A method for predicting copper alloy catalysts for carbon dioxide electroreduction

Abstract

A method for predicting copper alloy catalysts for carbon dioxide electroreduction belongs to the field of carbon dioxide electrochemical reduction. By applying density functional theory calculations and machine learning, the problems of low efficiency and poor selectivity are overcome. This method optimizes the surface structure of different types of CuM alloys and applies density functional theory to calculate the adsorption energies of key intermediates (CO, HCOO, COOH, H) in the CO2 reduction reaction on various surfaces. In order to reduce the spatial dimension of the characteristics, five characteristic parameters of the materials are selected, including work function (W), atomic number (AN), interplanar spacing (d), electronegativity (EN) and local electronegativity (χ i ), and through machine learning training, a gradient boosting regression (GBR) model with good prediction performance was obtained. The training results were close to the prediction performance of the model containing 13 features. This method not only quickly predicts the copper alloy catalyst for carbon dioxide electroreduction, but also provides ideas for the design of other catalysts.

Description

Technical Field

[0001] The present invention belongs to the field of catalytic material research and development, and in particular, relates to a method for predicting copper alloy catalysts for carbon dioxide electroreduction. Background Art

[0002] In recent years, in order to meet the needs of industrial development, people have consumed a large amount of fossil fuels, and the concentration of carbon dioxide in the atmosphere has increased, causing many environmental problems. In order to reduce the concentration of carbon dioxide in the atmosphere, the carbon dioxide emitted can be captured and the electricity generated by renewable energy such as wind energy, hydropower, and tidal energy can be used to convert the generated carbon dioxide into clean energy such as fuel or common chemicals. However, high reaction overpotential and poor product selectivity are the main problems of carbon dioxide electroreduction. Due to the stable chemical properties of carbon dioxide molecules, the bond energy of carbon-oxygen double bonds (C=O) is much greater than the bond energy of carbon-oxygen single bonds (CO) and carbon-hydrogen single bonds (CH), so carbon dioxide molecules are difficult to convert into chemical substances containing the above single bonds. Although copper electrodes can obtain deep reduction products, the generation of the main products is accompanied by a large number of by-products, which makes it difficult to separate the subsequent products. The catalytic performance is far from meeting the requirements of industrial applications, so accelerating the development of catalysts and achieving efficient and highly selective catalysts are currently difficult problems that people are trying to overcome.

[0003] Since different metal catalysts have different catalytic activities for CO2 reduction, alloy materials often show better electrochemical activity and selectivity for products. For metal catalysts, alloying is an effective method to improve reaction activity and selectivity. It can well adjust the adsorption capacity of the catalyst for intermediates, thereby guiding the CO2 reduction reaction kinetics to be enhanced, and can obtain ideal products and improve the selectivity of polyols. However, due to the wide variety of copper alloys, experimental methods to explore the catalytic performance of each alloy are not only time-consuming, but also costly. Therefore, there is an urgent need for a method to quickly and efficiently find catalysts.

[0004] The mechanism of CO2 reduction reaction can be studied through density functional theory calculations, but accurate theoretical calculations consume a lot of machine time. Therefore, we combine traditional theoretical calculations with machine learning (ML) to establish targeted or generalized models with good performance, which is expected to solve the two long-standing problems of finding high-quality materials and predicting reaction performance, and significantly reduce the computational cost. Summary of the invention

[0005] In view of the fact that there are many types of copper alloys and it takes a lot of time and cost to find an ideal catalyst using traditional trial and error methods, the present invention provides a method for predicting CO2 reduction Cu alloy catalysts by combining density functional theory calculations and machine learning.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is: a method for predicting carbon dioxide electroreduction copper alloy catalyst, the method comprising the following steps:

[0007] Step 1. Apply the density functional theory method to optimize the structure of the unit cells of CuAl, CuZn, CuSn, CuGa, CuIn, CuAu, CuPd, CuPt, CuNi, CuFe, CuIr, and CuRh alloys, and then use the optimized unit cells to construct surface structures with different crystal plane indices, which are (111), (110), (100), and (211), respectively. The energy convergence standard in the structural optimization process is 10 -5 eV, the force convergence criterion is And perform non-self-consistent calculations to obtain the work function.

[0008] Step 2. Since CO and COOH are key intermediates in the process of CO2 reduction to CO, HCOO is the key intermediate for the generation of HCOOH, and H is the key intermediate for the hydrogen evolution reaction, the adsorption energy of these four intermediates is calculated. Each intermediate is placed at the top site, bridge site, or hollow site on the surface of the copper alloy, and the structure is optimized. The adsorption energy of each intermediate at the corresponding site is calculated. The adsorption energy calculation formula is E adsorption =E absorbate / surface -(E surface +E adsorbate ), where E absorbate / surface is the energy of the catalyst surface and adsorbate as a whole, E surface is the energy of the clean surface, E adsorbate is the energy of a single adsorbate.

[0009] Step 3. The 13 selected features include work function (W), atomic number (AN), period number (P), group number (G), element ratio (rat), melting enthalpy (Δ fus H), ionic potential (IE), melting point (mp), boiling point (bp), interplanar spacing (d), d-electron number (d_E_cou), electronegativity (EN), and local electronegativity (χ i ), where the work function (W) is obtained by density functional theory (DFT) calculations and the local electronegativity (χ i ) is calculated as is the electronegativity of atom j, and N is the total number of atoms in the first nearest neighbor of adsorption site i. The Pearson correlation coefficient (P) was used to evaluate the correlation of the 13 features, and the calculation formula was: where f i and Fi are two features for comparison, and the range of P is from -1 to 1. The adsorption energies of CO, H, HCOO, and COOH are taken as target values ​​to establish the corresponding adsorption energy data sets.

[0010] Step 4. Randomly divide the adsorption energy dataset into training set and test set in a ratio of 9:1, with the training set accounting for 90% and the test set accounting for 10%.

[0011] Step 5. Apply different machine learning models including gradient boosting regression algorithm (GBR), random forest regression algorithm (RFR), support vector regression algorithm (SVR) and K nearest neighbor regression algorithm (KNR) to train the four adsorption energy data sets.

[0012] Step 6. Evaluation index is accuracy R 2 , mean square error (MSE). R 2 The closer it is to 1, the closer MSE is to 0, which means the accuracy of the prediction model is better.

[0013] Accuracy R 2 The calculation formula is:

[0014]

[0015] The formula for calculating the mean square error (MSE) is:

[0016]

[0017] Among them, Y i is the DFT calculated value, y i is the value predicted by the ML model, is the average of all DFT data.

[0018] Step 7. By comparing the evaluation indicators of the four machine learning algorithms, the best machine learning algorithm is selected. The prediction performance of the gradient boosting regression algorithm (GBR) is better than the other three, so the subsequent results are based on the GBR algorithm. The feature importance analysis of the four selected best adsorption energy models is performed, and the five features that are in the top four are selected, such as Figure 2 As shown, work function (W), local electronegativity (χ i ) and interplanar spacing (d) are in the top 4. In addition, atomic number (AN) appears more frequently in the top 4. Electronegativity (EN) is a commonly used descriptor for predicting adsorption energy. Therefore, the above 5 features are selected to retrain the data set and achieve the best prediction performance through hyperparameter adjustment. Performance of machine learning model using 5 features ( Figure 3 ) and the prediction performance of the machine learning model using 13 features ( Figure 1), which shows that these five features are sufficient to establish a prediction model. The prediction set is the data outside the training set and the test set. The trained GBR model containing five features is used for prediction, and the DFT calculation results are compared with the ML prediction results.

[0019] Step 8. When the CO2 reduction product is CO, the reaction intermediates involved are COOH and CO. When the reduction product is HCOOH, the reaction intermediates involved are HCOO. The related intermediate of the hydrogen evolution reaction is H. When the product is CO, the adsorption energy of CO needs to satisfy the range of (E CO >0.5eV, the H adsorption energy range is (E H =0.2±0.1eV); when the CO2 reduction product is HCOOH, the adsorption energy range of HCOO that needs to be satisfied is (E HCOO =3.2±0.1eV), the H adsorption energy range is (E H =0.2±0.1eV). Based on the range of related intermediates for CO and HCOOH production, the most promising catalysts for CO and HCOOH production were screened.

[0020] The present invention is based on density functional theory and machine learning. The adsorption energy data set is established through density functional theory. The data set uses only 5 features. Except for the work function, which needs to be calculated by DFT, the other features are simple and easy to obtain. After training, a machine learning model with good prediction performance is obtained, which is used to predict the catalytic performance of new materials. Compared with traditional theoretical calculations, a lot of time is saved. The model of this method is simple and has good generalization ability. It can not only be used to predict CO2 electroreduction materials, but also is applicable to other electrochemical fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Figure 2. Predictive performance of the machine learning model using 13 features in this approach.

[0022] Figure 2 Ranking of feature importance for the machine learning model using 13 features in this method.

[0023] Figure 3 Figure 2 is a graph of the predictive performance of the machine learning model using five features in this method. DETAILED DESCRIPTION

[0024] The present invention is further described in detail below with reference to the accompanying drawings and embodiments, but the embodiments of the present invention are not limited thereto.

[0025] The following uses an implementation as an example to illustrate the application of the method. Specifically, the following steps are included:

[0026] Step 1. Apply the density functional theory method to optimize the structure of the unit cells of CuAl, CuZn, CuSn, CuGa, CuIn, CuAu, CuPd, CuPt, CuNi, CuFe, CuIr, and CuRh alloys, and then use the optimized unit cells to construct surface structures with different crystal plane indices, which are (111), (110), (100), and (211), respectively. The energy convergence standard in the structural optimization process is 10 -5 eV, the force convergence criterion is And perform non-self-consistent calculations to obtain the work function.

[0027] Step 2. Since CO and COOH are key intermediates in the process of CO2 reduction to CO, HCOO is the key intermediate for HCOOH, and H is the key intermediate for hydrogen evolution reaction (HER), the adsorption energy of these four intermediates is calculated. Each intermediate is placed at the top site, bridge site or hollow site on the surface of the copper alloy, and the structure is optimized. The adsorption energy of each intermediate at the corresponding site is calculated. The adsorption energy calculation formula is E adsorption =E absorbate / surface -(E surface +E adsorbate ), where E absorbate / surface is the energy of the catalyst surface and adsorbate as a whole, E surface is the energy of the clean surface, E adsorbate is the energy of a single adsorbate.

[0028] Step 3. The 13 selected features include work function (W), atomic number (AN), period number (P), group number (G), element ratio (rat), melting enthalpy (Δ fus H), ionic potential (IE), melting point (mp), boiling point (bp), interplanar spacing (d), d-electron number (d_E_cou), electronegativity (EN), and local electronegativity (χ i ), where the work function (W) is obtained by density functional theory (DFT) calculations and the local electronegativity (χ i ) is calculated as is the electronegativity of atom j, and N is the total number of atoms in the first nearest neighbor of adsorption site i. The Pearson correlation coefficient (P) was used to evaluate the correlation of the 13 features, and the calculation formula was: where f i and F i are two features for comparison, and the range of P is from -1 to 1. The adsorption energies of CO, H, HCOO, and COOH are taken as target values ​​to establish the corresponding adsorption energy data sets.

[0029] Step 4. Randomly divide the adsorption energy dataset into training set and test set in a ratio of 9:1, with the training set accounting for 90% and the test set accounting for 10%.

[0030] Step 5. Apply different machine learning models including gradient boosting regression algorithm (GBR), random forest regression algorithm (RFR), support vector regression algorithm (SVR) and K nearest neighbor regression algorithm (KNR) to train the four adsorption energy data sets.

[0031] Step 6. Evaluation index is accuracy R 2 , mean square error (MSE). R 2 The closer it is to 1, the closer MSE is to 0, which means the accuracy of the prediction model is better.

[0032] Accuracy R 2 The calculation formula is:

[0033]

[0034] The formula for calculating the mean square error (MSE) is:

[0035]

[0036] Among them, Y i is the DFT calculated value, y i is the value predicted by the ML model, is the average of all DFT data.

[0037] Step 7. By comparing the evaluation indicators of the four machine learning algorithms, the best machine learning algorithm is selected. The prediction performance of the gradient boosting regression algorithm (GBR) is better than the other three, so the subsequent results are based on the GBR algorithm. The feature importance analysis of the four selected best adsorption energy models is performed, and the five features that are in the top four are selected, such as Figure 2 As shown, work function (W), local electronegativity (χ i ) and interplanar spacing (d) are in the top 4. In addition, atomic number (AN) appears more frequently in the top 4. Electronegativity (EN) is a commonly used descriptor for predicting adsorption energy. Therefore, the above 5 features are selected to retrain the data set and achieve the best prediction performance through hyperparameter adjustment. Performance of machine learning model using 5 features ( Figure 3 ) and the prediction performance of the machine learning model using 13 features ( Figure 1 ), which shows that these five features are sufficient to establish a prediction model. The prediction set is the data outside the training set and the test set. The trained GBR model containing five features is used for prediction, and the DFT calculation results are compared with the ML prediction results.

[0038] Step 8. When the CO2 reduction product is CO, the reaction intermediates involved are COOH and CO. When the reduction product is HCOOH, the reaction intermediates involved are HCOO. The related intermediate of the hydrogen evolution reaction is H. When the product is CO, the adsorption energy of CO needs to satisfy the range of (E CO >0.5eV, the H adsorption energy range is (E H =0.2±0.1eV); when the CO2 reduction product is HCOOH, the adsorption energy range of HCOO that needs to be satisfied is (E HCOO =3.2±0.1eV), the H adsorption energy range is (E H =0.2±0.1 eV). The most promising catalysts for CO and HCOOH production were screened out according to the range of related intermediates for CO and HCOOH production.

Claims

1. A method for predicting carbon dioxide electroreduction copper alloy catalyst, characterized in that: include: Step 1. Apply density functional theory method to optimize the structure of the unit cells of different types of CuM alloys, and then use the optimized unit cells to construct surface structures with different crystal plane indices; Step 2. In the electrochemical reduction reaction of carbon dioxide, COOH, CO and HCOO are key intermediates. Considering that the hydrogen evolution reaction is a competitive reaction of the electrochemical reduction reaction of carbon dioxide, H is the key intermediate of the hydrogen evolution reaction. Set these four adsorption intermediates and calculate their adsorption energies at different sites on different surfaces. Step 3. Taking the adsorption energy as the target value, five chemical and physical properties are selected as features, namely work function (W), atomic number (AN), interplanar spacing (d), electronegativity (EN) and local electronegativity (χ i ), the Pearson correlation coefficient (P) was used to evaluate the correlation of the features, and the CO adsorption energy dataset, H adsorption energy dataset, COOH adsorption energy dataset, and HCOO adsorption energy dataset were constructed respectively; Step 4. Perform a preliminary screening of the data set and divide it into training set and test set; Step 5. Apply the machine learning model, conduct training, and evaluate the prediction performance of different models using evaluation indicators; Step 6. Select the model with the best prediction performance and use it to predict potential new materials; Step 7. Screening the Cu alloy catalysts for producing CO or HCOOH based on the range of adsorption energy of each intermediate; In step 7, when the CO2 reduction product is CO, the reaction intermediates involved are COOH and CO, and when the reduction product is HCOOH, the reaction intermediates involved are HCOO. The relevant intermediate of the hydrogen evolution reaction is H; when the product is CO, the adsorption energy of CO needs to satisfy the range of E CO >0.5eV, and the H adsorption energy range is E H =0.2±0.1eV; When the CO2 reduction product is HCOOH, the adsorption energy range of HCOO that needs to be satisfied is E HCOO =3.2±0.1eV and the H adsorption energy range is E H =0.2±0.1eV; based on the range of related intermediates for CO and HCOOH production, screen the catalysts for CO or HCOOH production.

2. A method for predicting carbon dioxide electroreduction copper alloy catalyst according to claim 1, characterized in that: In step 1, the CuM alloys are CuAl, CuZn, CuSn, CuGa, CuIn, CuAu, CuPd, CuPt, CuNi, CuFe, CuIr or CuRh alloys, and the crystal plane indices include (111) crystal plane, (110) crystal plane, (100) crystal plane and (211) crystal plane. The obtained surface structure is structurally optimized, and the energy convergence standard in the structural optimization process is 10 -5 eV, the convergence criterion of the force is And perform non-self-consistent calculations to obtain the work function.

3. The method for predicting carbon dioxide electroreduction copper alloy catalyst according to claim 1, characterized in that: In step 2, each intermediate is placed at a top site, a bridge site or a hollow site on the surface of the copper alloy for structural optimization.

4. The method for predicting carbon dioxide electroreduction copper alloy catalyst according to claim 1, characterized in that: In step 5, the evaluation index is the accuracy R 2 , mean square error (MSE); Accuracy R 2 The calculation formula is: The formula for calculating the mean square error (MSE) is: Among them, Y i is the DFT calculated value, y i is the value predicted by the ML model, is the average of all DFT data.

5. A method for predicting carbon dioxide electroreduction copper alloy catalyst according to claim 4, characterized in that: In step 6, R 2 The closer it is to 1 and the closer MSE is to 0, the better the accuracy of the prediction model, that is, the better the prediction performance, and the gradient boosting regression model GBR is selected as the machine learning model.

Images