Pith. sign in

REVIEW 4 major objections 4 minor 44 references

Self-Refining Training for Amortized Density Functional Theory

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper argues that an amortized DFT model can train itself by alternating energy minimization with Langevin sampling from its own evolving energy surface, matching full-data accuracy with only 10% of the data.

desk verdict Solid variational framing and a genuinely new self-refining loop, but the alternating algorithm has an untested support-collapse risk and the empirical claims need error bars. read the letter →

arxiv 2506.01225 v1 pith:S5QM4EEB submitted 2025-06-02 cs.LG

classification cs.LG
keywords densityfunctionaltheoryamortizedDFTself-refiningtrainingLangevindynamicsvariationalenergyboundelectronicstatemodeldatascarcitymolecularconformations
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that a deep-learning model that predicts the electronic state of a molecule can be trained with far less pre-collected data by having the model generate its own training conformations. The target is the Boltzmann distribution of the true ground-state energy, and the method alternates between sampling conformations under the current energy model and minimizing that model's energy prediction on those samples. Both steps are derived from one variational upper bound on the KL divergence between the sample distribution and the target distribution. The experiments on MD17 molecules indicate that this self-refinement matches a full-dataset baseline using only 10% of the data and reaches chemical accuracy from as few as 25 conformations. A sympathetic reader would care because dataset construction, not model training, is the main cost in amortized DFT.

What carries the argument

The machinery is the variational upper bound $\Phi[q, f_\theta] = \mathbb{E}_{q}[\log q] + \mathbb{E}_{q}[E(R, f_\theta(R))] + \log Z^{*}$ from Proposition 1, which turns the unavailable target density $p^{*}$ into two tractable objectives. Minimizing $\Phi$ over the model parameters $\theta$ collapses to minimizing the implicit DFT energy $E(R, f_\theta(R))$ on whatever conformations are at hand (Corollary 1), which requires no labeled energies. Minimizing $\Phi$ over the sample density $q$ collapses to sampling from the Boltzmann density of the current energy model (Corollary 2), and the Wasserstein gradient of $\Phi$ is exactly Langevin dynamics (Proposition 2), so the paper generates new conformations by simulating $dR_t = -\nabla_R E(R_t, f_\theta(R_t))\, dt + \sqrt{2}\, dW_t$. The orthonormality constraint $f_\theta(R)^{\dagger} S(R) f_\theta(R) = 1$ is enforced by a QR-based reparameterization, and an asynchronous replay buffer lets the sampling and energy-minimization steps run on separate GPUs.

What would settle it

Sweep the number of Langevin steps per sample while holding everything else fixed: if longer chains do not reduce energy error on the held-out test set, the samples are not converging to the model's Boltzmann distribution and the claimed self-refinement mechanism is not doing the work.

Watch

Extended reading notes

Core claim

The paper's central claim is that the data-generation and energy-estimation problems in amortized DFT can be solved by the same learning loop, so that a pre-collected dataset is not a prerequisite. For any sample density $q(R)$ and any parameterized electronic state model $f_\theta(R)$ satisfying the DFT orthonormality constraint, the KL divergence between $q$ and the target Boltzmann density $p^{*}$ is upper-bounded by $\Phi[q, f_\theta] = \mathbb{E}_{q}[\log q] + \mathbb{E}_{q}[E(R, f_\theta(R))] + \log Z^{*}$. Minimizing this bound over $\theta$ is equivalent to minimizing the predicted energy on the current samples, and minimizing it over $q$ drives $q$ toward the Boltzmann distribution of the current energy model; the paper implements that second step as Langevin dynamics, which is the Wasserstein gradient flow of $\Phi$. The claim is that iterating these two steps creates a self-improving cycle: as the energy model improves, the sampled conformations become more relevant, and as the sampled conformations improve, the energy model gets better where it matters. Empirically, on ethanol, malondialdehyde, and uracil from MD17, self-refining models trained on 10% of the data match or beat the baseline trained on the full dataset, and the method reaches chemical accuracy with as few as 25 conformations.

Load-bearing premise

The whole loop depends on the assumption that short Langevin chains, run against a target density that keeps changing as the model updates, still produce samples representative enough of that density to train the model.

Editorial extensions

If this is right

  • Using 10% of the pre-collected conformations, the self-refined model matches or improves on the full-data baseline for the tested MD17 molecules, so dataset collection could be cut by an order of magnitude.
  • Models trained this way keep lower energy error along molecular-dynamics trajectories that leave the training distribution, where the baseline energy error grows quickly.
  • Because sampling and training run asynchronously, the total wall-clock time from a small starting dataset to a usable energy model can drop below the time needed to generate and label a full dataset.
  • The training objective is aligned with the downstream task of overdamped molecular dynamics, since that dynamics is the same Langevin process used for sampling.
  • The scheme is agnostic to the energy functional, so it transfers to other DFT variants such as orbital-free DFT or stochastic estimates of the energy quadrature.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that the same variational loop transfers to any amortized solver with a differentiable pointwise objective, such as orbital-free DFT, wavefunction ansatze, or classical force-field fitting.
  • The paper does not prove convergence of the alternating scheme; a plausible failure mode is that finite-length Langevin chains lag behind the changing energy model and amplify its errors, which the replay buffer and pretraining steps are practical hedges against.
  • A clean control experiment would replace the energy-gradient sampler with random perturbations of the buffer: if gains persist, the improvement comes from sample diversity rather than from the proposed Langevin mechanism.
  • The bound's tightness condition, $f_\theta = f^{*}$, suggests that model capacity, not just sample count, limits the ceiling; a learnable sampler trained on the evolving density could relax the mixing requirement the paper acknowledges.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes a self-refining training strategy for amortized density functional theory (DFT). The model, an electronic state (ES) model f_θ(R), predicts orbital coefficients that define an energy E(R, f_θ(R)) through an implicit DFT loss. Training alternates between sampling molecular conformations from the Boltzmann distribution of the current energy model via Langevin dynamics, and minimizing the energy on those sampled conformations to update θ. The authors derive a variational upper bound on the KL divergence between the sample distribution and the true Boltzmann distribution (Proposition 1), show that optimizing the bound over the sampling distribution yields the model's Boltzmann distribution (Corollary 2), and identify the Wasserstein gradient flow with the overdamped Langevin SDE (Proposition 2). They evaluate on MD17 (ethanol, malondialdehyde, uracil) with dataset sizes ranging from 25 to 25,000 conformations, reporting improvements in energy and Hamiltonian errors in low-data settings and improved behavior on simulated out-of-distribution trajectories. The implementation is open-sourced.

Significance. The paper addresses a real and expensive bottleneck in computational chemistry: the pre-collection and DFT labeling of large molecular conformation datasets. The variational bound in Proposition 1 is a genuine upper bound, not a circular reduction to the fitted objective, and the derivations of the sampling optimum and the Langevin dynamics (Corollary 2 and Proposition 2) are clean and correct. The held-out evaluation on MD17 conformations in Table 1 and Figure 2 is external to the training distribution, so the main data-scarcity claim is testable and not tautological. The open-source asynchronous implementation is a strength for reproducibility. The main risks are the lack of a guarantee that the alternating scheme drives the model toward the true ground-state energy, the risk of self-confirming distribution collapse, and an out-of-distribution evaluation that is partly circular. These concerns do not invalidate the core idea but require careful additional analysis and re-evaluation before the central claims can be accepted.

major comments (4)
  1. [Section 3.1–3.4, Eqs. (10) and (12)] Corollary 1 identifies f* as the minimizer of E_q[E(R, f_θ(R))] under the assumption that q(R) > 0 on all finite-energy configurations, but Algorithm 2 implements q as a replay buffer filled by Langevin chains targeting p_θ(R) ∝ exp(−E(R, f_θ(R))). Because E(R, f_θ(R)) ≥ E(R, f*(R)) pointwise by the variational principle, regions where f_θ is inaccurate are exponentially down-weighted and may never enter the buffer, violating the support condition. The energy-minimization step then fits f_θ only on this self-selected support. Nothing in the current objective prevents the loop from driving Φ downward while q moves away from p*, rather than toward it. The paper claims a 'virtuous cycle' (Section 1) but provides no support-preservation mechanism or convergence analysis for the joint (q, θ) iteration. The authors should either prove (under reasonable assumptions) that the alternating updates do not collapse the support, add a mechanism such as periodic injection of samples from a fixed base distribution, or explicitly reframe the theoretical justification as a heuristic motivation backed by experiments.
  2. [Section 4.1, Fig. 3] The out-of-distribution experiment is partly circular. Trajectories labeled 'Sim by SR-100' are generated by propagating the very energy model whose error is being measured, so low energy MAE along them demonstrates only that the model is self-consistent along its own simulated dynamics; it does not establish agreement with the true Boltzmann distribution or with trajectories produced by an independent physical simulator. The conclusion of 'improved robustness to out-of-distribution samples' should be supported by evaluating on trajectories from an independent reference (for example, DFT-quality MD or a converged classical force field) and by reporting errors on both the self-generated and independent trajectories, ideally with standard deviations across seeds.
  3. [Section 4.1, Table 1, and Abstract] The abstract and text claim that self-refinement 'often matches' the full-data baseline using only 10% of the data, but Table 1 shows that the total energy error for ethanol is 135.94 µEh with SR+D 10% versus 63.33 µEh for the full-data baseline, and for malondialdehyde it is 7089 µEh versus 773 µEh — an order of magnitude worse. The claim is therefore overstated and molecule- and metric-dependent. The manuscript should restrict the claim to the metrics and molecules where it holds (for example, Hamiltonian and orbital-energy errors for ethanol and most metrics for uracil), and it should discuss why malondialdehyde does not benefit, since the failure of the method on one of the three systems is essential information.
  4. [Section 3.4 and Appendix A] The asynchronous alternating scheme is presented as coordinate descent on Φ, but no convergence or stability result is given for the joint evolution of q and θ. Proposition 2 applies only for a fixed θ; in Algorithm 2 the target density of the Langevin sampler changes at every parameter update, and the finite chain length T and step size dt determine whether the samples are representative of that time-varying target. The theoretical framing in Section 3.1 goes beyond what is established. The authors should either provide a formal statement about the alternating minimization under idealized assumptions (e.g., slow-varying θ, enough samples per update) or state explicitly that the method is justified empirically and that the variational bound is a motivation rather than a convergence certificate.
minor comments (4)
  1. [Section 3.4] The sentence 'both steps minimize the variational lower bound' should refer to the variational upper bound from Proposition 1, since the derivation bounds D_KL from above.
  2. [Figures 2 and Table 1] Energy units are inconsistent: Table 1 reports errors in µEh, while Figure 2's caption and axis labels use ×10^−4 Hartree. Please unify the units to avoid confusion.
  3. [Section 5] There is a typo in 'we directly optimize the perfomance of our model'; it should be 'performance'.
  4. [Appendix C.2 and Table 4] The 'Pretrain Iterations' setting of 10K for all SR configurations is described in the text as using a small number of conformations for initialization, but it is not clear whether these pretraining iterations use the subset data, the replay buffer, or some combination, and how this interacts with the claim that self-refinement reduces dependence on pre-collected data. Please clarify.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the variational bound is a genuine upper bound, the energy-minimization and sampling updates are derived from it rather than fitted to test labels, and the main evaluation uses held-out MD17 conformations; the only self-citation concerns the architecture backbone and is not load-bearing.

full rationale

The central derivation is not circular. Proposition 1 states a genuine variational upper bound, DKL(q,p*) <= Eq log q + Eq E(R,f_theta) + log Z*, and the proof uses the pointwise inequality E(R,f*) <= E(R,f_theta) that follows from f* being the minimizer of the energy functional; this does not presuppose the prediction. Corollary 1 is a direct consequence of that pointwise inequality and therefore does not reduce to the training objective by construction. Corollary 2 and Proposition 2 are standard functional-derivative and Wasserstein-gradient calculations. The training signal, the implicit DFT energy E(R,f_theta(R)), is the same physical functional that defines the target f*, but the model never sees ground-truth DFT labels; test labels are computed externally with PySCF on held-out MD17 conformations. The self-refinement loop is coordinate descent on the same bound, which is a stated algorithmic choice rather than a hidden reuse of the evaluation target. The lack of a formal coverage guarantee for finite-time Langevin dynamics is a convergence risk, not a logical reduction of the output to the input. The only self-citation (Hassan et al. 2024, Appendix B) concerns the equivariant transformer backbone and is not load-bearing for the paper's main claim. The OOD evaluation in Fig. 3 partly uses trajectories generated by the self-refining model itself, which could be optimistic, but the paper also reports errors on baseline-generated trajectories and on held-out MD17 conformations, so this is an evaluation caveat rather than a circular derivation.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method does not introduce new physical entities or fitted constants. Its load-bearing premises are the availability of a differentiable DFT energy functional, sufficient model expressivity, the variational principle, and the mixing of the Langevin chain. Several sampling hyperparameters are left unspecified, which is a reproducibility gap.

free parameters (5)
  • Langevin step size dt
    Hyperparameter for the sampling SDE in Eq. (15); value not reported in the paper, required to reproduce the algorithm.
  • Number of Langevin steps T
    Length of the sampling chain in Alg. 2; not reported in the paper.
  • Mixing probability P
    Probability of initializing Langevin dynamics from replay buffer vs. initial distribution p0 (Alg. 2); not reported.
  • Replay buffer size = 2048
    Buffer capacity chosen by hand to balance diversity and freshness of samples.
  • Pretraining iterations = 10000
    Number of iterations on pre-collected data before self-refinement begins (Table 4).
assumptions (4)
  • domain assumption The analytic form of the energy functional E(R,C) and overlap matrix S(R) is known for arbitrary R (Section 2.2, end).
    The entire method depends on having a differentiable energy functional in the LCAO basis; this is true for fixed basis DFT but limits the method to basis sets where this functional is implemented.
  • domain assumption The electronic state model f_theta is expressive enough to represent the true minimizer f* of Eq. (7) (Cor. 1).
    The minimizer of the expected energy over q(R) is f* only if f_theta can achieve the pointwise minimum; limited model capacity would bias the result.
  • standard math The ground-state energy E(R, f*(R)) and the variational principle E(R,f*) <= E(R,C) for all feasible C (Section 2.2, Eq. 6 and Prop. 1 proof).
    This is the DFT/Hartree-Fock variational principle, a standard result the bound relies on.
  • domain assumption The Langevin dynamics in Eq. (15) mixes to the Boltzmann distribution exp(-E(R,f_theta)) in the time between model updates.
    The self-refining loop assumes the sampled chain is sufficiently close to the current model's stationary distribution; with finite steps and a moving target this is an approximation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-Refining Training for Amortized Density Functional Theory." pith.science (2026). https://pith.science/paper/S5QM4EEB

@misc{pith2026250601225,
  author       = {Pith},
  title        = {Pith review of: Self-Refining Training for Amortized Density Functional Theory},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S5QM4EEB}},
  note         = {Machine review of arXiv:2506.01225}
}
read the original abstract

Density Functional Theory (DFT) allows for predicting all the chemical and physical properties of molecular systems from first principles by finding an approximate solution to the many-body Schr\"odinger equation. However, the cost of these predictions becomes infeasible when increasing the scale of the energy evaluations, e.g., when calculating the ground-state energy for simulating molecular dynamics. Recent works have demonstrated that, for substantially large datasets of molecular conformations, Deep Learning-based models can predict the outputs of the classical DFT solvers by amortizing the corresponding optimization problems. In this paper, we propose a novel method that reduces the dependency of amortized DFT solvers on large pre-collected datasets by introducing a self-refining training strategy. Namely, we propose an efficient method that simultaneously trains a deep-learning model to predict the DFT outputs and samples molecular conformations that are used as training data for the model. We derive our method as a minimization of the variational upper bound on the KL-divergence measuring the discrepancy between the generated samples and the target Boltzmann distribution defined by the ground state energy. To demonstrate the utility of the proposed scheme, we perform an extensive empirical study comparing it with the models trained on the pre-collected datasets. Finally, we open-source our implementation of the proposed algorithm, optimized with asynchronous training and sampling stages, which enables simultaneous sampling and training. Code is available at https://github.com/majhas/self-refining-dft.

Figures

Figures reproduced from arXiv: 2506.01225 by the authors.

Figure 1
Figure 1. Overview of the self-refining training. From atom types [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Energy prediction error (MAE in ×10−4 Hartree) of the baseline and self-refining models across varying dataset sizes for three molecules. The self-refining approach consistently improves accuracy, especially in low-data regimes, often achieving chemical accuracy (red dashed line) with significantly fewer training points. 0 3 6 9 12 15 18 21 24 27 30 # Simulation Steps 0.01 0.1 1 10 100 Energy MAE (Hartree) 0.160 ×10… view at source ↗
Figure 3
Figure 3. Generalization of baseline and self-refining models evaluated on samples generated from simulation. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Runtime vs. performance comparison for baseline, self-refining, and supervised models on Ethanol [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 34 canonical work pages

  1. [1]

    J., Mittal, S., Lemos, P., Liu, C.-H., Sendera, M., Ravanbakhsh, S., Gidel, G., Bengio, Y., Malkin, N., and Tong, A

    Akhound-Sadegh, T., Rector-Brooks, J., Bose, A. J., Mittal, S., Lemos, P., Liu, C.-H., Sendera, M., Ravanbakhsh, S., Gidel, G., Bengio, Y., Malkin, N., and Tong, A. (2024). Iterated Denoising Energy Matching for Sampling from Boltzmann Densities . arXiv:2402.06121 [cs, stat]

  2. [2]

    and Savaré, G

    Ambrosio, L. and Savaré, G. (2007). Gradient Flows of Probability Measures . In Handbook of Differential Equations : Evolutionary Equations , volume 3, pages 1--136. Elsevier

  3. [3]

    Towards Foundational Models for Molecular Learning on Large-Scale Multi-Task Datasets

    Beaini, D., Huang, S., Cunha, J. A., Li, Z., Moisescu-Pareja, G., Dymov, O., Maddrell-Mander, S., McLean, C., Wenkel, F., Müller, L., Mohamud, J. H., Parviz, A., Craig, M., Koziarski, M., Lu, J., Zhu, Z., Gabellini, C., Klaser, K., Dean, J., Wognum, C., Sypetkowski, M., Rabusseau, G., Rabbany, R., Tang, J., Morris, C., Koutis, I., Ravanelli, M., Wolf, G.,...

  4. [4]

    J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., Vander P las, J., Wanderman- M ilne, S., and Zhang, Q

    Bradbury, J., Frostig, R., Hawkins, P., Johnson, M. J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., Vander P las, J., Wanderman- M ilne, S., and Zhang, Q. (2018). JAX : composable transformations of P ython+ N um P y programs

  5. [5]

    E., Poltavsky, I., Schütt, K

    Chmiela, S., Tkatchenko, A., Sauceda, H. E., Poltavsky, I., Schütt, K. T., and Müller, K.-R. (2017). Machine learning of accurate energy-conserving molecular force fields. Science Advances , 3(5):e1603015. Publisher: American Association for the Advancement of Science

  6. [6]

    T., and Vargas-Hern \'a ndez, R

    de Camargo, A., Chen, R. T., and Vargas-Hern \'a ndez, R. A. (2024). Leveraging normalizing flows for orbital-free density functional theory. Machine Learning: Science and Technology , 5(3):035061

  7. [7]

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. (2009). ImageNet : A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition , pages 248--255. ISSN: 1063-6919

  8. [8]

    W., Musaelian, A., Owen, C

    Falletta, S., Cepellotti, A., Johansson, A., Tan, C. W., Musaelian, A., Owen, C. J., and Kozinsky, B. (2024). Unified differentiable learning of electric response

Show all 44 references
  1. [9]

    J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y

    Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. (2014a). Generative adversarial nets. Advances in neural information processing systems , 27

  2. [10]

    J., Shlens, J., and Szegedy, C

    Goodfellow, I. J., Shlens, J., and Szegedy, C. (2014b). Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572

  3. [11]

    Hassan, M., Shenoy, N., Lee, J., Stark, H., Thaler, S., and Beaini, D. (2024). ET - Flow : Equivariant Flow - Matching for Molecular Conformer Generation . arXiv:2410.22388 [q-bio]

  4. [12]

    and Pople, J

    Head-Gordon, M. and Pople, J. A. (1988). Optimization of wave function and geometry in the finite basis Hartree - Fock method. The Journal of Physical Chemistry , 92(11):3063--3069. Publisher: American Chemical Society

  5. [13]

    J., Stewart, R

    Hehre, W. J., Stewart, R. F., and Pople, J. A. (1969). Self‐ Consistent Molecular ‐ Orbital Methods . I . Use of Gaussian Expansions of Slater ‐ Type Atomic Orbitals . The Journal of Chemical Physics , 51(6):2657--2664

  6. [14]

    and Fitzgibbon, A

    Helal, H. and Fitzgibbon, A. (2024). MESS : Modern Electronic Structure Simulations . arXiv:2406.03121 [cs]

  7. [15]

    and Kohn, W

    Hohenberg, P. and Kohn, W. (1964a). Inhomogeneous Electron Gas . Physical Review , 136(3B):B864--B871. Publisher: American Physical Society

  8. [16]

    and Kohn, W

    Hohenberg, P. and Kohn, W. (1964b). Inhomogeneous electron gas. Physical review , 136(3B):B864

  9. [17]

    F., Lehtola, S., and Vinko, S

    Kasim, M. F., Lehtola, S., and Vinko, S. M. (2022). DQC : a Python program package for Differentiable Quantum Chemistry . The Journal of Chemical Physics , 156(8):084801. arXiv:2110.11678 [physics]

  10. [18]

    A., Vassilev-Galindo, V., Cheng, B., Chmiela, S., Gastegger, M., Müller, K.-R., and Tkatchenko, A

    Keith, J. A., Vassilev-Galindo, V., Cheng, B., Chmiela, S., Gastegger, M., Müller, K.-R., and Tkatchenko, A. (2021). Combining Machine Learning and Computational Chemistry for Predictive Insights Into Chemical Systems . Chemical Reviews , 121(16):9816--9872. Publisher: America...

  11. [19]

    and Sham, L

    Kohn, W. and Sham, L. J. (1965a). Self- Consistent Equations Including Exchange and Correlation Effects . Physical Review , 140(4A):A1133--A1138. Publisher: American Physical Society

  12. [20]

    and Sham, L

    Kohn, W. and Sham, L. J. (1965b). Self-consistent equations including exchange and correlation effects. Phys. Rev. , 140:A1133--A1138

  13. [21]

    and Matthews, C

    Leimkuhler, B. and Matthews, C. (2015). Molecular Dynamics: With Deterministic and Stochastic Numerical Methods . Springer

  14. [22]

    Li, T., Lin, M., Hu, Z., Zheng, K., Vignale, G., Kawaguchi, K., Neto, A. H. C., Novoselov, K. S., and Yan, S. (2023). D4FT : A Deep Learning Approach to Kohn - Sham Density Functional Theory . arXiv:2303.00399 [cs]

  15. [23]

    Li, Y., Tang, Z., Chen, Z., Sun, M., Zhao, B., Li, H., Tao, H., Yuan, Z., Duan, W., and Xu, Y. (2024). Neural-network Density Functional Theory Based on Variational Energy Minimization . Physical Review Letters , 133(7):076401. arXiv:2403.11287 [physics]

  16. [24]

    Mathiasen, A., Helal, H., Balanca, P., Krzywaniak, A., Parviz, A., Hvilshøj, F., Banaszewski, B., Luschi, C., and Fitzgibbon, A. W. (2024). Reducing the Cost of Quantum Chemical Data By Backpropagating Through Density Functional Theory . arXiv:2402.04030 [cs]

  17. [25]

    and Shimazaki, T

    Nakata, M. and Shimazaki, T. (2017). PubChemQC Project : A Large - Scale First - Principles Electronic Structure Database for Data - Driven Chemistry . Journal of Chemical Information and Modeling , 57(6):1300--1308

  18. [26]

    Noé, F., Olsson, S., Köhler, J., and Wu, H. (2019). Boltzmann Generators -- Sampling Equilibrium States of Many - Body Systems with Deep Learning . arXiv:1812.01729 [stat]

  19. [27]

    Parr, R. G. (1980). Density Functional Theory of Atoms and Molecules . In Fukui, K. and Pullman, B., editors, Horizons of Quantum Chemistry , pages 5--15, Dordrecht. Springer Netherlands

  20. [28]

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Köpf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., and Chintala, S. (2019). PyTorch : ...

  21. [29]

    C., Teter, M

    Payne, M. C., Teter, M. P., Allan, D. C., Arias, T. A., and Joannopoulos, J. D. (1992). Iterative minimization techniques for ab initio total-energy calculations: molecular dynamics and conjugate gradients. Reviews of Modern Physics , 64(4):1045--1097. Publisher: American Phys...

  22. [30]

    P., Burke, K., and Ernzerhof, M

    Perdew, J. P., Burke, K., and Ernzerhof, M. (1996). Generalized Gradient Approximation Made Simple . Physical Review Letters , 77(18):3865--3868. Publisher: American Physical Society

  23. [31]

    and Wang, J

    Perez, L. and Wang, J. (2017). The effectiveness of data augmentation in image classification using deep learning. arXiv preprint arXiv:1712.04621

  24. [32]

    A., and Burke, K

    Pribram-Jones, A., Gross, D. A., and Burke, K. (2015). DFT : A Theory Full of Holes ? Annual Review of Physical Chemistry , 66(Volume 66, 2015):283--304. Publisher: Annual Reviews

  25. [33]

    O., Rupp, M., and von Lilienfeld, O

    Ramakrishnan, R., Dral, P. O., Rupp, M., and von Lilienfeld, O. A. (2014). Quantum chemistry structures and properties of 134 kilo molecules. Scientific Data , 1(1):140022. Publisher: Nature Publishing Group

  26. [34]

    Roothaan, C. C. J. (1951). New Developments in Molecular Orbital Theory . Reviews of Modern Physics , 23(2):69--89. Publisher: American Physical Society

  27. [35]

    T., Gastegger, M., Tkatchenko, A., Müller, K.-R., and Maurer, R

    Schütt, K. T., Gastegger, M., Tkatchenko, A., Müller, K.-R., and Maurer, R. J. (2019). Unifying machine learning and quantum chemistry -- a deep neural network for molecular wavefunctions. arXiv:1906.10033 [physics]

  28. [36]

    T., Sauceda, H

    Schütt, K. T., Sauceda, H. E., Kindermans, P.-J., Tkatchenko, A., and Müller, K.-R. (2018). SchNet - A deep learning architecture for molecules and materials. The Journal of Chemical Physics , 148(24):241722

  29. [37]

    C., Blunt, N

    Sun, Q., Berkelbach, T. C., Blunt, N. S., Booth, G. H., Guo, S., Li, Z., Liu, J., McClain, J. D., Sayfutyarova, E. R., Sharma, S., Wouters, S., and Chan, G. K.-L. (2018). PySCF : the Python -based simulations of chemistry framework. WIREs Computational Molecular Science , 8(1)...

  30. [38]

    and Fabritiis, G

    Thölke, P. and Fabritiis, G. D. (2022). TorchMD - NET : Equivariant Transformers for Neural Network based Molecular Potentials . arXiv:2202.02541 [cs]

  31. [39]

    T., Chmiela, S., Sauceda, H

    Unke, O. T., Chmiela, S., Sauceda, H. E., Gastegger, M., Poltavsky, I., Schütt, K. T., Tkatchenko, A., and Müller, K.-R. (2021). Machine Learning Force Fields . Chemical Reviews , 121(16):10142--10186. Publisher: American Chemical Society

  32. [40]

    Weber, V., VandeVondele, J., Hutter, J., and Niklasson, A. M. N. (2008). Direct energy functional minimization under orthogonality constraints. The Journal of Chemical Physics , 128(8):084113

  33. [41]

    and Sumita, M

    Yoshikawa, N. and Sumita, M. (2022). Automatic Differentiation for the Direct Minimization Approach to the Hartree - Fock Method . The Journal of Physical Chemistry A , 126(45):8487--8493. arXiv:2203.04441 [physics]

  34. [42]

    Yu, H., Xu, Z., Qian, X., Qian, X., and Ji, S. (2023). Efficient and Equivariant Graph Networks for Predicting Quantum Hamiltonian . arXiv:2306.04922 [cs]

  35. [43]

    Zhang, H., Liu, C., Wang, Z., Wei, X., Liu, S., Zheng, N., Shao, B., and Liu, T.-Y. (2024a). Self- Consistency Training for Density - Functional - Theory Hamiltonian Prediction . arXiv:2403.09560 [cs]

  36. [44]

    Zhang, H., Liu, S., You, J., Liu, C., Zheng, S., Lu, Z., Wang, T., Zheng, N., and Shao, B. (2024b). Overcoming the barrier of orbital-free density functional theory for molecular systems using deep learning. Nature Computational Science , 4(3):210--223

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.