pith. sign in

arxiv: 2504.20317 · v1 · pith:P4C6CZO2new · submitted 2025-04-29 · ❄️ cond-mat.str-el · physics.comp-ph· physics.data-an

Ant Colony Optimization for Density Functionals in Strongly Correlated Systems

G. M. Tonin , T. Pauletti , R. M. Dos Santos , V. V. Fran\c{c}a This is my paper

Pith reviewed 2026-05-22 19:13 UTC · model grok-4.3

classification ❄️ cond-mat.str-el physics.comp-phphysics.data-an
keywords ant colony optimizationdensity functionalstrongly correlated systemsground-state energymean relative errorparameter optimization
0
0 comments X

The pith

Ant colony optimization reduces error in the FVC density functional to 0.8 percent for strongly correlated systems.

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

This paper adapts the ant colony optimization algorithm to tune the parameters of the FVC density functional for calculating ground-state energies in strongly correlated quantum systems. It runs the algorithm across one to five dimensions, where each dimension corresponds to the number of parameters being adjusted simultaneously. With fifteen ants and a pheromone evaporation rate above 0.2, the approach reaches its lowest errors in the three- and five-dimensional cases, cutting the mean relative error from 2.4 percent to roughly 0.8 percent. A reader would care because density functionals offer a practical route to properties of materials with strong electron interactions, and this method provides an efficient search for better parameter values at modest computational cost.

Core claim

The paper establishes that the ant colony optimization algorithm, when applied to the FVC density functional, identifies parameter sets that lower the mean relative error of ground-state energy calculations to approximately 0.8 percent in three- and five-dimensional optimizations. This constitutes a 67 percent error reduction relative to the original functional's 2.4 percent mean relative error. Fifteen ants with pheromone evaporation rates superior to 0.2 prove sufficient across broad ranges of interaction strength, particle density, and spin magnetization, while computation time scales nearly linearly with dimension.

What carries the argument

The ant colony optimization algorithm, in which a swarm of artificial ants iteratively deposits and follows pheromone trails to locate optimal parameter values for the FVC functional.

If this is right

  • Three- and five-dimensional optimizations achieve the lowest mean relative error of approximately 0.8 percent.
  • Fifteen ants combined with a pheromone evaporation rate above 0.2 suffice to reach the minimum error across wide parameter regimes.
  • Simulation time increases almost linearly as the number of optimized parameters grows from one to five dimensions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The linear growth of runtime with dimension implies that the same algorithm could optimize functionals containing more than five parameters without a sharp rise in cost.
  • The superior performance of the three- and five-dimensional searches suggests that expanding the number of free parameters in the functional form itself may systematically improve accuracy within the tested regimes.

Load-bearing premise

Minimizing mean relative error over the specific regimes of interaction strength, particle density, and spin magnetization used in the optimization produces a functional that remains accurate for other strongly correlated systems and for observables not included in the training set.

What would settle it

Apply the optimized functional to compute ground-state energies for a strongly correlated system whose interaction strength, density, or magnetization lies outside the ranges used during the ant colony search and verify whether the mean relative error stays near 0.8 percent.

Figures

Figures reproduced from arXiv: 2504.20317 by G. M. Tonin, R. M. Dos Santos, T. Pauletti, V. V. Fran\c{c}a.

Figure 1
Figure 1. Figure 1: FIG. 1: Illustration of ant behavior during [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (b) reveals unstable results for Q < 40, what could then compromise the efficiency of the algorithm, especially for higher number of ants. This highlights the need to reinforce pheromone trails through good solutions, which is directly controlled by Q. For Q > 40, there is a substantial improvement in convergence, to￾wards stability, suggesting relative robustness once this parameter reaches a certain thre… view at source ↗
Figure 3
Figure 3. Figure 3: shows that for lower dimensions (1D, 2D) and for 4D, the Bestfval stabilizes at higher values (Bestfval ≥ 75), suggesting lim￾ited exploration of the search space. In con￾trast, the 3D and 5D optimizations stabilize FIG. 3: Optimal function value (Bestfval) and mean relative error (MRE) as a function of the dimensionality of the optimization, for distinct combinations of the optimized parameters P in Eqs. … view at source ↗
read the original abstract

The Ant Colony Optimization (ACO) algorithm is a nature-inspired metaheuristic method used for optimization problems. Although not a machine learning method per se, ACO is often employed alongside machine learning models to enhance performance through optimization. We adapt an ACO algorithm to optimize the so-called FVC density functional for the ground-state energy of strongly correlated systems. We find the parameter configurations that maximize optimization efficiency, while reducing the mean relative error ($MRE$) of the ACO functional. We then analyze the algorithm's performance across different dimensionalities ($1D-5D$), which are related to the number of parameters to be optimized within the FVC functional. Our results indicate that $15$ ants with a pheromone evaporation rate superior to $0.2$ are sufficient to minimize the $MRE$ for a vast regime of parameters of the strongly-correlated system -- interaction, particle density and spin magnetization. While the optimizations $1D$, $2D$, and $4D$ yield $1.5\%< MRE< 2.7\%$, the $3D$ and $5D$ optimizations lower the $MRE$ to $\sim0.8\%$, reflecting a $67\%$ error reduction compared to the original FVC functional ($MRE = 2.4\%$). As simulation time grows almost linearly with dimension, our results highlight the potential of ant colony algorithms for density-functional problems, combining effectiveness with low computational cost.

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

2 major / 2 minor

Summary. The manuscript applies the Ant Colony Optimization (ACO) metaheuristic to tune the free parameters of the FVC density functional for ground-state energies of strongly correlated systems. It reports that 3D and 5D optimizations (corresponding to the number of variational parameters) reduce the mean relative error (MRE) from 2.4% (original FVC) to ~0.8%, a 67% improvement, while finding that 15 ants and pheromone evaporation rates >0.2 suffice across ranges of interaction strength, particle density, and spin magnetization. Simulation time is stated to grow nearly linearly with dimension.

Significance. If the reported MRE reductions prove robust under out-of-sample testing and transfer to observables or regimes not used in the fit, the work would illustrate a computationally inexpensive route to refining approximate density functionals for strongly correlated regimes. The linear scaling with dimensionality and the identification of minimal ant-colony hyperparameters are potentially useful practical findings for similar optimization tasks in condensed-matter theory.

major comments (2)
  1. [Abstract] Abstract: The headline result that 3D/5D ACO optimizations lower MRE to ~0.8% is obtained by directly minimizing that same error metric on the interaction strengths, densities, and magnetizations under study. No information is supplied on whether these points constitute a training set, whether a held-out test set or cross-validation was used, or how the MRE was evaluated on independent observables or system sizes. This leaves open the possibility that the 67% reduction reflects ordinary in-sample parameter tuning rather than improved physical fidelity.
  2. [Abstract] Abstract: The performance claims are presented without concrete details on the Hamiltonian instances (e.g., lattice sizes, boundary conditions, or specific model parameters), baseline comparisons beyond the original FVC functional, or statistical error bars on the reported MRE values. These omissions make it impossible to assess the statistical significance or reproducibility of the stated improvements.
minor comments (2)
  1. [Abstract] The abstract states that simulation time grows 'almost linearly' with dimension but supplies no supporting timing data or scaling plot; a brief table or figure quantifying wall-clock time versus dimensionality would strengthen this claim.
  2. [Abstract] Notation for the FVC functional parameters and the precise definition of the mean relative error (MRE) should be introduced explicitly, even if only in the abstract, to allow readers to judge the optimization target.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful reading and constructive comments on our manuscript. We address each major comment below, providing clarifications on our methodology and indicating the revisions that will be made to strengthen the presentation and address the concerns.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The headline result that 3D/5D ACO optimizations lower MRE to ~0.8% is obtained by directly minimizing that same error metric on the interaction strengths, densities, and magnetizations under study. No information is supplied on whether these points constitute a training set, whether a held-out test set or cross-validation was used, or how the MRE was evaluated on independent observables or system sizes. This leaves open the possibility that the 67% reduction reflects ordinary in-sample parameter tuning rather than improved physical fidelity.

    Authors: We thank the referee for highlighting this important distinction. The ACO optimization minimizes the MRE directly over a dense grid of interaction strengths, particle densities, and spin magnetizations that define the regime of strongly correlated systems under study. This is intentional: the FVC functional parameters are meant to be broadly applicable rather than instance-specific, so the procedure searches for values that improve accuracy across the explored parameter space. The systematic improvement from 2.4% to ~0.8% only in the 3D and 5D cases, together with the robustness across wide ranges of U, n, and m, suggests the gain reflects better physical representation rather than simple in-sample fitting. We nevertheless agree that explicit discussion of generalization would be valuable. In the revised manuscript we will clarify the optimization as a global parameter search, add a dedicated paragraph on the procedure, and include preliminary checks on independent system sizes and observables not used in the minimization. revision: partial

  2. Referee: [Abstract] Abstract: The performance claims are presented without concrete details on the Hamiltonian instances (e.g., lattice sizes, boundary conditions, or specific model parameters), baseline comparisons beyond the original FVC functional, or statistical error bars on the reported MRE values. These omissions make it impossible to assess the statistical significance or reproducibility of the stated improvements.

    Authors: We agree that additional concrete information will improve reproducibility and allow readers to judge statistical significance. The calculations employ the Hubbard Hamiltonian on finite lattices with periodic boundary conditions; we will expand the methods section to list the specific lattice sizes, the exact ranges of interaction strength U, filling n, and magnetization m, and the number of independent ACO runs performed. We will also report standard deviations across these runs as error bars on the MRE values and make the comparison to the original (untuned) FVC functional explicit. These details and the associated statistical information will be added in the revised version. revision: yes

Circularity Check

0 steps flagged

No significant circularity: optimization results are direct outcomes of parameter tuning

full rationale

The paper applies ACO to tune parameters of the existing FVC density functional in order to minimize MRE on chosen regimes of interaction strength, density, and magnetization. The reported MRE values (∼0.8% for 3D/5D cases versus 2.4% for the original) are the measured performance after optimization on the same data distribution used for tuning. This is the expected and explicit goal of the optimization procedure rather than a claimed first-principles derivation or out-of-sample prediction that reduces to its inputs by construction. No self-definitional equations, load-bearing self-citations, uniqueness theorems, or smuggled ansatzes appear in the abstract or described chain. The study is self-contained as a demonstration of metaheuristic tuning and does not require external validation to report its optimization results.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the premise that the FVC functional form is adequate and that ACO can locate parameter values that generalize beyond the optimization data.

free parameters (1)
  • FVC functional parameters
    The adjustable coefficients inside the FVC density functional are the quantities being varied by the ACO search.
axioms (1)
  • domain assumption The FVC density functional provides a reasonable starting approximation whose accuracy can be improved by parameter tuning alone.
    The paper never alters the functional form, only its numerical coefficients.

pith-pipeline@v0.9.0 · 5816 in / 1311 out tokens · 54619 ms · 2026-05-22T19:13:42.032502+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages

  1. [1]

    Prezhdo, O. V.;J. Phys. Chem. Lett. 2020, 11,

    work page 2020

  2. [2]

    Karthikeyan, A.; Priyakumar, U. D.;J. Chem. Sci. 2022, 134, 1. [Crossref]

    work page 2022

  3. [3]

    Chen, J.; Gao, Q.; Huang, M.; Yu, K.;Phys. Chem. Chem. Phys. 2025, 27, 1. [Crossref]

    work page 2025

  4. [4]

    Tropsha, A.; Isayev, O.; Varnek, A.; Schneider, G.; Cherkasov, A.; Nat. Rev. Drug Discovery 2023, 23, 141. [Crossref]

    work page 2023

  5. [5]

    A.; Ersan, R

    Barcin, T.; Yucel, M. A.; Ersan, R. H.; Alagoz, M. A.; Dogen, A.; Burmaoglu, S.; Algul, O.;J. Mol. Struct. 2024, 1295, 136668. [Crossref]

    work page 2024

  6. [6]

    [Crossref]

    Qi, X.; Zhao, Y.; Qi, Z.; Hou, S.; Chen, J.; Molecules 2024, 29, 903. [Crossref]

    work page 2024

  7. [7]

    N.; Behler, J.; ACS Catal

    Omranpour, A.; Elsner, J.; Lausch, K. N.; Behler, J.; ACS Catal. 2025, 15, 1616. [Cross- ref]

    work page 2025

  8. [8]

    S.; Balyakin, I

    Pikalova, N. S.; Balyakin, I. A.; Yuryev, A. A.; Rempel, A. A.;Dokl. Phys. Chem. 2024, 514,

    work page 2024

  9. [9]

    M.; Saidi, W

    Andolina, C. M.; Saidi, W. A.;Digit. Discovery 2024, 3, 1421. [Crossref]

    work page 2024

  10. [10]

    Custódio, C.A.; Filletti, R.; França, V.V.;Sci. Rep. 2019, 9, 1. [Crossref]

    work page 2019

  11. [11]

    J.; Helm, D.; Arch

    Dornheim, J.; Morand, L.; Nallani, H. J.; Helm, D.; Arch. Comput. Methods Eng. 2023, 31, 1097. [Crossref]

    work page 2023

  12. [12]

    2024, 47, 1

    Habring, A.; Holler, M.; GAMM-Mitt. 2024, 47, 1. [Crossref]

    work page 2024

  13. [13]

    Dorigo, M.; Blum, C.; Theor. Comput. Sci. 2005, 344, 243. [Crossref]

    work page 2005

  14. [14]

    [Crossref]

    Dorigo, M.; Stützle, T.;Springer 2003, 0, 250. [Crossref]

    work page 2003

  15. [15]

    [Crossref]

    Chen,Wei-Neng.; Jun Zhang.; Yang Yu.;IEEE CEC., 2007, 3308. [Crossref]

    work page 2007

  16. [16]

    Sci Rep 2025 15, 12202

    Chen, D.; Jia, W.; Chen, Q.; et al. Sci Rep 2025 15, 12202. [Crossref]

    work page 2025

  17. [17]

    Dorigo, M.; Birattari, M.; Stutzle, T.; IEEE Comput. Intell. Mag. 2006, 1, 28. [Crossref]

    work page 2006

  18. [18]

    Jin, H.; Wang, W.; Cai, M.; Wang, G.; Yun, C.; Adv. Mech. Eng. 2017, 9, 168781401771370. [Crossref]

    work page 2017

  19. [19]

    Li, X.; Wang, L.;Math. Biosci. Eng. 2020, 17,

    work page 2020

  20. [20]

    V.; Zaika, O

    Riabko, A. V.; Zaika, O. V.; Kukharchuk, R. P.; Vakaliuk, T. A.; Hordiienko, I. V.;J. Phys.: Conf. Ser. 2022, 2288, 012001. [Crossref]

    work page 2022

  21. [21]

    H.; BMC Bioinf

    Shmygelska, A.; Hoos, H. H.; BMC Bioinf. 2005, 6, 1. [Crossref]

    work page 2005

  22. [22]

    H.; Pang, B.; Tan, Q.;IEEE Trans

    Wu, Y.; Low, K. H.; Pang, B.; Tan, Q.;IEEE Trans. Veh. Technol. 2021, 70, 7464. [Crossref]

    work page 2021

  23. [23]

    Goshen, Y.; Kraisler, E.;J. Phys. Chem. Lett. 2024, 15, 2337. [Crossref]

    work page 2024

  24. [24]

    C.;JPhysical Review

    Slater, J. C.;JPhysical Review. 1951, 81, 385. [Crossref]

    work page 1951

  25. [25]

    G.; Phys

    Valatin, J. G.; Phys. Rev. 1961, 122, 1012. [Crossref]

    work page 1961

  26. [26]

    Witzorky, C.; Wang, L.; Klamroth, T.;Phys. Rev. B 2024, 110, 134303. [Crossref]

    work page 2024

  27. [27]

    Nys, J.; Pescia, G.; Sinibaldi, A.; Carleo, G.; Nat. Commun. 2024, 15, 1. [Crossref]

    work page 2024

  28. [28]

    van Mourik, T.; Bühl, M.; Gaigeot, M.-P.;Phi- los. Trans. R. Soc., A 2014, 372, 20120488. [Crossref]

    work page 2014

  29. [29]

    Kohn, W.; Rev. Mod. Phys. 1999, 71, 1253. [Crossref]

    work page 1999

  30. [30]

    M.; Mazzola, G.; Mez- zacapo, A.; Millis, A.; Robledo Moreno, J.; Ne- upert, T.; Nomura, Y.; Nys, J.; Parcollet, O.; Pohle, R.; Romero, I.; Schmid, M.; Silvester, J

    Wu, D.; Rossi, R.; Vicentini, F.; Astrakhant- sev, N.; Becca, F.; Cao, X.; Carrasquilla, J.; Ferrari, F.; Georges, A.; Hibat-Allah, M.; Imada, M.; Läuchli, A. M.; Mazzola, G.; Mez- zacapo, A.; Millis, A.; Robledo Moreno, J.; Ne- upert, T.; Nomura, Y.; Nys, J.; Parcollet, O.; Pohle, R.; Romero, I.; Schmid, M.; Silvester, J. M.; Sorella, S.; Tocchio, L. F.;...

    work page 2024

  31. [31]

    J.; Phys

    Wang, J.; Baerends, E. J.; Phys. Rev. Lett. 2022, 128, 013001. [Crossref]

    work page 2022

  32. [32]

    Yu.; Streltsov, S

    Irkhin, V. Yu.; Streltsov, S. V.;J. Supercond. Nov. Magn. 2022, 35 2135. [Crossref]

    work page 2022

  33. [33]

    Hubbard, J.; Proc. R. Soc. London, Ser. A 1963, 276, 238. [Crossref]

    work page 1963

  34. [34]

    Mott, N. F.; Rev. Mod. Phys. 1968, 40, 677. [Crossref]

    work page 1968

  35. [35]

    A.; França, V

    Canella, G. A.; França, V. V.; Phys. Rev. B 2021, 104, 134201. [Crossref]

    work page 2021

  36. [36]

    M.; França, V

    Pauletti, T.; Sanino, M.; Gimenes, L.; Car- valho, I. M.; França, V. V.; J. Mol. Model. 2024, 30, 1. [Crossref]

    work page 2024

  37. [37]

    Anderson, P. W.;Phys. Rev. 1958, 109, 1492. [Crossref]

    work page 1958

  38. [38]

    Arisa, D.; França, V.;Phys. Rev. B 2020, 101, 214522. [Crossref]

    work page 2020

  39. [39]

    Fulde, P.; Ferrell, R. A.;Phys. Rev. 1964, 135, A550. [Crossref]

    work page 1964

  40. [40]

    Larkin, A.; Ovchinnikov, Y. N.; Sov. Phys. JETP 1965, 20, 762. [Crossref]

    work page 1965

  41. [41]

    Lima, N. A. d.; Silva, M. F.; Oliveira, L. N. d.; Capelle, K.;Phys. Rev. Lett. 2003, 90, 146402. [Crossref]

    work page 2003

  42. [42]

    V.; Vieira, D.; Capelle, K.;New J

    França, V. V.; Vieira, D.; Capelle, K.;New J. Phys. 2012, 14, 073021. [Crossref]

    work page 2012

  43. [43]

    H.; Wu, F.-Y.;Physica A 2003, 321,

    Lieb, E. H.; Wu, F.-Y.;Physica A 2003, 321,

    work page 2003

  44. [44]

    A.; Oliveira, L

    Lima, N. A.; Oliveira, L. N.; Capelle, K.;Eu- rophys. Lett. 2002, 60, 601. [Crossref]

    work page 2002

  45. [45]

    P.; Campo, V

    Xianlong, G.; Polini, M.; Tosi, M. P.; Campo, V. L.; Capelle, K.; Rigol, M.; Phys. Rev. B 2006, 73, 165120. [Crossref]

    work page 2006

  46. [46]

    [Crossref]

    Palamara, A.; Plastina, F.; Sindona, A.; and D’Amico, I.;Physical Review A , 2024 110, 062203. [Crossref]

    work page 2024

  47. [47]

    [Crossref]

    Zhang, Z.; Zhao, Z.; Wang, H.; Deng, K.; Liu, Y.; Chen, W.; and Hu, J.;Physical Review Re- search, 2025 7 013184. [Crossref]

    work page 2025

  48. [48]

    M.; Navimipour, N

    Mostafaie, T.; Khiyabani, F. M.; Navimipour, N. J.; Comput. Oper. Res. 2020, 120, 104850. [Crossref]

    work page 2020

  49. [49]

    Vittorio Maniezzo, L. M. G.; de Luigi, F.; Springer: London, UK, 2004. [Crossref]

    work page 2004