Charting the Landscape of Oxygen Ion Conductors: A 60-Year Dataset with Interpretable Regression Models

Abstract

Oxygen ion conductors are indispensable materials for such as solid oxide fuel cells, sensors, and membranes. Despite extensive research across diverse structural families, systematic data enabling comparative analysis remains scarce. Here, we present a curated dataset of oxygen ion conductors compiled from 84 experimental reports spanning 60 years, covering 483 materials. Each record includes activation energy (E_a) and prefactor (A) derived from Arrhenius plots, alongside detailed metadata on structure, composition, measurement method, and data source. When the original papers derive these using an erroneous Arrhenius equation, we replotted these using the correct one. To illustrate how the database can be used, we constructed interpretable regression models for predicting oxygen ionic conductivity. Two symbolic regression models for E_a and A suggest that oxygen ion transport is primarily governed by local coordination environment and the electrostatic interactions, respectively. This dataset establishes a reliable foundation for data-driven discovery and predictive modeling of next-generation oxygen ion conductors.

Background & Summary

Oxygen ion conductors are key functional materials underpinning technologies such as solid oxide fuel cells, oxygen sensors, and catalytic membranes, where efficient ionic transport is essential for performance and sustainability^1–8. Over the past six decades, a wide range of structural families, including perovskites/fluorites family^9–45, Bi-based family (such as $\delta -{\mathrm{Bi}}_2{\mathrm{O}}_3$^41,46–54, BIMEVOX^55–59, and ${\mathrm{Bi}}_2\mathrm{U}{\mathrm{O}}_6$ classes⁶⁰), complex framework family (such as scheelite^61–72, apatite^73–81, cuspidine^82–84, melilite^85–88, garnet^89,90, pyrochlore⁹¹, and columbite classes⁹²) and $\alpha$-$\mathrm{SrSi}{\mathrm{O}}_3$ family^93–96, have been reported to exhibit oxygen ion conductivity. Recent breakthroughs in ${\mathrm{Ba}}_4{\mathrm{Nb}}_4\mathrm{Mo}{\mathrm{O}}_{20}$- and α-$\mathrm{SrSi}{\mathrm{O}}_3$-based materials highlight the promise of structural diversity in enabling fast ion transport^36,96. However, the richness of these material families also poses a challenge: the structural and chemical space remains sparsely explored, and systematic design principles for new conductors are limited.

To address this gap, comprehensive data collection and curation are essential. Here, we present a manually curated dataset of oxygen ion conductors derived from $84$ experimental reports, encompassing $483$ unique oxides published over $60$ years^{10–39,41,43–96}. The dataset provides activation energies ($E_a$) and prefactors ($A$) extracted from Arrhenius-type analyses of ionic conductivity. While oxygen tracer diffusion measurements provide direct information on oxygen mobility, they usually require additional assumptions regarding carrier concentration and correlation effects. Therefore, to construct a coherent and internally large consistent dataset centered on a single transport metric, the present work focuses on oxygen ion conductivity derived from electrochemical measurements, which remains the most widely reported experimental probe of oxygen-ion transport.

When the original papers fit the conductivity using an erroneous Arrhenius-type formula, we replot the conductivity and reevaluated these (see Methods). To examine the effectiveness of our database, we constructed interpretable two symbolic regression models that predict $E_a$ and $A$. These models enable the identification of dominant descriptors and offering physical insights into the factors governing oxygen ion transport. This dataset not only consolidates the historical knowledge base but also offers a foundation for data-driven discovery and design of next-generation oxygen ion conductors. In addition, our systematic data will be utilized for the validation of machine learning potentials, which have been developing rapidly in recent years^97–99.

Methods

To construct a comprehensive dataset of oxygen ion conductors, we systematically surveyed the literature spanning the past $60$ years (Fig. 1a). We first used multiple academic search engines including Web of Science¹⁰⁰, Scopus¹⁰¹, and Google Scholar¹⁰² with keywords such as “oxygen ion conductivity”, “Arrhenius plots”, and “solid oxide fuel cells”. From the identified reports containing oxygen ion conductivities at multiple temperatures (${\sigma }_T$, where $T$ are measured temperatures), we then expanded our search by following two directions: (i) backward citations (references cited by these papers) and (ii) forward citations (subsequent papers citing them). This iterative process was repeated until no new experimental studies with relevant conductivity data were found. In this study, we targeted both purely oxygen ion conducting electrolytes and mixed ionic-electronic conductors; in the latter case, only the data in which the oxygen-ion conductivity was explicitly separated from the electronic contribution were included^11,12,22,31, and all others were excluded (Fig. 1b). Note that such mixed ionic–electronic conductors may include additional inaccuracies due to methodological assumptions/uncertainties in deriving oxygen-ion diffusivities.

Fig. 1.

Workflow for dataset construction and curation of oxygen ion conductors. (a) Initial literature search using multiple academic databases identified candidate reports containing multi-temperature oxygen ion conductivities. Forward and backward citation searches were then performed iteratively until no additional relevant studies were found. (b) Conductivities dominated by electronic contributions were excluded to ensure that the dataset represents oxygen ion transport. (c) To correct for the widespread use of an incorrect Nernst-Einstein equation, all reported plots were manually re-examined (see text for details). For example, activation energies ($E_a$) and prefactors ($A$) were recalculated from at least four data points in figures to eliminate systematic biases and ensure consistency across the dataset.

A common issue in the literature is the use of an incorrect form of the Arrhenius plots. The correct form of the relationship between ${\sigma }_T$ and $E_a$ is ${\sigma }_{T}T=A\exp \left(-\frac{E_a}{k_{\mathrm{B}}T}\right)$, which is derived from the Nernst-Einstein equation ${\sigma }_{T}T=D_T\frac{c{q}^2}{k_{\mathrm{B}}}$ and the Arrhenius equation $D_T=A_D\exp \left(-\frac{E_a}{k_{\mathrm{B}}T}\right)$, where $D_T$, $c$, $q$, $k_{\mathrm{B}}$, and $A_D$ denote the self-diffusion coefficient, the density of charge carriers, the ionic charge, the Boltzmann constant, and the prefactor for $A_D$, respectively. Using the same $E_a$ for both ${\sigma }_T$ and $D_T$ implicitly assumes that the concentration of mobile oxygen defects is temperature independent. However, many reports instead erroneously use ${\sigma }_T=A\exp \left(-\frac{E_a}{k_{\mathrm{B}}T}\right)$. This inconsistency results in systematic errors in $E_a$ and $A$. To address this, all conductivity plots were carefully re-examined (Fig. 1c); an example is provided in the Supplementary Information. For figures, at least four points were manually extracted across the measured temperature range; for tables, a minimum of two data points were used. From these, $E_a$ and $A$ were recalculated to ensure consistency across the dataset. This labor-intensive manual curation is the most important contribution of this study, which allowed us to standardize parameters and eliminate inherited biases.

Data Records

The dataset comprises oxygen ion conductivity parameters extracted from $84$ experimental reports spanning $60$ years, covering $483$ oxides. Each entry in the dataset corresponds to one oxide under a given measurement condition and includes $E_a$ and $A$ derived from the Arrhenius plots. For cases where the Nernst-Einstein relationship exhibits two distinct linear regions, both the low-temperature ($T<T^{\ast }$) and high-temperature ($T>T^{\ast }$) regimes are reported separately, with the break point temperature ($T^{\ast }$) explicitly recorded. To ensure consistency and reproducibility, each data entry is annotated with extensive metadata (Table 1). We note that, when available, bulk conductivity values separated from grain boundary contributions were preferentially used. However, in older studies where such separation was not reported, total conductivity values were included and annotated accordingly in the dataset metadata (“Type of ${\sigma }_T$” field). In addition, the dataset records the oxygen non-stoichiometry parameter $∆\mathrm{O}$, which quantifies deviations from ideal oxygen content and provides indirect information on defect chemistry. The curated dataset is provided as the file oxygen_ion_conductor_dataset.csv, which can be downloaded and used directly without additional processing (see Data Availability)¹⁰³.

Table 1.

Metadata fields for the curated dataset.

Field	Description
References	Family name of the first author, journal title, volume, initial page number, and published year.
DOI	Digital Object Identifier of the source publication, enabling traceability.
Year	Publication year of the experimental report.
Class	Structural class of the material (e.g., perovskite, scheelite, apatite).
Formula	Reported chemical composition.
Parsed formula	Standardized composition represented by $A_s$–$B_s$(–$C_s$) site classification. For example, in the perovskite oxide ${\mathrm{La}}_{0.95}{\mathrm{Sr}}_{0.05}{\mathrm{Ga}}_{0.95}{\mathrm{Mg}}_{0.05}{\mathrm{O}}_{3-\delta }$, La and Sr occupy the $A_s$-site, while Ga and Mg occupy the $B_s$-site of the perovskite lattice10.
$∆\mathrm{O}$	Relative oxygen non-stoichiometry defined as $∆\mathrm{O}=\left({\mathrm{O}}_{\mathrm{act}}-{\mathrm{O}}_{\mathrm{sto}}\right)/{\mathrm{O}}_{\mathrm{sto}};{\mathrm{O}}_{\mathrm{act}}$ and ${\mathrm{O}}_{\mathrm{sto}}$ denote actual and stoichiometric oxygen contents, respectively. For example, in the perovskite oxide ${\mathrm{La}}_{0.95}{\mathrm{Sr}}_{0.05}{\mathrm{Ga}}_{0.95}{\mathrm{Mg}}_{0.05}{\mathrm{O}}_{2.95}$, it is given that ${\mathrm{O}}_{\mathrm{act}}=2.95$, ${\mathrm{O}}_{\mathrm{sto}}=3$, and $∆\mathrm{O}=-0.0167$10.
$E_a$/$A$	Activation energy in $\mathrm{meV}$ and prefactor in $\mathrm{K}·\mathrm{S}·{\mathrm{cm}}^{-1}$. When there are two distinct linear regions, values for the low-temperature regime are shown.
$E_{a,\mathrm{HT}}$/$A_{\mathrm{HT}}$	Activation energy and prefactor values for the high-temperature regime (if applicable).
$T^{*}$	Transition temperature separating low- and high-temperature regimes (if applicable).
Type of ${\sigma }_T$	Conductivity type, distinguishing bulk conductivity from total conductivity (bulk and grain boundary contributions).
Measurement	Experimental technique employed (e.g., two-probe AC, impedance spectroscopy).
Measurement temperature range	The range of temperatures over which experimental measurements were performed.
Source	Origin of the data within the publication (figure or table reference).
Plot type	Axis configuration used in the original Arrhenius representation (e.g., ${\log }_e{\sigma }_T$-$1000/T$).

Data Overview

Figure 2a illustrates the distribution of data points across all oxides in the $A$-$E_a$ domain. For visualization, we used the low-temperature regime ($T<T^{\ast }$) when two-region behavior was reported. Dashed guidelines indicate the condition to achieve ${\sigma }_T\ge {10}^{-2}$ $\mathrm{S}·{\mathrm{cm}}^{-1}$ at different temperatures. This value corresponds to the widely cited target proposed for solid oxide fuel cell electrolytes to achieve an area-specific resistance of $\sim 0.15\mathrm{\Omega }\cdot {\mathrm{c}\mathrm{m}}^2$ for an electrolyte thickness of $~15\mathrm{\mu }m$¹⁰⁴. Notably, none of the investigated materials surpass this threshold at $473\mathrm{K}$, underscoring the persistent challenge of achieving high ionic conductivity at low-to-intermediate temperatures. Access to this regime may be achieved either by reducing $E_a$ (green-shaded region) or by increasing $A$ (pink-shaded region). However, the weak positive correlation between $E_a$ and $A$ complicates this effort, underscoring the need to decouple descriptors for $E_a$ and $A$ in order to guide efficient materials design. Recent studies have also emphasized this challenge and proposed physically motivated descriptors to disentangle migration barriers and attempt frequencies in ionic conductors^105,106. Based on our dataset, the most effective oxygen ion conductors exhibiting ${\sigma }_T\ge {10}^{-2}$ $\mathrm{S}·{\mathrm{cm}}^{-1}$ for $T=473-573$ K are $\alpha$-$\mathrm{SrSi}{\mathrm{O}}_3$-type ${\mathrm{Sr}}_{1.65}{\mathrm{Na}}_{1.35}{\mathrm{Si}}_3{\mathrm{O}}_{8.325}$ ($E_a=315$ meV and $A=3.59\times {10}^3$ $\mathrm{K}·\mathrm{S}·{\mathrm{cm}}^{-1}$)⁹⁵, apatite-type ${\mathrm{Nd}}_{9.5}{\mathrm{Si}}_{5.5}{\mathrm{Al}}_{0.5}{\mathrm{O}}_{26}$ ($E_a=658$ meV and $A=1.45\times {10}^7$ $\mathrm{K}·\mathrm{S}·{\mathrm{cm}}^{-1}$) and ${\mathrm{Nd}}_{9.83}{\mathrm{Si}}_{4.5}{\mathrm{Al}}_{1.5}{\mathrm{O}}_{26}$ ($E_a=779$ meV and $A=7.05\times {10}^7$ $\mathrm{K}·\mathrm{S}·{\mathrm{cm}}^{-1}$)⁷⁹, ${\mathrm{Bi}}_2\mathrm{U}{\mathrm{O}}_6$ ($E_a=856$ meV and $A=2.69\times {10}^8$ $\mathrm{K}·\mathrm{S}·{\mathrm{cm}}^{-1}$)⁶⁰, and BIMEVOX-type ${\mathrm{Bi}}_2{\mathrm{V}}_{0.9}{\mathrm{Mg}}_{0.1}{\mathrm{O}}_{5.35}$ ($E_a=911$ meV and $A=1.54\times {10}^9$ $\mathrm{K}·\mathrm{S}·{\mathrm{cm}}^{-1}$)⁹³.

Fig. 2.

Overview of the curated dataset of oxygen ion conductivities. (a) Distribution of activation energy ($E_a$) and prefactor ($A$) for $483$ oxides. We show $A$ in the low-temperature regime when two-region behavior was reported. Dashed lines indicate the condition required to achieve oxygen ionic conductivity ${\sigma }_T\ge {10}^{-2}$ $\mathrm{S}·{\mathrm{cm}}^{-1}$ at different temperatures ($T=473$–$873$ $\mathrm{K}$). The sharded region ($T<473$ $\mathrm{K}$) highlights the particularly challenging region, which may be accessed either by suppressing $E_a$ (green) or enhancing $A$ (pink). The inset shows the classification of the dataset into $14$ structural classes, along with the corresponding number of entries in each class. (b,c) Histograms of (b) $E_a$ and (c) ${\log }_{10}A$. In (b) and (c), statistical parameters (minimum, average, maximum, standard deviation, and skewness) are provided; for minima and maxima, the experimental reports with their formula are referred to^20,89,92.

The oxides in the dataset are classified into $14$ structural classes, namely perovskite, perovskite-related, fluorite, $\delta$-${\mathrm{Bi}}_2{\mathrm{O}}_3$, BIMEVOX, ${\mathrm{Bi}}_2\mathrm{U}{\mathrm{O}}_6$, scheelite, apatite, cuspidine, melilite, garnet, pyrochlore, columbite, and $\alpha$-$\mathrm{SrSi}{\mathrm{O}}_3$ classes. The counts of oxides per class are shown in the inset of Fig. 2a. Perovskites and related compounds constitute the largest fraction of reported data, reflecting the dominant research focus in the field, followed by scheelites, apatites, and other structures.

Figure 2b shows the distribution of $E_a$, with most values lying between 200 and 2000 meV and an average of $1020$ $\mathrm{meV}$. The lowest $E_a$ ($288\mathrm{meV}$) was observed for the columbite ${\mathrm{Mg}}_{0.9}{\mathrm{Cu}}_{0.1}{\mathrm{Nb}}_2{\mathrm{O}}_{6-\delta }$⁹², while the highest ($3470\mathrm{m}\mathrm{e}\mathrm{V}$) was reported for the garnet ${\mathrm{Y}}_3{\mathrm{Al}}_{4.99}{\mathrm{Fe}}_{0.01}{\mathrm{O}}_{12}$⁸⁹. Figure 2c presents the distribution of ${\log }_{10}A$, with most A values lying between 1 and 10¹⁰ K∙S∙cm⁻¹ and an average of 10^5.55 = 3.55 × 10⁵ K∙S∙cm⁻¹. The lowest $A$ (${10}^{-2.05}=8.91\times {10}^{-3}\mathrm{K}\cdot \mathrm{S}\cdot {\mathrm{c}\mathrm{m}}^{-1}$) was observed for the columbite ${\mathrm{Mg}}_{0.8}{\mathrm{Li}}_{0.2}{\mathrm{Nb}}_2{\mathrm{O}}_{6-\delta }$⁹², while the highest (${10}^{16.0}\mathrm{K}\cdot \mathrm{S}\cdot {\mathrm{c}\mathrm{m}}^{-1}$) was reported for the perovskite-related brownmillerite $\mathrm{S}\mathrm{r}\mathrm{F}\mathrm{e}{\mathrm{O}}_{2.5+\delta }$²⁰.

We also evaluated the reproducibility of the conductivities. The conductivities of $26$ oxides were reported in more than one publication, providing opportunities for cross-validation. The average standard deviations were $78.1\mathrm{m}\mathrm{e}\mathrm{V}$ for $E_a$ and ${10}^{0.41}=2.58\mathrm{K}\cdot \mathrm{S}\cdot {\mathrm{c}\mathrm{m}}^{-1}$ for A. Considering the distributions of A and E_a shown in Fig. 2b,c, respectively, a high degree of consistency among independent studies is achieved despite variations in synthesis and measurement conditions.

Technical Validation

To support the reliability and scientific utility of the dataset, we performed symbolic regression analyses on both $E_a$ and $A$ values by using GoodRegressor¹⁰⁷, developed to derive analytical models that link experimental datasets with the structural and compositional features of materials. A brief summary of the program package is included in the Supplementary Information, whereas the technical and numerical details are provided in ref. ¹⁰⁷. The descriptors used in the regressions are summarized in Table 2.

Table 2.

Structural, chemical, and physical properties of constituent elements of oxides for regression models, given as “features” for symbolic regression modeling.

Properties	Description	Unit
$\stackrel{\circledR }{O}$	Molar ratio of oxygen ions to metal ions	—
$M$	Atomic mass	$\mathrm{g}·{\mathrm{mol}}^{-1}$
$Z$	Valence	—
$r_{\mathrm{V}\mathrm{I}}$	Shannon ionic radius with sixfold coordination to oxygen108,109	$Å$
$r$	Shannon ionic radius depending on $n_c$ (see below)108,109	$Å$
$B$	Bulk modulus	$\mathrm{GPa}$
$G$	Shear modulus	$\mathrm{GPa}$
$\rho$	Density	$\mathrm{g}·{\mathrm{cm}}^{-3}$
${\rho }_{\mathrm{m}\mathrm{ol}}$	Molar density, defined as $\rho /M$	$\mathrm{mol}·{\mathrm{cm}}^{-3}$
${\eta }_f$	Ionic filling rate (per unit volume), defined as ${\eta }_f=N_{\mathrm{avo}}{\rho }_{\mathrm{mol}}\left(\frac{4}{3}\pi r_{\mathrm{VI}}^3\right)$ where $N_{\mathrm{avo}}$ is the Avogadro’s constant	—
$n$	Principal quantum number of valence electrons	—
$l$	Azimuthal quantum number of valence electrons	—
$\alpha$	Thermal expansion coefficient (linear not volumetric)	${\mathrm{K}}^{-1}$
$\kappa$	Thermal conductivity	$\mathrm{W}·{\mathrm{m}}^{-1}·{\mathrm{K}}^{-1}$
$\chi -{\chi }_{\mathrm{O}}$	Difference in electronegativity between metal ions and oxygen	—
$\nu$	Poisson’s ratio	—
${\theta }_D$	Debye temperature	$\mathrm{K}$
$n_c$	Coordination number to oxygen ions, which depends on the occupied sites ($A_s$, $B_s$, or $C_s$) in different crystal classes. For example, in perovskite oxides, $n_c$ for $A_s$ and $B_s$ are given as $12$ and $6$, respectively	—
$〈\cdots 〉$	Average over the constituent metal ions	Common to $\cdots$
$\sigma \left(\cdots \right)$	Standard deviation over the constituent metal ions	Common to $\cdots$
$r\left(\cdots \right)$	Skewness over the constituent metal ions	—

As shown in Fig. 3a, the symbolic regression model for $E_a$ achieved $R^2=0.831$, $\mathrm{RMSE}=162\mathrm{meV}$, and $\mathrm{M}\mathrm{A}\mathrm{E}=118\mathrm{m}\mathrm{e}\mathrm{V}$ across the full dataset. The key interactions distinguishing oxides with suppressed $E_a$ (green curve) from those with large $E_a$ (pink curve) were identified (see ref. ¹⁰⁷ for details). The most important interaction was found between the average metal-oxygen coordination number ($⟨n_c⟩$) and the molar ratio of oxygen to metal ions $\left(\overline{\mathrm{O}}\right)$ Note that the most important interaction does not necessarily exclude the contribution of the other features. The two-way partial dependence plot in Fig. 3b, obtained by varying $⟨n_c⟩$ and $\overline{\mathrm{O}}$ from the minimum to the maximum values of the full dataset while fixing other features at their mean values. As illustrated in Fig. 3c, a clear trend emerges: oxygen-rich environments with higher coordination numbers $⟨n_c⟩$ (such as cubic coordination) become more favorable than oxygen-poor environments with lower $⟨n_c⟩$ (such as octahedral or tetrahedral coordination), probably because the increased O-O repulsion in high-$⟨n_c⟩$ configurations expands and softens the local oxygen sublattice, thereby facilitating oxygen-ion migration.

Fig. 3.

Symbolic regression models governing factors of oxygen ion transport. (a) Experimental versus regressed values for the symbolic regression model for $E_a$, with the schematic illustrations of potential energy surfaces exhibiting low (green curve) and high $E_a$ (pink curve). (b) Two-way partial dependence plot of the most important interaction for E_a: (n_c)-$\overline{\mathrm{O}}$ (see text for details). (c) Illustration of high (pink panel) and low (green panel) $E_a$, which depends on $⟨n_c⟩$ and $\overline{\mathrm{O}}$ both. (d) Experimental versus regressed values for the symbolic regression model for ${\log }_{10}A$, with the schematic illustrations of potential energy surfaces with low (green arrows) and high ${\log }_{10}A$ (pink arrow). (e) Two-way partial dependence plot of the most important interaction for log₁₀A: 〈r_VI〉-〈Z〉. (f) Illustration of how $〈r_{\mathrm{V}\mathrm{I}}〉$ and 〈Z〉 modulate Coulombic interactions and oxygen ion jump attempts. In (b) and (e), histograms of the corresponding features for the full dataset are also shown along the upper and right axes.

As shown in Fig. 3d, the symbolic regression model for ${\log }_{10}A$ achieved $R^2=0.758$, $\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}={10}^{1.07}\mathrm{K}\cdot \mathrm{S}\cdot {\mathrm{c}\mathrm{m}}^{-1}$, and $\mathrm{MAE}={10}^{0.769}$ $\mathrm{K}·\mathrm{S}·{\mathrm{cm}}^{-1}$ for the entire dataset. The principal interactions distinguishing low and high $A$ values (green and pink arrows, respectively) were identified (see also ref. ¹⁰⁷ for details). The most important interaction involves the average Shannon ionic radius with sixfold coordination ($〈r_{\mathrm{V}\mathrm{I}}〉$)^108,109 and the average metal ion valence (〈Z〉), whose two-way dependence is shown in Fig. 3e. A straightforward interpretation follows; small $〈r_{\mathrm{V}\mathrm{I}}〉$ combined with large 〈Z〉 strengthens the metal-oxygen Coulombic attraction, which stabilizes oxygen ions at their lattice sites and reduces their configurational freedom. As a result, the rate of attempted ionic jumps is suppressed. Conversely, when the metal-oxygen attraction is weaker, oxygen ions experience less localization and can attempt jumps more frequently, leading to the higher ${\log }_{10}A$ (Fig. 3f).

Our regression models can also be utilized for materials design. By substituting or doping ions and adjusting compositions, the model enables the exploration of new compounds with desirable target metrics, characterized by reduced $E_a$ and enhanced $A$. In this study, the most promising candidate was identified as the apatite-type compound ${\mathrm{La}}_{9.5}{\mathrm{Si}}_{5.5}{\mathrm{Al}}_{0.5}{\mathrm{O}}_{26}$. The predicted $E_a$ and $A$ were $494\text{meV}$ and $7.41\times {10}^6\mathrm{K}\cdot \mathrm{S}\cdot {\mathrm{c}\mathrm{m}}^{-1}$, respectively. This composition is a modified analogue of the experimentally reported apatite ${\mathrm{Nd}}_{9.5}{\mathrm{Si}}_{5.5}{\mathrm{Al}}_{0.5}{\mathrm{O}}_{26}$, which exhibited $E_a=658\mathrm{meV}$ and $A=1.44\times {10}^7\mathrm{K}\cdot \mathrm{S}\cdot {\mathrm{c}\mathrm{m}}^{-1}$ ⁷⁹. This example demonstrates the utility of the curated dataset and regression framework as a quantitative guide for both experimental and theoretical materials discovery. We envision that such data-driven predictions will accelerate the development of next-generation oxygen-ion conductors.

Author contributions

Y.K. proposed the project. S.-H.J. curated all data under the supervision of H.T., developed the regression model, and prepared the initial draft of the manuscript. The manuscript was revised by S.-H.J. and Y.K. All authors contributed to discussions and reviewed the manuscript.

Data availability

The total dataset created in this study is openly available at https://github.com/JerryGarcia1995/OxygenIonConductor/blob/main/database/oxygen_ion_conductor_dataset.csv and 10.5281/zenodo.18947543.

Code availability

The source code supporting materials prediction and design in this study is openly available at https://github.com/JerryGarcia1995/OxygenIonConductor/tree/main/modelling_GoodRegressor.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41597-026-07100-x.