A method for screening binary alloy acidic oxygen evolution catalysts using machine learning models

Abstract

This invention discloses a method for screening machine learning models for binary alloy acidic oxygen evolution catalysts. This method sequentially completes four core steps: constructing a standardized machine learning dataset, multi-model training and cross-validation, progressive feature engineering optimization, and screening the optimal machine learning model. The result is a high-precision, high-reliability prediction model adapted to the scenario of predicting the catalytic performance of binary alloy acidic oxygen evolution catalysts. Simultaneously, based on the selected optimal model, a method for predicting the catalytic performance of binary alloy acidic oxygen evolution catalysts is designed. By obtaining the corresponding input features of the binary alloy to be predicted and inputting them into the optimized model, the method can quickly output accurate prediction results of the overpotential of the acidic oxygen evolution reaction, achieving efficient evaluation and ranking of catalytic performance. A supporting process for controllable catalyst preparation and electrochemical performance verification based on the prediction results is also provided.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of materials science and artificial intelligence, specifically to a method for screening binary alloy acidic oxygen evolution catalysts using machine learning models. Background Technology

[0002] Current technologies for developing acidic oxygen evolution reaction (OER) catalysts generally suffer from lengthy development cycles, hindering the rapid development and iteration of high-performance catalytic materials. This makes it difficult to meet the rapid deployment requirements of new energy industries such as proton exchange membrane water electrolysis for hydrogen production. Existing technologies also suffer from high experimental trial-and-error frequencies, resulting in significant waste of precious metal raw materials, experimental consumables, and equipment time, substantially increasing both the development and mass production costs of catalysts. Furthermore, they exhibit high dependence on precious metals, making it difficult to effectively optimize and reduce the amount of precious metals used, thus hindering the cost barrier to the large-scale deployment of high-performance acidic OER catalysts.

[0003] For binary alloy oxygen evolution catalysts, existing research and development technologies suffer from extremely low candidate material screening efficiency, making it impossible to accurately predict the catalytic performance of binary alloys upfront. Performance evaluation can only be completed through post-processing verification of material preparation and performance testing, resulting in a continuous and ineffective waste of research and development resources and hindering the targeted and efficient development of high-performance binary alloy catalysts. Moreover, it is difficult to achieve batch and rapid evaluation of the performance of binary alloy catalytic materials, and it is impossible to complete the performance ranking of all candidate systems and identify the optimal solution, which seriously restricts the development process and industrial application of binary alloy catalytic materials.

[0004] Although machine learning models have been widely used in various technical fields, existing machine learning technologies applied to the research and development of oxygen evolution catalysts suffer from poor model adaptability and insufficient prediction accuracy. They are unable to select the optimal machine learning model that is suitable for predicting the performance of binary alloy acidic oxygen evolution catalysts, making it difficult to output reliable performance prediction results and providing effective theoretical support and directional guidance for the development of catalytic materials. At the same time, there is also a serious disconnect between prediction and experimental processes. There is a lack of a complete and closed-loop performance prediction and experimental verification system, making it impossible to effectively connect theoretical prediction results with experimental preparation. The prediction results are difficult to directly guide the actual development and production of catalysts. Summary of the Invention

[0005] To address the aforementioned shortcomings of existing technologies, this invention provides a method for screening machine learning models for binary alloy acidic oxygen evolution catalysts, and a method for predicting the catalytic performance of binary alloy acidic oxygen evolution catalysts based on the machine learning models obtained from the screening method. This addresses the problems of long catalyst development cycles, high costs, and the lack of a closed loop between computational prediction and experimental verification in machine learning prediction, which fails to support the development of high-performance catalysts.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for screening machine learning models for binary alloy acidic oxygen evolution catalysts, the specific steps of which are as follows:

[0007] Step 1: Dataset Construction: Taking binary alloy catalysts formed by any two combinations of 26 transition metal elements as the research object, a surface catalytic structure model of the binary alloy catalyst is established, and the vertex sites of the structure model are designated as the catalytic active sites for the acidic oxygen evolution reaction. The target features corresponding to each catalytic structure are obtained by density functional theory calculation, and the input features of the corresponding binary alloy are extracted. The target features and input features are integrated to construct a dataset for training machine learning models.

[0008] Step 2: Model Training and Cross-Validation: Divide the dataset constructed in Step 1 into a training set and a test set. Based on the training set, perform initial training on multiple candidate machine learning models. At the same time, use the five-fold cross-validation method to evaluate the generalization ability and prediction accuracy of each candidate machine learning model, and initially select the model for subsequent training.

[0009] Step 3 Feature Engineering Optimization: Perform multiple rounds of screening and optimization on the input features obtained in Step 1 to remove redundant and low-contribution features, and obtain the feature set with the best correlation to the target features;

[0010] Step 4: Optimal Model Selection: Using the feature set obtained from Step 3 as input parameters, and based on the evaluation results of the initial training and five-fold cross-validation in Step 2, the optimal machine learning model suitable for predicting the performance of binary alloy acidic oxygen evolution catalysts is selected from the initially selected machine learning models, according to the preset accuracy evaluation index.

[0011] Preferably, in step 3, the target feature is the overpotential of the binary alloy catalyst in the acidic oxygen evolution reaction; the input features include two categories: atomic features and statistical features. The atomic features are the intrinsic properties of the two metal elements in the binary alloy, and the statistical features are the statistical values ​​calculated for the corresponding atomic properties of the two metal elements in the binary alloy.

[0012] Preferably, the atomic characteristics include atomic number, atomic mass, atomic radius, van der Waals radius, covalent radius, electronegativity, melting point, boiling point, density, first ionization energy, second ionization energy, first electron affinity, number of valence electrons, number of outermost d orbitals, number of outermost s orbitals, and work function; the statistical characteristics include the sum, difference, mean, variance, and standard deviation of each of the atomic characteristics; specifically, the sum of atomic numbers, the difference of atomic numbers, the mean of atomic numbers, the variance of atomic numbers, and the standard deviation of atomic numbers; other atomic characteristics follow the same pattern.

[0013] Preferably, in step 3, the multi-round screening and optimization sequentially includes three stages: feature importance analysis, correlation analysis, and recursive feature elimination. The specific steps are as follows:

[0014] The first round is feature importance analysis: features with an importance greater than 0.3% are selected and those that meet the requirements are retained.

[0015] In the second round, Pearson correlation analysis was performed: linear correlation was calculated on the features retained in the first round. For two features with an absolute linear correlation coefficient higher than 0.9, the feature with the lower feature importance was deleted, thus completing the removal of redundant features.

[0016] The third round, recursive feature elimination: through iterative model training, features that contribute the least to the model's predictive performance are gradually removed, using the model's coefficient of determination R0 as the criterion. 2 The maximum value is used as the termination condition, and the optimal feature set is finally obtained through screening.

[0017] Preferably, in step 2, the dataset constructed in step 1 is divided into a training set and a test set in an 8:2 ratio; the training set is used for initial model training and five-fold cross-validation, and the test set is used for final prediction accuracy verification of the model.

[0018] Preferably, in step 2, the five-fold cross-validation method is implemented as follows: the training set is divided into five subsets on average, and four subsets are used as training subsets and the remaining subset is used as validation subsets in turn to iteratively train and validate the machine learning model; based on the average prediction accuracy of the five rounds of validation results, the generalization ability of the machine learning model is evaluated, and the model hyperparameters are optimized.

[0019] Preferably, the alternative machine learning models include adaptive boosting models, gradient boosting regression models, K-nearest neighbor regression models, kernel ridge regression models, Lasso regression models, partial least squares regression models, random forest regression models, correlation vector regression models, extreme gradient boosting regression models, and support vector regression models.

[0020] Preferably, in step 4, the preset accuracy evaluation index is the coefficient of determination R. 2 The selection criteria, along with the mean squared error (MSE), are the coefficient of determination (R²). 2 It has the highest mean square error (MSE) and the lowest mean square error (MSE).

[0021] Secondly, the present invention provides a method for predicting the catalytic performance of a binary alloy acidic oxygen evolution catalyst, wherein the optimal machine learning model obtained by the above-mentioned screening method is used to perform the catalytic performance prediction, including the following steps:

[0022] S1: Obtain the input features corresponding to the binary alloy to be predicted, wherein the input features are the atomic features and statistical features defined above;

[0023] S2: Input the input features into the optimal machine learning model;

[0024] S3: Using the optimal machine learning model, the overpotential of the binary alloy to be predicted in the acidic oxygen evolution reaction is obtained, thus completing the prediction of catalytic performance.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. This invention achieves a breakthrough in high precision and high reliability for catalytic performance prediction models, solving the technical problems of poor adaptability, insufficient prediction accuracy, and weak generalization ability of existing machine learning models. This invention constructs a multi-dimensional initial feature set encompassing atomic intrinsic features and combinatorial statistical features. Combined with feature importance analysis, Pearson correlation analysis, and recursive feature elimination through multiple rounds of progressive feature engineering, it accurately identifies the core feature set strongly correlated with acidic oxygen evolution catalytic performance. Simultaneously, through horizontal comparison of 10 mainstream machine learning models and evaluation of generalization ability using 5-fold cross-validation, the optimal GBR model is finally selected, with a test set determination coefficient R0. 2 With an overpotential accuracy as high as 0.8935, the average error between the predicted overpotential and the calculated overpotential is only 5.03% for catalytic systems without DFT calculations, achieving accurate and reliable prediction of the oxygen evolution catalytic performance of binary alloys. This effect far exceeds conventional expectations in the field: existing machine learning prediction schemes for acidic OER of binary alloys generally suffer from problems such as blind feature selection and poor model adaptability to application scenarios, making it difficult to balance prediction accuracy and generalization ability. However, this invention achieves a leapfrog improvement in prediction accuracy by significantly reducing the number of core features through feature-model co-optimization throughout the entire process, providing reliable theoretical support for the targeted development of catalytic materials.

[0027] 2. This invention improves the R&D efficiency of acidic oxygen evolution catalysts, fundamentally breaking through the efficiency bottleneck of traditional trial-and-error methods and pure DFT calculations. Targeting 650 unique binary alloy catalytic structures formed by 26 transition metals, this invention only requires DFT calculations on 25% of the systems to train a high-precision prediction model. The trained model can quickly predict the catalytic performance of the remaining 75% of the systems and rank the performance of all candidate systems. This completely avoids the massive computational consumption of full-system DFT calculations in traditional R&D models, as well as the lengthy R&D cycle caused by blind experimentation. It compresses the R&D cycle, which originally required months, into days, achieving rapid and comprehensive screening of high-performance catalytic materials. This effect is something that those skilled in the art could not easily predict: existing binary alloy catalytic systems, due to their large component space and complex structure-activity relationships, are generally considered difficult to screen efficiently across a wide range of systems. This invention, through a data-driven machine learning approach, completely solves the technical problem of low screening efficiency for binary alloy catalysts, achieving an upgrade from "experience-based trial and error" to "rational design" in the R&D model.

[0028] 3. This invention achieves breakthrough catalytic performance in acidic OER reactions through targeted screening of binary alloy catalysts, simultaneously realizing a significant improvement in both catalytic activity and stability, far exceeding the level of commercial benchmark catalysts. Based on the accurate prediction results of the optimal model of this invention, a FeRu binary alloy catalyst was successfully prepared. This catalyst exhibits excellent performance in acidic OER reactions at 10 mA·cm⁻¹. -2 The overpotential at the current density is only 171 mV, and the Tafel slope is as low as 42.46 mV·dec. -1 The catalytic reaction kinetics and intrinsic activity are significantly superior to commonly used benchmark catalysts such as commercial RuO2 and pure Ru; at the same time, this catalyst exhibits industry-leading long-term operational stability at 100 mA·cm⁻¹. -2 It can operate stably for 500 hours at current density, with a potential decay rate of only 24.6 μV·h. -1 This significantly outperforms the stability levels of existing acidic OER catalysts of the same type. This effect is not readily apparent: conventional binary alloy modification research in this field struggles to simultaneously achieve a substantial increase in catalytic activity and ensure operational stability. However, this invention eliminates the need for extensive experimental trial and error. Through precise pre-emptive prediction using machine learning, it can precisely target the optimal alloy system that combines ultra-low overpotential, excellent reaction kinetics, and ultra-long durability. Furthermore, the predicted results match the experimental results, achieving a breakthrough upgrade in the catalytic performance of acidic OER.

[0029] 4. This invention significantly reduces the development and industrialization costs of catalysts from the research and development stage, and substantially reduces resource consumption and precious metal waste. Through algorithmic pre-prediction, this invention greatly reduces the resource consumption of high-computing-power DFT calculations, while also significantly reducing the number of invalid experimental trials. It avoids the ineffective consumption of precious metal raw materials, experimental consumables, and equipment time during catalyst preparation and performance testing, achieving efficient and precise utilization of precious metals. This fundamentally reduces the research and development and mass production costs of high-performance acidic oxygen evolution catalysts, breaking down the core cost barrier for the large-scale promotion of new energy scenarios such as proton exchange membrane water electrolysis for hydrogen production. This invention constructs a complete closed-loop research and development system of "data-algorithm-experiment," possessing strong universality and scalability, overcoming the problem of limited application scenarios in existing technologies. This invention is not limited to the development of a single binary alloy system, but forms a replicable and transferable rational design methodology for electrocatalytic materials. It is applicable not only to the rapid development of various transition metal binary alloy oxygen evolution catalysts, but also directly extended to complex catalytic systems such as ternary alloys and multi-element alloys, as well as catalyst development for other electrocatalytic reaction scenarios such as hydrogen evolution reaction and carbon dioxide reduction reaction. It has extremely high industry promotion value and reusability, and solves the technical problems of poor universality and inability to be reused across systems in existing catalyst development technologies. Attached Figure Description

[0030] Figure 1 This is a flowchart of the prediction system for binary alloy acidic oxygen evolution catalysts based on machine learning, as described in this invention.

[0031] Figure 2 This is a schematic diagram of the selected elements and structure of the binary alloy in this invention; where A and B represent the selected transition metals, and the numbers 1 and 2 in the circles represent adsorption sites.

[0032] Figure 3 This refers to the result of training the machine learning model in this invention; Figure 3 a and Figure 3 b represents the training and 5-fold cross-validation results of 10 machine learning models.

[0033] Figure 4 The importance analysis results were used to initially select the model; Figure 4 a and Figure 4 b represents the importance analysis results for the GBR and RFR models, respectively.

[0034] Figure 5 The results of the correlation analysis were used to initially select the model; Figure 5 a and Figure 5 b represents the correlation analysis results of the GBR and RFR models, respectively.

[0035] Figure 6 The results of recursive feature elimination for the initially selected model; Figure 6 a and Figure 6 b represents the feature recursive elimination results for the GBR and RFR models, respectively.

[0036] Figure 7 This represents the final fitting result of the model in this invention; wherein... Figure 7 'a' represents the GBR training result. Figure 7 b represents the RFR training result.

[0037] Figure 8 This is a comparison of the overpotentials predicted by optimal machine learning and calculated by DFT for VNi, CuRh, and ZnMo in this invention.

[0038] Figure 9 This is a scanning electron microscope image of the FeRu alloy in this invention.

[0039] Figure 10 The image shows the HAADF-STEM image of the FeRu alloy and the corresponding FeRu elemental mapping.

[0040] Figure 11 The OER polarization curves for the FeRu alloy are shown.

[0041] Figure 12 This is the Tafel slope diagram for the FeRu alloy.

[0042] Figure 13 For the stability test of FeRu alloy. Detailed Implementation

[0043] The technical solutions of the present invention will be clearly and completely described in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the present invention are within the scope of protection of the present invention.

[0044] Unless otherwise specified in the specific circumstances, the numerical ranges listed herein include upper and lower limits, as well as all integers and fractions within that range, but are not limited to the specific values ​​listed when the range is defined.

[0045] I. A method for screening binary alloy acidic oxygen evolution catalysts using machine learning models

[0046] To address the technical problems existing in current technologies, this invention proposes a solution from the following perspectives: First, addressing the issues of low efficiency, high cost, and excessive resource consumption in traditional R&D models involving full-scale computation and trial-and-error, this invention considers designing a standardized dataset construction stage. Using a binary alloy system formed by 26 transition metals as the research object, it clarifies the catalytic active sites and selects only a small number of catalytic structures for density functional theory calculations to obtain true overpotential labels. Simultaneously, it constructs a full-dimensional initial feature set covering atomic intrinsic characteristics and combinatorial statistical characteristics, forming a standardized machine learning dataset. This maximizes prediction coverage with a small amount of precisely labeled data, avoiding the problems associated with full-scale computation and trial-and-error. The massive computational power required for functional theory calculations provides a unified and reliable data foundation for subsequent model training, fundamentally breaking through the bottleneck of low efficiency in traditional R&D. Secondly, addressing the technical problems of poor model adaptability, insufficient prediction accuracy, and inadequate generalization ability in existing machine learning solutions, this invention considers designing a multi-dimensional, progressive model screening and validation system. It integrates multiple mainstream machine learning regression models covering different algorithm principles, completes initial training based on a unified dataset and optimized feature set, and then conducts a systematic evaluation of the generalization ability of each model through 5-fold cross-validation. Authoritative quantitative evaluation indicators are used to rank the adaptability of different models, and a preliminary selection of machine learning models is made. Furthermore, addressing the issue of blind and redundant feature selection in existing schemes that restrict model performance, this invention considers designing a three-round progressive feature engineering optimization process. This involves sequentially eliminating low-contribution features through feature importance analysis, eliminating multicollinearity among features through Pearson correlation analysis, and locking in the smallest feature subset that optimizes model prediction accuracy through recursive feature elimination. This precisely selects core features strongly correlated with the acidic oxygen evolution catalytic performance, significantly reducing the complexity of model training while significantly improving the model's prediction accuracy and generalization ability, laying a core foundation for reliable prediction. Ultimately, it accurately selects the core features relevant to the prediction scenario of binary alloy acidic oxygen evolution catalytic performance. The optimally matched machine learning model addresses the issues of poor adaptability and unreliable prediction results associated with single models. Finally, to address the disconnect between existing predictions and experiments, and the lack of closed-loop verification, this invention designs a closed-loop implementation process: "model prediction - targeted preparation - performance verification." Based on the selected optimal model, the performance of the entire binary alloy catalytic system is predicted and ranked, the optimal candidate alloy system is targeted, and standardized, controllable preparation and performance verification methods are provided. This not only verifies the accuracy of the model predictions but also achieves a complete connection from theoretical prediction to the practical application of high-performance catalysts, solving the problem of machine learning solutions "only predicting, not implementing." Therefore, this invention forms the following screening method:

[0047] Step 1: Dataset Construction: Taking binary alloy catalysts formed by any two combinations of 26 transition metal elements as the research object, a surface catalytic structure model of the binary alloy catalyst is established, and the vertex sites of the structure model are designated as the catalytic active sites for the acidic oxygen evolution reaction. The target features corresponding to each catalytic structure are obtained by density functional theory calculation, and the input features of the corresponding binary alloy are extracted. The target features and input features are integrated to construct a dataset for training machine learning models.

[0048] Step 2: Model Training and Cross-Validation: Divide the dataset constructed in Step 1 into a training set and a test set. Based on the training set, perform initial training on multiple candidate machine learning models. At the same time, use the five-fold cross-validation method to evaluate the generalization ability and prediction accuracy of each candidate machine learning model, and initially select the model for subsequent training.

[0049] Step 3 Feature Engineering Optimization: Perform multiple rounds of screening and optimization on the input features obtained in Step 1 to remove redundant and low-contribution features, and obtain the feature set with the best correlation to the target features;

[0050] Step 4: Optimal Model Selection: Using the feature set obtained from Step 3 as input parameters, and based on the evaluation results of the initial training and five-fold cross-validation in Step 2, the optimal machine learning model suitable for predicting the performance of binary alloy acidic oxygen evolution catalysts is selected from the initially selected machine learning models, according to the preset accuracy evaluation index.

[0051] The screening method described in this invention has achieved many unexpected technical effects after practical application: First, the model prediction accuracy and reliability have been significantly improved. Through the synergistic optimization of feature engineering and model screening, the coefficient of determination R of the gradient boosting regression model test set obtained by the final screening is significantly improved. 2 The first achievement is a high overpotential of 0.8935. For unknown catalytic systems without density functional theory calculations, the average error between the predicted and theoretical overpotential values ​​is only 5.03%. This significantly reduces the number of core features while achieving a leapfrog improvement in prediction accuracy, far exceeding the prediction results of conventional machine learning schemes in this field. Secondly, catalyst development efficiency is significantly improved. Only 130 sets of catalytic structures need to be calculated using density functional theory to predict and rank the catalytic performance of the remaining 520 binary alloy systems in a short time, compressing the development cycle from several months to several days, increasing efficiency by more than an order of magnitude. This also significantly reduces the number of ineffective experimental trials, fundamentally reducing the consumption of computing power, precious metal raw materials, experimental consumables, and equipment time, breaking through the cost barrier for the development and mass production of high-performance acidic oxygen evolution catalysts. Thirdly, the targeted catalyst development achieves a dual breakthrough in catalytic activity and stability. Based on the precise prediction and targeted preparation of the FeRu binary alloy catalyst of this invention, in the acidic oxygen evolution reaction, the overpotential at 10 mA·cm⁻¹…-2 The overpotential at current density is only 171 mV, and the Tafel slope is as low as 42.46 mV·dec. -1 The intrinsic catalytic activity and reaction kinetics are significantly superior to industry benchmark catalysts such as commercial RuO2 and pure Ru, while at 100 mA·cm -2 It can operate stably for 500 hours at current density, with a potential decay rate of only 24.6 μV·h. -1 This invention achieves a balance between ultra-low overpotential and ultra-long-term operational stability, a breakthrough performance that traditional trial-and-error methods cannot achieve in a short time. It also further proves the reliability of the machine learning model selected by the screening method described in this invention. Fourth, it constructs a scalable closed-loop R&D system with far greater universality than existing technologies. The core technology system formed by this invention is not limited to a single binary alloy system, but is a replicable and transferable rational design methodology for electrocatalytic materials. It is not only applicable to the rapid development of various transition metal binary alloy oxygen evolution catalysts, but can also be directly extended to complex catalytic systems such as ternary alloys and multi-element alloys, as well as other electrocatalytic reaction scenarios such as hydrogen evolution reaction and carbon dioxide reduction reaction. This breaks through the limitations of existing R&D technologies, which have poor universality and cannot be reused across systems, and possesses extremely high promotional value.

[0052] In some embodiments of the present invention, the OER process in the screening method of the present invention is generally divided into four steps:

[0053] *+H2O → *OH + H + + e - ΔG I ;

[0054] *OH → *O +H + +e - ΔG II ;

[0055] O + H₂O → OOH + H₂ + + e - ΔG III ;

[0056] *OOH → O2 + H + + e - +* ΔG IV .

[0057] Where * represents an adsorption active site on the substrate.

[0058] The formula for calculating overpotential is:

[0059] ΔG I = ΔG HO* ;

[0060] ΔG II = ΔG O* - ΔG HO* ;

[0061] ΔG III = ΔG HOO* - ΔG O* ;

[0062] ΔG Ⅳ = 4.92 - ΔG HOO* ;

[0063] η = max{ΔG I , ΔG II , ΔG III , ΔG Ⅳ} – 1.23V.

[0064] In some embodiments of the present invention, in step 3, the target feature is the overpotential of the binary alloy catalyst in the acidic oxygen evolution reaction; the input features include two categories: atomic features and statistical features. The atomic features are the intrinsic properties of the two metal elements in the binary alloy, and the statistical features are statistical values ​​calculated for the corresponding atomic properties of the two metal elements in the binary alloy. This embodiment, by clearly defining the prediction target and the core classification of the input features, lays a standardized and highly adaptable foundation for the entire technical solution from the data source. On the one hand, the overpotential in the acidic oxygen evolution reaction is precisely anchored as the target feature, directly aligning with the core application scenario of this invention. Overpotential is a recognized core indicator in the field that can be directly quantified to evaluate OER catalytic activity, completely solving the problem of ambiguous prediction targets and disconnection from actual application scenarios, ensuring that the core objectives of model training are fully matched with the actual needs of catalyst development. On the other hand, the input features are clearly divided into two categories: intrinsic atomic features of elements and statistical features of binary combinations. This covers both the intrinsic basic properties that determine the electronic structure, adsorption characteristics, and bonding ability of metal elements, and captures the combined features related to the synergistic effect of binary alloys. This solves the problem of the one-sided feature selection in existing technologies, which cannot cover the core factors affecting the catalytic performance of binary alloys, avoiding insufficient model fitting accuracy and inaccurate predictions due to the lack of core features. At the same time, the definition boundaries of the features are clearly defined, ensuring the reproducibility and standardization of the dataset construction.

[0065] In some embodiments of the present invention, the atomic features include atomic number, atomic mass, atomic radius, van der Waals radius, covalent radius, electronegativity, melting point, boiling point, density, first ionization energy, second ionization energy, first electron affinity, number of valence electrons, number of outermost d orbitals, number of outermost s orbitals, and work function; the statistical features include the sum, difference, mean, variance, and standard deviation of each of the atomic features; specifically, the sum of atomic numbers, the difference of atomic numbers, the mean of atomic numbers, the variance of atomic numbers, and the standard deviation of atomic numbers; other atomic features follow the same pattern. This embodiment further clarifies the specific dimensions and complete composition of the input features, achieving full coverage of the core factors affecting the OER catalytic performance of binary alloys, and ensuring the predictive potential of the model from the data source. On the one hand, the 16 defined atomic features cover the core intrinsic parameters that determine the electronic structure, intermediate adsorption capacity, crystal structure, and physicochemical stability of transition metals. Among them, the outermost d orbital number, electronegativity, ionization energy, and covalent radius are all key attributes directly related to the energy barrier and active site adsorption characteristics of the four-step OER reaction, with no core influencing factors omitted. On the other hand, the defined statistical features comprehensively cover the differences, average levels, and dispersion of the corresponding attributes of the two elements, which can accurately quantify the synergistic effect characteristics of the two elements in the binary alloy, solving the technical problem of blind and arbitrary feature selection in existing technologies. By clarifying the specific composition of the 97 initial features, the comprehensiveness of the feature set is ensured, and the standardization of feature definitions is achieved, providing a unified and reliable input benchmark for subsequent feature optimization and model training. This fundamentally avoids the problems of insufficient model accuracy and weak generalization ability caused by an imperfect feature system.

[0066] In some embodiments of the present invention, step 3, the multi-round screening and optimization sequentially includes three stages: feature importance analysis, correlation analysis, and recursive feature elimination. The specific steps are as follows:

[0067] The first round is feature importance analysis: features with an importance greater than 0.3% are selected and those that meet the requirements are retained.

[0068] In the second round, Pearson correlation analysis was performed: linear correlation was calculated on the features retained in the first round. For two features with an absolute linear correlation coefficient higher than 0.9, the feature with the lower feature importance was deleted, thus completing the removal of redundant features.

[0069] The third round, recursive feature elimination: through iterative model training, features that contribute the least to the model's predictive performance are gradually removed, using the model's coefficient of determination R0 as the criterion. 2The maximum value serves as the termination condition, ultimately selecting the optimal feature set. This implementation constructs a progressive, rule-defined, and reproducible feature optimization system, specifically addressing the core problems of feature redundancy, severe multicollinearity, and low-contribution features interfering with model training in existing technologies. It is a key improvement for achieving a dual enhancement of model accuracy and efficiency. First, through the first round of feature importance analysis, with a clear threshold of 0.3%, invalid features that contribute almost nothing to prediction performance are quickly eliminated, significantly reducing feature dimensionality and subsequent computational load, while fully preserving core influencing features. Second, through the second round of Pearson correlation analysis, highly linearly correlated redundant features are accurately eliminated, solving the problems of distorted model weight allocation and increased overfitting risk caused by multicollinearity. Third, through the third round of recursive feature elimination, using the model determination coefficient R... 2 The maximum value serves as an objective termination condition, achieving optimal simplification of the feature set and maximizing prediction accuracy with the fewest possible features, thus balancing model complexity, computational efficiency, and prediction performance. Overall, this progressive feature engineering design significantly improves the model's prediction accuracy, generalization ability, and computational efficiency. Furthermore, through explicit selection thresholds and execution rules, it ensures the standardization and reproducibility of the feature optimization process, avoiding arbitrariness in feature selection and ultimately locking in the optimal feature set.

[0070] In some embodiments of the present invention, in step 2, the dataset constructed in step 1 is divided into a training set and a test set in an 8:2 ratio; the training set is used for initial model training and five-fold cross-validation, and the test set is used for final prediction accuracy verification of the model. This clarifies the standardized partitioning rules of the dataset and the usage boundaries of each dataset, solving the problems of non-standard dataset partitioning, data leakage, and distorted model evaluation results in the prior art, and ensuring the rigor of model training and evaluation. On the one hand, the 8:2 ratio is the classic optimal ratio in structured data regression tasks that balances the sufficiency of model training with the reliability of validation. It ensures that there are enough labeled samples for model training, allowing the model to fully learn the structure-activity relationship between features and overpotentials, while reserving an independent test set that is not involved in training for unbiased validation of the model's true prediction accuracy and generalization ability. On the other hand, it clarifies the exclusive uses of the training and test sets, strictly distinguishing between the "model training and cross-validation stage" and the "final accuracy validation stage." This completely avoids the problem of inflated model accuracy and insufficient generalization ability caused by the early participation of test set data in training, ensuring that the final selected model has the ability to make true extrapolation predictions for unknown catalytic systems, rather than just overfitting the training data.

[0071] In some embodiments of the present invention, in step 2, the five-fold cross-validation method is executed as follows: the training set is divided into five equal subsets, and four subsets are used as the training subset and the remaining subset as the validation subset in turn to iteratively train and validate the machine learning model; based on the average prediction accuracy of the five rounds of validation results, the generalization ability of the machine learning model is evaluated, and the model hyperparameters are optimized. This specifically solves the problems of model overfitting, distorted generalization ability evaluation, and insufficient hyperparameter optimization in the small sample training scenario of the present invention. The labeled dataset of this invention contains only 130 DFT calculation samples, which is a typical small-sample training scenario. The five-fold cross-validation method in this implementation maximizes the use of limited labeled data by alternately dividing the training and validation subsets. This ensures that there are enough samples to complete model training in each iteration, and through multiple rounds of cross-validation, it comprehensively and objectively evaluates the model's generalization ability, avoiding evaluation bias caused by a single dataset partition. At the same time, this process simultaneously achieves global optimization of model hyperparameters, ensuring that the model hyperparameters are adapted to the full training data, rather than just fitting a specific training subset. This further improves the model's stability and predictive ability for unknown catalytic systems, solving the problem of insufficient validation of model generalization ability in existing technologies, which easily leads to "high training accuracy but poor actual prediction effect". This ensures that the finally selected optimal model has reliable engineering application value.

[0072] In some embodiments of this invention, the candidate machine learning models include Adaptive Boosting Model (ABR), Gradient Boosting Regression Model (GBR), K-Nearest Neighbor Regression Model (KNR), Kernel Ridge Regression Model (KR), LAS Regression Model (LAS), Partial Least Squares Regression Model (PLS), Random Forest Regression Model (RFR), Relevance Vector Regression Model (RDG), Extreme Gradient Boosting Regression Model (XGBR), and Support Vector Regression Model (SVR). This candidate model library, covering multiple algorithm principles, solves the core problem of existing technologies that directly apply a single model and cannot guarantee the model's adaptability to the application scenario, and is a prerequisite for achieving accurate selection of the optimal model. The 10 machine learning models included in this invention cover regression models with completely different algorithm principles, such as ensemble learning, linear regression, kernel methods, nearest neighbor algorithms, and support vector machines. Different models have different focuses on the fitting ability of data distribution and feature-target correlation, which can avoid the inherent defects and scenario adaptability limitations of single models. In fact, the model screening method of this invention has strong versatility and universality in its core screening logic and evaluation system. It can not only accurately screen these 10 machine learning models, but also accurately and reliably evaluate the suitability and rank the merits of other machine learning models adapted to this application scenario. The screening results have high credibility and will not cause inaccuracies due to adjustments in model type. All supervised learning regression machine learning models suitable for structured feature input and continuous value (overpotential) output can be accurately screened by this method, including but not limited to: ensemble learning models LightGBM, CatBoost, ExtraTrees, etc.; classic statistical learning models: elastic network regression, Bayesian regression, Gaussian process regression, etc.; deep learning models adapted to small sample structured data: shallow fully connected neural networks, multilayer perceptrons (MLP), etc.; graph machine learning models adapted to catalysis scenarios: graph neural networks (GNN) for alloy structures, etc. (only the standardization processing of input features needs to be adjusted accordingly, and the core screening and evaluation logic remains completely unchanged). By incorporating multiple types of candidate models, this invention enables unbiased horizontal comparisons, selecting the optimal model best suited to the specific scenario of "prediction of the catalytic performance of binary alloy acidic OER" from models with different algorithmic logics. This ensures the accuracy and reliability of the final prediction model from the source of model selection, and also verifies the compatibility and universality of the model selection system of this invention with different types of models.

[0073] In some embodiments of the present invention, in step 4, the preset accuracy evaluation index is the coefficient of determination R. 2 The selection criteria, along with the mean squared error (MSE), are the coefficient of determination (R²). 2The highest and lowest mean squared error (MSE) are achieved. This invention clarifies the objective and quantitative criteria for determining the optimal model, solving the problems of existing technologies lacking authoritative evaluation indicators and having subjectively unreliable selection results. On one hand, the selected coefficient of determination R... 2 The mean squared error (MSE) is a universally accepted and authoritative core evaluation indicator in regression prediction tasks, where R0 is the mean squared error. 2 The model's ability to explain the structure-activity relationship between features and overpotentials is measured; a value closer to 1 indicates a stronger model fit and generalization ability. MSE measures the average deviation between predicted and actual values; a value closer to 0 indicates higher prediction accuracy. Both metrics have clear industry consensus and are quantifiable. On the other hand, setting "R..." 2 The dual screening criteria of "highest performance and lowest MSE" take into account both the model's structure-activity relationship fitting ability and the numerical accuracy of actual predictions, avoiding the one-sidedness of evaluation by a single indicator. This ensures that the selected model is the scenario-adaptive model with the best overall performance, provides an objective and reproducible basis for model selection, and guarantees the fairness and reliability of comparisons between different models.

[0074] II. A method for predicting the catalytic performance of a binary alloy acidic oxygen evolution catalyst

[0075] The optimal machine learning model obtained using the above screening method is used to predict catalytic performance, including the following steps:

[0076] S1: Obtain the input features corresponding to the binary alloy to be predicted, wherein the input features are the atomic features and statistical features defined above;

[0077] S2: Input the input features into the optimal machine learning model;

[0078] S3: Using the optimal machine learning model, the overpotential of the binary alloy to be predicted in the acidic oxygen evolution reaction is obtained, thus completing the prediction of catalytic performance.

[0079] The method for predicting the catalytic performance of binary alloy acidic oxygen evolution catalysts described in this invention relies on the scenario-adapted optimal machine learning model obtained from the aforementioned screening method. For unknown binary alloy catalytic systems without theoretical calculations, the average error between the predicted overpotential and the density functional theory gold standard calculation is only 5.03%, achieving ultra-high accuracy prediction with a high degree of agreement with first-principles calculations. This method can directly replace the time-consuming and computationally expensive full-scale theoretical calculations. Simultaneously, it can complete batch prediction and performance ranking of hundreds of binary alloy catalytic structures formed by 26 transition metals in a short time, achieving an order-of-magnitude improvement in R&D efficiency compared to traditional techniques. This method completely solves the screening problem caused by the vast compositional space of binary alloy systems. This method addresses the technical problems of low efficiency and easy omission of high-performance catalytic combinations. Its predicted results are highly consistent with the actual experimental catalytic performance, and can directly guide the preparation and development of high-performance catalysts. It solves the long-standing problem of a serious disconnect between theoretical prediction and experimental implementation in this field. It also significantly reduces the ineffective consumption of supercomputing power, precious metal raw materials, experimental consumables and equipment time from the application end, and significantly lowers the R&D threshold and the overall cost of high-performance acidic oxygen evolution catalysts. At the same time, the method has strong universality and scalability, and can be directly transferred and adapted to complex catalytic systems such as ternary alloys and multi-element alloys, as well as the performance prediction of other electrocatalytic reaction scenarios such as hydrogen evolution and carbon dioxide reduction. It has extremely high promotion value and reusability.

[0080] III. A Catalytic Performance Prediction System for a Binary Alloy Acidic Oxygen Evolution Catalyst

[0081] The present invention also provides a prediction system for performing the above prediction method, specifically including the following:

[0082] Data collection module: used to acquire the atomic and statistical characteristics of the two metal elements in the binary alloy catalyst system, as well as the overpotential target characteristics of the binary alloy acid oxygen evolution reaction, and send the collected data to the data preprocessing module and the model building and training optimization module respectively.

[0083] Data preprocessing module: It receives raw data sent by the data collection module, completes missing value imputation and outlier removal operations through the built-in cleaning engine, integrates input features and target features to construct a dataset, and sends the processed dataset to the feature engineering optimization module.

[0084] Feature Engineering Optimization Module: This module receives the dataset sent by the data preprocessing module, performs feature engineering optimization on atomic and statistical features, removes redundant and low-contribution features, selects the optimal feature set that is strongly correlated with overpotential prediction, and sends the optimal feature set to the model building and training optimization module.

[0085] Model building and training optimization module: It receives the standardized dataset sent by the data preprocessing module and the optimal feature set sent by the feature engineering optimization module, builds multiple candidate machine learning models adapted to the catalytic performance prediction task, completes model training, five-fold cross-validation, hyperparameter optimization and accuracy evaluation, and finally selects and determines the optimal machine learning model adapted to the scenario, and sends the optimal model to the catalytic performance prediction and verification module.

[0086] Catalytic performance prediction and verification module: This module receives the optimal machine learning model sent by the model building and training optimization module, completes batch prediction and performance ranking of the overpotential of the acid oxygen evolution reaction of the binary alloy to be predicted, screens the optimal candidate binary alloy, and simultaneously completes the preparation of the target catalyst and the verification of its acid oxygen evolution performance.

[0087] IV. Examples

[0088] Reference Figure 1 The specific steps are as follows:

[0089] Step 1: Obtain the dataset for training the machine learning model;

[0090] Specifically, a binary alloy model is created in Materials Studio software. Figure 2For binary alloy material systems and structural models, the vertex sites of the two alloys were designated as catalytic active sites, thereby generating 650 unique catalytic structures. Density functional theory calculations were performed using VASP software. Overpotentials were obtained by DFT calculations on 25% of the material space (130 systems), while the activity of the remaining 75% was predicted by a well-trained machine learning model. The dataset contains 97 features, divided into two main categories: atomic features and statistical features. Atomic features include atomic number, atomic mass, atomic radius, van der Waals radius, covalent radius, electronegativity, melting point, boiling point, density, first ionization energy, second ionization energy, first electron affinity, number of valence electrons, number of outermost d orbitals, number of outermost s orbitals, and work function. Statistical features include the sum, difference, mean, variance, and standard deviation of atomic numbers; the sum, difference, mean, variance, and standard deviation of atomic masses; the sum, difference, mean, variance, and standard deviation of atomic radii; and the sum, difference, mean, variance, and standard deviation of van der Waals radii. The dataset includes the sum, difference, mean, variance, and standard deviation of covalent radii; the sum, difference, mean, variance, and standard deviation of electronegativity; the sum, difference, mean, variance, and standard deviation of melting points; the sum, difference, mean, variance, and standard deviation of boiling points; the sum, difference, mean, variance, and standard deviation of densities; the sum, difference, mean, variance, and standard deviation of first ionization energies; the sum, difference, mean, variance, and standard deviation of second ionization energies; the sum, difference, mean, variance, and standard deviation of first electron affinity energies; and the sum, difference, mean, variance, and standard deviation of valence electron numbers. This embodiment constructs a dataset containing 130 target features and 97 descriptors, which is then used to train a machine learning model.

[0091] Step 2: Train the model based on the dataset, and optimize the feature descriptors of the machine learning model using feature engineering techniques. Select the optimal model for predicting the catalytic performance of binary alloys. Specifically, first, select 10 different machine learning models, and divide the dataset into training and testing sets in an 8:2 ratio for training these 10 models. The coefficient of determination (R²) is then used to predict the performance of the binary alloy catalytic performance. 2 The stability and prediction accuracy of all models are evaluated using the mean squared error (MSE), and the corresponding results are as follows: Figure 3 As shown in a. The calculation formulas for these two indicators are as follows:

[0092] ;

[0093] ;

[0094] Where y i , and These are the predicted value, the actual value, and the average value, respectively. R 2 The score ranges from 0 to 1. R 2Models with a value (MSE) close to 1 (0) exhibit better model performance.

[0095] Then, five-fold cross-validation was used to train these ten machine learning models, and the results are as follows: Figure 3 As shown in b, based on the machine learning training results and 5-fold cross-validation results, GBR and RFR models were initially selected for subsequent steps.

[0096] In feature engineering, the first step is to perform feature importance analysis to identify the descriptors that have the most significant impact on model predictions. For example... Figure 4 a and Figure 4 As shown in b, using an importance greater than 0.3% as the criterion for feature retention, RFR retained the top 40 features, and GBR retained the top 22 features. Subsequently, correlation analysis was used to quantify the linear correlation between features. For two features with an absolute correlation coefficient higher than 0.9, the feature with the relatively lower importance was deleted. The results are shown in... Figure 5 a and Figure 5 As shown in b, RFR retains 30 features, while GBR retains 20 features. Next, recursive feature elimination is employed, iteratively removing features that contribute the least to prediction performance through model training. Figure 6 a and Figure 6 As shown in b, when the number of features is 10, the R of RFR is... 2 The maximum value is 0.773; when the number of features is 11, the R of GBR is... 2 The maximum value is 0.794.

[0097] After feature engineering, GBR retains 11 features: the number of outermost d orbitals of adsorbed atoms, boiling point of adsorbed atoms, variance of second ionization energy, standard deviation of van der Waals radius, density difference, difference of covalent radius, melting point of adsorbed atoms, standard deviation of atomic mass, difference of first ionization energy, difference of melting point, and covalent radius of coordinating atoms. RFR retains 10 features: the number of outermost d orbitals of adsorbed atoms, standard deviation of van der Waals radius, melting point of adsorbed atoms, boiling point, density difference, difference of first ionization energy, variance of valence electrons, difference of covalent radius, mean atomic radius, and standard deviation of atomic mass. These retained features are selected as optimized input data for subsequent model training, such as... Figure 7 As shown, the GBR model achieves R on the test set. 2 The value is 0.8935, which is better than the RFR test set R. 2 The value is 0.8810. This result indicates that, in the catalyst performance prediction system of this invention, the GBR model exhibits superior fitting performance and prediction accuracy.

[0098] Finally, the trained GBR model was saved, predicting the overpotentials of 520 catalytic units that failed density functional theory (DFT) calculations. By integrating these machine learning predictions with DFT calculations, a final binary alloy database was constructed. Simultaneously, to verify the reliability and accuracy of the machine learning predictions, three binary alloys were randomly selected, and their oxygen evolution reaction overpotentials were calculated using DFT. The overpotentials calculated by DFT and the machine learning predictions were systematically compared, and the results are as follows: Figure 8 The average error shown is only 5.03%, indicating that the deviation between the machine learning prediction and the density functional theory calculation is extremely small. This confirms that the GBR model can accurately predict the catalytic performance of binary alloys in the oxygen evolution reaction. Data from binary alloy databases show that the FeRu binary alloy has excellent catalytic activity in the oxygen evolution reaction. Based on this prediction, the FeRu alloy was ultimately selected as the target oxygen evolution reaction catalyst in this invention.

[0099] Step 3: Preparation of the target catalyst and testing of its catalytic performance.

[0100] Specifically, FeRu precursors were prepared by electrodeposition, and FeRu alloys were obtained by Joule heating. Figure 9 The image shows a scanning electron microscope image of the FeRu alloy. The characterization results show that the alloy has a rough specific surface area, which exposes more active sites. Figure 10 The aberration-corrected HAADF-STEM image and the corresponding FeRu element mapping were obtained, and the results showed that the FeRu elements were relatively uniformly distributed.

[0101] Next, using a three-electrode system, the OER performance of the FeRu alloy was measured in 0.5M H2SO4 solution. Figure 11 It can be seen that the overpotential of the FeRu alloy reaches 10 mA cm⁻¹ at 171 mV. -2 The current density is superior to that of the control sample. From Figure 12 As can be seen from this, the Tafel slope of the FeRu alloy (42.46 mV dec) -1 The smaller size directly indicates a more robust OER reaction kinetic in the FeRu alloy. Chronopotential measurements demonstrate its superior stability, such as... Figure 13 As shown, after long-term cycling, the potential loss rate of the FeRu alloy is only 24.6 μV h. -1 .

[0102] Analysis of the examples shows that:

[0103] The machine learning model screening method and catalytic performance prediction method for binary alloy acidic oxygen evolution catalysts constructed in this invention possess highly reliable and quantifiable verification capabilities and engineering application value. In the embodiments, the optimal GBR model was finally selected by conducting horizontal comparative training and 5-fold cross-validation on 10 machine learning models covering different algorithm principles, combined with three rounds of progressive feature engineering optimization, and finally achieving the coefficient of determination R on the test set. 2 With an overpotential as high as 0.8935, the average error between the predicted overpotential and the DFT gold standard calculated value for randomly selected unknown binary alloy catalytic systems was only 5.03%, directly verifying that the model screening system of this invention can accurately lock the optimal model suitable for predicting the performance of acidic oxygen evolution catalysis. The progressive feature engineering can significantly improve the prediction accuracy, generalization ability, and computational efficiency of the model while greatly simplifying the feature dimensions, achieving ultra-high accuracy and high reliability prediction for unknown catalytic systems. At the same time, the example only used 130 sets of labeled samples calculated by DFT to complete the model training, and could quickly complete the full performance prediction and ranking of the remaining 520 sets of binary alloy catalytic systems, fully demonstrating that this invention can significantly shorten the catalyst development cycle and significantly reduce the ineffective consumption of supercomputing power, precious metal raw materials, experimental consumables, and equipment time, fundamentally solving the problems of low efficiency and high cost of traditional R&D models. In addition, the FeRu binary alloy catalyst prepared based on the prediction results of the optimal model in the example showed comprehensive performance far exceeding the commercial benchmark in the acidic oxygen evolution reaction, with a performance of 10 mA·cm -2 The overpotential at the current density is only 171 mV, and the Tafel slope is as low as 42.46 mV·dec. -1 The catalytic activity and reaction kinetics are significantly better than those of commercial RuO2 and pure Ru catalysts, and at 100 mA·cm⁻¹, the reaction is also superior. -2 It can operate stably for 500 hours at current density, with a potential decay rate of only 24.6 μV·h. -1 It possesses both excellent intrinsic catalytic activity and long-term operational durability, fully verifying that the predicted results of this invention are highly consistent with the actual experimental performance. It has completely opened up a closed-loop R&D system of "data-algorithm-experiment" and solved the core problem of the long-standing disconnect between theoretical prediction and experimental implementation in this field.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A method for screening binary alloy acidic oxygen evolution catalysts using machine learning models, characterized in that, The method described above is used to screen models suitable for predicting the catalytic performance of binary alloy oxygen evolution catalysts. The specific steps are as follows: Step 1: Dataset Construction: Taking binary alloy catalysts formed by any two combinations of 26 transition metal elements as the research object, a surface catalytic structure model of the binary alloy catalyst is established, and the vertex sites of the structure model are designated as the catalytic active sites for the acidic oxygen evolution reaction. The target features corresponding to each catalytic structure are obtained by density functional theory calculation, and the input features of the corresponding binary alloy are extracted. The target features and input features are integrated to construct a dataset for training machine learning models. Step 2: Model Training and Cross-Validation: Divide the dataset constructed in Step 1 into a training set and a test set. Based on the training set, perform initial training on multiple candidate machine learning models. At the same time, use the five-fold cross-validation method to evaluate the generalization ability and prediction accuracy of each candidate machine learning model, and initially select the model for subsequent training. Step 3 Feature Engineering Optimization: Perform multiple rounds of screening and optimization on the input features obtained in Step 1 to remove redundant and low-contribution features, and obtain the feature set with the best correlation to the target features; Step 4: Optimal Model Selection: Using the feature set obtained from Step 3 as input parameters, and based on the evaluation results of the initial training and five-fold cross-validation in Step 2, the optimal machine learning model suitable for predicting the performance of binary alloy acidic oxygen evolution catalysts is selected from the initially selected machine learning models, according to the preset accuracy evaluation index.

2. The screening method according to claim 1, characterized in that, In step 3, the target feature is the overpotential of the binary alloy catalyst in the acidic oxygen evolution reaction; the input features include two categories: atomic features and statistical features. The atomic features are the intrinsic properties of the two metal elements in the binary alloy, and the statistical features are the statistical values ​​calculated for the corresponding atomic properties of the two metal elements in the binary alloy.

3. The screening method according to claim 2, characterized in that, The atomic characteristics include atomic number, atomic mass, atomic radius, van der Waals radius, covalent radius, electronegativity, melting point, boiling point, density, first ionization energy, second ionization energy, first electron affinity, number of valence electrons, number of outermost d orbitals, number of outermost s orbitals, and work function; the statistical characteristics include the sum, difference, mean, variance, and standard deviation of each of the atomic characteristics.

4. The screening method according to claim 1, characterized in that, In step 3, the multi-round screening and optimization process includes three stages: feature importance analysis, correlation analysis, and recursive feature elimination. The specific steps are as follows: The first round is feature importance analysis: features with an importance greater than 0.3% are selected and those that meet the requirements are retained. In the second round, Pearson correlation analysis was performed: linear correlation was calculated on the features retained in the first round. For two features with an absolute linear correlation coefficient higher than 0.9, the feature with the lower feature importance was deleted, thus completing the removal of redundant features. The third round, recursive feature elimination: through iterative model training, features that contribute the least to the model's predictive performance are gradually removed, using the model's coefficient of determination R0 as the criterion. 2 The maximum value is used as the termination condition, and the optimal feature set is finally obtained through screening.

5. The screening method according to claim 1, characterized in that, In step 2, the dataset constructed in step 1 is divided into a training set and a test set in an 8:2 ratio; the training set is used for initial model training and five-fold cross-validation, and the test set is used for final prediction accuracy verification of the model.

6. The screening method according to claim 5, characterized in that, In step 2, the five-fold cross-validation method is executed as follows: the training set is divided into five equal subsets, and four subsets are used as the training subset and the remaining subset is used as the validation subset in turn to iteratively train and verify the accuracy of the machine learning model. Based on the average prediction accuracy of the five rounds of validation results, the generalization ability of the machine learning model is evaluated, and the model hyperparameters are optimized.

7. The screening method according to claim 1, characterized in that, The alternative machine learning models include adaptive boosting model, gradient boosting regression model, K-nearest neighbor regression model, kernel ridge regression model, Lasso regression model, partial least squares regression model, random forest regression model, correlation vector regression model, extreme gradient boosting regression model, and support vector regression model.

8. The screening method according to claim 1, characterized in that, In step 4, the preset accuracy evaluation index is the coefficient of determination R. 2 The selection criteria, along with the mean squared error (MSE), are the coefficient of determination (R²). 2 It has the highest mean square error (MSE) and the lowest mean square error (MSE).

9. A method for predicting the catalytic performance of a binary alloy acidic oxygen evolution catalyst, characterized in that, The optimal machine learning model obtained by the screening method according to any one of claims 1 to 9 is used to predict catalytic performance, comprising the following steps: S1: Obtain the input features corresponding to the binary alloy to be predicted, wherein the input features are the atomic features and statistical features defined in claim 2; S2: Input the input features into the optimal machine learning model; S3: Using the optimal machine learning model, the overpotential of the binary alloy to be predicted in the acidic oxygen evolution reaction is obtained, thus completing the prediction of catalytic performance.

Images