Pith. sign in

REVIEW 3 major objections 4 minor 13 references

Physics-informed Machine Learning Prediction of Hubbard Interaction Parameters

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The paper claims machine learning on a few physical descriptors reproduces cRPA-quality Hubbard parameters for transition-metal oxides to within about 0.15 eV, with explicit formulas linking U_eff to electron localization and d-p hybridizat

desk verdict Useful ML surrogate for cRPA U/V/J, but the headline errors are likely optimistic because feature and model selection happened on the same 54 compounds before LOOCV. read the letter →

arxiv 2607.26422 v1 pith:EKG2ZRNN submitted 2026-07-29 cond-mat.mtrl-sci cond-mat.str-el

Jiyeon Kim , Indukuru Ramesh Reddy , Bongjae Kim , Sooran Kim This is my paper
classification cond-mat.mtrl-scicond-mat.str-el
keywords HubbardparameterscRPAtransition-metaloxidesmachinelearningensembleregressionHund'scouplinginter-siteCoulombinteractiondescriptor-based
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to show that cRPA-quality Hubbard parameters do not require a cRPA calculation for every new oxide: a small set of descriptors, including the transition-metal d-band width and the d-band/O-p band-center separation, suffices to predict U_eff, V, and J with root-mean-square errors near 0.15, 0.06, and 0.007 eV on 54 transition-metal oxides. The payoff is practical, because fast parameter estimates can accelerate DFT+U, DFT+U+V, and DMFT-style workflows, and physical, because explicit formulas assign distinct drivers to each interaction: localization and hybridization for U_eff, hybridization plus structural compactness for V, and atomic identity for J. A sympathetic reader would take the central claim to be that correlation parameters can be estimated cheaply and interpretably at cRPA-level accuracy within the oxide families studied.

What carries the argument

The central object is the pair of electronic-structure descriptors: W, the bandwidth of the transition-metal d-band, and Dd-p, the energy separation between the TM-d and O-p band centers, which quantify d-electron localization and d-p hybridization. These are supplemented by atomic, compositional, and structural features such as d-electron count, oxidation state, and volume per formula unit. The mechanism that carries the argument is the regression-based brute-force search (BFS): each primary feature is transformed through a predefined set of nonlinear functions, the transformed variables are combined multiplicatively into compound features, and linear regressions of the form g = w0 + sum wi

What would settle it

Compute cRPA-derived U_eff, V, and J for a handful of transition-metal oxides outside the 54, for example NiO, LaCrO3, or a 5d oxide such as SrIrO3, and plug their descriptors into Equations 2-6; if the U_eff errors systematically exceed roughly 0.15 eV, the selected formulas are specific to the training set rather than general physical relations.

Watch

Extended reading notes

Core claim

The paper's central claim is that the three cRPA-derived interaction parameters of transition-metal oxides can be predicted accurately from a compact set of twelve electronic, structural, and atomic descriptors. Leave-one-out cross-validation on the 54-compound dataset gives root-mean-square errors of 0.148 eV for the effective on-site U_eff, 0.062 eV for the inter-site V, and 0.007 eV for Hund's J with the best ensemble model. A regression-based brute-force search then produces explicit closed-form expressions (Eqs. 2-6) in which U_eff is written in terms of the TM-d bandwidth W and the TM-d/O-p band-center separation Dd-p, V in terms of structural compactness and Dd-p, and J essentially in

Load-bearing premise

The reported accuracies come from testing each compound one at a time with models whose best feature combination was chosen after looking at all 54 cRPA targets, so a new oxide outside this set could be predicted less accurately than the stated errors suggest.

Editorial extensions

If this is right

  • If the models are correct, cRPA calculations can be bypassed for screening: U_eff, V, and J for a new transition-metal oxide can be estimated from a few standard DFT or even purely compositional/structural quantities.
  • The analytical formulas separate the physical drivers: U_eff responds to d-band localization and TM-d/O-p hybridization, V to hybridization and structural compactness, and J mostly to the transition-metal's atomic identity.
  • The four-feature set nd, nox, Nu.c, and Vf.u already nearly saturates the best ensemble model for U_eff, suggesting that accurate predictions may be possible without electronic-structure descriptors in some regimes.
  • For Hund's coupling, a single descriptor such as the TM atomic number reaches R^2 near 0.95, and a simple linear formula using only the group number of the TM reproduces J within 0.015 eV.
  • The resulting equations are simple enough to be embedded directly in DFT+U or DFT+U+V codes, replacing empirical parameter choices with descriptor-based estimates.

Reading between the lines

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

  • A natural test the paper leaves implicit is to hold out entire structure families, say all spinels or all layered oxides, and fit on the rest; if errors remain near the reported LOOCV values, the feature selection generalizes to new geometries, and if not, the reported RMSEs are likely optimistic.
  • Because U_eff is written explicitly in terms of W and Dd-p, and both respond to strain and pressure, the formulas could be used to predict how Hubbard parameters change under epitaxial strain or compression without new cRPA runs.
  • Since J tracks the atomic number and group of the transition metal so closely, a periodic-table lookup or a one-line formula could replace the empirical Hund's coupling in many TMO calculations, removing one adjustable parameter from DFT+U workflows.
  • The same descriptor-plus-brute-force-search workflow could be exported to other correlated chemistries such as sulfides or fluorides, provided W and Dd-p are recomputed for those systems; the paper does not test this transfer.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper presents an ML workflow for predicting cRPA-derived Hubbard parameters (Ueff, V, J) for transition-metal oxides. Using 54 compounds and 12 features, including electronic descriptors W and Dd-p, the authors train RF, GBR, and XGB models; for each model, all 4,095 feature subsets are screened and LOOCV is used. They report best XGB RMSE/R2 values of 0.148/0.899 for Ueff, 0.062/0.784 for V, and 0.007/0.989 for J. A regression-based brute-force search (BFS) yields analytical expressions (Eqs. 2-6). The authors claim these enable efficient cRPA-quality parameter prediction and provide physical insight into localization, hybridization, and structural effects.

Significance. If the reported predictive errors were honest estimates of generalization, this would be a useful contribution: cRPA is computationally expensive, and an interpretable surrogate for three Hubbard parameters is valuable for oxide screening. The paper provides a multi-family cRPA dataset, uses physically motivated descriptors, and releases code. However, the central quantitative claim is not secured because feature selection and BFS model selection are performed on the same 54 targets before LOOCV. The reported RMSEs are selected-model scores rather than unbiased generalization estimates. This issue is load-bearing and requires additional validation before the headline accuracy can be accepted.

major comments (3)
  1. [Sec. 2.2; Sec. 3.2; Fig. 3] The reported LOOCV errors are optimistically biased. All 4,095 feature combinations are screened on the full dataset, and only the best subset per model is then evaluated by LOOCV. Each held-out sample has already influenced feature selection, so the LOOCV score is not an unbiased estimate of generalization. This is particularly sharp for J, where the three-feature model achieves R2=0.989 (Sec. 3.4). Please add nested CV in which feature selection is performed inside each training fold, or provide a holdout set or external cRPA benchmark, and report the distribution of RMSE across selection repeats.
  2. [Sec. 2.2; Eqs. 2-6] The BFS equations are also selected on the full data before LOOCV. For each mCnF configuration, all candidate compound features are screened by fitting the full dataset and the top-three models are then evaluated by LOOCV. Consequently Eqs. 2-6 are fits selected on all 54 points, and their reported errors do not account for model selection. The physical interpretations in Secs. 3.2-3.4 are post-hoc readings of fitted coefficients rather than tests of an independent derivation. A nested or split evaluation is required before these formulas can be claimed as transferable.
  3. [Table S1] The dataset is not i.i.d.: compounds are grouped by structure family (perovskite, olivine, spinel, layered, layered perovskite) and by 3d/4d TM series. Standard LOOCV can overstate accuracy for new materials of different families. To support the transferability claim, use grouped CV such as leave-one-family-out and report the resulting errors alongside the LOOCV numbers.
minor comments (4)
  1. [Eqs. 2-6] The typeset formulas are corrupted in the provided text, making them unverifiable. Please fix the rendering and list the equations in a machine-readable form in the SI.
  2. [Data Availability] The data are only available from the authors upon request. To enable reproduction of the selection protocol and independent validation, deposit the full target and feature table alongside the code.
  3. [Sec. 2.2] Hyperparameters for RF, GBR, and XGB are not stated. Clarify whether they are fixed and, if tuned, describe whether tuning was performed inside or outside the CV folds.
  4. [Fig. 3 caption] The x-axis indicates number of selected features, but the curves show the best subset of each cardinality rather than sequential addition. Please clarify this in the caption.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported LOOCV errors and analytical formulas are selected-model fits: feature-subset and BFS functional-form selection use all 54 cRPA targets before the same targets are 'predicted'.

  1. fitted input called prediction [Sec 2.2 (ML Algorithms), Sec 3.2 (Prediction of on-site Ueff), Fig 3 and BFS model selection]
    "All possible combinations of the 12 features were examined for different numbers of selected features, resulting in a total of 4,095 models for each regression model. Only the best-performing combination for each number of features is shown. [...] For a given mCnF configuration, all possible combinations of compound features were first screened by fitting the full dataset, after which the top-three performing candidate models were evaluated using LOOCV."

    The headline RMSE/R2 values (e.g., Ueff RMSE 0.148 eV, V 0.062 eV, J 0.007 eV) are computed after selecting the best feature subset or the top BFS expressions using the same 54 cRPA target values that are then used in LOOCV. Each left-out target has already influenced which features, functional forms, and coefficients were chosen, so the LOOCV error is not an unbiased estimate of prediction on new materials. Instead, the 'prediction' is statistically forced by the very labels used during model selection. With n=54, screening 4,095 feature combinations and thousands of BFS compound features can exploit noise, making the reported transferability a selected-model fit rather than an independent prediction.

full rationale

The cRPA calculations, feature definitions, and the physical motivation for using W and Dd-p are independent of the ML targets, and the paper does not derive Ueff, V, or J from their own definitions. The central circularity lies in the evaluation protocol: the BFS analytical expressions (Eqs. 2-6) and the ensemble feature sets are selected by fitting the full cRPA dataset, and the same dataset is then used to report LOOCV 'prediction' errors. The paper states this explicitly in Sec. 2.2. Consequently, the reported RMSEs and R2 values are optimistic selected-model estimates, and the physical interpretations drawn from the fitted formulas are post-hoc readings of coefficients of functions chosen to minimize error on those same targets. There is no nested CV, holdout set, or external cRPA benchmark provided, so the claimed transferability is not independently secured. The self-citations to the authors' earlier cRPA studies are not load-bearing in a definitional way, so the circularity is partial and methodological rather than a full reduction of the results to their inputs.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claims rest on the cRPA reference data, the representativeness of the 54-compound dataset, and the assumption that LOOCV after feature selection estimates generalization. The fitted regression coefficients, selected feature subsets, and energy-window choices are the main free parameters; no new physical entities are introduced.

free parameters (4)
  • BFS regression coefficients = e.g., Eq 2: 1.898, -9.978, 133.485, 11.337; Eq 4: 0.919, 1.673e-3, 8.342e4, 1.193; Eq 6: 3.001, 0.354
    Coefficients in analytical expressions are least-squares fits to the cRPA targets; they are the core fitted content of the BFS models.
  • Selected feature subset = e.g., Ueff XGB: nox, nd, Nu.c, Vf.u, ZTM, GTM; V XGB: nd, Nu.c, Vf.u, SG, ZTM; J XGB: GTM, PTM, SG
    Best feature combinations chosen by scanning all 4,095 subsets against LOOCV performance; the selection is fit to the data and biases reported errors.
  • BFS mCnF model configuration = Ueff: 3C2F; V set1: 3C2F; J set1: 1C2F
    Number of compound features and primary features per compound chosen by full-data screening; not an independent benchmark.
  • Band-center energy windows [Emin, Emax] and disentanglement windows = Per-compound values in Table S7, e.g., SrTiO3 [-8,9] eV
    Hand-set integration windows determine W and Dd-p, so the physics-informed features are not parameter-free.
assumptions (6)
  • domain assumption cRPA-derived Ueff, V, J are reliable reference values for Hubbard interactions
    The entire ML target depends on cRPA being the correct ground truth; the paper adopts the projector-based cRPA implementation in VASP (Sec 2.1).
  • domain assumption The 54-compound dataset is representative enough for transferable predictions
    Generalization claims rest on coverage of perovskite, olivine, spinel, layered, layered-perovskite TMOs; no external or grouped test set is provided.
  • ad hoc to paper LOOCV after best-of-4095 feature selection estimates generalization error
    Sec 2.2/3.2 select features on the same targets they evaluate; this assumes the selection does not bias LOOCV, which is violated in general.
  • ad hoc to paper The predefined 15-function basis and multiplicative compounds can express Ueff/V/J dependence
    BFS restricts formulas to functions like 1/x, e^x, ln(x+1); no guarantee these capture the true mapping (Sec 2.2).
  • domain assumption W and Dd-p as defined capture localization and hybridization
    Features depend on manually chosen energy windows; the physical interpretation of equations follows from this assumption (Sec 3.1, Note S2).
  • domain assumption Default ML hyperparameters are adequate
    No hyperparameter values or tuning details are reported; results may depend on defaults in scikit-learn/XGBoost (Sec 2.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics-informed Machine Learning Prediction of Hubbard Interaction Parameters." pith.science (2026). https://pith.science/paper/EKG2ZRNN

@misc{pith2026260726422,
  author       = {Pith},
  title        = {Pith review of: Physics-informed Machine Learning Prediction of Hubbard Interaction Parameters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EKG2ZRNN}},
  note         = {Machine review of arXiv:2607.26422}
}
abstract

Accurate determination of Hubbard interaction parameters is essential for beyond-DFT approaches such as DFT+$U$, DFT+DMFT, and DFT+$U$+$V$ in correlated materials. In practice, however, these parameters are often chosen empirically, limiting their transferability across materials. Advanced computational approaches such as the constrained random-phase approximation (cRPA) provide a rigorous route for evaluating Hubbard interactions, but their computational cost remains a bottleneck for large-scale materials screening. Here, we present machine-learning (ML) models for predicting cRPA-derived Hubbard interaction parameters: effective on-site $U_{\rm eff}$, inter-site $V$, and Hund's coupling $J$ for transition-metal oxides (TMOs). We combine ensemble-learning models with a regression-based brute-force search (BFS) approach to achieve both predictive accuracy and explicit analytical expressions. We construct features that capture electronic, structural, and atomic properties, including the TM-$d$ bandwidth and TM-$d$/O-$p$ band-center separation, as physically motivated descriptors of localization and screening. Our ensemble models achieve RMSEs of 0.148 eV, 0.062 eV, and 0.007 eV for $U_{\rm eff}$, $V$, and $J$, respectively. The derived analytical forms directly relate $U_{\rm eff}$ to electron localization and TM-$d$/O-$p$ hybridization, suggest the importance of hybridization and structural compactness in determining $V$, and indicate that $J$ is governed primarily by elemental descriptors of the TM ion. Together, the present study provides an efficient approach for predicting cRPA-derived $U_{\rm eff}$, $V$, and $J$, while offering physical insight into the factors underlying these Hubbard interactions.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

13 extracted references · 5 canonical work pages

  1. [2]

    RMSE (eV) R2 1 −0.055 2𝑊Z56 −1.287×1071√𝑊Z56*+1.666 0.022 0.929 2 −815.016 1√𝑊Z560−0.099ln(1+𝑊)2Z56 +2.148 0.023 0.922 3 −0.073ln(1+𝑊)2Z56−9.933×10*1ln(1+𝑊)Z56*)+1.678 0.024 0.918 Table S7. List of transition-metal oxide compounds, the corresponding [Emin, Emax] energy window (in eV) used for calculating the TM-d and O-p band centers, and the disentanglem...

    1996

  2. [4]

    This approach achieves high predictive accuracy while also providing analytical expressions that offer insight into the underlying physics of the interaction parameters

    CONCLUSIONS In this work, we developed ML models for predicting cRPA-derived Hubbard interaction parameters, Ueff, V, and Hund’s coupling J in TMOs by combining ensemble-learning models with a regression-based BFS analysis. This approach achieves high predictive accuracy while also providing analytical expressions that offer insight into the underlying ph...

    2025

  3. [8]

    (51) Mosey, N

    http://arxiv.org/abs/2505.03698. (51) Mosey, N. J.; Carter, E. A. Ab Initio Evaluation of Coulomb and Exchange Parameters for DFT+U Calculations. Phys. Rev. B 2007, 76 (15), 155123. https://doi.org/10.1103/PhysRevB.76.155123. (52) Kulik, H. J.; Cococcioni, M.; Scherlis, D. A.; Marzari, N. Density Functional Theory in Transition-Metal Chemistry: A Self-Con...

    arXiv 2007

  4. [12]

    A.; Yates, J

    (5) Mostofi, A. A.; Yates, J. R.; Lee, Y .-S.; Souza, I.; Vanderbilt, D.; Marzari, N. Wannier90: A Tool for Obtaining Maximally-Localised Wannier Functions. Comput. Phys. Commun. 2008, 178 (9), 685–699. https://doi.org/10.1016/j.cpc.2007.11.016. (6) Franchini, C.; Kováčik, R.; Marsman, M.; Sathyanarayana Murthy, S.; He, J.; Ederer, C.; Kresse, G. Maximall...

    doi:10.1016/j.cpc.2007.11.016 2008

  5. [13]

    (8) Reddy, I

    https://doi.org/10.1038/s41524-024-01454-9. (8) Reddy, I. R.; Kaltak, M.; Kim, B. Comparative Analysis of Methods for Calculating Hubbard Parameters Using CRPA. Phys. Rev. B 2025, 111 (19), 195144. https://doi.org/10.1103/PhysRevB.111.195144. (9) Wang, V .; Xu, N.; Liu, J.-C.; Tang, G.; Geng, W.-T. V ASPKIT: A User-Friendly Interface Facilitating High-Thr...

    arXiv 2025

  6. [19]

    (49) Yang, W.; Timrov, I.; Aquilante, F.; Son, Y.-W

    https://doi.org/10.1038/s41524-024-01501-5. (49) Yang, W.; Timrov, I.; Aquilante, F.; Son, Y.-W. Comparing Hubbard Parameters from Linear-Response Theory and Hartree-Fock-Based Approach. ArXiv

    work page doi:10.1038/s41524-024-01501-5

  7. [180]

    (45) Yu, W.; Zhang, Z.; Wan, X.; Guo, H.; Gui, Q.; Peng, Y.; Li, Y.; Fu, W.; Lu, D.; Ye, Y.; Guo, Y

    https://doi.org/10.1038/s41524-020-00446-9. (45) Yu, W.; Zhang, Z.; Wan, X.; Guo, H.; Gui, Q.; Peng, Y.; Li, Y.; Fu, W.; Lu, D.; Ye, Y.; Guo, Y. Active Learning the High-Dimensional Transferable Hubbard U and V Parameters in the DFT + U + V Scheme. J. Chem. Theory Comput. 2023, 19 (18), 6425–6433. https://doi.org/10.1021/acs.jctc.2c01116. (46) Cai, G.; Ca...

    doi:10.1038/s41524-020-00446-9 2023

  8. [183]

    (64) Tomczak, J

    https://doi.org/10.1038/s41524-025-01685-4. (64) Tomczak, J. M.; Miyake, T.; Aryasetiawan, F. Realistic Many-Body Models for Manganese Monoxide under Pressure. Phys. Rev. B 2010, 81 (11), 115116. https://doi.org/10.1103/PhysRevB.81.115116. (65) Tomczak, J. M.; Miyake, T.; Sakuma, R.; Aryasetiawan, F. Effective Coulomb Interactions in Solids under Pressure...

    work page doi:10.1038/s41524-025-01685-4 2010

Show all 13 references
  1. [263]

    (31) Aryasetiawan, F.; Karlsson, K.; Jepsen, O.; Schönberger, U

    https://doi.org/10.1038/s41524-022-00958-6. (31) Aryasetiawan, F.; Karlsson, K.; Jepsen, O.; Schönberger, U. Calculations of Hubbard U from First-Principles. Phys. Rev. B 2006, 74 (12), 125106. https://doi.org/10.1103/PhysRevB.74.125106. (32) Miyake, T.; Aryasetiawan, F. Scree...

    2006 doi

  2. [286]

    26 (15) Reddy, I

    https://doi.org/10.1038/s41524-024-01454-9. 26 (15) Reddy, I. R.; Ghosh, S.; Kim, B.; Kang, C.-J. Role of On-Site and Intersite Coulomb Interactions in KV3Sb5: A first-Principles DFT+U+V Study. Phys. Rev. B 2025, 112 (11), 115109. https://doi.org/10.1103/fnbn-65bd. (16) Naveas...

    2025 doi

  3. [2015]

    Efficient Iterative Schemes for Ab Initio Total-Energy Calculations Using a Plane-Wave Basis Set

    (54) Kresse, G.; Furthmüller, J. Efficient Iterative Schemes for Ab Initio Total-Energy Calculations Using a Plane-Wave Basis Set. Phys. Rev. B 1996, 54 (16), 11169–11186. https://doi.org/10.1103/PhysRevB.54.11169. (55) Kaltak, M.; Hampel, A.; Schlipf, M.; Reddy, I. R.; Kim, B...

    1996

  4. [2025]

    (50) Carta, A.; Timrov, I.; Beck, S.; Ederer, C

    http://arxiv.org/abs/2512.16803. (50) Carta, A.; Timrov, I.; Beck, S.; Ederer, C. Bridging Constrained Random-Phase Approximation and Linear Response Theory for Computing Hubbard Parameters. ArXiv

  5. [2395]

    (13) Di Sante, D.; Kim, B.; Hanke, W.; Wehling, T.; Franchini, C.; Thomale, R.; Sangiovanni, G

    https://doi.org/10.3390/app11052395. (13) Di Sante, D.; Kim, B.; Hanke, W.; Wehling, T.; Franchini, C.; Thomale, R.; Sangiovanni, G. Electronic Correlations and Universal Long-Range Scaling in Kagome Metals. Phys. Rev. Res. 2023, 5 (1), L012008. https://doi.org/10.1103/PhysRev...

    2023 doi

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.