acceptodds
Under review as a conference paper at ICLR 2027

Complete Neural Electronic Initialization Accelerates Materials DFT

Abstract

We present the first complete machine learning method for accelerating plane-wave density functional theory (DFT) in materials under the projector augmented wave (PAW) formalism. We formalize seven criteria that a Complete Neural Electronic Initializer must satisfy for practical end-to-end PAW DFT acceleration. Applying these to prior work reveals two structure-dependent components, augmentation occupancies and spin initialization, whose absence prevents existing acceleration methods from providing complete reference-free initialization. We show that omitting these components can eliminate or reverse the acceleration obtained via models that only predict the smooth valence density. We satisfy the missing requirements by introducing AugNet, a general equivariant model for PAW augmentation occupancies, and the first general spin density model for materials, which predicts the smooth spin-difference density and spin-difference PAW augmentation occupancies using predicted magnetic moments to constrain the global magnetic state. Combined with existing valence density models, our full method satisfies all seven criteria and forms a fully reference-free electronic initializer for materials DFT, requiring no electronic quantities from a converged target calculation. We show that perfect initialization could cut PAW DFT wall time by 40-52 %, and our method recovers up to 62 % of this saving, reducing end-to-end DFT wall time by up to % on unseen structures while preserving converged energies.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
electronic structure, charge density, charge, spin, equivariant, periodic, materials, crystals, dft, paw, machine learning potentials

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_mxzWOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work investigates machine learning for predicting initial electronic states to accelerate plane-wave DFT calculations under the PAW formalism. Existing ML acceleration schemes mostly target smooth valence charge densities, which alone cannot fully initialize electronic states in PAW. PAW augmentation occupation matrices and spin-resolved observables are structure-dependent and strongly affect self-consistent convergence. To tackle this, the authors propose the Complete Neural Electronic Initializer (CNEI) and define criteria for complete electronic initialization. Their implementation builds a multi-component PAW initialization pipeline by combining the existing ELECTRAFI charge-density model, a newly designed equivariant AugNet for augmentation occupation matrices, spin-ELECTRAFI, spin-AugNet, and CHGNet-based magnetic moment prediction. Ablation studies show that even highly accurate smooth valence charge density predictions may fail to speed up SCF iterations.

While the study has some scientific and practical merit, the empirical speedup is modest. This work is best framed as a empirical investigation of PAW electronic initialization and DFT workflows, lacking substantive machine learning methodological innovation. To meet ICLR’s acceptance threshold, the authors must demonstrate more transformative ML advances or substantially larger computational gains. Overall, the manuscript’s current contributions do not warrant acceptance.

Critical strengths
  1. This work identifies a critical limitation of existing ML-based electronic density initialization approaches. The authors emphasize that electronic structure within the PAW method is jointly determined by the smooth charge density and augmentation occupation matrices. Predicting only the smooth charge density, or benchmarking using converged augmentation data from the target structure, fails to yield a truly reference-independent initialization scheme. This distinction carries clear physical meaning and enables more objective evaluation of the practical speedup ML methods can bring to DFT calculations.

  2. The ablation studies offer valuable insights. The authors systematically examine how different electronic-state initialization components affect SCF convergence. They find that even supplying converged smooth valence charge densities may not improve convergence when other electronic-state components are inconsistent. This demonstrates that electron-density prediction error is not a reliable proxy for realized speedup in DFT self-consistent iterations.

Critical weaknesses
  1. The empirical speedup is modest. The full framework combines multiple ML components to predict and couple diverse electronic-state variables, yet yields only 15.04% runtime reduction on the MP dataset and 25.17% on GNoME. The method preserves DFT’s asymptotic scaling and does not eliminate SCF iterations.
  2. The reported end-to-end speedup omits training-data generation and model training overhead. Training the neural networks itself also carries non-negligible cost. For a per-run saving of merely 15%–25%, whether upfront training expenses can be sufficiently amortized across production calculations is critical for evaluating practical utility.
  3. The method's strong dependence on a specific PAW representation limits its transferability and generalizability. Predicted observables are tied to a fixed PAW dataset, projectors, partial waves and associated conventions and are not invariant to electronic-structure calculation settings. The MoS2 transfer test demonstrates this weakness: zero-shot performance degrades markedly when a different PAW dataset is used for Mo, requiring retraining or fine-tuning to restore predictive accuracy. Changing PAW datasets, pseudopotentials, exchange-correlation functionals or other calculation parameters will likely require model readaptation, further increasing training and maintenance costs.
  4. This work offers limited novelty in machine learning, as its contributions mainly consist of domain-specific improvements with limited broader applicability.
Questions for authors
  1. After accounting for the cost of generating DFT training data and training the neural networks, how many subsequent DFT calculations must be performed to reach the break-even point in overall computational cost?
Reviewer_wixtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work presents a Complete Neural Electronic Initializer (CNEI) designed to accelerate plane-wave density functional theory (DFT) calculations under the projector augmented wave (PAW) formalism. The work identifies and resolves two critical missing components in existing reference-free acceleration methods: PAW augmentation occupancies (addressed via the equivariant model AugNet) and spin initialization (addressed via a hybrid charge-informed model). Combined with existing valence density models, the proposed method reduces end-to-end DFT wall time by up to 25% on unseen structures without requiring any converged reference data, while fully preserving the accuracy of the final converged energy.

This work also has some limitations, such as limited acceleration (25% only, while MLFFs can bring about acceleration by several orders of magnitude), the model binding basis set and functionals for training, as well as the ability to accelerate in energy minimization and dynamics simulation using the initial guesses from the previous steps. All these significantly reduce the practical application value of this work. Overall, I don't think this is a very attractive ML work. Perhaps journals (such as JCTC?) or conferences on computational chemistry could be considered.

Critical strengths
  • Completeness of Framework: The paper formalizes and satisfies seven criteria for true reference-free initialization, making it the first comprehensive work to explicitly handle valence density, PAW augmentation, and spin initialization simultaneously.
  • Excellent Architecture Design: The proposed AugNet utilizes a shared equivariant backbone and a schema-conditioned readout, enabling it to robustly predict one-center augmentation coefficients across diverse chemical elements and PAW schemas.
  • Significant Practical Acceleration: The method successfully recovers up to 62% of the theoretical acceleration bound (Oracle), and even demonstrates higher reliability than VASP's default initialization in guiding magnetic systems toward the lowest energy ground state.
Critical weaknesses
  • Strict binding to specific functionals, basis sets, and pseudopotentials: The model's predictions are implicitly bound to the specific material distribution, PAW setup, and DFT settings used during training. As demonstrated in the appendix, when transferring to a calculation with a different underlying PAW dataset (e.g., the MoS2 test), the model cannot generalize zero-shot and requires fine-tuning on new data. The current architecture lacks the ability to explicitly encode functional types or pseudopotential parameters as inputs, which severely restricts its broad, out-of-the-box transferability.
  • Severely limited acceleration in core practical workflows (e.g., energy minimization and MD): The impressive acceleration metrics presented apply primarily to single-point static calculations or the very first ionic step of a relaxation. In the most common high-throughput materials science workflows—such as geometry optimization (relaxation) or ab initio molecular dynamics (MD)—subsequent ionic steps naturally restart using the already converged charge density from the immediately preceding step. In these multi-step iterative scenarios, the benefits of ML-based initialization will be heavily marginalized, likely yielding negligible overall wall-time savings for the complete computational pipeline.
Questions for authors

Nothing.

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

In this paper, the author proposed a new machine learning framework, CNEI. CNEI offers a machine learning-based initialization for PAW-DFT. Compared to the SAD initial guess, CNEI lowers the number of SCF cycles and speeds up PAW-DFT calculations.

Critical strengths

This paper demonstrates that ML based initialzation method can accelerate PAW-DFT calculation. For a PAW-DFT initial guess, as shown in Figure 1, four values (ρ~+\tilde\rho^+, ρ~−\tilde\rho^-, d+d^+, d−d^-) must be predicted at the same time The previous framework failed to predict all four at the same time. They either use a ground-truth value for the missing quantity, which is not available in practice, in SCF acceleration evaluation. Interestingly, the authors also demonstrate that a perfect prediction of one component, and using the default initialization for the rest, still fails to reduce the SCF cycles.

Critical weaknesses
  1. This paper demonstrates that existing ML approaches can be assembled to predict all the quantities, rather than a genuine theoretical breakthrough. The gap between the existing work and the full reference-free acceleration method is well known in the literature. For example, in the ICML 2026 ELECTRAFI paper (http://arxiv.org/abs/2601.19966), the gap between ELECTRAFI itself and the full reference-free acceleration method is explicitly acknowledged (dd prediction is missing): it would be necessary to also model the augmentation occupancies directly. In addition, the existing work, such as EdenGNN and CJM, shows that predicting dd is possible. Considering that ELECTRAFI is also adopted as an essential part of the author-proposed framework, and the similarity between EdenGNN and this paper's AugNet, this makes me question whether this paper's contribution is too limited.

  2. AugNet, proposed by this paper, is similar to the previous method EdenGNN. There are two major differences: the encoder changes from NequIP to MACE, and the augmentation readout is also different. The use of MACE over NequIP and AugNet's element-dependent equivariant readout need to be justified with ablation studies. In addition, this paper compares MAE error, showing AugNet is 0.0041, better than 0.0085, and also acknowledges that these two models are trained with different datasets and VASP settings. It's not clear if the advantage comes from the new method or purely from dataset differences. I didn't see a strong justification for replacing EdenGNN with AugNet. Since EdenGNN can predict dd as well, can we also extend EdenGNN's architecture to predict d+d^+ and d−d^-, then evaluate its performance as an SCF initializer?

  3. This paper didn't address one critical limitation in this area: the prediction fails to transfer across different pseudopotential settings. AugNet only works with the same PAW setup in the training set. Its learned function maps from fθ(structure)→df_\theta(\text{structure})\rightarrow d for a prescribed PAW library. However, in order to make the method generalizable enough, I would expect we can learn a map from fθ(structure, PAW setup)→df_\theta(\text{structure, PAW setup})\rightarrow d. Could the authors comment on the challenge and the possibility of learning across different PAW setups?

  4. It is not clear to me how diverse the training and testing sets are. The proposed model was trained with 118k structures. SAD, the VASP default initialization, can work in a wide range of structures and chemical formulas. Could the authors explore how the proposed method works with unseen structures and unseen atomic combinations (different chemical formulas)?

Questions for authors

See weakness

Comments on the paper

1 comment

Thank you for the detailed response. However, these points do not change my overall assessment.

  1. I agree that an initialization method cannot eliminate SCF iterations or change the asymptotic scaling of DFT. This is precisely my concern: whether the selected problem is sufficiently important to justify the proposed approach. The method primarily provides a local optimization of SCF initialization without addressing the main computational bottlenecks of DFT. Moreover, the practical gains are likely to be limited in routine DFT workflows, particularly structural relaxation and molecular dynamics, where the converged electronic state from the previous ionic step can often be reused or extrapolated, leaving less room for further improvement through ML-based initialization. The authors have not sufficiently demonstrated the advantages of their approach over existing initialization strategies in such practical DFT workflows.
  2. The authors argue that using existing Materials Project CHGCAR files makes the training-data generation cost zero. This argument confuses the marginal cost under a specific set of conditions with the intrinsic data requirements of the method. The availability of an existing database does not mean that extending the method to new material domains requires no additional data or computational resources. More importantly, the authors' own results demonstrate a clear dependence on the PAW representation. Changing the PAW dataset requires additional fine-tuning. However, the cost of generating the corresponding reference data for adapting to a new setup is not included in the break-even analysis. I therefore maintain my original assessment and score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.