Physics-informed Machine Learning Prediction of Hubbard Interaction Parameters

arXiv:2607.26422 · cond-mat.mtrl-sci, cond-mat.str-el · Submitted 2026-07-29 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Physics-informed Machine Learning Prediction of Hubbard Interaction Parameters".

Title and authors: Kai: We’re looking at "Physics-informed Machine Learning Prediction of Hubbard Interaction Parameters Jiyeon Kim Indukuru Ramesh Reddy Bongjae Kim Sooran Kim on page one of that work <ref:2607.26422#pg1>." The title itself tells you the main point right away: they are using physics to guide the machine learning.

Mira: It’s about predicting those three crucial interaction parameters—Ueff, V, and J—that are essential for things like DFT+U and DFT+DMFT in transition metal oxides.

Lev: So the authors are tackling a problem where these parameters are often chosen empirically by just looking at experimental data, which limits how well we can use them to predict new materials.

Kai: They’re presenting machine learning models that combine ensemble learning with a brute-force search method to achieve both accurate predictions and explicit analytical expressions about what drives the interaction parameters.

Mira: They are building features that pull from electronic structure, structural characteristics, and atomic properties as physical descriptors of localization and screening.

Lev: So they aren't just feeding it raw numbers; they are engineering specific features based on what we think should matter physically for these kinds of interactions in transition metal oxides.

Kai: They test different feature sets, like one focusing on composition and structure and another one specifically using those bandwidth and band center separations to see how well the prediction holds up.

The paper's summary: Kai: So the paper summarizes that they developed machine learning models to predict cRPA-derived Hubbard interaction parameters—Ueff, V, and J—for a diverse set of transition metal oxides.

Mira: They are combining ensemble learning models like Random Forest, Gradient Boosting Regression, and Extreme Gradient Boosting with a regression-based brute-force search approach to get both predictive accuracy and explicit analytical expressions.

Lev: This means they aren't just giving us a number; they are also trying to generate formulas that show the relationship between those parameters and the physical descriptors they chose.

Mira: The paper highlights that they tested different feature sets to see how well the prediction holds up across different physical interpretations.

Lev: It’s about checking if focusing on composition versus focusing on those specific electronic features actually gives a better result, which is important for model robustness.

Kai: They found that for Ueff, one feature set was best, and another set was selected because it allowed them to look more directly at the roles of electron localization and hybridization.

Mira: For V, they identified a primary feature set based on compositional descriptors as a starting point to analyze the inter-site interactions.

Lev: That’s smart because it gives us a baseline using the simplest properties before diving into the more complex physics captured by bandwidth or band center separations.

The paper's improvements: Kai: The improvements they suggest are focused on making sure the models are physically grounded, not just statistically accurate, by incorporating features that reflect how electrons actually behave in these materials.

Mira: They specifically point to the inclusion of TM-d bandwidth and the TM-d/O-p band center separation as essential descriptors because those directly quantify electron localization and hybridization.

Lev: So they are pushing for the AI to focus on things like how tightly localized electrons are, which is a good way to make sure its predictions aren't just fitting noise in the data.

Kai: They also improve it by using a multi-faceted approach where they compare results from feature sets based on composition against those that use physically motivated electronic features.

Mira: This allows them to provide a more direct analysis of how much those structural or compositional aspects influence the Hubbard U parameter versus how much localization and hybridization do.

Lev: That’s valuable because it lets us see which physical aspect is the main driver for a given parameter, instead of just getting a single score.

Conclusion: Kai: So to wrap up, this paper demonstrates that cRPA-derived Hubbard parameters can be predicted not only efficiently but also in a physically interpretable manner.

Mira: The selected descriptors were designed to capture electronic, structural, and atomic characteristics including physically motivated quantities associated with localization and TM-d/O-p hybridization.

Lev: This study provides a practical route for identifying descriptor-property relationships in correlated oxides and accelerating beyond-DFT workflows through physics-informed machine learning.

Kai: It’s about using AI to accelerate the workflow while ensuring the results have a clear connection to the underlying electronic structure of transition metal oxides.

Mira: It’s about building a system that can predict Ueff, V, and J in a way that lets us see exactly which physical factors are responsible for those interactions.

Lev: If this framework is reliable, it could really speed up the process of exploring new correlated materials before we even get to the expensive DFT+U calculations.

Kai: So that’s what we saw in "Physics-informed Machine Learning Prediction of Hubbard Interaction Parameters." It shows a path forward for using these tools responsibly in this field.

Mira: It’s a system that gives us both efficiency and physical insight into those critical interaction parameters.

Jiyeon Kim, Indukuru Ramesh Reddy, Bongjae Kim, Sooran Kim

Department of Physics Education, Kyungpook National University · Department of Physics, Kyungpook National University · KNU G-LAMP Project Group, KNU Institute of Basic Sciences

cond-mat.mtrl-sci, cond-mat.str-el

Submitted: 2026-07-29

Updated: 2026-07-29

Journal ref: J. Phys. Chem. Lett. (2026)

DOI: 10.1021/acs.jpclett.6c02654

Code: https://github.com/CMPL-KNU/cRPA-parameterprediction

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 89/100

The gist: 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 Here…

Key concepts

Hubbard Interaction Parameters
These are crucial energy parameters used in advanced electronic structure calculations (like DFT+U) to accurately model strongly correlated materials, such as transition-metal oxides. Predicting them correctly is essential for simulating their physical behavior.
Ensemble Learning Models
These are machine learning techniques (Random Forest, Gradient Boosting Regression, XGBoost) that combine multiple individual models to make more robust and accurate predictions. They are used here to predict the three interaction parameters based on material features.
Brute-Force Search (BFS)
This is a method used alongside the ML models to systematically test different combinations of input features. The BFS helps identify which specific physical descriptors have the strongest, explicit relationship with the target parameters (Ueff, V, J), making the model's predictions physically meaningful.

Terminology

Summary

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 Here, we present machine-learning (ML) models for predicting cRPA-derived Hubbard interaction parameters: effective on-site Ueff, inter-site V, and Hund’s coupling J for transition-metal oxides (TMOs). p. 1

Where the paper says it

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 Ueff, 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. p. 1
Where the paper says it

Here, we present machine-learning (ML) models for predicting cRPA-derived Hubbard interaction parameters: effective on-site Ueff, 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.

Predictive Framework

The study develops ML models to predict cRPA-derived Hubbard interaction parameters: effective on-site Ueff, inter-site V, and the Hund’s coupling J for a diverse set of transition-metal oxides (TMOs). The approach combines ensemble learning models with a regression-based brute-force search (BFS) method to achieve both predictive accuracy and physical interpretability. p. 4

Where the paper says it

In particular, we include electronic-structure descriptors associated with the competing effects of localization and screening. To achieve both predictive accuracy and physical interpretability, we combine two complementary approaches: ensemble learning models and a regression-based brute-force search (BFS) method.

The ensemble models are Random Forest (RF), Gradient Boosting Regression (GBR), and Extreme Gradient Boosting (XGB) models implemented using the scikit-learn Python library to predict the cRPA-derived Ueff, V, and J parameters. p. 6
Where the paper says it

For the inter-site V, 𝒲'(𝒓) and 𝒲)(𝒓) are centered on one TM site, while 𝒲((𝒓) and 𝒲*(𝒓) are centered on another TM site. 2 ML Algorithms We employed Random Forest Regression (RF), Gradient Boosting Regression (GBR), and Extreme Gradient Boosting (XGB) models implemented using the scikit-learn Python library to predict the cRPA-derived Ueff, V, and J parameters.

The BFS algorithm is applied to identify explicit relationships between primary features and target parameters Ueff, V, and J. p. 7
Where the paper says it

8 root-mean-square error (RMSE) and the coefficient of determination (R2 ), with R2 evaluated from the aggregated LOOCV predictions. We further applied a regression-based BFS algorithm59–62 to identify explicit relationships between the primary features and the target parameters Ueff, V, and J.

Feature Construction

The study constructs a dataset of cRPA-derived Hubbard interaction parameters (Ueff, V, and J) for TMOs with perovskite, olivine, spinel, layered, and layered perovskite structures, Table S1>. To identify physics-informed features, the focus is on two key factors expected to influence these interactions in TMOs: the localization of TM-d orbitals and the hybridization between TM-d and O-p states. p. 7

Where the paper says it

1 Feature Construction We first constructed a dataset of cRPA-derived Hubbard interaction parameters (Ueff, V, and J) for TMOs with perovskite, olivine, spinel, layered, and layered perovskite structures, as listed in Table S1. To identify physics-informed features, we focus on two key factors expected to influence these interactions in TMOs: the localization of TM-d orbitals and the hybridization

The selected electronic-structure features are the bandwidth of the TM-d band (W) and the energy separation between the TM-d and O-p band centers (Dd-p), as illustrated in Fig. p. 8
Where the paper says it

9 between TM-d and O-p states. Based on this physical picture, we selected two electronic-structure features, the bandwidth of the TM-d band (W) and the energy separation between the TM-d and O-p band centers (Dd-p), as illustrated in Fig.

2(a). These features quantify electron localization and p-d hybridization, respectively. p. 8
Where the paper says it

Based on this physical picture, we selected two electronic-structure features, the bandwidth of the TM-d band (W) and the energy separation between the TM-d and O-p band centers (Dd-p), as illustrated in Fig. These two features quantify the strength of electron localization and p-d hybridization, respectively: a larger W indicates weaker localization, whereas a larger Dd-p corresponds to weaker p-d hybridization.

Fundamental atomic properties (atomic number, group, and period of TM ion), compositional parameters (average d-electron number, average oxidation state, and the number of atoms and TM ions per unit cell), and structural characteristics are also considered. p. 9
Where the paper says it

These trends are consistent with previous cRPA studies, 14,35,41,43 in which an increase in U was attributed to enhanced localization of TM-d orbitals (smaller W) and reduced hybridization between TM-d and O-p orbitals (larger Dd-p). In addition to these two electronic-structure features, we also considered fundamental atomic properties (atomic number, group, and period of TM ion), compositional parameters (average d-electron number, average oxidation state, and the number of atoms and TM ions per unit cell), and structural characteristics (shortest TM-O bond length, volume per formula unit, and space group number), as summarized in Table 1.

Model Performance for Ueff

The ensemble models show strong predictive performance for Ueff. The XGB model achieved the highest overall performance with an RMSE of 0. p. 6

Where the paper says it

2 ML Algorithms We employed Random Forest Regression (RF), Gradient Boosting Regression (GBR), and Extreme Gradient Boosting (XGB) models implemented using the scikit-learn Python library to predict the cRPA-derived Ueff, V, and J parameters. 56–58 Model performance was evaluated using leave-one-out cross-validation (LOOCV), and the prediction performance was quantified using the

148 eV and an R2 of 0.899 using six features (nox, nd, Nu. p. 9
Where the paper says it

In addition to these two electronic-structure features, we also considered fundamental atomic properties (atomic number, group, and period of TM ion), compositional parameters (average d-electron number, average oxidation state, and the number of atoms and TM ions per unit cell), and structural characteristics (shortest TM-O bond length, volume per formula unit, and space group number), as summarized in Table 1. Among these features, the average oxidation state of the TM ion, nox, exhibits the strongest Pearson correlation coefficient (PCC) with Ueff (-0.

c, Vf.u, ZTM, GTM). For the best four-feature XGB model including W and Dd-p, the prediction accuracy was moderate with an RMSE of 0. p. 6
Where the paper says it

2 ML Algorithms We employed Random Forest Regression (RF), Gradient Boosting Regression (GBR), and Extreme Gradient Boosting (XGB) models implemented using the scikit-learn Python library to predict the cRPA-derived Ueff, V, and J parameters. 56–58 Model performance was evaluated using leave-one-out cross-validation (LOOCV), and the prediction performance was quantified using the

250 eV and an R2 of 0.756. The BFS analysis selected two four-feature sets; set 1 (nox, nd, Nu. p. 7
Where the paper says it

The primary features were selected from key descriptors through the ensemble-model analysis. Each primary feature was transformed using a predefined set of 15 nonlinear functions: 1, x, 1/x, x2 , 1/x2 , x3 , 1/x3 , √𝑥, 1/√𝑥, ex , e-x , e1/x , e-1/x , ln(x + 1), and 1/ln(x + 1).

c, Vf.u) was chosen as the best four-feature combination from the XGB model. p. 6
Where the paper says it

2 ML Algorithms We employed Random Forest Regression (RF), Gradient Boosting Regression (GBR), and Extreme Gradient Boosting (XGB) models implemented using the scikit-learn Python library to predict the cRPA-derived Ueff, V, and J parameters. 56–58 Model performance was evaluated using leave-one-out cross-validation (LOOCV), and the prediction performance was quantified using the

Set 2 (W, Dd-p, nd, and Vf. p. 4
Where the paper says it

5 analysis of cRPA data may also provide insight into DFT-based approaches for estimating the Hubbard interactions, particularly in their extension to low-energy model descriptions. 27,30,51,52 In this work, we develop ML models to predict cRPA-derived Hubbard interaction parameters: effective on-site Ueff, inter-site V, and the Hund’s coupling J for a diverse set of transition-metal oxides (TMOs).

u) was selected as a physically motivated set to enable a more direct analysis of the roles of electron localization and hybridization in determining the Hubbard U parameter. p. 1
Where the paper says it

The derived analytical forms directly relate Ueff 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 Ueff, V, and J, while offering physical insight into the factors underlying these Hubbard interactions.

Prediction of V and J

For the prediction of inter-site V, the XGB model achieved an RMSE of 0.062 eV using five features (nd, Nu. p. 1

Where the paper says it

Our ensemble models achieve RMSEs of 0. 062 eV, and 0.

c, Vf.u, SG, and ZTM). The BFS analysis selected set 1 (nd, Nu. p. 7
Where the paper says it

The primary features were selected from key descriptors through the ensemble-model analysis. Each primary feature was transformed using a predefined set of 15 nonlinear functions: 1, x, 1/x, x2 , 1/x2 , x3 , 1/x3 , √𝑥, 1/√𝑥, ex , e-x , e1/x , e-1/x , ln(x + 1), and 1/ln(x + 1).

c, Vf.u, and PTM) as the first primary feature set for V. p. 7
Where the paper says it

The primary features were selected from key descriptors through the ensemble-model analysis. Each primary feature was transformed using a predefined set of 15 nonlinear functions: 1, x, 1/x, x2 , 1/x2 , x3 , 1/x3 , √𝑥, 1/√𝑥, ex , e-x , e1/x , e-1/x , ln(x + 1), and 1/ln(x + 1).

For Hund’s coupling J, accurate prediction was achieved using only a very small number of elemental descriptors. p. 4
Where the paper says it

27,30,51,52 In this work, we develop ML models to predict cRPA-derived Hubbard interaction parameters: effective on-site Ueff, inter-site V, and the Hund’s coupling J for a diverse set of transition-metal oxides (TMOs). Using cRPA values as reference targets, this approach enables efficient estimation of first-principles Hubbard interactions and can accelerate DFT+U and DFT+U+V workflows while retaining a physical basis.

The best performing BFS model for set 1 yielded an RMSE of 0. p. 7
Where the paper says it

8 root-mean-square error (RMSE) and the coefficient of determination (R2 ), with R2 evaluated from the aggregated LOOCV predictions. We further applied a regression-based BFS algorithm59–62 to identify explicit relationships between the primary features and the target parameters Ueff, V, and J.

015 eV and an R2 of 0.954.

Conclusion

Overall, the present framework demonstrates that cRPA-derived Hubbard parameters can be predicted not only efficiently but also in a physically interpretable manner. The selected descriptors were designed to capture electronic, structural, and atomic characteristics including physically motivated quantities associated with localization and TMd/O-p hybridization. p. 1

Where the paper says it

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.

This study provides a practical route for identifying descriptor-property relationships in correlated oxides and accelerating beyond-DFT workflows through physics-informed machine learning. p. 0
Where the paper says it

1 Physics-informed Machine Learning Prediction of Hubbard Interaction Parameters Jiyeon Kim1* , Indukuru Ramesh Reddy2* , Bongjae Kim2† , Sooran Kim1,2,3† 1 Department of Physics Education, Kyungpook National University, Daegu 41566, South Korea 2 Department of Physics, Kyungpook National University, Daegu 41566, South Korea 3 KNU G-LAMP Project Group, KNU Institute of Basic Sciences, Kyungpook National University, Daegu, 41566, Korea. contributed equally to this work.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Incorporate a machine learning framework combining ensemble learning models like Random Forest, Gradient Boosting Regression, and Extreme Gradient Boosting to predict cRPA-derived Hubbard interaction parameters Ueff, V, and J for transition-metal oxides (TMOs). This allows the AI system to achieve high predictive accuracy while also providing analytical expressions that offer insight into the underlying physics of the interaction parameters.

  2. Develop a regression-based brute-force search (BFS) approach to identify explicit analytical relationships between key physical features and target parameters. This enables the AI system to generate explicit analytical formulas that improve model interpretability by relating Ueff, V, and J directly to descriptors such as electron localization and screening effects.

  3. Design feature construction modules that select physically motivated descriptors, such as the TM-d bandwidth W and the energy separation between TM-d and O-p band centers Dd-p for Ueff prediction. This allows the AI system to focus on factors like localization of TM-d orbitals and hybridization when predicting interaction parameters.

  4. Implement a multi-faceted approach where different feature sets are tested, such as Set 1 (compositional and structural descriptors) and Set 2 (physically motivated descriptors W and Dd-p). This allows the AI system to provide a more direct analysis of the roles of electron localization and hybridization in determining the Hubbard U parameter by comparing predictive performance across different physical interpretations.

  5. Create models that predict inter-site Coulomb interaction V using feature sets derived from both ensemble learning and BFS, specifically testing sets that include Dd-p and W. This allows the AI system to suggest structural insights, such as structural compactness, as a factor influencing inter-site interactions based on the results of the BFS analysis.

  6. Establish a minimal descriptor set approach for predicting Hund's coupling J, showing that accurate prediction was achieved using only a very small number of elemental descriptors. This allows the AI system to efficiently predict J using only intrinsic atomic characteristics of the TM ion, such as ZTM, GTM, and PTM.

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 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 eff, V, and J, respectively. The derived analytical forms directly relate U 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 eff, V, and J, while offering physical insight into the factors underlying these Hubbard interactions.

Sources

Related papers