Efficient low-temperature thermochemical CO₂ splitting enabled by Gibbs free energy engineering

Qi Wang^a, Yimin Xuan*^a, Zhonghui Zhu^a, Liang Teng^a, Xianglei Liu^a and Yunfei Gao^b
aSchool of Energy and Power Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China. E-mail: ymxuan@nuaa.edu.cn
^bInstitute of Clean Coal Technology, East China University of Science and Technology, Shanghai 200237, China

First published on 12th August 2025

---

---

Introduction

The over-consumption of fossil fuels triggered a global energy crisis and caused substantial CO₂ emissions.^1,2 The continued increase in CO₂ has led to the emergence of significant climate problems, including global warming, which hinders global societal progress. Developing effective strategies to mitigate energy and environmental challenges is critical. Two-step direct thermochemical CO₂ splitting, which converts CO₂ into CO through redox cycles, can address environmental and energy challenges.^3,4 In recent years, significant progress has been made in enhancing the performance of CO₂ splitting, particularly through the design and optimization of various material systems,^5,6 the development of innovative morphological structures,⁷ and improvements in reactor efficiency.^8,9 Nevertheless, discovering materials that balance high redox activity, long-term cyclability, and economic viability remains a key scientific challenge.

Various material systems have been extensively investigated, such as ZnO/Zn,^10,11 ferrites,¹² CeO₂-based materials,^13–15 and perovskite-based oxides.^16,17 Perovskites, represented by the formula ABO₃, exhibit broad compositional flexibility, high thermal stability, and the ability to tolerate oxygen non-stoichiometry while maintaining structural stability, making them promising candidates for thermochemical applications.¹⁸ The typical two-step thermochemical cycle comprises the following steps:

Step 1, reduction reaction (ΔG₁):

	(1)

Step 2, oxidation reaction (ΔG₂):

δCO2 + ABO3−δ → δCO + ABO3	(2)

where ABO₃ and ABO_3−δ represent the perovskite before and after the reduction process, respectively, δ denotes the non-stoichiometric amount of oxygen. During the reduction process, perovskite oxides release lattice oxygen at high temperatures (∼1350 °C), generating oxygen gas. During oxidation, the reduced perovskite is re-oxidized while CO₂ is converted to CO. LaMnO₃-based perovskites have been extensively studied for thermochemical CO₂ splitting due to their excellent oxygen release and absorption capabilities.¹⁹ However, their limited oxygen non-stoichiometry hinders practical efficiency.²⁰ To address this issue, strategies such as A-site/B-site doping or composite design have been employed. Notably, doping LaMnO₃ with Sr/Ca (A-site)^17,21 or Al/Cr (B-site),^22,23 as well as adopting dual-doping approaches (e.g., Sr/Al), has proven effective in enhancing both CO₂ conversion rates and cycling stability.²⁴ As reported by Bork A. H. et al.,²⁵ the synergy between La_0.65Sr_0.35MnO₃ and CeO₂ optimizes reaction thermodynamics, thereby improving CO₂ conversion. In a parallel study, Wang et al.²⁶ enhanced CO production efficiency by strategically designing perovskite materials with tunable configurational entropy. Despite these achievements, most of such thermochemical conversion systems still suffer from high operating temperatures (around 1350 °C) and a significant temperature difference between the two-step cycles, leading to energy inefficiency. To address this challenge, recent studies have also focused on lowering the reaction temperature. Kildahl et al., proposed a double perovskite that facilitates two-step CO₂ conversion at 700 °C; however, it requires very long reduction and oxidation half-cycles (24 hours), resulting in impractical operational durations.^27,28 Therefore, discovering materials capable of efficient CO₂ conversion at lower temperatures with rapid kinetics and minimal energy loss is critical. Nevertheless, conventional trial-and-error experimentation remains prohibitively slow and labor-intensive, underscoring the urgency for accelerated discovery frameworks.

Recently, the integration of high-throughput computing and machine learning (ML) technologies has opened new pathways for material discovery. For instance, A. Emery et al. proposed a high-throughput density functional theory (HT-DFT) method to investigate 5329 cubic and distorted perovskite ABO₃ compounds, successfully identifying materials with favorable thermodynamic properties for two-step thermochemical water splitting (TWS).²⁹ Similarly, Jiang et al. employed high-throughput computing to screen 21 out of 2018 double perovskite oxides for photocatalytic applications.³⁰ While high-throughput computing facilitates rapid material screening for potential applications, it is constrained by computational power and cost. In contrast, ML enhances screening efficiency and predictive accuracy.^31,32 For example, Wang et al. integrated high-throughput computing with ML to identify promising materials for chemical looping (CL) air separation and CL CO₂ splitting.³³ Additionally, Nguyen et al. combined high-throughput screening with nonlinear supervised machine learning to accurately predict C2 hydrocarbon production.³⁴ Overall, the integration of high-throughput computing, ML techniques, and physical mechanisms offers a more efficient approach to material discovery compared to traditional experimental trial-and-error methods.

Herein, we have successfully identified high-performance materials for isothermal low-temperature thermochemical CO₂ splitting by integrating high-throughput computing with machine learning (ML). The optimal material exhibits an impressive CO yield of 1834 μmol g⁻¹ at 1100 °C under isothermal conditions, representing the highest reported CO₂-to-CO conversion efficiency to date, to the best of our knowledge. Our methodology synergistically combines high-throughput DFT calculations with ML to systematically investigate A-site and/or B-site doped LaMnO₃-based perovskites and their extended combinations (Fig. 1). The initial screening of candidate materials is conducted based on perovskite structural stability, followed by DFT calculations of the Gibbs free energy (ΔG) for the qualified materials. Based on the Gibbs free energy formula ΔG = ΔH − TΔS, we establish the relationship between material composition and the two-step redox reaction by calculating enthalpy and entropy. By doping with various A- and/or B-site elements, we optimize thermodynamic properties and generate a comprehensive dataset of Gibbs free energy (ΔG) across different oxygen non-stoichiometries and temperatures. This approach enables the rapid identification of promising candidates through a screening criterion that requires negative ΔG values for both reduction and oxidation processes. Furthermore, a machine learning model trained on high-throughput data predicts the Gibbs free energy of oxygen vacancy formation across 279142 perovskite combinations, thereby accelerating material screening. Experimental validation confirms not only the model's accuracy but also elucidates the mechanisms behind the superior performance of selected materials. This work not only identifies novel materials for efficient CO₂ splitting but also establishes a scalable framework for accelerating material discovery, providing a practical pathway to realize low-temperature thermochemical processes.

Results and discussion

Material screening based on perovskite structural stability

Using LaMnO₃ as a prototypical perovskite material, we engineered the La_xA_1−xMn_yB_1−yO_3−δ structure by substituting A-site (La) and/or B-site (Mn) cations. In this study, the A-site positions were occupied by alkaline-earth or rare-earth elements, while transition metal elements occupied the B-site positions. The specific replacement elements for both cationic sites and their respective concentration ratios are detailed in Table 1. In perovskite materials, δ represents the oxygen non-stoichiometry, whose magnitude critically determines redox activity. When δ > 0.25, the reducing capacity declines with further increase in δ.¹⁸ In order to comprehensively examine the possible influence of δ on material properties, we set the value range of δ from 0 to 0.5 in this work.³⁵

---

Through the exploration of all feasible A-site and/or B-site dopants within the defined cation pool, we constructed 1681 distinct La_xA_x−1Mn_yB_y−1O₃ perovskite oxide configurations. However, certain combinations may not stabilize into a perovskite phase. To address this issue, we first eliminated non-charge-neutral configurations based on the charge neutrality criterion (see Section S1 in the SI). After this screening, all 1681 combinations remained charge-balanced. Furthermore, the Goldschmidt tolerance factor serves as a fundamental empirical descriptor for predicting the stability of perovskite structures. Recently, Bartel et al.³⁶ proposed an improved tolerance factor, denoted as τ , where r_O, r_A, and r_B represent the ionic radii of oxygen anions, A-site cations, and B-site cations, respectively, while n_A and n_B correspond to the oxidation states of the cations occupying the A- and B-sites (Fig. 2a). The modified τ incorporates additional chemical insights and demonstrates superior predictive accuracy compared to the Goldschmidt tolerance factor. In this study, τ was employed as the primary stability indicator. Given the preliminary nature of this screening, we applied a more flexible threshold (τ < 4.3) rather than the stricter limit (τ < 4.18) suggested by Bartel et al.³⁶

Fig. 2b compares the assessment of perovskite stability using the Goldschmidt tolerance factor (t) and the modified tolerance factor (τ). While the Goldschmidt tolerance factor identifies structurally stable perovskite phases in zones III and IV, the modified factor (τ) identifies them in zones I, III, and V. Consequently, τ expands the predictions of structural stability to include additional configurations in zone V, with yttrium-containing perovskites constituting the majority of these candidates. Notably, the clustering behavior observed in the data distribution of Fig. 2b arises from two distinct factors: fundamental differences in perovskite compositions and stoichiometric variations within identical perovskite systems. Furthermore, four perovskite combinations with high nickel content in Region IV were excluded after applying the modified tolerance factor. In total, 1677 perovskite compositions were identified as structurally stable candidates for subsequent high-throughput computational screening.

High-throughput DFT calculations

In the two-step thermochemical CO₂ splitting cycle, the redox thermodynamic properties of perovskite oxides are critical determinants of process efficiency. A significant challenge is the development of robust yet simplified descriptors that accurately quantify these redox characteristics. Although the oxygen vacancy formation energy of perovskites has been proposed as a potential indicator of redox performance, this parameter frequently demonstrates a weak correlation with experimental results.^37,38 These discrepancies primarily arise from the metastable oxygen vacancies that are inherently present in synthesized perovskites, whose concentrations dynamically fluctuate under operational redox cycling conditions.

To address these limitations, researchers have adopted the Gibbs free energy change of oxygen vacancy formation (ΔG(δ)) within a specific range of δ to quantify the redox performance of oxygen carriers in chemical looping (CL) processes.³³ ΔG(δ) can be directly compared with the oxygen partial pressure and has proven effective in predicting material behavior.^35,39 In this study, we utilize ΔG(δ) to evaluate the redox properties of perovskite oxides in the two-step thermochemical CO₂ splitting cycle. This descriptor facilitates a systematic investigation of the redox thermodynamics of La_xA_1−xMn_yB_1−yO_3−δ across varying δ ranges, bridging theoretical predictions with experimental performance.

	(3)

where δ₁ → δ₂ represents the change in the value of from δ₁ to δ₂. The slope of δ and G represents the oxygen concentration across a range of vacancy concentrations at a specific temperature. This information allows for the assessment of a material's suitability to absorb and release oxygen within a defined temperature range. In the two-step thermochemical CO₂ splitting reaction, the reduction capability of the perovskite material varies with different δ values. At low δ values, the driving force for reduction is strong; however, this driving force decreases rapidly as δ increases.¹⁸ When δ exceeds 0.25, the material becomes increasingly difficult to reduce. Therefore, we analyze the change in Gibbs free energy during two specific intervals: when δ transitions from 0 to 0.125 and from 0.125 to 0.25.

High-throughput DFT calculations were performed on all candidate materials using ΔG as the descriptor. Fig. 3a and b summarize the variation of ΔG as δ changes from 0 to 0.125 and from 0.125 to 0.25. The choice of doping elements significantly influences the ΔG values. For example, titanium (Ti) doping consistently elevates ΔG, exceeding 6 eV in certain configurations, while yttrium (Y) doping induces a marked reduction. Notably, a systematic increase in ΔG occurs when δ shifts from 0.125 to 0.25. This phenomenon may be attributed to the increase in the nonstoichiometric oxygen ratio, which causes the oxygen in the lattice to become less easily dissociated, thereby requiring more energy and resulting in a higher ΔG.

Verification of the high-throughput DFT results

Efficient thermochemical CO₂ splitting through redox cycling necessitates thermodynamic favorability in both half-reactions: oxygen uncoupling and CO₂ reduction. This requirement entails a negative Gibbs free energy change (ΔG) for each reaction step. The energy of the reduction half-reaction (ΔG₁) was computed using high-throughput DFT, while the oxidation half-reaction energy (ΔG₂) at 1100 °C for various perovskite compositions was calculated according to the methodology outlined in SI Section S2.

As illustrated in Fig. 3c and d, these results encompass two distinct ranges of oxygen non-stoichiometry: 0–0.125 and 0.125–0.25. ΔG₁ and ΔG₂ exhibit a strong inverse linear correlation, with each data point representing a unique perovskite composition. Materials in zones I, II, and IV exhibit ΔG > 0 in either the reduction and/or oxidation, violating the thermodynamic spontaneity criterion. In contrast, zone III contains perovskite candidates with ΔG < 0 for both reactions, thereby satisfying the dual thermodynamic requirements for efficient cyclic operation. Among the 1677 screened compositions, 11 candidates that meet this criterion were identified. These selected perovskites were synthesized using the sol–gel method and characterized by X-ray diffraction (XRD) (Fig. S2), with their redox properties systematically quantified. Most synthesized perovskites exhibited phase-pure structures (referencing the perovskite standard card PDF#77-0182), although minor metal oxide impurities were observed in a few instances. Given their extremely low concentration and the fact that the main structure of the synthesized materials remains the ABO₃-type perovskite phase, the presence of these trace impurities does not compromise the main conclusions of the experiment. To avoid differences in thermochemical cycling conditions, CeO₂ was added as a blank control and tested under the same experimental procedures. Notably, SmMn_0.25Fe_0.75O₃ (SMF) demonstrated exceptional performance, achieving a CO yield of 1277.69 μmol g⁻¹—1.93 times higher than CeO₂ (Tables 2 and S3). Importantly, the initial cycle was discarded to eliminate activation effects, and redox performance was evaluated based on the average CO production from subsequent cycles.²⁶ The experimental results show remarkable consistency with high-throughput DFT predictions.

---

Notably, the DFT screening was restricted to perovskites with dual A- and B-site doping due to computational constraints. In contrast, existing studies on high-entropy perovskites suggest that more extensive multi-site doping configurations are feasible.²⁶ To address this limitation, we developed machine learning (ML) models that leverage the high-throughput DFT dataset, allowing for the exploration of compositions with up to three dopants per crystallographic site. The discoveries enabled by these ML models are systematically analyzed in the following section.

ML-guided material discovery

The high-throughput DFT dataset served as the foundation for the development of machine learning (ML) models. Initial structural screening, utilizing the Goldschmidt tolerance factor, eliminated 2819 unstable candidates from a total of 281961 possible perovskite combinations. As established in previous sections, the DFT calculations quantified the oxygen vacancy formation energy (ΔG), which was designated as the target output for ML training.

Fig. 4a compares the mean absolute errors (MAEs) of various algorithms across two δ regimes: 0–0.125 and 0.125–0.25. Notably, the 0.125–0.25 range exhibited lower MAEs than the 0–0.125 range, likely attributable to its narrower data distribution. Linear algorithms, including linear regression (LR) and ridge regression, displayed substantially higher MAEs (2.508–2.752), indicating limitations in capturing the composition-ΔG relationships through linear approximations. In contrast, nonlinear models effectively addressed this constraint. Decision tree (DT) algorithms reduced MAEs to 1.746–2.396, while ensemble learning achieved even greater precision; extreme gradient boosting (XGB) lowered MAEs to 1.004–1.593. The implementation of AutoML via AutoGluon yielded further improvements,^40,41 particularly with its 1-layer XGBoost_BAG_L1 model (as shown in Fig. 4b and c and other ML models in Fig. S3–S5). This optimized model demonstrated consistent predictive accuracy across both δ ranges (0–0.125 and 0.125–0.25), establishing it as the preferred framework for subsequent analyses.

Leveraging the optimized ML model, we systematically evaluated the redox thermodynamics of 279142 multi-cation perovskites, each containing 5 to 6 cations. This screening process identified 8451 viable candidates for low-temperature thermochemical conversion (Fig. 4d and e, Region III). From this pool, eight representative compositions were experimentally validated for their CO₂ splitting performance at reduced temperatures, accompanied by structural characterization through X-ray diffraction (XRD) (refer to Tables 3 and S4; crystallographic data available in Fig. S6). Remarkably, La_0.5Sm_0.125Sr_0.375Co_0.25Fe_0.125Ti_0.625O₃ (LSSCFT) exhibited unprecedented CO productivity of 1833.52 μmol g⁻¹, outperforming all materials evaluated in this study. We conducted long-term cycling stability tests (20 cycles) on the LSSCFT material. As shown in Fig. S7, after an initial decay in CO yield during the early cycles, the yield stabilized from the 4th cycle. By the 20th cycle, the CO yield stabilized at 1135.58 μmol g⁻¹, corresponding to 70.7% of the 4th-cycle value, demonstrating retained activity during prolonged cycling. The initial rapid decay in CO yield arises from the high-entropy effect,²⁶ which introduces abundant intrinsic oxygen vacancies during synthesis, but only a fraction of these vacancies are reversibly cyclable. Early cycles rapidly deplete irreversibly “non-cyclic” oxygen vacancies, ultimately leaving only cyclable oxygen vacancies. This results in an initial rapid decay of CO production followed by stabilization. Notably, post-cycling SEM reveals perovskite grain coarsening (Fig. S8). Material sintering further contributes to a ∼29% decline in late-cycle stability.⁴²

---

Gibbs free energy and redox properties of materials

To establish design principles for high-performance materials, we conducted a systematic thermodynamic analysis of the screened candidates. The reduction and oxidation Gibbs free energies (ΔG) were quantitatively mapped (Fig. 5a and b). While most materials exhibited near-zero reduction ΔG values (ΔG ≈ 0), a small subset demonstrated significant positive deviations. Notably, the top-performing materials displayed systematically lower oxidation ΔG values, indicating that the oxidation step—rather than the reduction thermodynamics—governs overall redox efficiency. This inverse correlation between oxidation ΔG and CO yield arises because the oxidation half-reaction acts as both the thermodynamic driver and the kinetic bottleneck for CO generation.

Notably, the oxidation ΔG of the screened materials was lower compared to the reduction ΔG (Fig. 5b). This thermodynamic observation implies that oxidation energetics—rather than reduction parameters—should be prioritized in subsequent material screening protocols. While most high-potential candidates demonstrate comparable ΔG values for both half-reactions (with less than a 10% differential), critical exceptions underscore the necessity for balanced optimization. Ideal candidates should maintain an oxidation ΔG below −0.8 eV while ensuring that the absolute value of the difference between the reduction ΔG and the oxidation ΔG remains within 1.5 eV, thereby achieving optimal energy coupling between redox steps.

The top-performing LSSCFT material underwent comprehensive characterization using X-ray diffraction (XRD), scanning electron microscopy (SEM), high-resolution transmission electron microscopy (TEM), and energy-dispersive spectroscopy (EDS). The XRD patterns confirming perovskite structure retention throughout redox cycling (Fig. 5c and d). Meanwhile, SEM and TEM images show that the perovskite lattice retains its orderly structure, with no phase changes observed, confirming the material's stability during cycling (Fig. 5e). EDS elemental mapping confirmed a homogeneous cation distribution both before and after stability testing. Notably, varying degrees of line defects (white lines, Fig. 5f) were observed on the perovskite surface. Although the exact origins of the line defects in LSSCFT after redox cycling are not fully understood, they may be related to the high configurational entropy.²⁶ The mechanisms underlying the effects of these line defects will be explored in future studies.

We compare the CO yields of perovskite materials in this study with literature-reported values at various temperatures (Fig. 5g). The reported CO yields for most materials are below 1000 μmol g⁻¹, typically requiring reaction temperatures above 1200 °C. Although materials like LaCo_0.7Zr_0.3O₃ achieve yields above 1000 μmol g⁻¹, their reaction temperature of 1300 °C results in significant energy loss. Notably, LSSCFT achieve both high CO yields and exceptional isothermal cycling stability at 1100 °C, validating the ML predictions and demonstrating the material's industrial applicability.

Conclusion

In summary, this study presents a systematic approach to discovering low-temperature materials suitable for isothermal two-step thermochemical CO₂ splitting, combining high-throughput density functional theory (DFT) calculations with machine learning (ML) techniques. Through DFT analysis, we evaluated the Gibbs free energy of oxygen vacancy formation in a dataset of 1677 perovskite oxides, considering various cation doping types and concentrations. This investigation identified 11 promising candidates for CO₂ splitting applications, and the DFT predictions were subsequently validated through experiments. By incorporating high-throughput data into ML models utilizing advanced AutoML techniques, we expanded the screening range and accurately predicted the Gibbs free energy for approximately 279142 perovskite combinations. By applying the criteria of negative Gibbs free energies for both oxygen decoupling and CO₂ reduction reactions, we identified approximately 8451 potential materials, of which 8 were selected for experimental validation. Among these, the novel material La_0.5Sm_0.125Sr_0.375Co_0.25Fe_0.125Ti_0.625O₃ achieved a CO yield of 1833.52 μmol g⁻¹ under isothermal cycling conditions at 1100 °C. This temperature is 250 °C lower than that required by conventional thermochemical two-step methods. The material also demonstrates long-term cycling stability. From the 4th cycle, it entered a stabilized phase with an initial stabilized yield of 1605.39 μmol g⁻¹. By the 20th cycle, it maintained a CO yield of 1135.58 μmol g⁻¹, retaining 70.7% of its initial stabilized yield. This approach significantly improves the discovery efficiency of perovskite materials, accelerating the development of CO₂ reduction technologies. The methodology facilitates low-temperature operation with exceptional yield and efficiency while providing mechanistic insights critical for bridging the gap between fundamental research and industrial-scale applications.

Experimental section

Computational method

First-principles calculations were conducted using the Vienna Ab initio Simulation Package (VASP),^47,48 employing density functional theory (DFT) based on the generalized gradient approximation (GGA) of Perdew–Burke–Ernzerhof (PBE).⁴⁹ The kinetic energy cutoff for the plane-wave basis was set to 450 eV. The convergence criteria for force and energy were 0.01 eV Å⁻¹ and 10⁻⁵ eV, respectively. To accurately model the disordered atomic configurations, the Monte Carlo Special Quasirandom Structure (MCSQS) method was used to determine the spatial distribution of A-site and B-site dopants along with oxygen vacancies.⁵⁰ A 2 × 2 × 2 perovskite supercell was constructed, and each perovskite model contained 40-8δ atoms. For the perovskite model containing 40-8δ atoms, the gamma K point was adopted to reduce computational intensity, set to 1 × 1 × 1. The strong Coulomb interactions of d-orbital electrons at the Ti, Mn, Fe, Co, and Ni sites were addressed using the GGA+U method,⁵¹ with U values of 3, 3.9, 4, 3.4, and 6, respectively. Given that the influence of magnetic order on oxygen vacancy formation is relatively minor, we adopted ferromagnetic (FM) phase magnetic ordering for all doping structures to facilitate the simulation. The initial spin magnetic moments of Mn, Fe, Co, and Ni were assigned values of 5, 4, 5, and 5, respectively, to define their magnetic configurations in the computational model. The CBS-QB3 method implemented in Gaussian16 was used to calculate the enthalpy and entropy of O₂.³³ Vibration-related properties, including zero-point energy (ZPE) and the contribution to entropy from phonon vibrations (ΔS_vib), were calculated using Phonopy⁵² under the harmonic approximation. Theoretically, the Gibbs free energy of a system can be expressed as follows.

ΔG = ΔH − T(ΔSvib + ΔSconf) ≈ ΔU − T(ΔSvib + ΔSconf)	(4)

H = U + PV, PV of solid materials is negligible, then

ΔG = ΔU − TΔSvib − TΔSconf	(5)

In the formula, ΔU – TΔS_vib ≈ ΔE + ΔF_corr, ΔE represents the energy difference between the reactants and products, which is obtained from VASP after performing a DFT calculation for geometry optimization. ΔF_corr is the difference in the Helmholtz free energy correction of the reactants and products, which can be obtained from Phonopy. The change in configurational entropy is given by ΔS_conf = aR[2δln(2δ) + (1 − 2δ)ln(1 − 2δ)],⁵³ where R is the ideal gas constant, δ represents the nonstoichiometric amount of oxygen in the solid, and a is set to 2, indicating an ideal solid solution without defect interactions.

Material synthesis and characterization

A modified Pechini method was utilized to synthesize all materials in this study. In the synthesis of perovskite materials, stoichiometric amounts of metal nitrates were initially dissolved in 30 mL of deionized water. Citric acid was then added to the solution and stirred at room temperature for 60 minutes, maintaining a molar ratio of citric acid to metal ions of 2.5:1. During the synthesis of the titanium-containing material, with titanium positioned at the B site, a stoichiometric ratio of titanium butoxide was precisely incorporated. Ethanol was added at a mass ratio of 10:1 relative to titanium butoxide. Subsequently, ethylene glycol was introduced as an auxiliary complexing agent, with its molar ratio controlled at 1.5:1 relative to citric acid. The prepared solution was placed in a water bath at 90 °C, continuously stirred for 3 hours until a gel was formed. The resulting gel was then dried in an oven at 120 °C for 24 hours. Following this step, the dried sample was ground and subjected to calcination in an air atmosphere at 1100 °C for 8 hours to eliminate organic residues. The crystal phase of the synthesized materials was characterized by powder X-ray diffraction using a Bruker D8 ADVANCE X-ray diffractometer. The instrument operated at a current of 40 mA and a voltage of 40 kV. All specimens were analyzed via X-ray diffraction (XRD) within a 2θ angular range of 10–80° at a scan rate of 0.02° per second. The microstructural morphology of the perovskite powder samples was characterized using a thermal field emission scanning electron microscope (GeminiSEM 300). The chemical composition and bonding states were further investigated through X-ray photoelectron spectroscopy (XPS) measurements performed on a Thermo Fisher K-Alpha instrument.

ML methods

In this work, all ML algorithms are implemented using scikit-learn. Various ML models are employed to predict the Gibbs free energy of oxygen vacancy formation in materials. They include linear models such as linear regression (LR) and ridge regression (Ridge), nonlinear models such as support vector regression (SVR) and decision tree (DT), and ensemble learning models such as random forest (RF), gradient boosting decision tree (GBDT), and extreme gradient boosting (XGBoost). To further reduce the “overfitting” issue, AutoGluon is also used as an automatic machine learning (AutoML) approach.⁵⁴ The features used in the ML include oxygen vacancy formation energy, element type, element ratio, and the perovskite tolerance factor, while the target variable is the free energy of oxygen vacancy formation.

A total of 8385 datasets were randomly divided into two parts: 6708 datasets were randomly selected for training, while the remaining 1677 datasets were used for testing. To quantitatively analyze the regression prediction model, the mean squared error (MSE) was employed to evaluate the model's predictive performance.

	(6)

In the formula, f_i is the predicted value and y_i is the actual value.

Sample evaluation

The low-temperature thermochemical CO₂ splitting performance was evaluated in a vertical tube furnace (Model TFV-1200-12-200) under controlled atmosphere conditions. The synthesized sample was thoroughly ground prior to testing, and 1.5 g of the sample was placed in a quartz tube. The thermal protocol employed a linear heating rate of 10 °C per minute until reaching the isothermal operating temperature of 1100 °C, which was maintained throughout both the reduction and oxidation phases under distinct gas environments: argon flow (5 mL min⁻¹) during the 120-minute reduction phase, followed by CO₂ flow (5 mL min⁻¹) during the subsequent 60-minute oxidation phase. A total of four cycles were performed to ensure that the reaction reached a steady state. The tail gas produced by each oxidation reaction was collected using an aluminum foil gas collection bag. The gases produced from the oxidation reaction were sampled using a vacuum-tight syringe and analyzed using a gas chromatography (GC9720 Plus, Zhejiang Fuli Analytical Instrument Co., Ltd, China) equipped with a thermal conductivity detector (TCD) and a flame ionization detector (FID). To eliminate the influence of instability in the first cycle, this work uses the average CO yield of the subsequent three cycles as the criterion for judging the material's cycling performance.

Conflicts of interest

There are no conflicts of interest to declare.

Data availability

The data supporting this article have been included as part of the SI.

Supplementary information is available. See DOI: https://doi.org/10.1039/d5ta04855a.

Acknowledgements

This work is supported by the Basic Science Center Program for Ordered Energy Conversion of the National Natural Science Foundation of China (No. 52488201 and No. 52076106) and Natural Science Foundation of Jiangsu Province (No. BK20232022).

References

1. J. Wang, P. Zhu, H. Qin, K. Zuo, H. Zhao and Z. B. Zhang, Electrochemical reactors for the utilization of liquid-phase carbon species, Energy Environ. Sci., 2025, 18, 6438–6455 RSC .
2. P. Du, T. Liu, T. Chen, M. Jiang, H. Zhu, Y. Shang, H. H. Goh, H. Zhao, C. Huang, F. Kong, T. A. Kurniawan, K. C. Goh, Y. Du and D. Zhang, Enhancing green mobility through vehicle-to-grid technology: potential, technological barriers, and policy implications, Energy Environ. Sci., 2025, 18, 4496–4520 RSC .
3. R. Schappi, D. Rutz, F. Dahler, A. Muroyama, P. Haueter, J. Lilliestam, A. Patt, P. Furler and A. Steinfeld, Drop-in fuels from sunlight and air, Nature, 2022, 601, 63–68 CrossRef CAS PubMed .
4. S. Zoller, E. Koepf, D. Nizamian, M. Stephan, A. Patane, P. Haueter, M. Romero, J. Gonzalez-Aguilar, D. Lieftink, E. de Wit, S. Brendelberger, A. Sizmann and A. Steinfeld, A solar tower fuel plant for the thermochemical production of kerosene from H₂O and CO₂, Joule, 2022, 6, 1606–1616 CrossRef CAS PubMed .
5. L. Yang, Z. Huang, J. Fang, J. Zhang and J. Wei, Efficient oxygen exchange and performance of Fe-substituted cobalt-based perovskites for solar thermochemical CO₂ splitting, Chem. Eng. J., 2025, 519, 165504 CrossRef CAS .
6. B. Chen, W. Huang, W. Guo, L. Tong, Y. Ding and L. Wang, Experimental and computational investigations of Ba₂Ca_0.66Nb_1.34-xFe_xO_6-δ perovskites in thermochemical CO₂ splitting cycle: Reducibility, chemical stability, and oxidation kinetics, J. Alloys Compd., 2025, 1020, 179355 CrossRef CAS .
7. F. A. C. Oliveira, M. A. Barreiros, M. Sardinha, M. Leite, J. C. Fernandes and S. Abanades, Thermochemical performance of ceria coated-macroporous 3D-printed black zirconia structures for solar CO/H₂ fuels production, Int. J. Hydrogen Energy, 2025, 100, 477–490 CrossRef .
8. S. Abanades, X. Wang and S. Chuayboon, Chemical Looping CH₄ Reforming Through Isothermal Two-Step Redox Cycling of SrFeO₃ Oxygen Carrier in a Tubular Solar Reactor, Molecules, 2025, 30, 1076 CrossRef CAS PubMed .
9. L. Wei, Z. Pan, S. Sun, Z. Yi, G. Li and L. An, A novel electro-assisted thermochemical reactor for conversion of CO₂/H₂O into solar fuels, Int. J. Heat Mass Transfer, 2024, 229, 125742 CrossRef CAS .
10. S. Abanades and M. Chambon, CO₂ dissociation and upgrading from two-step solar thermochemical processes based on ZnO/Zn and SnO₂/SnO redox pairs, Energy Fuels, 2010, 24, 6667–6674 CrossRef CAS .
11. R. R. Bhosale, Solar hydrogen production via ZnO/Zn based thermochemical water splitting cycle: Effect of partial reduction of ZnO, Int. J. Hydrogen Energy, 2021, 46, 4739–4748 CrossRef CAS .
12. J. Huang, Y. Fu, S. Li, W. Kong, J. Zhang and Y. Sun, Enhanced activity of Mg-Fe-O ferrites for two-step thermochemical CO₂ splitting, J. CO₂ Util., 2018, 26, 544–551 CrossRef CAS .
13. D. Arifin and A. W. Weimer, Kinetics and mechanism of solar-thermochemical H₂ and CO production by oxidation of reduced CeO₂, Sol. Energy, 2018, 160, 178–185 CrossRef CAS .
14. A. de la Calle and A. Bayon, Annual performance of a thermochemical solar syngas production plant based on non-stoichiometric CeO₂, Int. J. Hydrogen Energy, 2019, 44, 1409–1424 CrossRef CAS .
15. S. S. Naghavi, A. A. Emery, H. A. Hansen, F. Zhou, V. Ozolins and C. Wolverton, Giant onsite electronic entropy enhances the performance of ceria for water splitting, Nat. Commun., 2017, 8, 285 CrossRef PubMed .
16. D. R. Barcellos, M. D. Sanders, J. Tong, A. H. McDaniel and R. P. O'Hayre, BaCe_0.25Mn_0.75O_3−δ—a promising perovskite-type oxide for solar thermochemical hydrogen production, Energy Environ. Sci., 2018, 11, 3256–3265 RSC .
17. Y. Kim, S. J. Jeong, B. Koo, S. Lee, N. W. Kwak and W. Jung, Study of the surface reaction kinetics of (La,Sr)MnO₃- oxygen carriers for solar thermochemical fuel production, J. Mater. Chem. A, 2018, 6, 13082–13089 RSC .
18. B. Bulfin, J. Vieten, C. Agrafiotis, M. Roeb and C. Sattler, Applications and limitations of two step metal oxide thermochemical redox cycles: A review, J. Mater. Chem. A, 2017, 5, 18951–18966 RSC .
19. A. Demont and S. Abanades, High redox activity of Sr-substituted lanthanum manganite perovskites for two-step thermochemical dissociation of CO₂, RSC Adv., 2014, 4, 54885–54891 RSC .
20. A. Riaz, P. Kreider, F. Kremer, H. Tabassum, J. S. Yeoh, W. Lipinski and A. Lowe, Electrospun manganese-based perovskites as efficient oxygen exchange redox materials for improved solar thermochemical CO₂ splitting, ACS Appl. Energy Mater., 2019, 2, 2494–2505 CrossRef CAS .
21. X. Liu, T. Wang, K. Gao, X. Meng, Q. Xu, C. Song, Z. Zhu, H. Zheng, Y. Hao and Y. Xuan, Ca- and Ga-doped LaMnO₃ for solar thermochemical CO₂ splitting with high fuel yield and cycle stability, ACS Appl. Energy Mater., 2021, 4, 9000–9012 CrossRef CAS .
22. A. H. McDaniel, E. C. Miller, D. Arifin, A. Ambrosini, E. N. Coker, R. O'Hayre, W. C. Chueh and J. Tong, Sr- and Mn-doped LaAlO_3−δ for solar thermochemical H₂ and CO production, Energy Environ. Sci., 2013, 6, 2424–2428 RSC .
23. H. Sawaguri, N. Gokon, N. Ito, S. Bellan, T. Kodama and H.-s. Cho, Thermochemical two-step CO₂ splitting using La_0.7Sr_0.3Mn_0.9Cr_0.1O₃ of perovskite oxide for solar fuel production, AIP Conf. Proc., 2020, 2303, 170013 CrossRef CAS .
24. M. M. Nair and S. Abanades, Cation synergy in Sr and Al substituted LaMnO₃ during solar thermochemical CO₂ splitting, Energy Adv., 2022, 2, 137–147 RSC .
25. A. H. Bork, A. J. Carrillo, Z. D. Hood, B. Yildiz and J. L. M. Rupp, Oxygen exchange in dual-phase La_0.65Sr_0.35MnO₃-CeO₂ composites for solar thermochemical fuel production, ACS Appl. Mater. Interfaces, 2020, 12, 32622–32632 CrossRef CAS PubMed .
26. Q. Wang, Y. Xuan, K. Gao, C. Sun, Y. Gao, J. Liu, S. Chang and X. Liu, High-entropy perovskite oxides for direct solar-driven thermochemical CO₂ splitting, Ceram. Int., 2023, 50, 1564–1573 CrossRef .
27. H. Kildahl, Z. Li, H. Cao, P. Slater and Y. Ding, Carbon dioxide decomposition through gas exchange in barium calcium iron niobates, Catal. Today, 2021, 364, 211–219 CrossRef CAS .
28. H. Kildahl, L. Wang, L. Tong, H. Cao and Y. Ding, Industrial carbon monoxide production by thermochemical CO₂ splitting – A techno-economic assessment, J. CO₂ Util., 2022, 65, 102181 CrossRef CAS .
29. A. A. Emery, J. E. Saal, S. Kirklin, V. I. Hegde and C. Wolverton, High-throughput computational screening of perovskites for thermochemical water splitting applications, Chem. Mater., 2016, 28, 5621–5634 CrossRef CAS .
30. X. Jiang and W.-J. Yin, High-throughput computational screening of oxide double perovskites for optoelectronic and photocatalysis applications, J. Energy Chem., 2021, 57, 351–358 CrossRef CAS .
31. M. Zhong, K. Tran, Y. Min, C. Wang, Z. Wang, C. T. Dinh, P. De Luna, Z. Yu, A. S. Rasouli, P. Brodersen, S. Sun, O. Voznyy, C. S. Tan, M. Askerka, F. Che, M. Liu, A. Seifitokaldani, Y. Pang, S. C. Lo, A. Ip, Z. Ulissi and E. H. Sargent, Accelerated discovery of CO₂ electrocatalysts using active machine learning, Nature, 2020, 581, 178–183 CrossRef CAS PubMed .
32. L. H. Mou, T. Han, P. E. S. Smith, E. Sharman and J. Jiang, Machine Learning Descriptors for Data-Driven Catalysis Study, Adv. Sci., 2023, 10, e2301020 CrossRef PubMed .
33. X. Wang, Y. Gao, E. Krzystowczyk, S. Iftikhar, J. Dou, R. Cai, H. Wang, C. Ruan, S. Ye and F. Li, High-throughput oxygen chemical potential engineering of perovskite oxides for chemical looping applications, Energy Environ. Sci., 2022, 15, 1512 RSC .
34. T. N. Nguyen, T. T. P. Nhat, K. Takimoto, A. Thakur, S. Nishimura, J. Ohyama, I. Miyazato, L. Takahashi, J. Fujima, K. Takahashi and T. Taniike, High-throughput experimentation and catalyst informatics for oxidative coupling of methane, ACS Catal., 2019, 10, 921–932 CrossRef .
35. X. Wang, E. Krzystowczyk, J. Dou and F. Li, Net electronic charge as an effective electronic descriptor for oxygen release and transport properties of SrFeO₃-based oxygen sorbents, Chem. Mater., 2021, 33, 2446–2456 CrossRef CAS .
36. C. J. Bartel, C. Sutton, B. R. Goldsmith, R. Ouyang, C. B. Musgrave, L. M. Ghiringhelli and M. Scheffler, New tolerance factor to predict the stability of perovskite oxides and halides, Sci. Adv., 2019, 5, eaav0693 CrossRef CAS PubMed .
37. A. Mishra, T. Li, F. Li and E. E. Santiso, Oxygen vacancy creation energy in Mn-containing perovskites: an effective indicator for chemical looping with oxygen uncoupling, Chem. Mater., 2018, 31, 689–698 CrossRef .
38. D. Maiti, Y. A. Daza, M. M. Yung, J. N. Kuhn and V. R. Bhethanabotla, Oxygen vacancy formation characteristics in the bulk and across different surface terminations of La_(1−x)Sr_xFe_(1−y)Co_yO_(3−δ) perovskite oxides for CO₂ conversion, J. Mater. Chem. A, 2016, 4, 5137–5148 RSC .
39. E. Krzystowczyk, X. Wang, J. Dou, V. Haribal and F. Li, Substituted SrFeO₃ as robust oxygen sorbents for thermochemical air separation: correlating redox performance with compositional and structural properties, Phys. Chem. Chem. Phys., 2020, 22, 8924–8932 RSC .
40. X. He, K. Zhao and X. Chu, AutoML: A survey of the state-of-the-art, Knowl. Base Syst., 2021, 212, 106622 CrossRef .
41. M.-A. Zoeller and M. F. Huber, Benchmark and survey of automated machine learning frameworks, J. Artif. Intell. Res., 2021, 70, 409–472 CrossRef .
42. K. Gao, X. Liu, Q. Wang, Z. Jiang, C. Tian, N. Sun and Y. Xuan, Remarkable solar thermochemical CO₂ splitting performances based on Ce- and Al-doped SrMnO₃ perovskites, Sustain. Energy Fuels, 2023, 7, 1027–1040 RSC .
43. L. Wang, T. Ma, S. Dai, T. Ren, Z. Chang, L. Dou, M. Fu and X. Li, Experimental study on the high performance of Zr doped LaCoO₃ for solar thermochemical CO production, Chem. Eng. J., 2020, 389, 124426 CrossRef CAS .
44. C. N. R. Rao and S. Dey, Generation of H₂ and CO by solar thermochemical splitting of H₂O and CO₂ by employing metal oxides, J. Solid State Chem., 2016, 242, 107–115 CrossRef CAS .
45. K. Gao, X. Liu, Z. Jiang, H. Zheng, C. Song, X. Wang, C. Tian, C. Dang, N. Sun and Y. Xuan, Direct solar thermochemical CO₂ splitting based on Ca- and Al- doped SmMnO₃ perovskites: Ultrahigh CO yield within small temperature swing, Renew. Energy, 2022, 194, 482–494 CrossRef CAS .
46. K. Gao, X. Liu, T. Wang, Z. Zhu, P. Li, H. Zheng, C. Song, Y. Xuan, Y. Li and Y. Ding, Sr-doped SmMnO₃ perovskites for high-performance near-isothermal solar thermochemical CO₂-to-fuel conversion, Sustain. Energy Fuels, 2021, 5, 4295–4310 RSC .
47. G. Kresse and D. Joubert, From ultrasoft pseudopotentials to the projector augmented-wave method, Phys. Rev. B: Condens. Matter Mater. Phys., 1999, 59, 1758–1775 CrossRef CAS .
48. P. E. Blöchl, C. J. Först and J. Schimpl, Projector augmented wave method:ab initio molecular dynamics with full wave functions, Bull. Mater. Sci., 2003, 26, 33–41 CrossRef .
49. J. P. Perdew, K. Burke and M. Ernzerhof, Generalized gradient approximation made simple, Phys. Rev. Lett., 1997, 78, 1396 CrossRef CAS .
50. A. van de Walle, P. Tiwary, M. de Jong, D. L. Olmsted, M. Asta, A. Dick, D. Shin, Y. Wang, L. Q. Chen and Z. K. Liu, Efficient stochastic generation of special quasirandom structures, Calphad, 2013, 42, 13–18 CrossRef CAS .
51. V. V. Anisimov, J. Zaanen and O. K. Andersen, Band theory and Mott insulators: Hubbard U instead of Stoner I, Phys. Rev. B: Condens. Matter Mater. Phys., 1991, 44, 943–954 CrossRef CAS PubMed .
52. A. Togo and I. Tanaka, First principles phonon calculations in materials science, Scr. Mater., 2015, 108, 1–5 CrossRef CAS .
53. B. Bulfin, J. Vieten, D. E. Starr, A. Azarpira, C. Zachäus, M. Hävecker, K. Skorupska, M. Schmücker, M. Roeb and C. Sattler, Redox chemistry of CaMnO₃ and Ca_0.8Sr_0.2MnO₃ oxygen storage perovskites, J. Mater. Chem. A, 2017, 5, 7912–7919 RSC .
54. N. Erickson, J. Mueller, A. Shirkov, H. Zhang, P. Larroy, M. Li and A. Smola, AutoGluon-Tabular: robust and accurate AutoML for structured data, arXiv, 2020, preprint, arXiv:2003.06505,  DOI:10.48550/arXiv.2003.06505.

This journal is © The Royal Society of Chemistry 2025