ML and AI for density functional theory: different priorities for Kohn-Sham and orbital-free DFT, for electronic and nuclear DFT

Sergei Manzhos; Xin-Hui Wu

REVIEW 2 major objections 7 minor 6 references

ML and AI for density functional theory: different priorities for Kohn-Sham and orbital-free DFT, for electronic and nuclear DFT

T0 review · 2 major / 7 minor · reviewed 2026-07-11 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Deep neural networks suit Kohn-Sham DFT, but their cost and rough error surfaces undercut orbital-free kinetic energy functionals; symbolic regression is the more promising route for both electronic and nuclear DFT.

desk verdict Solid cross-field Perspective that maps why ML priorities differ for KS vs OF and electronic vs nuclear DFT; the comparative framing and transfer list are the real contribution. read the letter →

arxiv 2607.04095 v1 pith:7BV4LA6N submitted 2026-07-05 physics.chem-ph cond-mat.mtrl-scinucl-thquant-ph

Xin-Hui Wu , Sergei Manzhos This is my paper

classification physics.chem-phcond-mat.mtrl-scinucl-thquant-ph

keywords densityfunctionaltheorymachinelearningorbital-freeDFTKohn-Shamkineticenergysymbolicregressionnuclearexchange-correlation

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

This perspective argues that machine learning for density functional theory is not one-size-fits-all. Kohn-Sham and orbital-free formulations, and electronic versus nuclear systems, face different accuracy demands and computational bottlenecks, so the right ML tool depends on what is being learned: the exchange-correlation functional, the kinetic energy functional, the density, or the basis. Deep neural networks remain useful when the main cost is diagonalizing a Hamiltonian, as in Kohn-Sham calculations of exchange-correlation. They become counterproductive once the functional itself is the dominant cost, as in orbital-free kinetic energy functionals that must recover a large fraction of the total energy without error cancellation. The authors highlight lighter models and especially symbolic regression—automatic discovery of formulas—as conceptually advantageous for both electronic and nuclear DFT, and they map which successful electronic-DFT ML strategies can transfer to the less mature nuclear case.

What carries the argument

A four-way comparison of bottlenecks and accuracy criteria (electronic vs nuclear; Kohn-Sham vs orbital-free) that maps each ML target—XC functional, kinetic energy functional, density, basis, or pseudopotential—onto suitable algorithmic choices, especially the contrast between deep networks and symbolic regression for kinetic energy functionals.

What would settle it

A production-scale orbital-free calculation in which a large neural-network kinetic energy functional simultaneously matches or beats a symbolic-regression or gradient-expansion formula on energy-volume curves or nuclear potential-energy surfaces and on wall-clock time per energy evaluation, without extra SCF or geometry-optimization failures.

Watch

Extended reading notes

Core claim

The central claim is that promising ML and AI choices for DFT follow from the target and the paradigm: deep neural networks remain potent for Kohn-Sham exchange-correlation functionals, but their overhead, corrugated errors, and poor error cancellation become disadvantages when building kinetic energy functionals for orbital-free DFT, where lighter methods and symbolic regression offer conceptual advantages that are still largely unrealized, with useful transfer paths from electronic to nuclear DFT.

Load-bearing premise

That large neural networks will systematically cost too much and converge too poorly for orbital-free kinetic energy functionals once those functionals are evaluated repeatedly, while symbolic-regression formulas will stay accurate and portable enough to replace them.

Editorial extensions

If this is right

KS exchange-correlation models can keep using relatively large neural networks without killing overall cost; orbital-free kinetic energy models should prefer small networks, kernels, trees, or formulas.
Symbolic regression of kinetic energy functionals and of nuclear energy density functionals becomes a high-priority research direction for both electronic and nuclear orbital-free DFT.
Electronic-DFT techniques—gradient-expansion features, ML local pseudopotentials, density prediction, and basis optimization—can be ported to accelerate nuclear DFT mass tables, fission paths, and superheavy systems.
Precomputed ML densities can remove functional-derivative stability problems and cut self-consistency cost in both communities.

Reading between the lines

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

If symbolic regression keeps producing stable, interpretable kinetic energy formulas, orbital-free DFT may become production-ready for systems far beyond simple metals before deep-network KEFs do.
Nuclear DFT’s milder relative-energy tolerances and routine use of 1D/2D symmetry-constrained densities make it a faster proving ground for ML orbital-free methods than full 3D molecular chemistry.
The same cost argument that disfavors oversized networks for KEFs also applies to ML-optimized bases and per-orbital state representations: calling a heavy model once per basis function can erase the gain from a smaller basis.

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.

Desk Editor's Note

read the letter

This is a clean Perspective, not a methods paper. What is new is the side-by-side map of accuracy and cost priorities across the four flavors (electronic/nuclear × KS/OF) and the explicit list of what can transfer from the more mature electronic ML-DFT literature into nuclear work. That framing is useful and, as far as I can tell, not already sitting in one place.

They do the basics well. The DFT summaries are accurate without being textbook padding. The central cost argument is grounded: XC is a few percent of the energy and is not the KS bottleneck, while KE is a large fraction of the energy and is the OF bottleneck, so large NNs that are fine for XC become a real prefactor problem for KEFs. The DM21 timing figure and the nuclear PES example (their Fig. 2) make that concrete. Citations are dense and on-point; self-cites point to prior numerical work rather than circular claims. They correctly flag that symbolic regression is still under-realized.

Soft spots are minor and mostly the ones the authors already own. The overhead/corrugation argument for large NNs in OF-DFT is still partly prospective (one strong timing comparison plus the energy-fraction logic). The symbolic-regression pitch is conceptual, not demonstrated at scale. Sub-D models get more practical weight in nuclear DFT than in electronic DFT; that is fair, but the electronic side of that claim is thinner. None of this breaks the synthesis.

Who it is for: people already working on ML functionals, OF-DFT, or nuclear EDF who need a clear priority list rather than another method paper. It deserves a serious referee. I would accept it as a Perspective and would cite the comparative sections when writing about method choice for KEFs or nuclear OF-DFT.

Referee Report

2 major / 7 minor

Summary. This Perspective synthesizes similarities and differences in computational bottlenecks and accuracy requirements for machine learning and AI applied to four flavors of DFT: electronic vs nuclear, and Kohn–Sham vs orbital-free. From a comparative analysis of XC functionals, kinetic energy functionals (KEFs), density and basis representations, and pseudopotentials, the authors argue that deep neural networks remain attractive for KS XC work, but that their cost and error-surface properties can become bottlenecks for KEFs in OF-DFT, where functional evaluation dominates the cost. They highlight conceptual advantages—explicitly hedged as “yet to be fully realized”—of symbolic regression and lighter models for both electronic and nuclear OF-DFT, and identify transferable methods from the more mature electronic ML-DFT literature to nuclear DFT. The argument is supported by extensive citation of the literature and by concrete illustrations (DM21 timing vs PBE0/CCSD(T), nuclear PES with an ML KEF, BigDFT basis optimization).

Significance. If the comparative framing holds, the paper offers a useful organizing map for a rapidly growing and fragmented literature, with concrete methodological priorities that differ by DFT flavor rather than a one-size-fits-all endorsement of deep networks. The cross-field transfer recommendations (electronic → nuclear) and the emphasis on lighter models and symbolic regression for OF KEFs are actionable for practitioners and for method developers. Strengths include the breadth of correctly cited prior work, the explicit hedging of prospective claims, and the use of published numerical illustrations (timing, nuclear PES, basis examples) rather than new unverified fits. As a Perspective rather than a derivation or benchmark study, its value is organizational and programmatic rather than a single new result.

major comments (2)

§3.1 and the central claim about NN disadvantages for KEFs: the main quantitative cost illustration (DM21 vs PBE0/CCSD(T), Fig. 3 / Ref. 165) is for a hybrid ML XC functional in the KS regime, not for a KEF evaluated repeatedly in OF-DFT. The transfer of the overhead/corrugation argument to OF KEFs is therefore partly by analogy (plus the energy-fraction argument that KE is ~60–70% of total energy). Please either (i) add OF-specific timing or iteration-count evidence from the cited ML KEF literature, or (ii) state more explicitly that the DM21 comparison is an existence proof of ML-functional overhead in DFT workflows and that the OF-KEF case remains partly prospective. This does not overturn the recommendation, but it is load-bearing for how strongly the “defeats the purpose” language can be read.
§2.3.2 and §3.1: the quantitative accuracy and energy-fraction statements that underwrite the “more stringent requirements for KEFs” claim (chemical accuracy ballparks; KE ~60–70% vs XC ~5–10% of total energy; nuclear MeV-scale tolerances) are used as free-standing facts. Please anchor each with a standard reference or a short derivation/estimate (e.g., typical KS energy decomposition for a representative molecule/solid, and a standard nuclear EDF accuracy survey). Without that, the comparative priority ranking remains plausible but not fully checkable from the manuscript alone.

minor comments (7)

Figure numbering is inconsistent: the text refers to “Figure 5” for the DM21 timing comparison (around p. 23), but the caption is labeled Figure 3; a later figure is also labeled Figure 5 (BigDFT basis). Please renumber all figures and cross-references consistently.
§2.2: equation numbers (2.2.3) and (2.2.4) are reused for both the Skyrme interaction energy / functional derivative and the orbital density / KS equations. Renumber so each equation has a unique label.
Abstract and §1: “artificial intelligence (AI)” is used for symbolic regression / automatic formula discovery; a one-sentence working definition early on would help readers who reserve “AI” for broader agentic systems.
Fig. 1 is described as a visual summary of similarities, differences, and ML opportunities; ensure the figure is self-contained (legend/key for electronic vs nuclear and KS vs OF) so it can be read without the surrounding prose.
Occasional typos and wording: e.g. “reviewers” → “reviews” (p. 4), “afferent issues” → “associated issues” (§4), “move involved algorithms” → “more involved” (§4), “in in ground-state” (double “in”, §2.3.2).
§3.1.1: the claim that sub-D models are of “very limited use” for real 3D electronic systems is fair, but a brief pointer to which featurization ideas (e.g., grid/PCA) fail to scale would help non-specialists.
References: a few entries appear truncated or have formatting artifacts in the compiled text (e.g., nuclear mass-table titles with residual MathML-like fragments). Clean for production.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: perspective synthesis with recommendations grounded in external benchmarks and standard cost analyses, not self-referential derivations.

full rationale

This is a Perspective/overview paper that surveys computational bottlenecks and accuracy requirements across electronic/nuclear and KS/OF DFT, then recommends methodological priorities (e.g., lighter models and symbolic regression for KEFs). It does not claim a new functional form, uniqueness theorem, or first-principles derivation whose output is forced by its own inputs. Timing and accuracy arguments rest on external literature (DM21 cost comparison of Ref. 165 / Fig. 3; energy-fraction estimates standard in the field; gradient-expansion baselines). Self-citations (e.g., authors’ prior ML-KEF and nuclear OF-DFT papers) point to independent numerical tests already published elsewhere and are used only as illustrative examples of the state of the art, not as load-bearing uniqueness or definitional premises. No fitted parameter is renamed a prediction, no ansatz is smuggled via self-citation, and no known empirical pattern is merely re-labeled. The strongest claim is explicitly hedged (“yet to be fully realized”). The derivation chain is therefore self-contained against external benchmarks; circularity score is zero.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

A review paper inherits the standard axioms of DFT (Hohenberg–Kohn, Kohn–Sham mapping, existence of an exact but unknown universal functional) and the empirical status of nuclear EDFs. No free parameters are fitted and no new physical entities are postulated; the only ‘assumptions’ are the authors’ judgments about relative CPU cost and accuracy requirements drawn from the cited literature.

assumptions (3)

domain assumption Hohenberg–Kohn theorem: ground-state energy is a unique functional of the density
Invoked throughout §§2.1–2.2 as the foundation of both electronic and nuclear DFT.
domain assumption Kohn–Sham non-interacting kinetic energy can be approximated by a density functional (orbital-free route)
Central premise of all OF-DFT discussion in §§2–3.
domain assumption Nuclear energy density functionals are phenomenological (Skyrme, Gogny, covariant) rather than ab initio
Stated in §2.3.1; explains why ML XC analogues do not exist in nuclear DFT.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ML and AI for density functional theory: different priorities for Kohn-Sham and orbital-free DFT, for electronic and nuclear DFT." pith.science (2026). https://pith.science/paper/7BV4LA6N

@misc{pith2026260704095,
  author       = {Pith},
  title        = {Pith review of: ML and AI for density functional theory: different priorities for Kohn-Sham and orbital-free DFT, for electronic and nuclear DFT},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7BV4LA6N}},
  note         = {Machine review of arXiv:2607.04095}
}

read the original abstract

We overview similarities and, importantly, differences in computational bottlenecks and accuracy requirements that can be addressed with machine learning (ML) and artificial intelligence (AI) techniques in electronic and nuclear DFT. From these follow different promising methodological and algorithmic choices depending on whether one machine learns the exchange correlation (XC) functional, the kinetic energy functional (KEF), the density or the basis functions. In particular, while the popular deep neural networks remain a potent choice in the context of KS DFT, we highlight their disadvantages when building KEFs and highlight conceptual advantages - yet to be fully realized - of symbolic regression for both electronic and nuclear DFT. We point out promising approaches that can be carried from the more extensively investigated ML-enhanced electronic DFT to nuclear DFT.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

6 extracted references · 3 canonical work pages

[1]

(84) Li, L.; Snyder, J

https://doi.org/10.1021/acs.jctc.5c01995. (84) Li, L.; Snyder, J. C.; Pelaschier, I. M.; Huang, J.; Niranjan, U.-N.; Duncan, P.; Rupp, M.; Müller, K.-R.; Burke, K. Understanding Machine -Learned Density Functionals. Int. J. Quantum Chem. 2016, 116 (11), 819–833. https://doi.org/10.1002/qua.25040. (85) Akashi, R.; Sogal, M.; Burke, K. Can Machines Learn De...

arXiv doi:10.1021/acs.jctc.5c01995 2016
[2]

(151) Behler, J

https://doi.org/10.1103/PhysRevB.29.6443. (151) Behler, J. Representing Potential Energy Surfaces by High -Dimensional Neural Network Potentials. J. Phys.: Condens. Matter 2014, 26 (18), 183001. https://doi.org/10.1088/0953-8984/26/18/183001. (152) Behler, J. Four Generations of High-Dimensional Neural Network Potentials. Chem. Rev. 2021, 121 (16), 10037–...

doi:10.1103/physrevb.29.6443 2014
[3]

(186) Kameda, K.; Ariga, T.; Ito, K.; Ihara, M.; Manzhos, S

https://doi.org/10.1021/acs.jpca.3c02949. (186) Kameda, K.; Ariga, T.; Ito, K.; Ihara, M.; Manzhos, S. Machine Learning the Screening Factor in the Soft Bond Valence Approach for Rapid Crystal Structure Estimation. Digital Discovery 2024, 3 (10), 1967–1979. https://doi.org/10.1039/D4DD00152D. (187) Tang, Y .; Xiao, B.; Ihara, M.; Manzhos, S.; Liu, Y . Eff...

work page doi:10.1021/acs.jpca.3c02949 2024
[4]

(219) Pathrudkar, S.; Taylor, S.; Keripale, A.; Gangan, A

https://doi.org/10.1038/s41524-022-00863-y. (219) Pathrudkar, S.; Taylor, S.; Keripale, A.; Gangan, A. S.; Thiagarajan, P.; Agarwal, S.; Marian, J.; Ghosh, S.; Banerjee, A. S. Electronic Structure Prediction of Medium and High Entropy Alloys across Composition Space. arXiv August 19, 2025. https://doi.org/10.48550/arXiv.2410.08294. (220) Qiao, Z.; Christe...

work page doi:10.1038/s41524-022-00863-y 2025
[5]

Surprisingly High Redundancy in Electronic Structure Data Across Materials Explained by Low Intrinsic Dimensionality

https://doi.org/10.48550/arXiv.2507.09001. (222) Li, C.; Sharir, O.; Yuan, S.; Chan, G. K. -L. Image Super-Resolution Inspired Electron Density Prediction. Nat. Commun. 2025, 16 (1), 4811. https://doi.org/10.1038/s41467- 025-60095-8. (223) Pathrudkar, S.; Thiagarajan, P.; Agarwal, S.; Banerjee, A. S.; Ghosh, S. Electronic Structure Prediction of Multi-Mil...

work page Pith review arXiv doi:10.48550/arxiv.2507.09001 2025
[6]

(228) De Vries, H.; De Jager, C

https://doi.org/10.1098/rspa.1969.0037. (228) De Vries, H.; De Jager, C. W.; De Vries, C. Nuclear Charge -Density-Distribution Parameters from Elastic Electron Scattering. Atomic Data and Nuclear Data Tables 1987, 36 (3), 495–536. https://doi.org/10.1016/0092-640X(87)90013-1. (229) Dreher, B.; Friedrich, J.; Merle, K.; Rothhaas, H.; Lührs, G. The Determin...

doi:10.1098/rspa.1969.0037 1969

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.

[1] [1]

(84) Li, L.; Snyder, J

https://doi.org/10.1021/acs.jctc.5c01995. (84) Li, L.; Snyder, J. C.; Pelaschier, I. M.; Huang, J.; Niranjan, U.-N.; Duncan, P.; Rupp, M.; Müller, K.-R.; Burke, K. Understanding Machine -Learned Density Functionals. Int. J. Quantum Chem. 2016, 116 (11), 819–833. https://doi.org/10.1002/qua.25040. (85) Akashi, R.; Sogal, M.; Burke, K. Can Machines Learn De...

arXiv doi:10.1021/acs.jctc.5c01995 2016

[2] [2]

(151) Behler, J

https://doi.org/10.1103/PhysRevB.29.6443. (151) Behler, J. Representing Potential Energy Surfaces by High -Dimensional Neural Network Potentials. J. Phys.: Condens. Matter 2014, 26 (18), 183001. https://doi.org/10.1088/0953-8984/26/18/183001. (152) Behler, J. Four Generations of High-Dimensional Neural Network Potentials. Chem. Rev. 2021, 121 (16), 10037–...

doi:10.1103/physrevb.29.6443 2014

[3] [3]

(186) Kameda, K.; Ariga, T.; Ito, K.; Ihara, M.; Manzhos, S

https://doi.org/10.1021/acs.jpca.3c02949. (186) Kameda, K.; Ariga, T.; Ito, K.; Ihara, M.; Manzhos, S. Machine Learning the Screening Factor in the Soft Bond Valence Approach for Rapid Crystal Structure Estimation. Digital Discovery 2024, 3 (10), 1967–1979. https://doi.org/10.1039/D4DD00152D. (187) Tang, Y .; Xiao, B.; Ihara, M.; Manzhos, S.; Liu, Y . Eff...

work page doi:10.1021/acs.jpca.3c02949 2024

[4] [4]

(219) Pathrudkar, S.; Taylor, S.; Keripale, A.; Gangan, A

https://doi.org/10.1038/s41524-022-00863-y. (219) Pathrudkar, S.; Taylor, S.; Keripale, A.; Gangan, A. S.; Thiagarajan, P.; Agarwal, S.; Marian, J.; Ghosh, S.; Banerjee, A. S. Electronic Structure Prediction of Medium and High Entropy Alloys across Composition Space. arXiv August 19, 2025. https://doi.org/10.48550/arXiv.2410.08294. (220) Qiao, Z.; Christe...

work page doi:10.1038/s41524-022-00863-y 2025

[5] [5]

Surprisingly High Redundancy in Electronic Structure Data Across Materials Explained by Low Intrinsic Dimensionality

https://doi.org/10.48550/arXiv.2507.09001. (222) Li, C.; Sharir, O.; Yuan, S.; Chan, G. K. -L. Image Super-Resolution Inspired Electron Density Prediction. Nat. Commun. 2025, 16 (1), 4811. https://doi.org/10.1038/s41467- 025-60095-8. (223) Pathrudkar, S.; Thiagarajan, P.; Agarwal, S.; Banerjee, A. S.; Ghosh, S. Electronic Structure Prediction of Multi-Mil...

work page Pith review arXiv doi:10.48550/arxiv.2507.09001 2025

[6] [6]

(228) De Vries, H.; De Jager, C

https://doi.org/10.1098/rspa.1969.0037. (228) De Vries, H.; De Jager, C. W.; De Vries, C. Nuclear Charge -Density-Distribution Parameters from Elastic Electron Scattering. Atomic Data and Nuclear Data Tables 1987, 36 (3), 495–536. https://doi.org/10.1016/0092-640X(87)90013-1. (229) Dreher, B.; Friedrich, J.; Merle, K.; Rothhaas, H.; Lührs, G. The Determin...

doi:10.1098/rspa.1969.0037 1969