Perovskite oxide OER catalyst screening method combining deep learning and multi-task optimization

Abstract

The invention discloses a perovskite oxide OER catalyst screening method combining deep learning and multi-task optimization, and the method comprises the steps: 1, screening out the density functional theory DFT data of perovskite oxide for OER research from a literature database, cleaning, and constructing a data set; 2, extracting features by using the central environment feature model to form a comprehensive descriptor data set; 3, training a deep neural network multi-task learning regression model based on the extracted feature data set, and optimizing the generalization ability and prediction precision of the model by combining a leave-one-out method cross validation with a loss function; 4, applying the trained model to predict and analyze a large-scale perovskite oxide structure, and implementing a multi-layer screening process, including an electric neutrality principle, a tolerance factor, an octahedral factor, overpotential evaluation and safety and economical efficiency consideration; and 5, experimental verification. The method provided by the invention has the advantages of high efficiency, accuracy, systematicness and the like, and opens up a new way for discovery and application of electro-catalytic materials.

Description

Technical Field

[0001] The present invention relates to a screening method for perovskite oxide OER catalysts combining deep learning and multi-task optimization. Background Art

[0002] In the context of responding to the global energy crisis and environmental challenges, electrocatalytic water splitting technology, as a clean way to convert electrical energy into chemical energy, has received great attention from the scientific research community. Among them, the oxygen evolution reaction (OER), as the key rate-limiting step in the water splitting process, is crucial to improving the efficiency of the entire system.

[0003] Perovskite oxides have shown great potential in the field of electrocatalysis due to their unique electronic properties and adjustable chemical composition. Their ABO3 structure provides a broad chemical space for diverse replacements of the cation A site and B site and for exploring new materials with excellent physicochemical properties.

[0004] However, there are many obvious limitations in the development and screening of perovskite oxide OER catalysts. First, the experimental methods of traditional material research and development are time-consuming and costly, and it takes weeks or even months from material synthesis, characterization to performance testing. Second, generating data through first-principles calculations requires huge computing resources and time investment, which limits the rapid screening and performance prediction of a large number of materials. Third, data collection and utilization are inefficient. Existing data collection and analysis methods are difficult to handle large-scale data sets, which hinders the application of machine learning in materials science. Fourth, there is a lack of efficient screening strategies. Existing screening methods mostly rely on experience and trial and error, lack of systematicity and efficiency, resulting in a slow process of discovery and optimization of new materials. Fifth, the generalization ability of existing machine learning models is limited, and they often rely on specific data sets, making it difficult to apply to a wider range of materials and conditions. Sixth, the accuracy and reliability of experimental verification of potential high-performance materials screened out through theoretical calculations and machine learning predictions are still challenges and need further optimization and verification.

[0005] Therefore, it is of great significance to develop a more efficient, accurate and systematic method to accelerate the discovery and optimization of new materials for perovskite oxide OER catalysts. Summary of the invention

[0006] In view of this, the object of the present invention is to provide a screening method for perovskite oxide OER catalysts combining deep learning and multi-task optimization, which has the advantages of high efficiency, accuracy, and systematicity.

[0007] In order to achieve the above invention purpose, the present invention adopts the following technical solutions:

[0008] The present invention provides a method for screening perovskite oxide OER catalysts combining deep learning and multi-task optimization, the method comprising the following steps:

[0009] 1: Data collection and preprocessing

[0010] 1.1 Screening DFT data of perovskite oxides for OER research from the literature database;

[0011] 1.2 Clean the data to remove erroneous and inconsistent data points;

[0012] 1.3 Build a database containing chemical composition, structural properties and electronic structure properties;

[0013] 2: Feature extraction

[0014] 2.1 Using the central environment (CE) feature model, key features are extracted from the collected data;

[0015] 2.2 Integrate the structure factors of perovskite oxides, the oxidation states of cations, and the ionic radius information to form a comprehensive descriptor dataset;

[0016] 3: Model training

[0017] 3.1 Use the PyTorch framework to train a deep neural network multi-task learning regression model based on the extracted feature dataset;

[0018] 3.2 Use leave-one-out cross validation (LOOCV) in combination with the loss function to optimize the generalization ability and prediction accuracy of the model;

[0019] 4: High-throughput screening

[0020] 4.1 Apply the trained model to predict and analyze large-scale perovskite oxide structures;

[0021] 4.2 Implement a multi-layer screening process, including the principle of electroneutrality, tolerance factor, octahedral factor, overpotential assessment, safety considerations, and economic feasibility analysis;

[0022] 5: Experimental verification

[0023] 5.1 Select potential perovskite oxide structures based on model prediction results;

[0024] 5.2 Experimentally synthesize the selected structure and verify the selection of perovskite oxide OER catalysts through physicochemical property characterization and electrochemical property testing.

[0025] Preferably, in step 1.3, when constructing a database containing chemical composition, structural properties and electronic structure properties, the data needs to be standardized to ensure that the data has a unified unit and format.

[0026] Preferably, in step 2.1, when extracting key features using the central environment (CE) feature model, a variance threshold is applied for filtering, and features with a variance greater than 0.02 are retained.

[0027] Preferably, in step 3.1, the deep neural network multi-task learning regression model consists of an input layer, four hidden layers and an output layer, wherein the first hidden layer has 64 neurons, the second hidden layer has 128 neurons, the third hidden layer has 32 neurons, and the fourth hidden layer has 16 neurons.

[0028] Preferably, in step 3.2, the loss function uses the mean absolute error.

[0029] Preferably, in step 3.2, the learning rate is set to 0.001.

[0030] Preferably, in step 4.2, the tolerance factor screening range is 0.8-1.0.

[0031] Preferably, in step 4.2, the octahedron factor screening range is 0.45-0.7.

[0032] Preferably, in step 4.2, the overpotential screening range is no more than 400 mV.

[0033] Of course, the perovskite oxide OER catalyst screened by the method of the present invention is also within the protection scope of the present invention.

[0034] The beneficial effects of the present invention are as follows: the present invention uses a multi-task machine learning technology combined with an experimental verification method to efficiently screen perovskite oxide OER catalysts, which has the following significant advantages over the prior art:

[0035] 1. Efficient data-driven screening process

[0036] The method uses a multi-task machine learning model that can quickly process and analyze large-scale data sets. Compared with traditional experimental methods, the method significantly reduces the time and economic cost required to screen potential perovskite oxide catalysts.

[0037] 2. Optimized computing resource utilization

[0038] By building a high-quality database and adopting efficient data preprocessing steps, the present invention reduces the reliance on computing resources. Compared with methods that rely on first-principles calculations, the present invention reduces the demand for expensive computing resources and improves the efficiency of data processing.

[0039] 3. Enhanced prediction accuracy and model generalization capabilities

[0040] The present invention adopts a deep neural network multi-task learning model and optimizes the model performance through cross-validation combined with a loss function. The model has good prediction accuracy and good generalization ability, and can be widely used in the prediction of different perovskite oxide structures.

[0041] 4. Systematic high-throughput screening strategy

[0042] The present invention implements a multi-layer screening process including the principle of electrical neutrality, tolerance factor, octahedral factor, overpotential, etc. The accuracy of the screening process is improved, ensuring that the screened materials have excellent catalytic performance in theory.

[0043] 5. Accuracy of experimental verification

[0044] After machine learning prediction, the present invention verifies the selected materials through experimental synthesis and electrochemical testing, thereby ensuring the reliability of theoretical prediction and improving the success rate of practical application of new materials.

[0045] 6. Innovative material property analysis

[0046] The present invention uses the central environment (CE) characteristic model to deeply analyze the structural characteristics of perovskite oxides, which can provide a deeper understanding of the OER activity of the material and provide new insights for material design.

[0047] 7. Environmentally friendly and cost-effective

[0048] The invention takes into account the safety and economic feasibility of materials during the screening process, and excludes materials that are toxic or too costly, which not only helps protect the environment, but also ensures the market competitiveness and practicality of the final materials.

[0049] In summary, the present invention provides an efficient, accurate and systematic method for screening perovskite oxide OER catalysts by combining advanced machine learning technology and experimental verification, which overcomes the limitations of the existing technology and opens up a new path for the discovery and application of electrocatalytic materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Provides a model structure framework for multi-task learning of deep neural networks.

[0051] Figure 2 This is the consistency analysis between the DNN predicted value and the DFT calculated value.

[0052] Figure 3 101251 A predicted by DNN x B 2-x Correlation between the adsorption free energy difference [ΔGO*-ΔGHO*] of OER intermediates and their OER overpotential (ηOER) of O3 perovskite oxides.

[0053] Figure 4 For multi-layer screening processes and results.

[0054] Figure 5 The XRD patterns of the four predicted perovskite structures synthesized experimentally and the electrochemical performance test results; a is the XRD pattern of the experimentally synthesized CrFeO3, LaFeO3, MnFeO3, and YFeO3; b is the XRD pattern of the CrFeO3, LaFeO3, MnFeO3, YFeO3, and RuO2 catalysts in an O2-saturated 1.0 M KOH solution at 5 mV s -1 IR corrected LSV polarization curves measured at a scan rate of 1.3 %; c is the Tafel curve of the prepared electrocatalyst; d is the electric double layer capacitance (C) of CrFeO3, LaFeO3, MnFeO3, YFeO3 and RuO2 catalysts measured at open circuit potential. dl ) changes with the scan rate; e is the EIS Nyquist diagram of different electrocatalyst samples. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with examples. It should be understood that the following examples are only exemplary descriptions and explanations of the present invention, and should not be construed as limiting the scope of protection of the present invention. All technologies realized based on the above content of the present invention are included in the scope that the present invention is intended to protect. Unless otherwise stated, the raw materials and reagents used in the following examples are all commercially available products, or can be prepared by known methods.

[0056] Example 1

[0057] 1: Data collection and preprocessing

[0058] 1.1 Density functional theory (DFT) data of perovskite oxides for OER research were screened from the literature database.

[0059] Specifically, 171 DFT calculation data of perovskite oxides were collected from the scientific literature database, involving 171 different perovskite structures. The data include the adsorption energy of OH, O and OOH intermediates on the surface of each structure, as well as the corresponding overpotential.

[0060] 1.2 Clean the data to remove erroneous and inconsistent data points.

[0061] 1.3 Build a database containing chemical composition, structural properties and electronic structure properties, standardize the data format, and make the data specific and unified in unit and form.

[0062] 2: Feature extraction

[0063] 2.1 Utilize the Central Environment (CE) feature model to extract key features from the collected data.

[0064] Specifically, 112 features were extracted from each perovskite structure, including chemical composition, crystal structure, and electronic structure properties. Variance threshold filtering was applied, and 104 features with variance greater than 0.02 were retained.

[0065] 2.2 Integrate the structure factors of perovskite oxides, the oxidation states of cations, and the ionic radius information to form a comprehensive descriptor dataset;

[0066] 3: Model training

[0067] 3.1 Using the PyTorch framework, a deep neural network multi-task learning regression model is trained based on the extracted feature dataset; the deep neural network multi-task learning regression model consists of an input layer, four hidden layers and an output layer (see Figure 1 ), ensuring that the complex nonlinear relationship between perovskite oxide properties and OER performance can be captured.

[0068] Input layer: Receives 104 features extracted from the feature engineering stage.

[0069] Hidden Layer:

[0070] First hidden layer: 64 neurons

[0071] Second hidden layer: 128 neurons

[0072] Third hidden layer: 32 neurons

[0073] Fourth hidden layer: 16 neurons

[0074] Output layer: Predicted OER intermediate adsorption energies and overpotentials for each perovskite oxide structure.

[0075] 3.2 The leave-one-out cross-validation (LOOCV) is combined with the loss function to optimize the generalization ability and prediction accuracy of the model.

[0076] The Leave-One-Out Cross-Validation (LOOCV) strategy is adopted. This method is particularly effective for small sample data sets and can ensure that the model is fully validated on independent data.

[0077] The loss function uses the mean absolute error (MAE), which measures the average absolute value of the difference between the predicted value and the actual value, is insensitive to outliers, and is suitable for the application scenario of the present invention.

[0078] The learning rate is set to 0.001, and this parameter value is adjusted after multiple experiments to ensure that the model can converge stably.

[0079] In each iteration, the network computes predictions via forward propagation and adjusts the weights via backpropagation to minimize the loss function.

[0080] During training, monitor the performance of the model on the validation set to avoid overfitting and ensure that the model has good generalization ability.

[0081] 4: High-throughput screening

[0082] 4.1 Apply the trained model to predict and analyze large-scale perovskite oxide structures;

[0083] The trained deep learning model was used to predict 101,251 theoretical perovskite structures, generating OER intermediate adsorption energy and overpotential data for each structure (see Figure 2 and 3 ).in, Figure 3 The scatter plots of the model predictions for OH, O and OOH intermediates and overpotentials are presented, as well as the average determination coefficients predicted by the model. <R 2 > Up to 0.99, 0.98, 0.99 and 0.96 respectively, mean absolute error <mae>They are 0.047eV, 0.071eV, 0.048eV and 0.034eV respectively. These indicators fully demonstrate the prediction accuracy of the model.

[0084] 4.2 Implement a multi-layer screening process, including the principle of electroneutrality, tolerance factor, octahedral factor, overpotential assessment, safety and economic considerations (see Figure 4 ).

[0085] Screening based on the principle of electroneutrality: All predicted structures were preliminarily screened based on the principle of electroneutrality. The principle of electroneutrality requires that the oxidation states of the elements in the material be balanced to ensure the charge balance of the material. By analyzing the oxidation states of the elements in each structure, structures that meet the electroneutrality condition were screened, and a total of 43,435 structures that meet the conditions were obtained.

[0086] Structural factor screening: Based on the preliminary screening, the structural factor of perovskite was further investigated. The tolerance factor is an important indicator for evaluating the stability of the material's crystal structure, which helps to screen out structurally stable perovskite oxides. By setting the tolerance factor (t) range from 0.8 to 1.0, candidate materials with ideal crystal structures were screened out, reducing the number of candidate structures to 9752.

[0087] Octahedral factor screening: The structures were further screened using the octahedral factor (μ). The octahedral factor range was set to 0.45 to 0.7, which helps to optimize the coordination geometry of the B-site cations, thereby improving the OER activity of the material. After octahedral factor screening, the number of structures was further reduced to 9752.

[0088] Overpotential screening: Based on the overpotential predicted by the model (η OER ) for screening. Structures with an overpotential of no more than 400 mV were selected. This criterion helps to identify catalytic materials with lower overpotential and high efficiency. After overpotential screening, a total of 56 potential high-performance catalysts were obtained.

[0089] Safety and economic considerations: Considering the practical application potential of the material, structures containing radioactive, highly toxic or costly metal elements are excluded. This step ensures that the screened materials not only have excellent catalytic performance in theory, but are also feasible and safe in practical applications. After the final screening, 17 perovskite oxides with potential OER catalytic application value were identified.

[0090] 5: Experimental verification

[0091] 5.1 Select potential perovskite oxide structures based on model prediction results;

[0092] Based on the model prediction results, four perovskite structures with the best performance were selected for experimental synthesis, including CrFeO3, LaFeO3, MnFeO3 and YFeO3.

[0093] 5.2 Experimentally synthesize the selected structure and verify the selection of perovskite oxide OER catalysts through physicochemical property characterization and electrochemical property testing.

[0094] The selected CrFeO3, LaFeO3, MnFeO3 and YFeO3 were synthesized by the sol-gel method. The specific synthesis method is: the corresponding metal salt precursor is dissolved in deionized water, and an appropriate complexing agent (such as citric acid or ethylene glycol) is added to form a stable sol. Subsequently, the sol is transformed into a gel by controlling the pH value and temperature. The gel is dried and calcined to finally form a powder with a perovskite structure.

[0095] XRD scanning was performed on the synthesized CrFeO3, LaFeO3, MnFeO3 and YFeO3. The results are as follows: Figure 5 a.

[0096] The electrochemical performance test of synthesized CrFeO3, LaFeO3, MnFeO3 and YFeO3 was carried out. The electrochemical performance test is a key step in verifying the activity of OER catalysts. The test was carried out in a standard three-electrode system.

[0097] The test conditions are as follows: the synthesized material is used as the working electrode, the graphite rod is used as the counter electrode, the mercury / mercury oxide (Hg / HgO) electrode is used as the reference electrode, and 1 M KOH solution is used as the electrolyte.

[0098] The test method is: OER performance is tested by linear sweep voltammetry (LSV) with a scan rate of 5 mV / s and a potential window set to 0 to 0.8 V vs. Hg / HgO.

[0099] The LSV polarization curve after IR correction reflects the actual catalytic performance of the catalyst, and the Tafel slope curve can be used to evaluate the reaction kinetic parameters of the catalyst. Figure 5 b, the Tafel slope curve is Figure 5 c. The results show that CrFeO3, LaFeO3, MnFeO3 and YFeO3 all have excellent catalytic performance.

[0100] The electric double layer capacitance (C dl ) versus scan rate can be used to evaluate the electrode surface area and measure the electrical double layer capacitance (C) of CrFeO3, LaFeO3, MnFeO3, and YFeO3 at open circuit potential. dl ) changes with the scanning rate, the results are as follows Figure 5 As shown in d.

[0101] By measuring the response to small AC signals of different frequencies, the Nyquist diagram is plotted. Figure 5 e.

[0102] The above results show that the screened CrFeO3, LaFeO3, MnFeO3 and YFeO3 have excellent OER catalytic activity, which also proves the effectiveness of the screening strategy of the present invention.

[0103] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. All modifications, equivalent substitutions, improvements, etc. within the spirit and principles of the present invention should be included in the protection scope of the present invention.< / mae>

Claims

1. A method for screening perovskite oxide OER catalysts combining deep learning and multi-task optimization, characterized in that the method comprises the following: 1: Data collection and preprocessing 1.1 Density functional theory (DFT) data of perovskite oxides for OER research were screened from the literature database; 1.2 Clean the data to remove erroneous and inconsistent data points; 1.3 Build a database containing chemical composition, structural properties and electronic structure properties; 2: Feature extraction 2.1 Using the central environment feature model, extract key features from the collected data; 2.2 Integrate the structure factors of perovskite oxides, the oxidation states of cations, and the ionic radius information to form a comprehensive descriptor dataset; 3: Model training 3.1 Use the PyTorch framework to train a deep neural network multi-task learning regression model based on the extracted feature dataset; 3.2 Use leave-one-out cross-validation combined with loss function to optimize the generalization ability and prediction accuracy of the model; 4: High-throughput screening 4.1 Apply the trained model to predict and analyze large-scale perovskite oxide structures; 4.2 Implement a multi-layer screening process, including the principle of electroneutrality, tolerance factor, octahedral factor, overpotential assessment, safety and economic considerations; 5: Experimental verification 5.1 Select potential perovskite oxide structures based on model prediction results; 5.2 Experimentally synthesize the selected structure and verify the selection of perovskite oxide OER catalysts through physicochemical property characterization and electrochemical property testing.

2. The method according to claim 1, characterized in that In step 1.3, when constructing a database containing chemical composition, structural properties, and electronic structure properties, the data needs to be standardized to ensure that the data has a unified unit and format.

3. The method according to claim 1, characterized in that In step 2.1, when the central environment feature model is used to extract key features, the variance threshold is applied for filtering, and features with variance greater than 0.02 are retained.

4. The method according to claim 1, characterized in that: In step 3.1, the deep neural network multi-task learning regression model consists of an input layer, four hidden layers and an output layer, where the first hidden layer has 64 neurons, the second hidden layer has 128 neurons, the third hidden layer has 32 neurons, and the fourth hidden layer has 16 neurons.

5. The method according to claim 1, characterized in that In step 3.2, the loss function uses the mean absolute error.

6. The method according to claim 1, characterized in that In step 4.2, the tolerance factor screening range is 0.8-1.

0.

7. The method according to claim 1, characterized in that In step 4.2, the octahedron factor screening range is 0.45-0.

7.

8. The method according to claim 1, characterized in that In step 4.2, the overpotential screening range is no more than 400 mV.

9. The perovskite oxide OER catalyst obtained by screening according to any one of claims 1 to 8, characterized in that: The screened perovskite oxide OER catalysts include: CrFeO3, LaFeO3, MnFeO3 and YFeO3.

Images