computer-aided materials design. Here, we propose AIDEN, an Atomic-Interaction Density Equivariant Network for solving real-space charge density. AIDEN separates the element-dependent one-center density from environment-induced density redistribution and represents the latter through complementary atom- and edge-centered tensor correlations. A continuous low-rank Gaussian decoder then reconstructs the density at arbitrary spatial coordinates while reusing atomic encodings independently of the evaluation grid. AIDEN achieves state-of-the-art accuracy on periodic crystal benchmarks while remaining competitive for molecular systems, and further demonstrates zero-shot transferability across several structurally distinct out-of-distribution case studies. Furthermore, AIDEN provides substantially faster inference than both baseline models and full SCF calculations, enabling efficient charge density reconstruction for large-scale electronic structure calculations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Graph Neural Network, Density Functional Theory, AI for Science, Computational Physics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_rhxnOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces a novel charge density prediction model, AIDEN, and evaluates it against the state-of-the-art on crystal and molecular benchmarks. The new architecture is constructed via a combination of Cartesian atomic cluster expansion and tensor edge cluster expansion message passing, as well as a physically motivated split between atomic and (chemical) environment density contributions. The model is then first evaluated on the ECD and QM9 benchmarks, beating SOTA on the crystal benchmark and achieving second place on the molecular one. Further testing is conducted via non-SCF calculations from the predicted charge density, showing that AIDEN generally matches or outperforms ChargE3Net on density and downstream energy/force accuracy. The model also demonstrates faster inference than ChargE3Net.

Critical strengths

Density split and physical interpretations:

The authors' decision to split the density is both intuitive and well argued for throughout the paper. The ablation performed to test the individual contributions of the architecture shows that GTOs from the local, atomic embeddings provide a useful bias that the chemical environment GTOs cannot capture. This, and other architectural choices, are well discussed from a physical perspective, giving the reader a clear understanding of why certain choices were made.

Inference efficiency:

The inference speedups over ChargE3Net are meaningful, but not novel in themselves (see below)

Critical weaknesses

Overall, the main weakness of the paper is the limited significance given the current state of the field. Marginal improvements in benchmark accuracy is not a sufficiently compelling contribution, particularly when density reconstruction and measurable SCF acceleration have already been demonstrated more convincingly in other work. At this point, new methods should instead demonstrate either substantially new capabilities or meaningful improvements in practical electronic-structure workflows, rather than primarily optimizing benchmark accuracy. This paper does not convincingly establish either, making its practical value and significance difficult to justify.

Lacking evaluations:

While the authors test on both ECD & QM9 and perform some non-SCF calculations, the results as presented are not entirely convincing. AIDEN is shown to outperform ChargE3Net by around 10–20% on PBE & HSE, but testing the models on Materials Project data, possibly even on the new r2SCAN data, would have provided more substantial evidence that AIDEN is a stronger model. This is compounded by the fact that ELECTRAFI [1] reports an HSE NMAE of 1.35%, outperforming AIDEN when trained from scratch (1.4307%). AIDEN does achieve a better result through PBE fine-tuning (1.3012%), but these are different training protocols and should be distinguished when discussing SOTA performance.

Furthermore, the OOD comparison uses the public Materials Project PBE-trained ChargE3Net checkpoint, whereas AIDEN is trained on the ECD PBE data. As the models have been trained on different datasets, this is not a controlled comparison of the two architectures, making it difficult to attribute the observed improvements to AIDEN itself.

The non-SCF experiments help, but they are also limited to a few structure types and are not a very applicable benchmark. Earlier models such as ChargE3Net and ELECTRA/ELECTRAFI studied the SCF savings when using the model initialization, a test lacking in this paper. While it is not the only density application, it is still one of the main motivations of charge density prediction for large-scale DFT applications. An experimental study of SCF iterations and total wall-clock savings, including inference time, would have allowed the reader to understand the practical impact of AIDEN.

Therefore, as the model is not tested broadly enough on different materials or in end-to-end DFT workflows, it is difficult for a potential user to justify using AIDEN over well-established and broadly tested models such as ChargE3Net.

Limited practical novelty or usefulness:

The paper's contribution is primarily architectural, improving density prediction accuracy and inference efficiency without substantially extending the applications enabled by previous models. Neither spin-difference densities nor augmentation occupancies are modeled, which limits the practical impact.

While most previous papers have done the same (ChargE3Net or ELECTRA), both mentioned quantities are known limitations of density models, and this is now a known limitation which has been mentioned in prior work [1,2]. In particular, the authors state in Appendix C.4.2 that the remaining VASP restart quantities are taken from the matched PBE-SCF reference calculation. This means that the downstream results still rely on information from a previously converged calculation. Adding augmentation occupancies would have enabled non-spin-polarized PAW calculations without requiring reference occupancies, making the model significantly more useful as an independent electronic-structure initialization.

The architectural improvements are interesting, but without demonstrating additional practical capabilities or end-to-end computational savings, the contribution is incremental, and recent work such as [1] and [2] have discussed these issues. I also find it generally concerning that [1] is omitted from citations in the related work and experimental comparisons. Particularly because ELECTRAFI has already demonstrated extremely fast full-grid periodic density reconstruction on its reported benchmarks (in the hundreths of a second), as well as end-to-end DFT wall-clock savings when used for SCF initialization. In contrast, AIDEN reports full-grid inference times of 4.34-23.70 seconds on its OOD systems. Faster inference relative to ChargE3Net alone is therefore insufficient to establish an advance or enough applicable novelty over the current state of the art in efficient density prediction.

[1] Elsborg, Jonas, et al. "Global Plane Waves From Local Gaussians: Periodic Charge Densities in a Blink." (ICML 2026), arXiv preprint arXiv:2601.19966 (2026).

[2] Li, Xiwen, et al. "Efficient E (3)-equivariant framework for universal charge density prediction." arXiv preprint arXiv:2510.00788 (2025).

Questions for authors
  1. Appendix C.4.2 states that the remaining VASP restart quantities are taken from the matched PBE-SCF reference calculation. Does this include the PAW augmentation occupancies? Could you clarify exactly which reference quantities are reused, and whether the non-SCF results could be reproduced without them?

  2. Have you evaluated the SCF iteration savings and total wall-clock savings when using AIDEN as an initialization, compared with SAD and ChargE3Net? Given the substantial inference speedups, this would be particularly interesting for establishing the practical value of the model.

5 comments

We thank the reviewer for carefully reading our work. We have added relevant experiments and address the comments below.

Comment #1: AIDEN's overall contribution is limited.

Reply: AIDEN's contributions span scientific insight, practical computational value, model design, and transferability:

  1. Scientific insight: AIDEN examines not only overall charge-density prediction errors but also differences between models in downstream electronic properties. NMAE and downstream property errors need not have a monotonic relationship. We analyze this using local errors, electron–ion interactions, and error cancellation; see Appendix C.6.5.
  2. Practical computational value: Across several systems excluded from training, the predicted densities can be used for fixed-density energy and force calculations while substantially reducing density-reconstruction costs. High accuracy and efficiency in generalization to complex systems are practically important. Consider charge-density-driven molecular dynamics (MD): deep-learning surrogate densities have already been used to drive MD [J. Chem. Phys. 161, 024109 (2024)], calculate infrared spectra [J. Chem. Phys. 163, 174104 (2025)], and simulate electrochemical interfaces [JACS Au (2025) 5 (12): 5939–5947]. For 10,000 MD configurations of 64-atom a-Si, the inference speeds reported here imply approximately 12 hours for AIDEN versus approximately 38 days for ChargE3Net. These timings were measured on an L20 GPU; the original manuscript mistakenly stated A100, and we have corrected this. The L20 results also illustrate the importance of efficiency: efficient deep-learning models may reduce users' dependence on high-performance computing GPUs and bring the cost of complex tasks within an acceptable range. ELECTRAFI is certainly a relevant example, which we discuss in more detail below.
  3. Methodological insight: The ablations show that the physically motivated atomic-density decomposition and continuous basis-function design have a more pronounced effect on accuracy than simply increasing network complexity. The central feature of the EAC-Net baseline is its "equivariant atomic contributions" (EAC), while we go further by physically motivating and modeling the environment-dependent term.
  4. Transfer value: PBE pretraining improves HSE density learning, providing evidence that data from inexpensive functionals can support learning with more expensive functionals.

Comment #2: AIDEN should be evaluated on the MP/r2SCAN datasets.

Reply: The MP dataset is also calculated using the PBE functional. We have added an evaluation of AIDEN on r2SCAN; see Appendix C.2. There, we briefly describe the different functional calculations and evaluate zero-shot inference using AIDEN's PBE-pretrained weights, training from scratch, and fine-tuning from the PBE weights on r2SCAN. Interestingly, the r2SCAN results follow a similar trend to HSE transfer learning: zero-shot inference directly using the PBE-pretrained weights has limited accuracy, while training from scratch and fine-tuning from PBE achieve best validation NMAEs of 0.8515% and 0.7855%, respectively. Fine-tuning gives the lower error, further demonstrating the value of transferring models pretrained on inexpensive functionals.

Comment #3: The HSE state-of-the-art claim should explicitly distinguish fine-tuning.

Reply: We thank the reviewer for their careful reading! In lines 367-369 of the revised manuscript, we explicitly state that the HSE state-of-the-art result is achieved through fine-tuning.

Comment #4: The OOD comparison uses a publicly released ChargE3Net checkpoint trained on Materials Project PBE data, whereas AIDEN is trained on ECD-PBE data.

Reply: The original description of the ChargE3Net weights' training source was incorrect. The weights used in the OOD experiments were also trained on ECD-PBE. We have clarified this in Appendix C.6.2, lines 2017-2018, of the revised manuscript.

Comment #5: The non-self-consistent-field experiments are useful, but they cover only a few types of structures and do not constitute a benchmark with broad practical applicability.

Reply: Our chosen test systems cover disordered molecular systems, an amorphous elemental material, an alloy, and twisted structures. Our choices of water, AlMg, and a-Si follow the examples in EAC-Net, while TBG is also widely used in electronic-structure studies. We first evaluate on general datasets with established benchmarks and then examine generalization to representative complex systems. These experiments demonstrate AIDEN's generalization capability and practical value in complex systems, including integration with MD as discussed above.

Comment #6: AIDEN does not predict spin-difference densities or augmentation occupancies.

Reply: We regard spin densities and PAW augmentation occupancies as interesting future directions, but they are not the current focus or main bottleneck of this work. The first issue is the data and the physical calculations used to generate them. These labels depend on the DFT settings: changing Hubbard-U for the same structure can substantially alter its calculated spin density, while there is still no unified way to determine appropriate U values across different materials. The key issue therefore lies in the physical calculations and label definitions; changing the network architecture alone does not resolve it.

Existing studies have already explored learning these quantities. Earlier spin-density studies mainly focused on specific systems, such as fcc Fe [J. Chem. Theory Comput. 22, 4813–4821 (2026)] and NaxV2(PO4)3\mathrm{Na}_x\mathrm{V}_2(\mathrm{PO}_4)_3 [Energy Storage Materials 63, 103023 (2023)]. More recently, AugNet arXiv:2609.21759 has extended prediction of spin densities and PAW augmentation occupancies to a broader range of materials, while EdenGNN also learns PAW augmentation occupancies DijaD^a_{ij}. This direction is still new, so we do not include these latest methods in our quantitative comparisons in this revision; the ICLR reviewer guidelines also do not require comparisons with contemporaneous work or works available only as preprints. We will continue to follow these developments and explore their integration with AIDEN.

Finally, smooth pseudo-charge densities themselves have clear physical and computational value. Beyond the fixed-density energy, force, and band calculations and SCF initialization demonstrated here, they support electrostatic dipole calculations, electron-transfer analysis, and bonding analysis [J. Phys. Chem. C (2023) 127 (48): 23459–23466]. Improving their prediction accuracy and reconstruction efficiency directly serves these practical applications, which are central to AIDEN's current work.

Comment #7: ELECTRAFI demonstrates faster inference than AIDEN.

Reply: ELECTRAFI is a valuable contribution. We have reproduced the model and evaluated its performance; see Appendix C.4. However, we also find that ELECTRAFI's KeOps version is indeed substantially faster. This probably reflects better engineering optimization and a more efficient computational-library implementation. We therefore do not directly compare it with Torch-based AIDEN to assess architectural efficiency. Instead, we reproduce its architecture with Torch, ensuring a forward-pass discrepancy below 1e-7, and analyze the models under the same computational library. Under this setting, ELECTRAFI appears less efficient than AIDEN, with inference cost growing more rapidly as the number of atoms increases. Most systems used for inference in the ELECTRAFI study are relatively small. Its engineering optimization is informative, and our own attempt to reconstruct AIDEN with KeOps did not yield a substantial efficiency improvement. We therefore cite and discuss ELECTRAFI in both the main text and the appendix as a useful reference for future work.

Comment #8: Regarding SCF acceleration and practical computational savings, have the authors evaluated how many SCF iterations and how much total wall-clock time can be saved by using AIDEN as an initial guess, relative to SAD and ChargE3Net? Given AIDEN's substantial inference speedup, such experiments would be particularly useful for demonstrating its practical value.

Reply: We thank the reviewer for this insightful question. We have added a subsection, "Deep-learning charge densities as initial guesses for SCF," to systematically evaluate SCF iteration counts and end-to-end computational costs. We consider 97 conventional materials containing 1–132 atoms from ECD-PBE (three of the 100 selected structures are excluded due to convergence failures) and three TBG systems containing 28, 76, and 148 atoms. We compare SAD, converged self-consistent (SC) densities, AIDEN, ChargE3Net, and ELECTRAFI (Torch and KeOps), recording model loading, density inference, PAW merging, VASP execution, and SCF iterations. Across all 100 evaluated systems, AIDEN achieves the lowest aggregate computational cost among the predictive models, reducing total time by 6.51% relative to SAD, whereas ChargE3Net and ELECTRAFI-KeOps increase it by 7.02% and 13.91%, respectively.

Average computational costs for the 97 conventional materials:

Initial guessLoad [s]Infer. [s]PAW merge [s]VASP [s]Total [s]vs. SAD
SAD0.0000.0000.000812.381812.3810.00%
SC0.0000.0000.181469.155469.336-42.23%
AIDEN0.6062.7990.399730.423734.226-9.62%
ChargE3Net14.287161.7650.395700.649877.096+7.97%
ELECTRAFI-Torch1.8844.4240.407721.753728.467-10.33%
ELECTRAFI-KeOps1.6770.4230.407721.753724.260-10.85%

Average computational costs for the three TBG systems:

Initial guessLoad [s]Infer. [s]PAW merge [s]VASP [s]Total [s]vs. SAD
SAD0.0000.0000.00011185.66711185.6670.00%
SC0.0000.0000.3737015.3337015.706-37.28%
AIDEN0.6066.6591.28411265.66711274.216+0.79%
ChargE3Net14.287305.9711.22811401.00011722.485+4.80%
ELECTRAFI-Torch1.88419.2411.33919241.66719264.131+72.22%
ELECTRAFI-KeOps1.6771.1591.33919241.66719245.842+72.06%

On conventional materials, AIDEN reduces total time by 9.62%, close to ELECTRAFI-KeOps at 10.85%, whereas ChargE3Net incurs a 7.97% slowdown due to its higher inference cost. On TBG, none of the predictive models achieves a net speedup, but AIDEN shows the smallest overhead (+0.79%), compared with ChargE3Net (+4.80%) and ELECTRAFI (+72%). Notably, ELECTRAFI's slowdown arises primarily from VASP execution rather than inference, illustrating that fast density prediction alone does not guarantee SCF acceleration.

We further compare the corresponding SCF iteration counts:

Initial guessConventional: steps ↓ΔNvs. SADTBG: steps ↓ΔNvs. SAD
SAD19.0930.0000.00%24.0000.0000.00%
SC10.608-8.485-44.44%13.333-10.667-44.44%
AIDEN17.454-1.639-8.59%22.667-1.333-5.56%
ChargE3Net16.866-2.227-11.66%22.000-2.000-8.33%
ELECTRAFI-Torch17.433-1.660-8.69%30.333+6.333+26.39%
ELECTRAFI-KeOps17.433-1.660-8.69%30.333+6.333+26.39%

AIDEN reduces SCF iterations by 8.59% on conventional materials and 5.56% on TBG, compared with 11.66% and 8.33% for ChargE3Net. ELECTRAFI achieves an 8.69% reduction on conventional materials but a 26.39% increase on TBG, while converged SC densities reduce iterations by approximately 44.44% in both groups. These findings agree with previous observations that density accuracy and SCF convergence are not necessarily monotonically related (ELECTRAFI; Chen et al., 2026). SCF is a nonlinear fixed-point iteration rather than simple energy-gradient descent, and its convergence depends on electronic response, density mixing, and the underlying eigensolver. Our results therefore highlight two complementary applications of AIDEN: accelerating conventional SCF through efficient density initialization and directly predicting approximate ground-state densities for fixed-density NSCF calculations. The latter avoids dependence on SCF iteration trajectories, with its reliability assessed through the energy, force, and band-structure errors reported in our manuscript.

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

The paper introduces AIDEN for predicting continuous real-space electronic charge densities from atomic configurations. It combines Cartesian atomic cluster expansion with tensor edge cluster expansion and radial rotary attention, then reconstructs densities using a low-rank Gaussian decoder with an element-dependent baseline and an environment-dependent correction. Evaluation includes ECD-PBE/HSE, QM9, and zero-shot case studies involving water, AlMg, amorphous silicon, and twisted bilayer graphene, including downstream fixed-density calculations and inference timings.

Critical strengths
  • Meaningful improvements on a chemically diverse crystal benchmark. AIDEN reduces ECD-PBE NMAE from 0.523% to 0.4542% relative to ChargE3Net at (L=4), and achieves 1.3012% on HSE after fine-tuning, compared with 1.534% for ChargE3Net. These are useful improvements on a substantial materials dataset and constitute the strongest evidence supporting acceptance.

  • Practical evaluation beyond density reconstruction. The paper evaluates full-grid inference and downstream energies, forces, and band structures. Reusing atomic encodings across spatial queries is computationally attractive, and the reported implementation demonstrates substantial acceleration on grids containing millions of points. The downstream calculations also provide a more informative assessment than density error alone.

Critical weaknesses
  • The incremental contribution over closely related methods is insufficiently established. Cartesian ACE follows TACE, while TECE and RRA originate in the cited TECE work. More consequentially, BOA already combines a quadratic GTO density representation with a pretrained element-dependent density initial guess that remains trainable, as well as learned radial corrections. AIDEN’s accumulation of atomic fields before multiplication differs from BOA’s atom-pair factorization, but the resulting representational or computational advance is not sufficiently demonstrated. The component ablations establish usefulness, yet do not isolate the claimed ACE–TECE complementarity against ACE-only and TECE-only encoders sharing the same decoder. Integration can constitute a valuable contribution; here, the evidence does not clearly establish what new capability or insight requires this particular combination.

  • The comparative evidence is too limited to compensate for this incremental novelty. On ECD-PBE, the accuracy comparison includes only eqDeepDFT and ChargE3Net; HSE includes only ChargE3Net. Close alternatives such as BOA, SCDP, ELECTRA, and EAC-Net are absent from the central crystal comparison. EAC-Net is particularly relevant because it already supports atom-centered reconstruction with cached atomic descriptors. On QM9, AIDEN trails BOA, with NMAE of 0.1628% versus 0.1339%. The OOD examples demonstrate transferability but provide mixed evidence of comparative advantage: ChargE3Net gives lower water density error and substantially lower AlMg energy and force errors despite worse density NMAE. Finally, the measured speedups against ChargE3Net’s official implementation do not isolate architectural gains from caching and implementation choices. Post-acceptance code release further limits independent verification of this implementation-sensitive result.

Questions for authors
  • What specific advance does the decoder provide over BOA’s quadratic representation and pretrained atomic density initial guess? In particular, what benefit follows from multiplying globally accumulated atomic fields rather than using atom-pair factors?

  • Can the claimed encoder complementarity be assessed using ACE-only and TECE-only variants with the same decoder and comparable parameter budgets? Reversing the layer order does not establish whether combining both interaction types is necessary.

  • Can you provide a comparison against at least one close atom-centered competitor on ECD, together with inference timings at comparable accuracy and a breakdown of atomic encoding versus grid evaluation?

4 comments

We thank the reviewer for carefully reading our manuscript and address the main questions below.

Comment #1: There is no quantitative assessment of the roles of TACE, TECE, RRA, GTO, and other components. It remains unclear which new capabilities or insights specifically require this combination.

Reply: Table 3 in Appendix C.1 of the original manuscript already quantifies the effects of these components, specifically how AIDEN's parameter count and error change when a component is removed. It includes the settings without RRA and with fixed GTOs mentioned by the reviewer. In the revised manuscript, we have added results for removing Cartesian ACE or TECE and replacing GTOs with a pure MLP; see Table 3 in Appendix C.1 of the revised manuscript.

Comment #2: ChargE3Net has a higher overall charge-density NMAE, yet it has a lower charge-density error for water and substantially lower energy and force errors for AlMg. Why?

Reply: This issue is discussed in Appendix C.4.5 of the original manuscript and has also been reported in related studies. NMAE quantifies average performance over a general dataset of crystal unit cells. It indicates which model performs better on average over a large sample, without guaranteeing the same ranking for every structure. Water and AlMg are disordered, complex OOD systems not seen during training, so a reversal in their NMAE ranking is not surprising. More importantly, a lower charge-density NMAE does not necessarily imply lower energy and force errors, as discussed in [Nat. Commun., 16(1):4811, 2025.] and [J. Chem. Theory. Comput., 17(11):7203–7214, 10 2021.]. NMAE measures only the global integrated absolute charge-density error, whereas energy and force errors also depend on its spatial distribution, sign, and cancellation among different physical contributions. As analyzed in Appendix C.4.5, although AIDEN has a lower global NMAE for AlMg, ChargE3Net has a lower integrated density error within 0.6 Å of the atoms (0.1520 vs. 0.2285 e), consistent with its lower energy and force errors. For water, we further resolve force errors by element and find that AIDEN has a substantially lower error for H atoms (0.08370 vs. 0.10303 eV/Å). This improvement is sufficient to offset its slight disadvantage for O atoms, resulting in a lower average force error despite its slightly higher overall density NMAE. Global density accuracy, local density error, and downstream property accuracy therefore characterize different aspects of prediction quality and need not have a monotonic relationship.

Comment #3: The reported speedup over the official ChargE3Net implementation does not distinguish architectural gains from the effects of caching and implementation choices.

Reply: The ChargE3Net results come from our own reproduction, rather than directly adopting the published numbers. Reviewer #1 (K8Jp) raised a similar question. In the additional experiments, we again use inference on TBG structures with three twist angles. Besides ChargE3Net, we include an efficient KeOps-based charge-density baseline. We divide inference into graph construction, encoding, decoding, and full-grid density output, quantifying each stage separately. Incidentally, it is interesting that most deep-learning models appear to admit this general four-stage description. The results are shown in Figure 15 of Appendix C.4. However, we also find that ELECTRAFI's KeOps version is indeed substantially faster, probably because of better engineering optimization and a more efficient computational-library implementation. We therefore do not directly compare it with Torch-based AIDEN to assess architectural efficiency; our discussion of architectural efficiency primarily uses results obtained with the same Torch library.

omment #4: Code release.

Reply: Thank you for the reminder! The code and data supporting this study are publicly available at https://anonymous.4open.science/r/F10T/.

Comment #5: BOA already has a quadratic density representation and a pretrained initial estimate of atomic densities. In particular, what is the benefit of globally summing the atomic fields before multiplication, compared with using atomic-pair factors?

Reply: We thank the reviewer for this technically informed question! AIDEN incorporates a modified design because the original BOA-style implementation did not perform satisfactorily in our experiments, as demonstrated by the additional experiment in Appendix C.1, lines 1493-1526. We design three decoders that retain the same GNN encoder and pretrained initial atomic-density estimate, changing only how the environment-dependent density is constructed:

  1. Global field (AIDEN): sum globally before multiplication, ρ=ρ0+∑kA(∑iLik)∑jRjk\rho=\rho_0+\sum_k A(\sum_i L_{ik})\sum_j R_{jk}.
  2. Atom multiply → sum: multiply atom by atom before summation, ρ=ρ0+∑i,kA(Lik)Rik\rho=\rho_0+\sum_{i,k}A(L_{ik})R_{ik}.
  3. Pair factor (BOA-style): explicitly construct atomic-pair factors, ρ=ρ0+∑i<jA(uij)vij∑pϕipϕjp\rho=\rho_0+\sum_{i < j}A(u_{ij})v_{ij}\sum_p\phi_{ip}\phi_{jp}.

Here, AA is the smooth absolute-value function, LikL_{ik} and RikR_{ik} are atom-centered GTO expansion fields, and uiju_{ij} and vijv_{ij} are independently predicted atomic-pair factors. We train from scratch for 100,000 steps using 5,000 training structures and 1,000 validation structures from the PBE dataset, keeping the model parameter counts nearly identical.

DecoderParametersBest val. NMAE ↓Training time (s/step) ↓Peak GPU memory (GiB) ↓
Ours2.159M0.0099550.6244.71
Atom multiply → sum2.159M0.0102450.5384.70
BOA-style2.159M0.0286370.85638.95

The results show that changing only the order of summation and multiplication further reduces NMAE, supporting the effectiveness of cross-atom coupling introduced by the global-field product. In contrast, the explicit atomic-pair formulation has 2.88 times AIDEN's NMAE and 8.26 times its GPU memory cost. These results indicate that global-field aggregation can effectively describe quadratic interatomic coupling while avoiding the substantial computational and storage costs of explicitly constructing atomic pairs, thereby providing substantial advantages in prediction accuracy and computational efficiency. Incidentally, we regard BOA primarily as a physics-inspired GNN. Its distinctive feature is the use of atomic-orbital basis coefficients to represent each atom, whereas conventional atomistic GNNs typically use one-hot elemental encodings. It performs message passing and attention computation through orbital overlaps, which incurs substantial computational costs. Nevertheless, we believe BOA-style models may have broader value for learning basis-dependent electronic-structure quantities such as density matrices and Hamiltonians.

Comment #6: Can the claimed encoder complementarity be assessed using pure ACE and pure TECE models with the same decoder and comparable parameter counts? Merely reversing the order of the two modules does not establish that combining them is necessary.

Reply: We address this comment together with Comment #1. The additional experiments in Appendix C.1 give the following results.

Module ablation results. Parameter counts and best validation NMAE ερ\varepsilon_{\rho} for models trained on the same 2,000-structure PBE subset. Relative changes are calculated with respect to the full ACE→TECE model.

Ablation settingParameters ↓Best val. ερ\varepsilon_{\rho} (%) ↓Relative to Full (%) ↓
Full ACE→TECE4.459M1.13151—
ACE-only4.503M1.11814-1.18
TECE-only4.564M1.32051+16.70
TECE→ACE4.459M1.14572+1.26
w/o GIE4.393M1.14622+1.30
w/o RRA4.205M1.16120+2.62
Fixed GTO4.454M1.21505+7.38
w/o ρinit\rho_{\rm init}4.457M1.53550+35.70

We further evaluate the training cost of using either ACE or TECE alone as the encoder under comparable parameter counts. As shown in the module ablation results, ACE-only appears to be the only configuration that achieves a lower NMAE than the original ACE→TECE design. However, the training efficiency results show that, despite having a similar parameter count, ACE-only requires approximately four times as many channels and more than three times the training time. In comparison, the ACE→TECE configuration adopted in AIDEN offers a more favorable trade-off between accuracy and computational cost.

Training efficiency of encoder variants. Timings include data loading, forward and backward passes, optimizer and EMA updates, and exclude validation and checkpoint saving.

EncoderChannelsTraining time (s/step) ↓
ACE→TECE480.604
ACE-only1802.000
TECE-only260.509

Comment #7: Could at least one closely related atom-centered model be added for comparison on the ECD dataset? It would also be helpful to compare inference times at similar prediction accuracy and report atomic-encoding and grid-evaluation times separately.

Reply: The original ECD and QM9 benchmarks reported published results using the same datasets. We have now added reproduction experiments for EAC-Net. We choose EAC-Net because its physically motivated atom-centered decomposition of charge density is a central feature, although it lacks our ρ(r)env\rho(\mathbf{r})_{\rm env} term. We have completed training on the ECD dataset and evaluated its inference efficiency on three TBG structures. Its ECDBench performance is shown in Figure 9, and its inference efficiency is quantified in Figure 15 and the table below.

Full-grid charge-density inference times for TBG. Timings include graph construction, encoding, full-grid density decoding, and output serialization. Total time is the sum of these four stages. Speedup is calculated relative to ChargE3Net on the same TBG structure.

ModelTBG (θ\theta, NaN_a)Graph [s]Encode [s]Decode [s]Output [s]Total [s]Speedup
AIDEN21.79∘21.79^\circ (28)1.0171.0183.6290.9716.63552.50×
13.17∘13.17^\circ (76)5.5670.2349.8992.50118.20047.91×
9.43∘9.43^\circ (148)20.4390.43919.4394.91545.23236.94×
ChargE3Net21.79∘21.79^\circ (28)56.5402.257289.1140.434348.3461.00×
13.17∘13.17^\circ (76)166.3411.801702.7111.111871.9651.00×
9.43∘9.43^\circ (148)340.0521.7931326.8962.2911671.0321.00×
ELECTRAFI (Torch)21.79∘21.79^\circ (28)N/A0.2814.0181.4235.72160.89×
13.17∘13.17^\circ (76)N/A1.10331.8983.87436.87523.65×
9.43∘9.43^\circ (148)N/A1.437133.5527.194142.18311.75×
ELECTRAFI (KeOps)21.79∘21.79^\circ (28)N/A0.2830.2051.4001.888184.51×
13.17∘13.17^\circ (76)N/A1.1081.5193.8186.446135.27×
9.43∘9.43^\circ (148)N/A1.4095.8667.39314.668113.92×
EAC-Net21.79∘21.79^\circ (28)553.8006.240392.7004.110956.8500.36×
13.17∘13.17^\circ (76)2173.01016.0101124.07011.0303324.1200.26×
9.43∘9.43^\circ (148)2065.05032.0402058.00020.4904175.5800.40×
Reviewer_vzFyOpenReview ↗

After rebuttal

Thank you for the detailed responses and additional experiments. These additions clarify the design choices, but the architectural novelty and experimental evidence remain insufficient to establish a substantial methodological advance. The demonstrated computational savings over conventional SCF calculations are also modest. After considering the manuscript, the rebuttal, and the other reviewers’ assessments, I believe my original rating best reflects my overall evaluation of the paper, and I therefore retain it.

Reviewer_K8JpOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AIDEN, an equivariant neural network for predicting ground-state charge densities as continuous real-space fields. It decomposes the density into an element-dependent one-center component and an environment-induced redistribution component. The model combines ACE and TECE-based equivariant encoders with a continuous Gaussian-type-orbital decoder, reusing atomic representations across spatial query points. Experiments cover periodic crystals and molecules under PBE and HSE, as well as out-of-distribution systems including water, AlMg, amorphous silicon, and twisted bilayer graphene. The evaluation includes density errors, fixed-density energies and forces, band structures, and inference time.

Critical strengths
  1. The one-center/environment-induced decomposition is physically motivated and provides a meaningful inductive bias. The continuous decoder is a natural way to reuse expensive equivariant features across many grid points.\n\n2. The evaluation is broader than a standard density-prediction benchmark, covering periodic and molecular systems, PBE and HSE, out-of-distribution structures, downstream quantities, and full-grid inference.
Critical weaknesses
  1. The overall motivation is important but well established, and the methodological novelty appears mainly to come from combining existing components. The paper should more clearly identify which architectural contribution is fundamentally new rather than an effective system integration.\n\n2. Charge conservation is not explicitly enforced. The reported electron-count deviations are small for the TBG examples, but it is unclear whether this remains reliable for larger, charged, or out-of-distribution systems. Data splits, grid sampling, and timing protocols also need more detail. Downstream results are fixed-density calculations rather than full self-consistent electronic-structure calculations.
Questions for authors
  1. How are ECD and QM9 split at the structure level, and how are duplicate or closely related structures handled?\n\n2. How sensitive are results to the choice of 5,000 training grid points and to alternative spatial sampling strategies?\n\n3. Can the authors report results with an explicit charge-conservation constraint or normalization?\n\n4. Do the timing comparisons include neighbor construction, model loading, preprocessing, and full-grid output generation for all baselines?
2 comments

We sincerely thank the reviewer for their contribution and positive assessment! We have summarized the comments and address them point by point below.

Comment #1: The methodological novelty appears to come mainly from combining existing components. The architectural contribution needs to be clarified.

Reply: We thank the reviewer for this objective assessment. When designing the architecture, we need to describe many-body interactions of different orders while preserving rotational equivariance, and also assign attention weights to contributions from different edges. We therefore use the established Cartesian ACE and TECE/RRA approaches to address these two requirements, respectively. These modules were originally proposed for machine-learning force fields and effectively improved their performance, which motivated us to explore their transfer to electronic-structure tasks. In the revised manuscript, we place greater emphasis on the methodological innovation arising from the physically motivated decomposition of charge density, while the established architectural components support AIDEN's expressive capacity. In addition, following Reviewer #2's suggestion, we have conducted more detailed component ablations to quantify their respective contributions, as shown below and described in Appendix C.1.

Module ablation results. Parameter counts and best validation NMAE ερ\varepsilon_{\rho} for models trained on the same 2,000-structure PBE subset. Relative changes are calculated with respect to the full ACE→TECE model.

Ablation settingParameters ↓Best val. ερ\varepsilon_{\rho} (%) ↓Relative to Full (%) ↓
Full ACE→TECE4.459M1.13151—
ACE-only4.503M1.11814-1.18
TECE-only4.564M1.32051+16.70
TECE→ACE4.459M1.14572+1.26
w/o GIE4.393M1.14622+1.30
w/o RRA4.205M1.16120+2.62
Fixed GTO4.454M1.21505+7.38
w/o ρinit\rho_{\rm init}4.457M1.53550+35.70

We further evaluate the training cost of using either ACE or TECE alone as the encoder under comparable parameter counts. As shown in the module ablation results, ACE-only appears to be the only configuration that achieves a lower NMAE than the original ACE→TECE design. However, the training efficiency results show that, despite having a similar parameter count, ACE-only requires approximately four times as many channels and more than three times the training time. In comparison, the ACE→TECE configuration adopted in AIDEN offers a more favorable trade-off between accuracy and computational cost.

Training efficiency of encoder variants. Timings include data loading, forward and backward passes, optimizer and EMA updates, and exclude validation and checkpoint saving.

EncoderChannelsTraining time (s/step) ↓
ACE→TECE480.604
ACE-only1802.000
TECE-only260.509

Comment #2: How are the ECD and QM9 datasets split at the structure level? How are duplicate or highly similar structures handled?

Reply: We use a common random seed to split the datasets without deliberately distinguishing particular structural types or elemental compositions. Like other related studies, we do not use cross-validation. The datasets are large, with approximately 140,000 PBE samples and 7,000 HSE samples, and cover a broad chemical space; in our experience, fluctuations caused by the particular split are relatively small. Training equivariant GNNs for either machine-learning force fields or electronic-structure modeling is also computationally demanding. Appendix C.4 quantitatively compares the parameter counts and inference costs of several baselines. Although AIDEN represents a relatively balanced choice in terms of efficiency, training its L=4L=4 model on a single A100 GPU still takes approximately 13 days. A fixed dataset split is therefore an acceptable and stable compromise. Regarding the second question, we remove potential duplicate structures during preprocessing, identifying them from the CIF files using pymatgen. In practice, such duplicates are rare and have little effect. Structurally similar cases include structures with the same elements but different space groups, or compounds with the same structural prototype but different elemental compositions. Following common practice, we do not deliberately separate these cases, with the aim of ensuring generalization across different systems.

Comment #3: How sensitive are the results to the 5,000 grid points selected during training? What would be the effects of other spatial sampling strategies?

Reply: We added a comparison of different numbers of sampled points. AIDEN is supervised using ρ(r)\rho(\mathbf{r}) values at 5,000 randomly sampled real-space points during training. This practice can be traced back to DeepDFT [npj Comput. Mater. 8, 183 (2022)], whose Methods section reports that only 1,000 sampled points per QM9 molecular structure, approximately 0.14% of the grid, are sufficient for accurate charge-density learning. The subsequent JLCDM study [npj Comput. Mater. 9, 87 (2023)] states more explicitly that "we can construct an accurate model trained with just about 0.1% of the available training points." We also note a recent study [npj Comput. Mater. (2026). https://doi.org/10.1038/s41524-026-02361-x] that examines information redundancy in electronic structure in detail, suggesting that essential charge-density information may lie on a low-dimensional nonlinear manifold and that many samples provide repeated or highly similar local-environment information. These studies provide a reference for sparse sampling. To examine the effect of the number of sampled points, we retrained models from scratch on the PBE dataset with {1000,2000,5000,10000} sampled points. Increasing the number of points produces only limited performance improvements while substantially increasing the training burden. Our original choice of 5,000 points is an empirical compromise; see Appendix C.1. This additional experiment also provides further support for our choice of the number of sampled points.

Comment #4: Can the authors report results with an explicit charge-conservation constraint or normalization?

Reply: The original manuscript used parity plots to show the near-conservation of charge that emerges naturally in AIDEN. We added an experiment with an explicit charge-conservation loss. The explicit constraint reduces the relative error in the total electron count by 0.010956 percentage points, but increases the charge-density NMAE by 0.011755 percentage points; see Appendix C.4, lines 1658-1668.

Comment #5: Do inference-time comparisons with all baselines include neighbor construction, model loading, preprocessing, and full-grid output generation?

Reply: We thank the reviewer for this suggestion. Reviewer #2 (vzFy) raised a similar question. We now divide inference into four stages: graph construction, encoding, decoding, and scalar-density output. We compare the costs of these stages for AIDEN, ChargE3Net, and ELECTRAFI, the additional baseline requested by Reviewer #3 (rhxn), on three TBG structures; see Appendix C.4, lines 1843-1902. The results show that AIDEN achieves high inference efficiency when the models use the same Torch library. However, we also find that ELECTRAFI's KeOps version is indeed substantially faster. This probably reflects better engineering optimization and a more efficient computational-library implementation. We therefore do not directly compare it with Torch-based AIDEN to assess architectural efficiency, and instead primarily refer to results obtained with the same computational library.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.