Construction and application of general descriptor for predicting catalyst activity by machine learning
By constructing a method that combines weighted atomic center symmetric function with machine learning, the problem of catalyst activity prediction across materials systems is solved, and efficient and accurate catalyst screening is achieved, which is suitable for multiple types of materials.
Patent Information
- Application Number
- CN202510454397.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-11
- Publication Date
- 2025-07-25
AI Technical Summary
The prior art cannot effectively predict catalyst activity across material systems, traditional methods have low computational efficiency and rely on a single material category, and machine learning models ignore microkinetic processes, resulting in large prediction errors.
Weighted atomic center symmetry function (wACSF), combined with machine learning and microkinetic modeling, calculate the radial and angular symmetry function values by extracting the inherent attribute data of the material, and multiplying it by atomic chemical environment characteristics, build a general descriptor, and apply an extreme gradient lift regression algorithm to make predictions.
It realizes efficient screening of metal alloys, metal oxides, perovskites and single-atom catalysts, improves prediction accuracy and versatility, simplifies feature engineering, and reduces computational complexity.
Smart Images
Figure CN120375967A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the cross - technical field of materials calculation and artificial intelligence, and specifically relates to the construction and application of a general descriptor for predicting catalyst activity by machine learning. Background Art
[0002] Catalytic science plays an irreplaceable role in fields such as energy conversion, chemical production, and environmental governance. However, the inherent challenges of precise design of catalyst active sites and mass transfer regulation at the surface and interface have always restricted the efficiency breakthrough of catalytic systems. Research on the electrocatalytic synthesis of hydrogen peroxide (H2O2) as a model system shows that by constructing an electron - proton synergistic channel and dynamic active sites, the target product can be synthesized under mild conditions, which provides a new idea for the design of a new generation of catalysts.
[0003] However, the development of highly efficient catalysts still faces many problems. Although the traditional atomic - centered symmetry function (ACSF) can characterize the local atomic environment, it does not consider the electronegativity and valence electron weight of coordinated atoms, resulting in limited mobility across material systems (such as metal oxides and single - atom catalysts). The literature shows that the prediction error (RMSE>0.54 eV) of ACSF for the adsorption energy of perovskites is significantly higher than that of the weighted improvement scheme. Although density functional theory (DFT) calculations can reveal key parameters such as adsorption free energy (ΔG OH* 、ΔG O* ), the calculation efficiency is low, making large - scale screening infeasible. In addition, existing machine - learning models mostly focus on a single material category (such as metal oxides), with complex feature engineering and reliance on the thermodynamic "limiting potential model", ignoring the micro - kinetic process, resulting in an overpotential prediction error>0.3 V, which shows a significant deviation from experimental data. And the existing technologies cannot cover multiple material categories such as metal alloys, perovskites, and single - atom catalysts at the same time.
[0004] Therefore, there is a need in this field to develop a general descriptor for predicting catalyst activity by machine learning, which can effectively solve the above problems. Summary of the Invention
[0005] The purpose of the present invention is to provide the construction and application of a general descriptor for predicting catalyst activity by machine learning, which combines the construction of an improved local atomic environment feature descriptor (i.e., weighted atomic - centered symmetry function, wACSF) with machine learning, micro - kinetic modeling, and computational high - throughput screening, realizing efficient screening of four material categories (i.e., metal alloys, metal oxides, perovskites, and single - atom catalysts) in a machine - learning (ML) model.
[0006] To achieve the above purpose, the present invention provides the construction of a general descriptor for predicting catalyst activity by machine learning, including the following steps:
[0007] Step S1: Extract the intrinsic property data of four materials in the database and calculate the radial and angular symmetry function values;
[0008] Step S2: Construct the atomic chemical environment characteristics by summing the electronegativity and valence electrons of the central atom of the catalyst;
[0009] Step S3: Multiply the calculated radial and angular symmetry function values by the atomic chemical environment characteristics to obtain the general descriptor.
[0010] Preferably, in step S1, the four materials include metal alloys, metal oxides, perovskites, and single-atom catalysts.
[0011] Preferably, in step S1, the intrinsic properties include electronegativity, valence electrons, first ionization energy, electron affinity, atomic number, and atomic radius.
[0012] Preferably, in step S3, the general descriptor is the weighted atomic center symmetry function, which is calculated by weighting the radial and angular symmetry functions with the electronegativity and valence electron numbers of the coordinated atoms;
[0013] Among them, the calculation formula is:
[0014] A i = G i × ΣNm i ;
[0015] In the formula, A i is the weighted symmetry function value; G i is the radial and angular symmetry function value; Nm i is the electronegativity;
[0016] B i = G i × ΣV i ;
[0017] In the formula, B i is the weighted symmetry function value; G i is the radial and angular symmetry function value; V i is the number of valence electrons.
[0018] The present invention also provides the application of the constructed general descriptor in machine learning prediction of catalyst activity.
[0019] Preferably, the machine learning prediction of catalyst activity includes the following steps:
[0020] Step S1: Using the adsorption free energy data of hydroxyl and oxygen as the target values, perform feature selection on the calculated data through Pearson correlation and recursive elimination;
[0021] Step S2: Conduct machine learning training and testing to obtain the optimal machine learning model; use the optimal machine learning model to predict the adsorption free energy.
[0022] Step S3: Use the microkinetic model to quickly identify the prediction results and screen out high-performance catalysts.
[0023] Preferably, the machine learning training and testing in Step S2 are specifically as follows:
[0024] a. Divide the adsorption free energy dataset into a training set and a test set at a ratio of 8:2, and optimize the hyperparameters by 10-fold cross-validation.
[0025] b. Compare the training models of different machine learning algorithms, and obtain the optimal machine learning model based on the coefficient of determination and the root mean square error.
[0026] c. Adopt the extreme gradient boosting regression algorithm and input the screened weighted atom-centered symmetry function descriptors and elemental inherent properties.
[0027] Preferably, the machine learning algorithm training models include random forest regression, kernel ridge regression, extreme gradient boosting regression, ridge regression, k-nearest neighbor, gradient boosting regression, support vector regression, and extra tree regression.
[0028] Preferably, in Step S3, catalyst screening is performed through the microkinetic volcano model; the microkinetic volcano model predicts the activity of the catalyst based on the adsorption free energy.
[0029] The screening criteria are 1.41 eV ≤ ΔG OH* < 2.4 eV and ΔG O* ≥ 3.52 eV; where ΔG OH* is the hydroxyl group, and ΔG O* is oxygen.
[0030] Preferably, in Step S3, the feature importance is analyzed through the Shapley additive explanation value to explain the prediction results of the machine learning model; among them, the electronegativity and the weighted symmetry function are the key influencing factors.
[0031] The present invention adopts the construction and application of the above-mentioned general descriptors for machine learning prediction of catalyst activity, and the beneficial effects are as follows:
[0032] (1) A general descriptor for predicting catalyst activity in machine learning constructed by the present invention integrates local geometric features, chemical environment weights, and cross-scale correlation characteristics. Taking four types of materials, namely metal alloys, metal oxides, perovskites, and single-atom catalysts, as the research objects, a weighted atom-centered symmetry function (wACSF) descriptor is constructed by combining the atom-centered symmetry function with chemical environment characteristics (electronegativity and valence electron numbers of the central atom and coordinated atoms), which can accurately reflect atomic environment information; realizing the unified characterization of multiple types of materials such as metal alloys, metal oxides, perovskites, and single-atom catalysts (SACs), thereby efficiently screening electrochemical catalysts;
[0033] (2) From the perspective of the relationship between adsorption characteristics and atomic environment, the present invention comprehensively considers geometric and chemical features, and uses the extreme gradient boosting regression (XGBR) algorithm in machine learning to fit the quantitative relationship between the adsorption free energy (ΔG OH* and ΔG O* ) and the descriptor for different material types and different catalyst surfaces; the relevant parameters in this relationship reflect the intrinsic characteristics of the catalyst, and according to the obtained quantitative relationship, the adsorption free energy of a new catalyst can be predicted;
[0034] (3) The general descriptor in the present invention eliminates the limitations of traditional methods on a single material type and the drawbacks of complex feature engineering, making the prediction results mainly reflect the activity of the catalyst itself, rather than being affected by the material type and feature engineering method, greatly improving the accuracy, generality, and screening efficiency of the prediction.
[0035] The technical solution of the present invention will be further described in detail below through the accompanying drawings and embodiments. Brief Description of the Drawings
[0036] Figure 1 is the flow chart of the general descriptor in the embodiment of the present invention for predicting catalyst activity in machine learning;
[0037] Figure 2 is the Pearson correlation matrix heat map drawn after the feature engineering in the embodiment of the present invention after removing highly correlated features (|r|>0.8); among them, a is the heat map of the Pearson correlation coefficient of the descriptor predicting ΔG OH* ; b is the heat map of the Pearson correlation coefficient of the descriptor predicting ΔG O* ;
[0038] Figure 3 is the machine learning regression performance evaluation chart for the adsorption free energy prediction in the embodiment of the present invention; among them, a is the R OH* for predicting ΔG 2 ; b is the RMSE for predicting ΔG OH* ; c is the R for predicting ΔG O* 2 ; d is the predicted ΔG O* with RMSE; e is the ΔG calculated by DFT in the training set OH* and the ΔG predicted by the ML model OH* for comparison; f is the ΔG calculated by DFT in the test set OH* and the ΔG predicted by the ML model OH* for comparison; g is the ΔG calculated by DFT in the training set O* and the ΔG predicted by the ML model O* for comparison; h is the ΔG calculated by DFT in the test set O* and the ΔG predicted by the ML model O* for comparison;
[0039] Figure 4 This is the SHAP analysis diagram of the feature importance of various materials based on the extreme gradient boosting regression model in the embodiments of the present invention; among them, a is the feature importance based on the average SHAP value of the ΔG OH* model; b is the SHAP sensitivity analysis of a single feature of the ΔG OH* model, indicating the influence of each feature on the model output; c is the feature importance based on the average SHAP value of the ΔG O* model; d is the SHAP sensitivity analysis of a single feature of the ΔG O* model, indicating the influence of each feature on the model output;
[0040] Figure 5 This is the microkinetic and thermodynamic volcano model and catalyst screening diagram of 2e-WOR in Embodiment 2 of the present invention; among them, a is the microkinetic volcano model of 2e - WOR simulated at 2.4V versus the reversible hydrogen electrode (RHE); b is the microkinetic and thermodynamic volcano model, illustrating the theoretical limiting potential and simulated exchange current density of various catalysts; c is the scaling relationship between ΔG OH* and ΔG O* on the surfaces of 962 catalysts; d is the workflow of the data screening process. Detailed implementation manners
[0041] The technical solutions of the present invention will be further described below with reference to the drawings and embodiments.
[0042] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those of ordinary skill in the art to which the present invention belongs.
[0043] Embodiment
[0044] As Figure 1As shown, a general descriptor is used for machine learning to predict catalyst activity, including the following steps:
[0045] Step S1: Extract the inherent property data of four materials in the database and calculate the radial and angular symmetry function values (G i ).
[0046] The four materials include metal alloys, metal oxides, perovskites, and single-atom catalysts (SACs).
[0047] The inherent properties include electronegativity, valence electrons, first ionization energy, electron affinity, atomic number, and atomic radius.
[0048] Step S2: Construct the atomic chemical environment feature by summing the electronegativity (Nm) and the number of valence electrons (V) of the central atom of the catalyst.
[0049] Step S3: Multiply the calculated radial and angular symmetry function values (Gi) by the atomic chemical environment feature to obtain the general descriptor.
[0050] The general descriptor is the weighted atomic-centered symmetry function (wACSF), which is calculated by weighting the radial and angular symmetry functions with the electronegativity and the number of valence electrons of the coordinated atoms.
[0051] Among them, the calculation formula is:
[0052] A i = G i × ΣNm i .
[0053] In the formula, A i is the weighted symmetry function value. G i is the radial and angular symmetry function value. Nm i is the electronegativity.
[0054] B i = G i × ΣV i .
[0055] In the formula, B i is the weighted symmetry function value. G i is the radial and angular symmetry function value. V i is the number of valence electrons.
[0056] Step S4: Using the adsorption free energy data of hydroxyl (ΔG OH* ) and oxygen (ΔG O* ) as the target values, perform feature selection on the calculated data through Pearson correlation and recursive elimination.
[0057] Step S5: Machine learning training and testing to obtain the optimal machine learning model. Use the optimal machine learning model to predict the adsorption free energy.
[0058] The machine learning training and testing are specifically as follows:
[0059] a. Divide the adsorption free energy dataset into a training set and a test set at a ratio of 8:2, and optimize the hyperparameters by 10-fold cross-validation.
[0060] b. Compare the training models of different machine learning algorithms, and obtain the optimal machine learning model based on the coefficient of determination (R 2 ) and the root mean square error (RMSE).
[0061] c. Adopt the Extreme Gradient Boosting Regression (XGBR) algorithm, and input the screened weighted atomic center symmetry function (wACSF) descriptors and elemental intrinsic properties.
[0062] The machine learning algorithm training models include Random Forest Regression (RFR), Kernel Ridge Regression (KRR), Extreme Gradient Boosting Regression (XGBR), Ridge Regression (RR), k-Nearest Neighbor (KNN), Gradient Boosting Regression (GBR), Support Vector Regression (SVR), and Extra Trees Regression (ETR).
[0063] Step S6: Use the microkinetic model to quickly identify the prediction results and screen out high-performance catalysts.
[0064] Perform catalyst screening through the microkinetic volcano model, and the microkinetic volcano model predicts the activity of the catalyst based on the adsorption free energy.
[0065] The screening criteria are 1.41 eV ≤ ΔG OH* < 2.4 eV and ΔG O* ≥ 3.52 eV. Among them, ΔG OH* is the hydroxyl group, and ΔG O* is oxygen.
[0066] Among them, analyze the feature importance through the Shapley Additive Feature Explanation method (SHAP) to explain the prediction results of the machine learning model. Among them, electronegativity and weighted symmetry function are the key influencing factors.
[0067] Specifically in this embodiment, download the adsorption free energy dataset of 962 materials from the DigCat database, including 426 metal alloys, 160 metal oxides (excluding perovskites), 250 perovskites, and 126 single-atom catalysts. Use custom python code to parse the CIF file and calculate the weighted atomic center symmetry function (wACSF) of each active site, including the radial symmetry function and the angular symmetry function. Concatenate the wACSF values with the elemental intrinsic properties to generate the initial feature matrix.
[0068] Through Pearson correlation analysis and recursive feature elimination (RFE), 21 key descriptors were determined for ΔG OH* and 19 descriptors were determined for ΔG O* Using the XGBoost regression algorithm (XGBR), the training set (80%) and test set (20%) were divided, and the hyperparameters were optimized by 10-fold cross-validation. For the test set, the ΔG OH* prediction R 2 = 0.84 and RMSE = 0.52 eV. For ΔG O* the prediction R 2 = 0.91 and RMSE = 0.65 eV. As Figure 2 shown, the darker colors correspond to a more positive Pearson correlation coefficient between the two features of the corresponding row and column.
[0069] A correlation value of 1 indicates the highest correlation between two descriptors. The internal relationships between the parameters within wACSF may lead to high correlations, which may result in increased correlations between the resulting features. The design of symmetric functions may lead to features with similar properties. However, these features can still provide unique information that can improve the performance of the model, which is crucial for improving the prediction accuracy of the adsorption free energy of the model.
[0070] To address the potential impact of highly correlated features on model complexity and overfitting, the feature set was further refined by removing descriptors with a Pearson correlation coefficient greater than 0.8, retaining descriptors with strong independence and complementary information, and the model was retrained and compared with the original model. As shown in Table 1, the reduced feature set led to a slight decrease in prediction accuracy (R 2 and RMSE). The retained descriptors continued to provide sufficient prediction for accurate information while simplifying the model and reducing the risk of overfitting.
[0071] Table 1
[0072] Type <![CDATA[R 2 (ΔG OH* )]]> <![CDATA[RMSE(eV)(ΔG OH* )]]> <![CDATA[R 2 (ΔG O* )]]> <![CDATA[RMSE(eV)(ΔG OH* )]]> Before refinement 0.93 0.32 0.97 0.36 After refinement 0.89 0.41 0.90 0.39
[0073] As Figure 3 shown, compared with the models trained using RFR, XGBR, GBR, and KRR, the models trained using ETR, SVR, KNN, and RR showed significantly lower prediction performance. In addition, the predictions of the ML models trained using XGBR and GBR were in close agreement with the target values, indicating that these models have superior performance. Based on the training performance and prediction performance, XGBR was determined as the optimal method for predicting the adsorption free energy. Therefore, in this embodiment, an ML model for ΔG OH* and ΔG O* was finally developed using the XGBR algorithm.
[0074] Among them, Figure 3 in e- Figure 3 Figure h shows the comparison between the machine learning predicted adsorption free energy of the optimal XGBR model and the adsorption free energy calculated by density functional theory (DFT). For ΔG OH* and ΔG O* , in the training set and the test set, the R 2 between DFT and machine learning prediction are 0.93 and 0.84 respectively. The root mean square error (RMSE) is used as the loss function to evaluate the optimization error, which is 0.32 eV for the training set and 0.52 eV for the test set ( Figure 3 in e- Figure 3 in f and ΔG OH* ), for R 2 , the values in the training set are 0.97 and 0.91 in the test set. The corresponding root mean square error values are 0.36 eV for the training set and 0.65 eV for the test set ( Figure 3 in g- Figure 3 in h).
[0075] Although the RMSE of the model used in this example in the test set is slightly higher than the lowest RMSE achieved by using intrinsic and electronic / geometric descriptors in existing studies, it has significant advantages. This example provides a general descriptor applicable to various materials in a machine learning model, which is not restricted by the material type during one training and testing process. It not only overcomes the limitations of traditional methods but also provides a more flexible tool through a simplified and automated process to screen various types of materials simultaneously.
[0076] As Figure 4 shown, to understand the influence of features on the predicted adsorption free energy, analyze the importance of each feature and its contribution to the prediction of ΔG OH* and ΔG O* , and represent it with SHAP values. As Figure 4 in b and Figure 4 in d shown, the horizontal axis represents the SHAP value, and the vertical axis represents the intrinsic property and the weighted symmetric descriptor. The more positive the SHAP value, the greater the positive contribution of the feature to the adsorption free energy. The wider the distribution of the SHAP value, the higher the sensitivity of the corresponding feature to the target adsorption free energy. For ΔG OH* , the three descriptors with the widest distribution of SHAP values are Nm (electronegativity of the central atom), B6 (radial symmetric function weighted by coordination valence electrons), and B21 (angular symmetric function weighted by coordination valence electrons). For ΔG O* , the three descriptors with the widest distribution of SHAP values are Nm (electronegativity of the central atom), B19 (angular symmetric function weighted by coordination valence electrons), and B4 (radial symmetric function weighted by coordination valence electrons).
[0077] ΔG OH* of B6 and B21 and ΔG O* The B4 and B19 of O* reflect the interaction strength and electron distribution between the central atom and its coordinating atoms in the radial and angular directions. This property provides necessary information for describing the complex angle-dependent interactions during the adsorption process and is crucial for simulating the local chemical perspective of atoms.
[0078] As Figure 5 shown, a microkinetic model of 2e-WOR established based on the general descriptor of this embodiment predicts the reaction exchange current density (j0) as a function of the electrode potential and ΔG OH* According to the relationship between ΔG OH* and j0, a volcano plot can be obtained, which is important for evaluating the catalytic activity of electrocatalysts. Figure 5 In Figure 5 , a represents that the microkinetic volcano simulated in this embodiment is in good agreement with the experimental observations. Figure 5 In Figure 5 , b represents the current density values calculated based on the microkinetic model, describing a dual-model plot of the microkinetic volcano of metal oxide catalysts and 24 thermodynamic volcanoes of existing studies. At a potential of 2.4 V relative to the reversible hydrogen electrode (RHE), the peak of the thermodynamic volcano plot appears at ΔG OH* = 1.76 eV, while the peak of the kinetic volcano plot is at 1.77 eV. In the left branch of the thermodynamic volcano plot, the rate-determining step is the oxidation of OH* to H2O2, and in the right branch, the rate-determining step is the activation of H2O to OH*. Figure 5 In Figure 5 , c shows the proportional relationship between ΔG OH* and ΔG O* (ΔG O* = 2ΔG OH* + 0.71). Therefore, considering the binding free energies of OH and O, a system with high selectivity for the production of hydrogen peroxide (H2O2) through the 2e-water oxidation reaction (2e-WOR) should meet the following criteria: ΔG O* ≥ 3.52 eV, and 1.41 eV ≤ ΔG OH* < 2.4 eV. Figure 5 In Figure 5 , d represents the workflow of the high-efficiency catalyst screening process.
[0079] Initially, 962 catalysts in the DigCat database were preliminarily screened, and 145 candidate catalysts meeting the criteria were selected. The OH adsorption energies of these catalysts are in the optimal range of 1.41 to 2.4 electron volts, and their ΔG OH* values are consistent with the conditions required for the high-selectivity production of H2O2. Applying a more stringent criterion, i.e., log(J) ≥ 0 mA / cm 2Afterwards, the number of candidate catalysts was reduced to 49. Subsequently, further microkinetic modeling analysis was performed on these 145 catalysts.
[0080] Therefore, the present invention adopts the construction and application of the above-mentioned general descriptors for machine learning prediction of catalyst activity, combines the construction of improved local atomic environment feature descriptors (i.e., weighted atomic center symmetry function, wACSF) with machine learning, microkinetic modeling, and computational high-throughput screening, and realizes the efficient screening of four material categories (i.e., metal alloys, metal oxides, perovskites, and single-atom catalysts) in the machine learning (ML) model.
[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. Construction of a general descriptor for predicting catalyst activity in machine learning, characterized in that, It includes the following steps: Step S1: Extract the intrinsic property data of four materials in the database and calculate the radial and angular symmetry function values; Step S2: Construct the atomic chemical environment characteristics by summing the electronegativity and valence electron number of the central atom of the catalyst; Step S3: Multiply the calculated radial and angular symmetry function values by the atomic chemical environment characteristics to obtain the general descriptor.
2. Construction of the general descriptor for predicting catalyst activity in machine learning, characterized in that: In step S1, the four materials include metal alloys, metal oxides, perovskites, and single-atom catalysts.
3. The construction of the general descriptor for predicting the activity of a catalyst in machine learning according to claim 1, characterized in that: In step S1, the intrinsic properties include electronegativity, valence electrons, first ionization energy, electron affinity, atomic number, and atomic radius.
4. The construction of the general descriptor for predicting catalyst activity in machine learning according to claim 1, characterized in that: In step S3, the general descriptor is the weighted atomic center symmetry function, which is calculated by weighting the radial and angular symmetry functions with the electronegativity and valence electron number of the coordinated atoms; Among them, the calculation formula is: A i = G i × ΣNm i ; Where A i is the weighted symmetric function value; G i is the radial and angular symmetric function value; Nm i is the electronegativity; B i = G i ×ΣV i ; where B i is the weighted symmetric function value; G i is the radial and angular symmetric function value; V i is the number of valence electrons.
5. The application of the general descriptor constructed as described in any one of claims 1-4 in machine learning for predicting the activity of catalysts.
6. Use of the general descriptor for predicting catalyst activity in machine learning according to claim 5, characterized in that, It includes the following steps: Step S1: Using the adsorption free energy data of hydroxyl and oxygen as the target values, perform feature selection on the calculation data through Pearson correlation and recursive elimination; Step S2: Conduct machine learning training and testing to obtain the optimal machine learning model; use the optimal machine learning model to predict the adsorption free energy; Step S3: Use the microkinetic model to quickly identify the prediction results and screen out high-performance catalysts.
7. Use of the general descriptor for predicting catalyst activity in machine learning according to claim 6, characterized in that, In step S2, the machine learning training and testing are specifically as follows: a. Divide the adsorption free energy data set into a training set and a testing set at a ratio of 8:2, and optimize the hyperparameters through 10-fold cross-validation; b. Compare the training models of different machine learning algorithms, and obtain the optimal machine learning model based on the coefficient of determination and root mean square error; c. Adopt the extreme gradient boosting regression algorithm, and input the screened weighted atomic center symmetry function descriptor and element intrinsic properties.
8. Use of the general descriptor for predicting catalyst activity in machine learning according to claim 7, characterized in that: The machine learning algorithm training models include random forest regression, kernel ridge regression, extreme gradient boosting regression, ridge regression, k-nearest neighbor, gradient boosting regression, support vector regression, and extra tree regression.
9. Use of the general descriptor for predicting catalyst activity in machine learning according to claim 6, characterized in that: In step S3, catalyst screening is carried out through the microkinetic volcano model; the microkinetic volcano model predicts the activity of the catalyst based on the adsorption free energy; The screening criteria are 1.41 eV ≤ ΔG OH* < 2.4 eV and ΔG O* ≥ 3.52 eV; where, ΔG OH* is a hydroxyl group, and ΔG O* is oxygen.
10. Use of the general descriptor for predicting catalyst activity in machine learning according to claim 6, characterized in that: In step S3, the Shapley additive feature explanation method is used to analyze the feature importance for explaining the prediction results of the machine learning model; among them, electronegativity and the weighted symmetry function are the key influencing factors.