arxiv: 2605.10331 · v1 · submitted 2026-05-11 · ⚛️ physics.chem-ph · physics.comp-ph

Recognition: 2 theorem links

· Lean Theorem

Constraint-aware functional cloning for stable and transferable machine-learned density functional theory

Alec Wills, Kimberly J. Daas, Mar\'ia Camarasa-G\'omez, Marivi Fern\'andez-Serra, Sara Navarro-Rodr\'iguez

Pith reviewed 2026-05-12 03:41 UTC · model grok-4.3

classification ⚛️ physics.chem-ph physics.comp-ph

keywords functional cloningneural exchange-correlation functionalsmachine-learned DFTself-consistent calculationstransferabilitymolecular to solidconstraint-aware training

0 comments

The pith

Constraint-aware neural networks clone known density functionals more accurately in self-consistent calculations and transfer from molecules to solids.

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

This paper examines functional cloning as a diagnostic test for neural exchange-correlation models, training them to reproduce an established semilocal functional inside full Kohn-Sham self-consistent calculations rather than on fixed densities alone. The authors compare constrained and unconstrained neural architectures and demonstrate that constraints improve fidelity in molecular self-consistent runs while supplying stronger initial parameters for later optimization against correlated energies. Clones trained exclusively on molecular densities also reproduce lattice constants and bulk moduli for a wide variety of solids. The work positions functional cloning as a practical step to verify stability and transferability before expensive training on higher-accuracy reference data.

Core claim

Functional cloning tests whether a neural model can reproduce a known semilocal exchange-correlation functional when it participates in generating the self-consistent density. Constrained models achieve higher accuracy than unconstrained ones in molecular self-consistent calculations and yield better starting points for optimization against correlated molecular energies. Models trained only on molecular densities transfer to solids, reproducing reference lattice constants and bulk moduli across metallic, covalent, ionic, oxide, and layered systems. Energy differences prove relatively robust across computational codes, while total energies depend on whether cloning descriptors derive from all

What carries the argument

The constraint-aware neural architecture for functional cloning, which trains models to reproduce a target functional while enforcing physical constraints to promote stable self-consistent behavior and cross-system transfer.

Load-bearing premise

That successful self-consistent reproduction of a semilocal functional reliably indicates the model will perform well when optimized against more accurate correlated energies and will retain transferability in that setting.

What would settle it

A result in which unconstrained models, after full optimization against correlated energies, match or surpass constrained clones on molecular energy errors or solid equations of state would falsify the claimed advantage of constraint-aware cloning.

Figures

Figures reproduced from arXiv: 2605.10331 by Alec Wills, Kimberly J. Daas, Mar\'ia Camarasa-G\'omez, Marivi Fern\'andez-Serra, Sara Navarro-Rodr\'iguez.

**Figure 1.** Figure 1: Scheme of the complete workflow for the construction and optimization of the machine-learned exchange–correlation functional. It [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Figure 2: Evaluation of self-consistent energies obtained with the [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: Mean absolute total-energy difference (MAE [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Evaluation of self-consistently optimized NN-XC [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

**Figure 5.** Figure 5: Comparison of (a) equilibrium lattice constants [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

read the original abstract

We study a simple but useful test for neural exchange-correlation (XC) functionals: can a neural model reproduce an established XC functional when it is used self-consistently? We call this test functional cloning. The model is trained at the GGA level to reproduce a known semilocal functional, using either a constrained or an unconstrained architecture. The motivation is that an XC functional is not used on a fixed input. In a Kohn-Sham self-consistent-field calculation it contributes to the potential, and the resulting density is part of the outcome of the same calculation. A good pointwise fit to sampled density descriptors is therefore not by itself enough. Because the target functional is known, the error can be measured directly. We compare the clones on sampled descriptors, molecular total energies, energy differences, transfer between PySCF and SIESTA, and equations of state for crystalline solids. The constrained models reproduce the reference functional more accurately in molecular self-consistent calculations. They also give better initial parameters for later optimization against correlated molecular energies. An additional observation is that the constrained architecture already gives a reasonable solid-state baseline before cloning, as seen from randomly initialized constrained models. Clones trained only on molecular densities transfer well to solids, reproducing reference lattice constants and bulk moduli across metallic, covalent, ionic, oxide, and layered systems. Cross-code tests show that energy differences are relatively robust, while total energies depend strongly on whether the cloning descriptors come from all-electron or pseudopotential densities. These results make functional cloning a useful diagnostic before full self-consistent training of neural XC functionals.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Functional cloning is a practical diagnostic for checking self-consistent behavior of ML XC functionals, and the constrained architecture shows clear gains in fidelity and molecular-to-solid transfer, but only when cloning a known semilocal GGA.

read the letter

The paper's core contribution is a straightforward test called functional cloning: train a neural XC model to match a known functional and then check how well it holds up when used self-consistently in Kohn-Sham calculations. They compare constrained and unconstrained networks on pointwise fits, molecular energies, energy differences, cross-code transfer between PySCF and SIESTA, and equations of state for solids. The constrained versions do better at reproducing the reference in self-consistent runs and provide stronger starting points for later optimization against correlated energies. Clones trained only on molecular data also transfer reasonably to solids, matching lattice constants and bulk moduli across several bonding types. Randomly initialized constrained models already give a decent solid-state baseline before any training. These results are internally consistent and the tests are orthogonal enough to make the diagnostic useful. The work is honest about limitations, such as total energies being sensitive to all-electron versus pseudopotential densities. The main soft spot is that everything is demonstrated only for reproducing a semilocal GGA. We do not see whether the initialization advantage survives when the target is a more accurate functional derived from correlated data, nor whether the molecular-to-solid descriptor overlap persists outside the semilocal regime. Quantitative error tables and data-selection details are not visible in the abstract, though the full text presumably supplies them. This is solid method-development work aimed at people building or testing neural XC functionals. A reader already working on ML-DFT will get immediate value from the cloning protocol and the architecture comparison. It is coherent on its own terms and deserves a serious referee rather than a desk reject, even if the final claims about downstream optimization need more evidence.

Referee Report

2 major / 3 minor

Summary. The manuscript introduces functional cloning as a diagnostic test for neural exchange-correlation functionals: models are trained at the GGA level to reproduce a known semilocal functional self-consistently, using either constrained or unconstrained architectures. Performance is assessed on pointwise descriptor accuracy, self-consistent molecular total energies and differences, transfer between PySCF and SIESTA, and equations of state for solids (lattice constants and bulk moduli across metallic, covalent, ionic, oxide, and layered systems). The central findings are that constrained models achieve better self-consistent reproduction of the reference and supply improved initial parameters for subsequent optimization against correlated energies, while molecular-trained clones transfer reasonably to solids; cross-code tests indicate that energy differences are more robust than total energies.

Significance. If the results hold, functional cloning provides a practical, low-cost diagnostic for stability and transferability prior to full self-consistent training of neural XC functionals against correlated data. The demonstration of molecular-to-solid transfer and the observation that the constrained architecture yields a reasonable solid-state baseline even from random initialization are potentially useful for the field. The cross-code robustness findings also highlight practical issues in descriptor consistency.

major comments (2)

Abstract and main text: the claim that constrained clones 'give better initial parameters for later optimization against correlated molecular energies' is not supported by any reported results; the manuscript does not present the outcomes of such optimization steps, final accuracies, or comparisons of optimized functionals, leaving the practical benefit unverified.
Abstract and results sections on solids: while molecular-to-solid transfer is shown for the semilocal GGA reference (reproducing lattice constants and bulk moduli), the paper does not repeat the cloning-plus-optimization protocol with a target functional derived from correlated methods, so it remains unclear whether the observed descriptor overlap and transferability persist when the target has different form or statistics.

minor comments (3)

The manuscript would benefit from explicit tables or figures reporting quantitative error metrics (e.g., MAE on energies, lattice constants) with error bars and details on data selection and sampling for the descriptor fits.
Notation for the constrained vs. unconstrained architectures should be defined more clearly early in the text, including any specific constraints on the neural network outputs or inputs.
A brief discussion of computational cost for the cloning procedure versus direct self-consistent training would help contextualize the diagnostic value.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful reading of the manuscript and for the constructive comments. We address each of the major comments below.

read point-by-point responses

Referee: Abstract and main text: the claim that constrained clones 'give better initial parameters for later optimization against correlated molecular energies' is not supported by any reported results; the manuscript does not present the outcomes of such optimization steps, final accuracies, or comparisons of optimized functionals, leaving the practical benefit unverified.

Authors: We acknowledge that the manuscript does not report the results of subsequent optimization steps against correlated energies. The claim was based on the observation that constrained models achieve lower self-consistent errors when cloning the reference functional, which we hypothesized would translate to better starting points for optimization. However, since no such optimization results are presented, we agree that the statement is not fully supported. We will revise the abstract and main text to remove this specific claim, focusing instead on the demonstrated benefits in cloning accuracy and transferability. revision: yes
Referee: Abstract and results sections on solids: while molecular-to-solid transfer is shown for the semilocal GGA reference (reproducing lattice constants and bulk moduli), the paper does not repeat the cloning-plus-optimization protocol with a target functional derived from correlated methods, so it remains unclear whether the observed descriptor overlap and transferability persist when the target has different form or statistics.

Authors: The referee is correct that the cloning and transferability tests are conducted using a known GGA functional as the target. The purpose of the study is to use functional cloning as a controlled diagnostic where the reference is exact, allowing direct measurement of self-consistent errors. We do not claim that the transferability results directly apply to correlated targets, but rather that the constrained architecture shows promise in maintaining stability and transfer from molecules to solids even when trained on molecular data. We will add a paragraph in the discussion section to explicitly state this scope limitation and suggest that future work should test the approach with correlated reference data. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical tests of cloning and transfer are independent of training inputs

full rationale

The paper trains neural XC models to reproduce a known semilocal functional on sampled density descriptors (pointwise fit). It then evaluates the clones via full self-consistent Kohn-Sham calculations on molecules, energy differences, cross-code transfer (PySCF/SIESTA), and equations of state on solids. These SCF and transfer tests are not equivalent to the training data by construction, as they involve iterative density-potential feedback and out-of-distribution systems. No equations reduce reported accuracies to fitted quantities, and no self-citation chain or uniqueness theorem is invoked to force the results. The derivation chain is self-contained against external reference functional and independent benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard neural-network approximation assumptions and the existence of a reference semilocal functional. No new physical entities are postulated.

free parameters (1)

neural network weights and biases
Parameters fitted during cloning to match the target functional on sampled density descriptors.

axioms (1)

domain assumption A neural network with local density descriptors can approximate the XC functional sufficiently for cloning purposes.
Invoked in the training and self-consistent evaluation setup.

pith-pipeline@v0.9.0 · 5607 in / 1301 out tokens · 79255 ms · 2026-05-12T03:41:05.375910+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
constrained output mappings ... F_x(˜x2)=1+I_1.804[˜x2 F_x(˜x2;ω_x)] ... guarantees ... uniform-electron-gas limit ... Lieb–Oxford bound
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear
cloned NN-XC functionals reproduce PBE lattice constants and bulk moduli across metallic, covalent, ionic, oxide, and layered systems

Reference graph

Works this paper leans on

44 extracted references · 44 canonical work pages · 1 internal anchor

[1]

Hohenberg and W

P. Hohenberg and W. Kohn. Inhomogeneous Electron Gas. Physical Review, 136(3B):B864–B871, November 1964. ISSN 0031-899X. doi:10.1103/PhysRev.136.B864. URLhttps: //link.aps.org/doi/10.1103/PhysRev.136.B864

work page doi:10.1103/physrev.136.b864 1964
[2]

Kohn and L

W. Kohn and L. J. Sham. Self-Consistent Equations In- cluding Exchange and Correlation Effects.Physical Review, 140(4A):A1133–A1138, November 1965. ISSN 0031-899X. doi:10.1103/PhysRev.140.A1133. URLhttps://link.aps. org/doi/10.1103/PhysRev.140.A1133

work page doi:10.1103/physrev.140.a1133 1965
[3]

Cohen, Paula Mori-Sánchez, and Weitao Yang

Aron J. Cohen, Paula Mori-Sánchez, and Weitao Yang. In- sights into Current Limitations of Density Functional The- ory.Science, 321(5890):792–794, August 2008. ISSN 0036- 8075, 1095-9203. doi:10.1126/science.1158722. URLhttps: //www.science.org/doi/10.1126/science.1158722

work page doi:10.1126/science.1158722 2008
[4]

Physical Review Letters , author =

John P. Perdew, Kieron Burke, and Matthias Ernzerhof. Gener- alized Gradient Approximation Made Simple.Physical Review Letters, 77(18):3865–3868, October 1996. ISSN 0031-9007, 1079-7114. doi:10.1103/PhysRevLett.77.3865. URLhttps: //link.aps.org/doi/10.1103/PhysRevLett.77.3865

work page doi:10.1103/physrevlett.77.3865 1996
[5]

AIP Conference Proceedings577(1), 1–20 (2001) https://doi.org/10.1063/1.1390175

John P. Perdew and Karla Schmidt. Jacob’s ladder of den- sity functional approximations for the exchange-correlation en- ergy. InAIP Conference Proceedings, volume 577, pages 1– 20, Antwerp (Belgium), 2001. AIP. doi:10.1063/1.1390175. URLhttps://pubs.aip.org/aip/acp/article/577/1/ 1-20/573973

work page doi:10.1063/1.1390175 2001
[6]

2012 , doi =

Aron J. Cohen, Paula Mori-Sánchez, and Weitao Yang. Chal- lenges for Density Functional Theory.Chemical Reviews, 112 (1):289–320, January 2012. ISSN 0009-2665, 1520-6890. doi: 10.1021/cr200107z. URLhttps://pubs.acs.org/doi/10. 1021/cr200107z

work page doi:10.1021/cr200107z 2012
[7]

Axel D. Becke. Perspective: Fifty years of density- functional theory in chemical physics.The Journal of Chemical Physics, 140(18):18A301, May 2014. ISSN 0021- 9606, 1089-7690. doi:10.1063/1.4869598. URLhttps: //pubs.aip.org/jcp/article/140/18/18A301/149389/ Perspective-Fifty-years-of-density-functional

work page doi:10.1063/1.4869598 2014
[8]

Pragya Verma and Donald G. Truhlar. Status and Chal- lenges of Density Functional Theory.Trends in Chem- istry, 2(4):302–318, April 2020. ISSN 25895974. doi: 10.1016/j.trechm.2020.02.005. URLhttps://linkinghub. elsevier.com/retrieve/pii/S2589597420300411

work page doi:10.1016/j.trechm.2020.02.005 2020
[9]

M. K. Horton, S. Dwaraknath, and K. A. Persson. Promises and perils of computational materials databases.Nature Computa- tional Science, 1(1):3–5, January 2021. ISSN 2662-8457. doi: 10.1038/s43588-020-00016-5. URLhttps://www.nature. com/articles/s43588-020-00016-5

work page doi:10.1038/s43588-020-00016-5 2021
[10]

Teale, Trygve Helgaker, Andreas Savin, Carlo Adamo, Bálint Aradi, Alexei V

Andrew M. Teale, Trygve Helgaker, Andreas Savin, Carlo Adamo, Bálint Aradi, Alexei V . Arbuznikov, Paul W. Ayers, Evert Jan Baerends, Vincenzo Barone, Patrizia Calaminici, Eric Cancès, Emily A. Carter, Pratim Kumar Chattaraj, Henry Chermette, Ilaria Ciofini, T. Daniel Crawford, Frank De Proft, John F. Dobson, Claudia Draxl, Thomas Frauenheim, Em- manuel F...

work page doi:10.1039/d2cp02827a 2022
[11]

Nature Communications 2020 11:1 11(1), 1–10 (2020) https://doi.org/10.1038/s41467-020-17265-7

Sebastian Dick and Marivi Fernandez-Serra. Machine learning accurate exchange and correlation functionals of the electronic density.Nature Communications, 11(1):3509, July 2020. ISSN 2041-1723. doi:10.1038/s41467-020-17265-7. URLhttps: //www.nature.com/articles/s41467-020-17265-7

work page doi:10.1038/s41467-020-17265-7 2020
[12]

Nature Reviews Physics 2022 4:64(6), 357–358 (2022) https://doi.org/ 10.1038/s42254-022-00470-2

Ryan Pederson, Bhupalee Kalita, and Kieron Burke. Ma- chine learning and density functional theory.Nature Reviews Physics, 4(6):357–358, May 2022. ISSN 2522-5820. doi: 10.1038/s42254-022-00470-2. URLhttps://www.nature. com/articles/s42254-022-00470-2

work page doi:10.1038/s42254-022-00470-2 2022
[13]

Roadmap on Machine learn- ing in electronic structure.Electronic Structure, 4(2):023004, June 2022

H J Kulik, T Hammerschmidt, J Schmidt, S Botti, M A L Marques, M Boley, M Scheffler, M Todorovi ´c, P Rinke, C Oses, A Smolyanyuk, S Curtarolo, A Tkatchenko, A P Bartók, S Manzhos, M Ihara, T Carrington, J Behler, O Isayev, M Veit, A Grisafi, J Nigam, M Ceriotti, K T Schütt, J West- ermayr, M Gastegger, R J Maurer, B Kalita, K Burke, R Na- gai, R Akashi, ...

work page doi:10.1088/2516-1075/ac572f 2022
[14]

Nonlocal machine-learned exchange functional for molecules and solids.Physical Re- view B, 110(7):075130, August 2024

Kyle Bystrom and Boris Kozinsky. Nonlocal machine-learned exchange functional for molecules and solids.Physical Re- view B, 110(7):075130, August 2024. ISSN 2469-9950, 2469-

work page 2024
[15]

URLhttps:// link.aps.org/doi/10.1103/PhysRevB.110.075130

doi:10.1103/PhysRevB.110.075130. URLhttps:// link.aps.org/doi/10.1103/PhysRevB.110.075130. 12

work page doi:10.1103/physrevb.110.075130
[16]

CIDER: An Expressive, Nonlocal Feature Set for Machine Learning Density Func- tionals with Exact Constraints.Journal of Chemical Theory and Computation, 18(4):2180–2192, April 2022

Kyle Bystrom and Boris Kozinsky. CIDER: An Expressive, Nonlocal Feature Set for Machine Learning Density Func- tionals with Exact Constraints.Journal of Chemical Theory and Computation, 18(4):2180–2192, April 2022. ISSN 1549- 9618, 1549-9626. doi:10.1021/acs.jctc.1c00904. URLhttps: //pubs.acs.org/doi/10.1021/acs.jctc.1c00904

work page doi:10.1021/acs.jctc.1c00904 2022
[17]

Accurate and scalable exchange-correlation with deep learning

Giulia Luise, Chin-Wei Huang, Thijs V ogels, Derk P. Kooi, Se- bastian Ehlert, Stephanie Lanius, Klaas J. H. Giesbertz, Amir Karton, Deniz Gunceler, Megan Stanley, Wessel P. Bruinsma, Lin Huang, Xinran Wei, José Garrido Torres, Abylay Katba- shev, Rodrigo Chavez Zavaleta, Bálint Máté, Sékou-Oumar Kaba, Roberto Sordillo, Yingrong Chen, David B. Williams- Y...

work page internal anchor Pith review Pith/arXiv arXiv 2025
[18]

Can ma- chines learn density functionals? Past, present, and future of ML in DFT, 2025

Ryosuke Akashi, Mihira Sogal, and Kieron Burke. Can ma- chines learn density functionals? Past, present, and future of ML in DFT, 2025. URLhttps://arxiv.org/abs/2503. 01709. Version Number: 1

work page 2025
[19]

Lars Goerigk, Andreas Hansen, Christoph Bauer, Stephan Ehrlich, Asim Najibi, and Stefan Grimme. A look at the density functional theory zoo with the advanced GMTKN55 database for general main group thermochemistry, kinetics and non- covalent interactions.Physical Chemistry Chemical Physics, 19(48):32184–32215, 2017. ISSN 1463-9076, 1463-9084. doi:10.1039/...

work page doi:10.1039/c7cp04913g 2017
[20]

“Slim” Benchmark Sets for Faster Method Development.Journal of Chemical Theory and Computation, 21(13):6517–6527, July 2025

Tim Gould and Stefan Vuckovic. “Slim” Benchmark Sets for Faster Method Development.Journal of Chemical Theory and Computation, 21(13):6517–6527, July 2025. ISSN 1549- 9618, 1549-9626. doi:10.1021/acs.jctc.5c00512. URLhttps: //pubs.acs.org/doi/10.1021/acs.jctc.5c00512

work page doi:10.1021/acs.jctc.5c00512 2025
[21]

font-variant:small- caps

Xing Zhang and Garnet Kin-Lic Chan. Differentiable quantum chemistry with <span style="font-variant:small- caps;">PySCF</span> for molecules and materials at the mean-field level and beyond.The Journal of Chemical Physics, 157(20):204801, November 2022. ISSN 0021- 9606, 1089-7690. doi:10.1063/5.0118200. URLhttps:// pubs.aip.org/jcp/article/157/20/204801/2...

work page doi:10.1063/5.0118200 2022
[22]

Highly accu- rate and constrained density functional obtained with dif- ferentiable programming.Physical Review B, 104(16): L161109, October 2021

Sebastian Dick and Marivi Fernandez-Serra. Highly accu- rate and constrained density functional obtained with dif- ferentiable programming.Physical Review B, 104(16): L161109, October 2021. ISSN 2469-9950, 2469-9969. doi: 10.1103/PhysRevB.104.L161109. URLhttps://link.aps. org/doi/10.1103/PhysRevB.104.L161109

work page doi:10.1103/physrevb.104.l161109 2021
[23]

Data-efficient learning of exchange-correlation functionals with differentiable DFT.Ma- chine Learning: Science and Technology, 7(2):025001, April

Antonius V on Strachwitz, Karim K Alaa El-Din, Ana C C Dutra, and Sam M Vinko. Data-efficient learning of exchange-correlation functionals with differentiable DFT.Ma- chine Learning: Science and Technology, 7(2):025001, April

work page
[24]

doi:10.1088/2632-2153/ae3c5a

ISSN 2632-2153. doi:10.1088/2632-2153/ae3c5a. URLhttps://iopscience.iop.org/article/10.1088/ 2632-2153/ae3c5a

work page doi:10.1088/2632-2153/ae3c5a
[25]

James Kirkpatrick, Brendan McMorrow, David H. P. Turban, Alexander L. Gaunt, James S. Spencer, Alexander G. D. G. Matthews, Annette Obika, Louis Thiry, Meire Fortunato, David Pfau, Lara Román Castellanos, Stig Petersen, Alexander W. R. Nelson, Pushmeet Kohli, Paula Mori-Sánchez, Demis Hassabis, and Aron J. Cohen. Pushing the frontiers of density functiona...

work page doi:10.1126/science.abj6511 2021
[26]

Ryan Pederson and Kieron Burke. The difference between molecules and materials: Reassessing the role of exact con- ditions in density functional theory.The Journal of Chemical Physics, 159(21):214113, December 2023. ISSN 0021- 9606, 1089-7690. doi:10.1063/5.0172058. URLhttps:// pubs.aip.org/jcp/article/159/21/214113/2926418/ The-difference-between-molecul...

work page doi:10.1063/5.0172058 2023
[27]

Perdew and Stefan Kurth

John P. Perdew and Stefan Kurth. Density Functionals for Non-relativistic Coulomb Systems in the New Century. In R. Beig, B. G. Englert, U. Frisch, P. Hänggi, K. Hepp, W. Hille- brandt, D. Imboden, R. L. Jaffe, R. Lipowsky, H. V . Löhney- sen, I. Ojima, D. Sornette, S. Theisen, W. Weise, J. Wess, J. Zittartz, Carlos Fiolhais, Fernando Nogueira, and Miguel...

work page doi:10.1007/3-540-37072-2_1 2003
[28]

Jianwei Sun, Adrienn Ruzsinszky, and John P. Perdew. Strongly Constrained and Appropriately Normed Semilo- cal Density Functional.Physical Review Letters, 115(3): 036402, July 2015. ISSN 0031-9007, 1079-7114. doi: 10.1103/PhysRevLett.115.036402. URLhttps://link. aps.org/doi/10.1103/PhysRevLett.115.036402

work page doi:10.1103/physrevlett.115.036402 2015
[29]

Machine- learning-based exchange correlation functional with physi- cal asymptotic constraints.Physical Review Research, 4 (1):013106, February 2022

Ryo Nagai, Ryosuke Akashi, and Osamu Sugino. Machine- learning-based exchange correlation functional with physi- cal asymptotic constraints.Physical Review Research, 4 (1):013106, February 2022. ISSN 2643-1564. doi: 10.1103/PhysRevResearch.4.013106. URLhttps://link. aps.org/doi/10.1103/PhysRevResearch.4.013106

work page doi:10.1103/physrevresearch.4.013106 2022
[30]

Cubuk, Patrick Riley, and Kieron Burke

Li Li, Stephan Hoyer, Ryan Pederson, Ruoxi Sun, Ekin D. Cubuk, Patrick Riley, and Kieron Burke. Kohn-Sham Equations as Regularizer: Building Prior Knowledge into Machine-Learned Physics.Physical Review Letters, 126 (3):036401, January 2021. ISSN 0031-9007, 1079-7114. doi:10.1103/PhysRevLett.126.036401. URLhttps://link. aps.org/doi/10.1103/PhysRevLett.126.036401

work page doi:10.1103/physrevlett.126.036401 2021
[31]

SoftwareX7, 1–5 (2018) https://doi.org/10.1016/J.SOFTX.2017.11.002

Susi Lehtola, Conrad Steigemann, Micael J.T. Oliveira, and Miguel A.L. Marques. Recent developments in libxc — A comprehensive library of functionals for density functional the- ory.SoftwareX, 7:1–5, January 2018. ISSN 23527110. doi: 10.1016/j.softx.2017.11.002. URLhttps://linkinghub. elsevier.com/retrieve/pii/S2352711017300602

work page doi:10.1016/j.softx.2017.11.002 2018
[32]

Soler, Alberto García, Micael J

José M. Soler, Alberto García, Micael J. T. Oliveira, Yan Pouil- lon, Nick R. Papior, and Luis C. Balbás. {libGridXC}: A library for the computation of exchange and correlation ener- gies and potentials in radial and {3D} grids. URLhttps: //gitlab.com/siesta-project/libraries/libgridxc

work page
[33]

A novel single-crystal & single-pass source for polarisation- and colour-entangled photon pairs,

Alexander Ryabov, Iskander Akhatov, and Petr Zhilyaev. Ap- plication of two-component neural network for exchange- correlation functional interpolation.Scientific Reports, 12(1): 14133, August 2022. ISSN 2045-2322. doi:10.1038/s41598- 022-18083-1. URLhttps://www.nature.com/articles/ s41598-022-18083-1

work page doi:10.1038/s41598- 2022
[34]

Furness, Yi Yao, V olker Blum, Tom J

Kanun Pokharel, James W. Furness, Yi Yao, V olker Blum, Tom J. P. Irons, Andrew M. Teale, and Jianwei Sun. Ex- act constraints and appropriate norms in machine-learned exchange-correlation functionals.The Journal of Chemical Physics, 157(17):174106, November 2022. ISSN 0021- 9606, 1089-7690. doi:10.1063/5.0111183. URLhttps:// pubs.aip.org/jcp/article/157/...

work page doi:10.1063/5.0111183 2022
[35]

Journal of Chemical Physics 136(15) (2012) https://doi.org/10.1063/1.4704546

Kieron Burke. Perspective on density functional theory.The 13 Journal of Chemical Physics, 136(15):150901, April 2012. ISSN 0021-9606, 1089-7690. doi:10.1063/1.4704546. URL https://pubs.aip.org/jcp/article/136/15/150901/ 941589/Perspective-on-density-functional-theory

work page doi:10.1063/1.4704546 2012
[36]

Lieb and Stephen Oxford

Elliott H. Lieb and Stephen Oxford. Improved lower bound on the indirect Coulomb energy.International Journal of Quantum Chemistry, 19(3):427–439, March 1981. ISSN 0020-7608, 1097-461X. doi:10.1002/qua.560190306. URL https://onlinelibrary.wiley.com/doi/10.1002/qua. 560190306

work page doi:10.1002/qua.560190306 1981
[37]

Perdew, Adrienn Ruzsinszky, Jianwei Sun, and Kieron Burke

John P. Perdew, Adrienn Ruzsinszky, Jianwei Sun, and Kieron Burke. Gedanken densities and exact constraints in density functional theory.The Journal of Chemi- cal Physics, 140(18):18A533, May 2014. ISSN 0021- 9606, 1089-7690. doi:10.1063/1.4870763. URLhttps: //pubs.aip.org/jcp/article/140/18/18A533/149384/ Gedanken-densities-and-exact-constraints-in

work page doi:10.1063/1.4870763 2014
[38]

Lieb, and Robert Seiringer

Mathieu Lewin, Elliott H. Lieb, and Robert Seiringer. Im- proved Lieb–Oxford bound on the indirect and exchange en- ergies.Letters in Mathematical Physics, 112(5):92, October

work page
[39]

doi:10.1007/s11005-022- 01584-5

ISSN 0377-9017, 1573-0530. doi:10.1007/s11005-022- 01584-5. URLhttps://link.springer.com/10.1007/ s11005-022-01584-5

work page doi:10.1007/s11005-022-
[40]

xcquinox-clone

Sara Navarro-Rodríguez and Alec Wills. xcquinox-clone. URL https://gitlab.com/saru1799/xcquinox-clone

work page
[41]

{xcquinox}: A machine learning package using the Equinox library for learning exchange–correlation func- tionals with JAX

Alec Wills. {xcquinox}: A machine learning package using the Equinox library for learning exchange–correlation func- tionals with JAX. URLhttps://github.com/alecpwills/ xcquinox/tree/b893ab8. GitHub repository, commit b893ab8, accessed 6 May 2026

work page 2026
[42]

Bryantsev, Mamadou S

Vyacheslav S. Bryantsev, Mamadou S. Diallo, Adri C. T. Van Duin, and William A. Goddard. Evaluation of B3LYP, X3LYP, and M06-Class Density Functionals for Predicting the Binding Energies of Neutral, Protonated, and Deprotonated Water Clusters.Journal of Chemical Theory and Computation, 5(4):1016–1026, April 2009. ISSN 1549-9618, 1549-9626. doi: 10.1021/ct...

work page doi:10.1021/ct800549f 2009
[43]

New accurate bench- mark energies for large water clusters: DFT is better than expected.Journal of Computational Chemistry, 35(8):634– 643, March 2014

Tony Anacker and Joachim Friedrich. New accurate bench- mark energies for large water clusters: DFT is better than expected.Journal of Computational Chemistry, 35(8):634– 643, March 2014. ISSN 0192-8651, 1096-987X. doi: 10.1002/jcc.23539. URLhttps://onlinelibrary.wiley. com/doi/10.1002/jcc.23539

work page doi:10.1002/jcc.23539 2014
[44]

Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, Shreyas Cholia, Dan Gunter, David Skinner, Gerbrand Ceder, and Kristin A. Persson. Commentary: The Materials Project: A materials genome approach to accelerating materi- als innovation.APL Materials, 1(1):011002, July 2013. ISSN 2166-532X. doi:10.1063/1.481...

work page doi:10.1063/1.4812323 2013