CHEMICAL SCIENCE

Descriptors and machine learning enable effective prediction of cathodic activity in solid oxide fuel cells


On September 5, 2022, Academician Xie Heping of Shenzhen University and his doctoral student Zhai Shuo were the corresponding and first authors, respectively, and Professor Ni Meng of the Hong Kong Polytechnic University and Professor Shao Zongping of Nanjing University of Technology published an article in the journal Nature Energy entitled “A combined ionic Lewis acid descriptor and machine-learning approach to prediction of” as co-corresponding authors efficient oxygen reduction electrodes for ceramic fuel cells” research results.

Combining machine learning, theoretical calculations and ceramic solid oxide development, the study developed an experimentally validated machine learning screening technique for cathode materials, enabling rapid and efficient screening of highly active solid oxide fuel cell cathode materials from large perovskite components.

Background

Realizing the clean and efficient utilization of coal and promoting the coal revolution are of great strategic significance to the reform of China’s energy structure! At present, modern coal-fired power plants are limited by the Carnot cycle, and the coal consumption per unit of power generation remains high, and it is difficult to solve the technical bottleneck of the large amount of CO2 emissions inherent in coal power generation, and it is impossible to truly achieve clean utilization of coal. The “near-zero carbon emission direct coal fuel cell (DCFC) power generation technology” (CN114284533A), which was first proposed by Academician Xie Heping’s team and is currently tackling the problem, can break the limitation of the Carnot cycle, not through combustion, but directly convert the chemical energy of modified coal into electrical energy through the electrochemical oxidation process, and at the same time realize the secondary energy utilization of CO2 in situ in the system. The technical system is environmentally friendly with high energy conversion efficiency and near-zero carbon emissions. Among them, DCFC is based on a solid oxide fuel cell, and its cathode provides the oxygen ions required for carbon oxidation reaction, and its intrinsic activity has a decisive effect on the kinetic reaction rate of the oxygen reduction reaction. However, traditional material design, characterization, and testing rely on inefficient trial-and-error processes that often require lengthy research cycles.

Obtaining potential information hidden behind existing data sets through data-driven methods and establishing material development trends are hot topics of research in recent years. For example, machine learning (ML) techniques, with their powerful ability to recombine information structures and support multidimensional features (referred to as descriptors in this study), are widely used in material informatics. After decades of development, a large amount of experimental data on the cathodes of solid oxide fuel cells in the literature is worth collecting, collating, and serving as the cornerstone for deriving new cathode candidate materials. To build accurate ML models for fast and efficient screening of highly active solid oxide fuel cell cathode materials, high-quality data sets, accurate perovskite oxide descriptors, and appropriate regression models are required. However, there is a lack of representative physical descriptors that accurately reflect the orr mechanism at high temperatures. To fill this gap, it is urgent to find an effective physical descriptor and identify a reliable regression model for data mining and refactoring.

Based on the above thinking, Academician Xie Heping’s team carried out research on the screening of highly active cathode materials for solid oxide fuel cells.

Graphic and text analysis

The research paper introduced as descriptors the Louis acidic intensity (ISA), which is strongly correlated with the reaction rate of perovskite oxide ORR kinetics at high temperatures, and verified the validity of eight different regression models. An experimentally validated machine learning screening technique for cathode materials has been developed to enable rapid and efficient screening of highly active cathode materials from large perovskite components. Experimental characterization and density functional theory (DFT) calculations elucidated the mechanism of enhancing the eigen-active mechanism of the Acidic Regulation Strategy of perovskite oxide Louis, and revealed the mechanism of electron pair shift caused by the polarization distribution of Louis acidity in the A and B position ions, thereby reducing the mechanism of oxygen vacancy generation energy and migration energy barrier.

Figure 1: Overall workflow diagram of cathode material development based on machine learning.

Regression model evaluation and analysis

In the literature of material informatics based on machine learning, linear regression methods are widely used because of their advantages of efficiency and intuitiveness. However, it is less capable of modeling nonlinear relationships. This study conducted a comprehensive experiment with a variety of different regression methods, including four nonlinear regression methods: ordinary least squares (OLS), Lasso algorithm (Lasso), ridge regression (Ridge) and elastic networks (Elastic Nets), and four nonlinear regression methods: Support Vector Regression (SVR), Random Forest (RF), Gaussian Process Regression (GPR) and Artificial Neural Networks (ANN). To measure the performance of different regression models, this study uses mean squared error (MSE) as an evaluation index for training and test sets (Table 1). Among all the regression methods, the MSE values of the training set and the test set based on the ANN model were 0.009 and 0.013 Ω cm2, respectively, which achieved the best fitting effect among all the regression models.

Table 1: Performance of different regression models based on training and test sets

Sensitivity analysis is used to assess the importance of individual descriptors in model building to gain a deeper understanding of the “black box” that generates an ANN model. The 9 different ion descriptors include: ion Louis acidic intensity at positions A and B (AISA, BISA), ion electronegativity at positions A and B (AIEN, BIEN), ion radius at positions A and B (RA, RB), ion energy at positions A and B (AIE, BIE), and tolerance factor

Figure 2: Model evaluation versus descriptor importance analysis

Cathode material screening and validation

The study screened four perovskite cathodes from 6871 different perovskite oxides automatically generated and predicted by the machine and successfully synthesized them. In particular, SCCN (Sr0.9Cs0.1Co0.9Nb0.1O3), BSCCFM (Ba0.4Sr0.4Cs0.2Co0.6Fe0.3Mo0.1O3) at 700 °C with an area ratio of only 0.0101 and 0.0113 Ω cm2, which is close to the predictions of machine learning. In the three-dimensional visualization schematic, the intrinsic activity Lg (ASR) of the four cathode materials conforms to the approximate linear trend with the ISA descriptor. By analyzing and quantifying the electrochemical AC impedance spectra of the relaxation time distribution (DRT) model and the equivalent circuit model, the intermediate frequency resistance of the four perovskite oxide cathodes showed great differences and had significant thermal activation characteristics, that is, the surface oxygen transfer related process was the determining step of the ORR reaction kinetics.

Figure 3: Structure and electrochemical properties of cathode materials with high activity perovskites

Figure 4: A three-dimensional visualization of the correlation between the intrinsic ORR activity of perovskite cathode materials and the AISA+BISA and RA+RB functions

Figure 5: Electrochemical properties of scCAN, a perovskite cathode material

Morphology is associated with processes related to oxygen delivery

In this study, BSCCFM and cornerstone material Ba0.5Sr0.5Co0.7Fe0.3O3 (BSCF73) were used as examples to explore the influence of Louis acidic regulation strategies on the process related to oxygen transfer on the surface of cathode materials. The position of the two main peaks is almost the same, revealing that the lattice expansion effect caused by Cs+ and the lattice shrinkage effect caused by Mo6+ compensate for each other. The EIS atlas reflects a similar partial pressure dependence of both Rp∝PO2-m, andIt is a speed limiting step of the kinetic reaction rate, that is, the ORR process is affected by the concentration of oxygen vacancies on the surface. XPS spectra of TG and O1s confirm the higher oxygen vacancy concentrations in BSCCFM samples and can therefore be seen as signals with better ORR activity. The conductivity measured by BSCCFM at 550 – 900 °C air is approximately 120 S cm-1, which is nearly 4 times that of BSCF73 at the corresponding temperature. Interestingly, Cs+ and Mo6+ present a polarized distribution in the bar graph of isa, and it can be speculated that this polarity distribution on isa may alter the coordination environment of the active site of B site, promoting more oxygen vacancy generation.

Figure 6: STEM image of perovskite cathode material BSCCFM and STEM-EDX results

Figure 7: Characterization of oxygen transport-related processes

Molecular simulation analysis of DFT for the evolution of electronic structures

In order to clarify whether the regulation of Lewis acidity is related to optimized electron configuration, the research team conducted a DFT molecular simulation analysis based on the established three-dimensional model. After the polarized Cs+ and Mo6+ on the ISA enter the lattice, Co and Fe in BSCCFM-m exhibit reduced Bader charges due to their variable valence states, while Ba and Sr remain unchanged, reflecting their intrinsically active enhancement due to changes in the coordination environment of the B-site active site. The study thoroughly analyzed the different O loci in the two model configurations, with the lowest ΔEvac of BSCF-m and BSCCFM-m being 0.26 eV and 0.74 eV, respectively. In particular, the lowest ΔEvac in BSCCFM-m is located at the O1 locus near Mo and Cs, revealing the effectiveness of Lewis’s acidic regulation strategy. In addition, the BSCCFM-m migration barrier energy (O1 locus to O2 locus) is significantly lower than BSCF-m. Under the influence of Lewis acidic regulation, the O2P and Co3d orbits of BSCCFM-m are offset from BSCF-m to near EF, which strengthens the migration of the carrier guide band and promotes the improvement of conductivity.

Figure 8: DFT calculation of the evolution of electronic structures

Summary and outlook

The study successfully combined machine learning with the development of cathode materials for highly active solid oxide fuel cells. Compared with the high-throughput DFT calculation method, this method does not need to build a molecular model, and can predict the material properties only by training the regression model with molecular formula, breaking the technical barrier of low cathode material development efficiency. As a data-driven approach, dataQuantity and quality directly affect the accuracy of ML, and the future development of ML needs to accelerate the construction of material databases. In order to clarify the role of isa descriptors in the ORR process, this study obtains electronic structure information through DFT calculations. Thus, DFT calculations can complement the material properties that ML cannot provide, and the two are mutually reinforcing.

Based on the above method, the study achieves rapid and effective screening of highly active cathode materials from large perovskite components. Experimental characterization and DFT calculation elucidated the mechanism of enhancing the intrinsic activity of the perovskite oxide Louis acidicity regulation strategy, and revealed the polarization distribution of The Lewis acid in the A and B position ions caused by the shift of the electron pair, thereby reducing the mechanism of oxygen vacancy generation energy generation and migration energy barrier. This achievement provides a theoretical basis and technical support for the “near-zero carbon emission DCFC power generation technology” that Academician Xie Heping’s team is tackling. (Source: Science Network)

Related paper information:https://doi.org/10.1038/s41560-022-01098-3



Source link

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button