pith. sign in

arxiv: 2512.25061 · v3 · pith:2FYLZTZ5new · submitted 2025-12-31 · ❄️ cond-mat.mtrl-sci · physics.geo-ph

Physics-Constrained Self-Energy Warm Starts for Charge-Self-Consistent DFT+DMFT: Application to Iron at Core Conditions

Rishi Rao , Li Zhu This is my paper

Pith reviewed 2026-05-21 16:56 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci physics.geo-ph
keywords DFT+DMFTself-energygraph neural networkwarm startiron meltingcore pressureselectronic correlationscharge self-consistency
0
0 comments X

The pith

A physics-constrained graph neural network predicts self-energy components to initialize DFT+DMFT cycles and cut iterations by two to four times.

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

The paper introduces a machine-learning warm start for charge-self-consistent DFT+DMFT that uses an E(3)-equivariant graph neural network to forecast a compact representation of the local self-energy and Fermi level. This initialization respects the known high-frequency and analytic properties of the self-energy, allowing the full self-consistency loop to converge with substantially fewer DMFT steps. The approach is demonstrated on metallic iron, correlated iron oxide, and Mott-insulating nickel oxide. It then enables generation of correlated energies and forces at core pressures, training of an interatomic potential, and direct simulation of the hcp-iron melting curve in large cells. The resulting melting temperature at 330 GPa aligns with experimental bounds and highlights the contribution of dynamical correlations.

Core claim

An E(3)-equivariant graph neural network can learn a compact, real-valued representation {Σ(∞), Σ_ℓ, E_f} of the local self-energy and Fermi level that is tied to the analytic structure of Σ(iω_n) and sufficient to initialize the charge-self-consistent DFT+DMFT cycle. Across metallic Fe, correlated FeO, and Mott-insulating NiO this warm start reduces the number of DMFT iterations needed for self-consistency by a factor of two to four. The same capability supplies correlated energies and forces for iron at core pressures, allowing an equivariant machine-learned potential to be trained and the hcp-Fe melting curve to be mapped via solid-liquid coexistence in 9216-atom NVE cells, producing a 0.

What carries the argument

E(3)-equivariant graph neural network that outputs the compact representation {Σ(∞), Σ_ℓ, E_f} of the local self-energy and Fermi level, constrained to match the high-frequency and analytic structure of Σ(iω_n) and used directly to seed the DFT+DMFT self-consistency loop.

Load-bearing premise

The neural-network prediction of the self-energy representation is close enough to the true converged value that starting the cycle from it reaches the same physical fixed point without introducing systematic bias.

What would settle it

Run identical DFT+DMFT calculations for the same iron configuration once from the neural-network warm start and once from a conventional initial guess; the two runs must converge to self-energies, total energies, and derived melting temperatures that agree within numerical tolerance.

Figures

Figures reproduced from arXiv: 2512.25061 by Li Zhu, Rishi Rao.

Figure 1
Figure 1. Figure 1: FIG. 1. Top row: Predicted real (red) and imaginary (blue) parts of the self energy vs DMFT (black) self-energies for 4 [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Density profile along the long axis of simulation cell [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
read the original abstract

Charge self-consistent DFT+DMFT quantitatively captures dynamical electronic correlations in real materials, but its cost precludes the large-scale thermodynamic sampling required for phase boundaries and equations of state. Here, we develop a physics-constrained machine-learning warm start for realistic DFT+DMFT: E(3)-equivariant graph neural networks predict a compact, real-valued representation of the local self-energy and Fermi level -- \{\,$\Sigma(\infty),\,\Sigma_\ell,\,E_f\,$\} -- tied to the known high-frequency and analytic structure of $\Sigma(i\omega_n)$, and used to initialize the full DFT+DMFT self-consistency cycle. Across metallic Fe, correlated FeO, and Mott-insulating NiO, the scheme yields a 2--4 times reduction in the number of DMFT iterations required to reach self-consistency. As a demanding application, we leverage this capability to generate correlated energies and forces for Fe at core pressures, train an equivariant machine-learned interatomic potential, and determine the hcp-Fe melting curve by solid--liquid coexistence simulations in the NVE ensemble in 9216-atom cells. We obtain a melting temperature of 6225 K at 330 GPa, in agreement with recent experimental constraints and consistent with the view that dynamical electronic correlations contribute to the discrepancy between DFT-based predictions and experiment.

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.

Referee Report

1 major / 2 minor

Summary. The manuscript develops a physics-constrained warm-start procedure for charge-self-consistent DFT+DMFT in which an E(3)-equivariant graph neural network predicts the compact real-valued representation {Σ(∞), Σ_ℓ, E_f} that respects the high-frequency asymptotics and analytic structure of the self-energy. This initialization reduces the number of DMFT iterations to convergence by a factor of 2–4 across metallic Fe, correlated FeO, and Mott-insulating NiO. The method is then applied to generate correlated energies and forces for hcp-Fe at core pressures; these data train an equivariant MLIP that is used in 9216-atom NVE solid–liquid coexistence simulations, yielding a melting temperature of 6225 K at 330 GPa that is stated to be consistent with recent experimental constraints.

Significance. If the central methodological claim holds, the work supplies a practical route to large-scale thermodynamic sampling with dynamical correlations, which has been a long-standing bottleneck for equations of state and phase boundaries in strongly correlated materials. The concrete melting-temperature prediction obtained from direct NVE coexistence in large cells constitutes a falsifiable output that can be confronted with experiment, and the use of an equivariant GNN that enforces known analytic properties of Σ(iω_n) is a clear technical strength.

major comments (1)
  1. [Application section / melting-curve paragraph] The central application to core-pressure Fe (the 9216-atom coexistence runs and subsequent MLIP training) rests on the assumption that the warm-started DFT+DMFT fixed point is identical, within statistical tolerance, to the cold-started fixed point. No explicit comparison of the converged self-energy Σ(iω_n), total energy, or forces between the two initialization protocols is reported for the high-pressure configurations that enter the MLIP training set. This comparison is load-bearing for the claim that the reported melting temperature of 6225 K at 330 GPa is free of systematic bias from the GNN warm start.
minor comments (2)
  1. [Methods / computational details] The abstract and main text should state the precise convergence criterion (e.g., maximum change in Σ or total energy) used to declare self-consistency for both warm- and cold-started runs so that the reported 2–4× iteration reduction can be reproduced.
  2. [Figures] Figure captions for the self-energy or iteration-count plots should include the precise definition of the compact representation {Σ(∞), Σ_ℓ, E_f} and the loss function used to train the GNN.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their positive assessment of the methodological advance and for highlighting the importance of validating the warm-start procedure in the demanding core-pressure application. We address the single major comment below.

read point-by-point responses

  1. Referee: The central application to core-pressure Fe (the 9216-atom coexistence runs and subsequent MLIP training) rests on the assumption that the warm-started DFT+DMFT fixed point is identical, within statistical tolerance, to the cold-started fixed point. No explicit comparison of the converged self-energy Σ(iω_n), total energy, or forces between the two initialization protocols is reported for the high-pressure configurations that enter the MLIP training set. This comparison is load-bearing for the claim that the reported melting temperature of 6225 K at 330 GPa is free of systematic bias from the GNN warm start.

    Authors: We agree that an explicit demonstration of equivalence between warm-started and cold-started fixed points for the high-pressure Fe configurations is necessary to confirm the absence of systematic bias in the MLIP training data. While the DFT+DMFT self-consistency loop is deterministic and the converged fixed point is independent of initialization once convergence is reached (as already verified for Fe, FeO, and NiO at ambient and moderate pressures), we acknowledge that this check was not reported for the specific core-pressure structures. In the revised manuscript we will add a direct comparison—either in the main text or as a supplementary figure—of the converged Σ(iω_n), total energies, and forces obtained from both initialization protocols on a representative subset of the high-pressure configurations used for MLIP training. Differences will be shown to lie within numerical tolerances, thereby supporting the reliability of the 6225 K melting temperature. revision: yes

Circularity Check

0 steps flagged

No significant circularity in warm-start initialization or MLIP-based melting curve

full rationale

The paper presents a GNN-based warm start for DFT+DMFT that predicts a compact self-energy representation to accelerate convergence, with the reduction in iterations (2-4x) demonstrated by direct comparison on metallic Fe, FeO, and NiO. The demanding application generates correlated energies/forces for high-pressure Fe using this initialization, trains an equivariant MLIP on those data, and extracts the melting temperature via independent NVE solid-liquid coexistence simulations in 9216-atom cells. The reported Tm = 6225 K at 330 GPa is obtained from explicit simulation and compared to external experimental constraints rather than fitted or derived from the warm-start parameters themselves. No step reduces by construction to its inputs, no fitted quantity is relabeled as a prediction of the target observable, and the chain contains no load-bearing self-citations or uniqueness theorems; the final result remains falsifiable against experiment and independent of the initialization details beyond convergence to the same fixed point.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The method rests on the assumption that the local self-energy possesses a known high-frequency expansion and analytic structure that can be compactly represented by a finite set of real numbers; the GNN is trained on data generated by conventional DFT+DMFT runs, so its parameters are learned rather than hand-tuned, but the training data themselves depend on prior DFT+DMFT calculations.

free parameters (1)
  • GNN model parameters
    Weights of the equivariant graph neural network trained to map atomic configurations to the compact self-energy representation; these are fitted to data from conventional DFT+DMFT runs.
axioms (1)
  • domain assumption The self-energy Σ(iω_n) admits a known high-frequency expansion and analytic continuation properties that allow a compact real-valued representation {Σ(∞), Σ_ℓ, E_f}.
    Invoked in the abstract when the authors state that the GNN predicts a representation 'tied to the known high-frequency and analytic structure of Σ(iω_n)'.

pith-pipeline@v0.9.0 · 5782 in / 1700 out tokens · 40407 ms · 2026-05-21T16:56:14.762158+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

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

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

50 extracted references · 50 canonical work pages · 1 internal anchor

  1. [1]

    Abrahams and G

    E. Abrahams and G. Kotliar, The metal-insulator transi- tion in correlated disordered systems, Science274, 1853 (1996)

    work page 1996

  2. [2]

    Imada, A

    M. Imada, A. Fujimori, and Y. Tokura, Metal-insulator transitions, Rev. Mod. Phys.70, 1039 (1998)

    work page 1998

  3. [3]

    Broholm, R

    C. Broholm, R. J. Cava, S. A. Kivelson, D. G. Nocera, M. R. Norman, and T. Senthil, Quantum spin liquids, Science367, eaay0668 (2020)

    work page 2020

  4. [4]

    T. E. Lee, S. Gopalakrishnan, and M. D. Lukin, Uncon- ventional magnetism via optical pumping of interacting spin systems, Phys. Rev. Lett.110, 257204 (2013)

    work page 2013

  5. [5]

    Dagotto, Correlated electrons in high-temperature su- perconductors, Rev

    E. Dagotto, Correlated electrons in high-temperature su- perconductors, Rev. Mod. Phys.66, 763 (1994)

    work page 1994

  6. [6]

    Keimer, S

    B. Keimer, S. A. Kivelson, M. R. Norman, S. Uchida, and J. Zaanen, From quantum matter to high-temperature superconductivity in copper oxides, Nature518, 179 (2015)

    work page 2015

  7. [7]

    G. R. Stewart, Unconventional superconductivity, Adv. Phys.66, 75 (2017)

    work page 2017

  8. [8]

    V. I. Anisimov, F. Aryasetiawan, and A. I. Lichtenstein, First-principles calculations of the electronic structure and spectra of strongly correlated systems: the LDA+U method, J. Phys. Condens. Matter9, 767 (1997)

    work page 1997

  9. [9]

    Tokura and N

    Y. Tokura and N. Nagaosa, Orbital physics in transition- metal oxides, Science288, 462 (2000)

    work page 2000

  10. [10]

    L. V. Pourovskii, J. Mravlje, A. Georges, S. I. Simak, and I.A.Abrikosov,Electron–electronscatteringandthermal conductivity ofϵ-iron at earth’s core conditions, New J. Phys.19, 073022 (2017)

    work page 2017

  11. [11]

    L. V. Pourovskii, Electronic correlations in dense iron: from moderate pressure to earth’s core conditions, J. Phys. Condens. Matter31, 373001 (2019)

    work page 2019

  12. [12]

    O. Y. Vekilova, L. V. Pourovskii, I. A. Abrikosov, and S. I. Simak, Electronic correlations in Fe at earth’s inner core conditions: Effects of alloying with Ni, Phys. Rev. B91, 245116 (2015)

    work page 2015

  13. [13]

    Anzellini, A

    S. Anzellini, A. Dewaele, M. Mezouar, P. Loubeyre, and G.Morard,Meltingofironatearth’sinnercoreboundary based on fast x-ray diffraction, Science340, 464 (2013)

    work page 2013

  14. [14]

    R. G. Kraus, R. J. Hemley, S. J. Ali, J. L. Belof, L. X. Benedict, J. Bernier, D. Braun, R. E. Cohen, G. W. Collins, F. Coppari, M. P. Desjarlais, D. Fratanduono, S. Hamel, A. Krygier, A. Lazicki, J. Mcnaney, M. Millot, P. C. Myint, M. G. Newman, J. R. Rygg, D. M. Ster- bentz, S.T.Stewart, L.Stixrude, D.C.Swift, C.Wehren- berg, and J. H. Eggert, Measuring...

    work page 2022

  15. [15]

    Balugani, J

    S. Balugani, J. A. Hernandez, N. Sévelin-Radiguet, O. Mathon, V. Recoules, J. J. Kas, D. E. Eakins, H. Doyle, A. Ravasio, and R. Torchio, New constraints on the melting temperature and phase stability of shocked iron up to 270 gpa probed by ultrafast x-ray absorption spectroscopy, Phys. Rev. Lett.133, 254101 (2024)

    work page 2024

  16. [16]

    D. Alfè, G. D. Price, and M. J. Gillan, Iron under earth’s core conditions: Liquid-state thermodynamics and high- pressure melting curve from ab initio calculations, Phys. Rev. B65, 165118 (2002)

    work page 2002

  17. [17]

    Alfè, Temperature of the inner-core boundary of the earth: Melting of iron at high pressure from first- principles coexistence simulations, Phys

    D. Alfè, Temperature of the inner-core boundary of the earth: Melting of iron at high pressure from first- principles coexistence simulations, Phys. Rev. B79, 060101 (2009)

    work page 2009

  18. [18]

    A. B. Belonoshko, J. Fu, and G. Smirnov, Free energies of iron phases at high pressure and temperature: Molecular dynamics study, Phys. Rev. B104, 104103 (2021)

    work page 2021

  19. [19]

    A. B. Belonoshko, R. Ahuja, and B. Johansson, Quasi–ab initio molecular dynamic study of Fe melting, Phys. Rev. Lett.84, 3638 (2000)

    work page 2000

  20. [20]

    González-Cataldo and B

    F. González-Cataldo and B. Militzer, Ab initio determi- nation of iron melting at terapascal pressures and super- earths core crystallization, Phys. Rev. Res.5, 033194 (2023)

    work page 2023

  21. [21]

    Sola and D

    E. Sola and D. Alfè, Melting of iron under earth’s core conditions from diffusion monte carlo free energy calcu- lations, Phys. Rev. Lett.103, 078501 (2009)

    work page 2009

  22. [22]

    Stixrude, Melting in super-earths, Philos

    L. Stixrude, Melting in super-earths, Philos. Trans. R. Soc. A372, 20130076 (2014)

    work page 2014

  23. [23]

    Y. Sun, F. Zhang, M. I. Mendelev, R. M. Wentzcovitch, and K.-M. Ho, Two-step nucleation of the earth’s inner core, Proc. Natl. Acad. Sci. U.S.A.119, e2113059119 (2022)

    work page 2022

  24. [24]

    Y. Sun, M. I. Mendelev, F. Zhang, X. Liu, B. Da, C.- Z. Wang, R. M. Wentzcovitch, and K.-M. Ho, Ab ini- tio melting temperatures of bcc and hcp iron under the earth’s inner core condition, Geophys. Res. Lett.50, e2022GL102447 (2023)

    work page 2023

  25. [25]

    F. Wu, S. Wu, C.-Z. Wang, K.-M. Ho, R. M. Wentzcov- 7 itch, and Y. Sun, Melting temperature of iron under the earth’s inner core condition from deep machine learning, Geosci. Front.15, 101925 (2024)

    work page 2024

  26. [26]

    Kotliar, S

    G. Kotliar, S. Y. Savrasov, K. Haule, V. S. Oudovenko, O. Parcollet, and C. A. Marianetti, Electronic struc- ture calculations with dynamical mean-field theory, Rev. Mod. Phys.78, 865 (2006)

    work page 2006

  27. [27]

    Paul and T

    A. Paul and T. Birol, Applications of DFT + DMFT in materials science, Annu. Rev. Mater. Res.49, 31 (2019)

    work page 2019

  28. [28]

    Mitra and H

    P. Mitra and H. Banerjee, Deep learning-based predic- tion of self-energies from *ab initio* dynamical mean- field theory for real materials with minimal data sets, ChemRxiv 10.26434/chemrxiv-2025-dp7rd (2025)

    work page doi:10.26434/chemrxiv-2025-dp7rd 2025

  29. [29]

    X. Dong, E. Gull, and L. Wang, Equivariant neural net- work for green’s functions of molecules and materials, Phys. Rev. B109, 075112 (2024)

    work page 2024

  30. [30]

    E. J. Sturm, M. R. Carbone, D. Lu, A. Weichselbaum, and R. M. Konik, Predicting impurity spectral functions usingmachinelearning,Phys.Rev.B103,245118(2021)

    work page 2021

  31. [31]

    Arsenault, A

    L.-F. Arsenault, A. Lopez-Bezanilla, O. A. von Lilienfeld, andA.J.Millis,Machinelearningformany-bodyphysics: The case of the anderson impurity model, Phys. Rev. B 90, 155136 (2014)

    work page 2014

  32. [32]

    X. Wang, H. T. Dang, and A. J. Millis, High-frequency asymptotic behavior of self-energies in quantum impurity models, Phys. Rev. B84, 073104 (2011)

    work page 2011

  33. [33]

    Boehnke, H

    L. Boehnke, H. Hafermann, M. Ferrero, F. Lechermann, and O. Parcollet, Orthogonal polynomial representation of imaginary-time Green’s functions, Phys. Rev. B84, 075145 (2011)

    work page 2011

  34. [34]

    Parcollet, M

    O. Parcollet, M. Ferrero, T. Ayral, H. Hafermann, I. Krivenko, L. Messio, and P. Seth, TRIQS: A toolbox for research on interacting quantum systems, Comput. Phys. Commun.196, 398 (2015)

    work page 2015

  35. [35]

    e3nn: Euclidean neural net- works,

    M. Geiger and T. Smidt, e3nn: Euclidean neural net- works, arXiv:2207.09453

    work page arXiv

  36. [36]

    Kondor, Z

    R. Kondor, Z. Lin, and S. Trivedi, Clebsch–gordan nets: a fully fourier space spherical convolutional neural net- work, inAdv. Neural Inf. Process. Syst., Vol. 31 (Curran Associates, Inc., 2018)

    work page 2018

  37. [37]

    Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds

    N. Thomas, T. Smidt, S. Kearnes, L. Yang, L. Li, K. Kohlhoff, and P. Riley, Tensor field networks: Rotation- and translation-equivariant neural networks for 3d point clouds, arXiv:1802.08219

    work page internal anchor Pith review Pith/arXiv arXiv

  38. [38]

    Weiler, M

    M. Weiler, M. Geiger, M. Welling, W. Boomsma, and T. S. Cohen, 3D steerable CNNs: Learning rotationally equivariant features in volumetric data, inAdv. Neu- ral Inf. Process. Syst., Vol. 31 (Curran Associates, Inc., 2018)

    work page 2018

  39. [39]

    L. Zhu, M. Amsler, T. Fuhrer, B. Schaefer, S. Faraji, S. Rostami, S. A. Ghasemi, A. Sadeghi, M. Grauzinyte, C.Wolverton,andS.Goedecker,Afingerprintbasedmet- ric for measuring similarities of crystalline structures, J. Chem. Phys.144, 034203 (2016)

    work page 2016

  40. [40]

    Rao and L

    R. Rao and L. Zhu, Phase transitions of correlated sys- tems from graph neural networks with quantum embed- ding techniques, Phys. Rev. B110, 245111 (2024)

    work page 2024

  41. [41]

    C. W. Tan, M. L. Descoteaux, M. Kotak, G. de Mi- randa Nascimento, S. R. Kavanagh, L. Zichi, M. Wang, A. Saluja, Y. R. Hu, T. Smidt, A. Johansson, W. C. Witt, B. Kozinsky, and A. Musaelian, High-performance training and inference for deep equivariant interatomic potentials, arXiv:2504.16068

    work page arXiv

  42. [42]

    Batzner, A

    S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials, Nat. Commun.13, 2453 (2022)

    work page 2022

  43. [43]

    Huang and B

    C. Huang and B. M. Rubenstein, Machine Learning Dif- fusion Monte Carlo Forces, J. Phys. Chem. A127, 339 (2023)

    work page 2023

  44. [44]

    Lechner and C

    W. Lechner and C. Dellago, Accurate determination of crystal structures based on averaged local bond order pa- rameters, J. Chem. Phys.129, 114707 (2008)

    work page 2008

  45. [45]

    Bemerkungen zur schmelzdrukkurve, Z. Anorg. Allg. Chem.178, 309 (1929)

    work page 1929

  46. [46]

    Sinmyo, K

    R. Sinmyo, K. Hirose, and Y. Ohishi, Melting curve of iron to 290 GPa determined in a resistance-heated diamond-anvil cell, Earth Planet. Sci. Lett.510, 45 (2019)

    work page 2019

  47. [47]

    A. V. Ruban, A. B. Belonoshko, and N. V. Skorodumova, Impact of magnetism on Fe under earth’s core conditions, Phys. Rev. B87, 014405 (2013)

    work page 2013

  48. [48]

    A. G. Gavriliuk, V. V. Struzhkin, A. G. Ivanova, V. B. Prakapenka, A. A. Mironovich, S. N. Aksenov, I. A. Troyan, and W. Morgenroth, The first-order structural transition in NiO at high pressure, Commun. Phys.6, 23 (2023)

    work page 2023

  49. [49]

    W.-G. D. Ho, P. Zhang, K. Haule, J. M. Jackson, V. Do- brosavljević, and V. V. Dobrosavljevic, Quantum critical phase of FeO spans conditions of earth’s lower mantle, Nat. Commun.15, 3461 (2024)

    work page 2024

  50. [50]

    J. Li, Q. Wu, J. Li, T. Xue, Y. Tan, X. Zhou, Y. Zhang, Z. Xiong, Z. Gao, and T. Sekine, Shock melting curve of iron: A consensus on the temperature at the earth’s inner core boundary, Geophys. Res. Lett.47, e2020GL087758 (2020)

    work page 2020