Machine learning based anti-hot sintering supported metal catalyst screening method

Abstract

This invention discloses a machine learning-based method for screening heat-resistant sintering supported metal catalysts, involving the interdisciplinary field of catalyst development and machine learning. The specific steps of this method are as follows: Using the interface system of metal nanoparticles and oxide supports as the screening object, and the interface adhesion energy as the prediction target, a wetting contact angle close to 90° is set as the criterion for heat resistance to sintering; first, a standardized dataset is constructed, then two types of interface descriptors are constructed as input features, and a prediction model is trained using the SISSO algorithm to predict the interface adhesion energy. Finally, the contact angle is calculated and ranked accordingly, outputting candidate material combinations. This invention sets evaluation criteria, constructs a standardized dataset, and innovatively designs element-level interface descriptors to achieve precise and scientific catalyst screening; it uses an adaptation algorithm to train the model, combines an association algorithm to transform parameters, and designs ranking rules to achieve efficient screening and performance quantification, combining machine learning and interface science to provide technical support for catalyst development.

Description

Technical Field

[0001] This invention relates to the field of catalyst development and machine learning, specifically a machine learning-based method for screening thermally sintering-resistant supported metal catalysts. Background Technology

[0002] Supported metal catalysts are core materials for high-temperature catalytic reactions in chemical synthesis, energy conversion, and environmental remediation. Their resistance to thermal sintering during high-temperature service directly determines the stability of catalytic activity and service life, becoming a core consideration in the research and optimization of this type of catalyst. The interfacial interaction characteristics between metal nanoparticles and metal oxide supports are key to regulating the catalyst's resistance to thermal sintering, directly reflecting the bonding state between the metal and the support, and thus determining whether metal nanoparticles are prone to agglomeration, migration, or other sintering behaviors under high-temperature conditions. With the rapid application of machine learning technology in materials science, high-throughput screening of materials using machine learning has become an important direction for breaking through traditional research and development models. Combining machine learning with the physical essence of interface science to develop screening methods for thermally sintering-resistant supported metal catalyst systems has become an important requirement in the field of catalysis research and development, and also provides a new technical path for the rapid discovery of high-performance catalysts.

[0003] Traditional methods for screening the heat resistance of supported metal catalysts rely on experimental trial and error, involving extensive manual sample preparation and performance testing. This approach is not only time-consuming and costly in terms of manpower and resources, but also struggles to achieve high-throughput screening of a vast number of metal-oxide support candidate systems, failing to quickly uncover potential high-performance combinations. Furthermore, traditional methods lack standardized processing procedures in data processing, failing to perform targeted screening, normalization, and outlier / missing value handling on various types of experimental data, resulting in inconsistent data quality and hindering reliable foundational support for subsequent analysis and modeling. Additionally, the parameter design used to characterize interface properties in traditional methods does not align with the physical nature of interface interactions, requiring additional testing or simulation parameters for some characterization methods, leading to insufficient accuracy. Moreover, model construction often employs algorithms lacking physical interpretability, limiting the reference value of prediction results. Performance evaluation and ranking also lack unified scientific rules, making it difficult to accurately screen catalysts that meet the heat resistance requirements. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a machine learning-based method for screening heat-resistant sintering supported metal catalysts. Using interfacial adhesion energy as the prediction target and a wetting contact angle close to 90° as the criterion for heat resistance to sintering, the method first collects and processes various types of relevant data into a standardized dataset. Then, it designs two types of interface descriptors—metal oxyphilicity and metal-support metal interaction—as input features. An interfacial adhesion energy prediction model is trained using the SISSO algorithm. After predicting the adhesion energy using the model, the wetting contact angle is calculated. Finally, candidate systems are ranked according to scientific rules. This method deeply integrates machine learning with the scientific and physical essence of interfaces. Both the descriptor and model construction rely on inherent thermodynamic properties, preserving physical interpretability and replacing the traditional experimental trial-and-error mode. This achieves efficient and high-throughput screening of catalyst systems, providing a precise and reliable technical method for the development of heat-resistant sintering supported metal catalysts, and can be widely adapted to the research and development needs in the catalysis field.

[0005] To solve the above-mentioned technical problems, this invention provides the following technical solution: a machine learning-based method for screening thermally sintering-resistant supported metal catalysts, the method comprising: S1, set target criteria: take the interface system composed of metal nanoparticles and metal oxide carrier as the screening object, take the interface adhesion energy as the prediction target, calculate the wetting contact angle of metal on the carrier surface through the interface adhesion energy, and take the wetting contact angle close to 90° as the criterion for evaluating the heat resistance of the system. S2, Construct a standardized dataset: Collect experimental interface adhesion energy data, metal surface energy data, oxide structural information, and basic thermodynamic parameters of metals and oxides at the metal-oxide carrier interface. Filter, normalize, remove outliers, and fill in missing values ​​on the collected data to form a standardized dataset. S3, Constructing Interface Descriptors: Based on the elemental thermodynamic parameters in the standardized dataset, for the loaded metal M and the metal cation M′ in the metal oxide support, the interaction strength between the metal and oxygen is calculated using the oxyphilicity descriptor algorithm, and the interaction strength between the metal and the metal cation is calculated using the metal interaction descriptor algorithm. Unique metal oxyphilicity descriptors and metal-support metal interaction descriptors are generated respectively, which serve as input features for the interface adhesion energy prediction model. The specific process of calculating the interaction strength between metal and oxygen using the oxyphilicity descriptor algorithm and the interaction strength between metal and metal cations using the metal interaction descriptor algorithm to generate two types of elemental-level interface descriptors is as follows: First, extract all elemental-level thermodynamic fundamental parameters corresponding to the loaded metal M and the metal cation M′ in the metal oxide support from the standardized dataset. The extraction process retains the original measurement accuracy and standard state labeling information of the parameters. Then, calculate the oxyphilicity descriptor based on the metal-oxygen interface interaction mechanism. The entire calculation uses the elemental-level thermodynamic fundamental parameters under the standard state of 298K and 101.325kPa from the standardized dataset. The calculation process relies only on the inherent thermodynamic properties of the loaded metal and the support metal cation, without introducing additional experimental test parameters or simulation calculation parameters. After completing the calculation of a single system, a unique metal oxyphilicity descriptor characterizing the metal-oxygen interaction strength is obtained. Then, based on the cross-interface metal... The direct affinity mechanism between the metal and metal cations was used to calculate the metal interaction descriptors. The parameters used in the calculations were the elemental thermodynamic fundamental parameters under the standard conditions of 298 K and 101.325 kPa in the standardized dataset. The calculation process relied solely on the intrinsic thermodynamic parameters of the loaded metal and the support metal cations, without performing additional crystal structure simulations or interface configuration calculations. After completing the calculation of a single system, a unique metal-support metal interaction descriptor characterizing the interaction strength between the metal and the metal cations was obtained. After completing the calculations of both types of descriptors for all the systems under test, the dimensions of all the calculated descriptor values ​​were uniformized to the general dimensions of the thermodynamic fundamental parameters. The dimension processing followed the conversion rules of thermodynamic parameters without changing the original size and relative differences of the descriptor values. After processing, two types of elemental interface descriptors were formed, which served as input features for the interface adhesion energy prediction model. S4, Training and Prediction of Adhesion Energy: Using the interfacial adhesion energy in the standardized dataset as the target variable and two types of interface descriptors as input features, the SISSO interpretable symbolic regression algorithm is used to train the model and fit it to obtain the interfacial adhesion energy prediction model. The interfacial adhesion energy prediction model is then used to predict the interfacial adhesion energy of the metal-oxide support system under test. S5, Calculate and sort the contact angle: Based on the interfacial adhesion energy predicted in step S4, combined with the metal surface energy data, the wetting contact angle is calculated using the adhesion energy-contact angle correlation algorithm. Based on the criterion that the wetting contact angle is close to 90°, the candidate metal-oxide support systems are sorted according to their resistance to thermal sintering, and the sorted candidate material combinations are output.

[0006] Furthermore, the metal nanoparticles are single-component transition metal nanoparticles selected from one of Pt, Pd, Rh, Au, Ag, Ni, Co, and Cu. The metal oxide support is a single-component metal oxide selected from one of Al2O3, ZrO2, CeO2, SiO2, MgO, TiO2, and ZnO. The metal nanoparticles and the metal oxide support form a binary interface system.

[0007] Furthermore, the experimental interface adhesion energy data were collected from SCI core journals in the fields of catalysis and interface science, proceedings of international authoritative conferences, and peer-reviewed public experimental databases. During the collection process, the corresponding test methods, test environment temperature, test vacuum degree, sample preparation process, and surface cleanliness control methods were recorded simultaneously. The oxide structure information was extracted from the MaterialsProject database and the ICSD database. The extracted content includes the stable crystal form of the oxide, crystal structure parameters, and atomic occupancy information. The thermodynamic fundamental parameters of the metal and oxide were collected from the FactSage database and the JANAF database. The collected content includes the metal formation enthalpy, metal sublimation enthalpy, metal mixing formation enthalpy, and oxide thermodynamic stability parameters.

[0008] Furthermore, the screening of collected data is only carried out on experimental interface adhesion energy data. Experimental interface adhesion energy data with consistent measurement methods, clearly controlled or discussed in detail surface cleanliness interference factors, and no less than 3 test repetitions are retained. Metal surface energy data, oxide structural information, and thermodynamic basic parameters of metals and oxides are all retained as original measured and extracted results without screening.

[0009] Furthermore, Z-score normalization was used to normalize all collected data, resulting in a mean of 0 and a variance of 1. Grubbs' criterion was used to remove outliers, with a significance level of 0.05. Data identified as outliers were directly removed from the dataset. Linear interpolation based on the periodic law was used to complete missing values ​​for the fundamental thermodynamic parameters of the metals, and a Miedema semi-empirical thermodynamic model was used to complete missing values ​​for the enthalpy of metal mixing. The Miedema semi-empirical thermodynamic model used for missing value completion is an alloy thermodynamic calculation model based on the intrinsic physicochemical properties of metallic elements. Its core application is the quantitative calculation of the enthalpy of mixing between different metals under infinite dilution conditions. This model uses short-range interactions between metal atoms as its core calculation logic, and the entire calculation relies solely on the intrinsic parameters of the metallic elements. The model eliminates the need for additional control parameters such as crystal structure and interface configuration, using a standardized benchmark of 298 K and 101.325 kPa. It selects three inherent elemental parameters—electronegativity, atomic electron density, and atomic radius—as core calculation bases. Before calculation, the units and accuracy of these three parameters are uniformly calibrated to fully match the parameter specifications of the standardized dataset. In this application, the model only completes the calculation of the mixing enthalpy of the carrier metal cation under infinite dilution in the supported metal, which is missing in the standardized dataset. During calculation, the three standard parameters corresponding to the supported metal and carrier metal cation are directly extracted from the standardized dataset without the introduction or correction of additional parameters. The calculated results retain significant figures and dimensions according to the requirements of the standardized dataset and are directly incorporated into the dataset as the basic thermodynamic parameters for subsequent interface descriptor construction.

[0010] Furthermore, the mathematical expression for the oxygen affinity descriptor algorithm is: ;in This is a descriptor for metal oxyphilicity. The molar enthalpy of formation for the supported metal M to form its thermodynamically most stable oxide is given by values ​​at standard conditions of 298 K and 101.325 kPa. The molar enthalpy of sublimation of the metal cation M' in the metal oxide support is taken as the value under standard conditions of 298 K and 101.325 kPa.

[0011] Furthermore, the mathematical expression for the metal interaction descriptor algorithm is: ;in This is a descriptor for metal-carrier metal interactions. Let be the molar enthalpy of formation of metal cation M' in an infinitely diluted state supported by metal M, and take the value under standard conditions of 298 K and 101.325 kPa. The molar enthalpy of sublimation of the metal cation M' Let M be the molar enthalpy of sublimation of the loaded metal M, and both are taken as values ​​under standard conditions of 298 K and 101.325 kPa.

[0012] Furthermore, the specific process of using the SISSO interpretable symbolic regression algorithm to train and fit the interfacial adhesion energy prediction model is as follows: The standardized dataset is divided into a training set, a validation set, and a test set using stratified random sampling; the interfacial adhesion energy in the training set is used as the target variable, and two types of interface descriptors are used as input features to import into the SISSO algorithm training framework. The algorithm's maximum feature dimension is set to 2, the sparsity parameter to be 1e-3 to 1e-2, and the number of iterations to be no less than 1000. Candidate prediction models are generated through the algorithm's built-in feature independence screening and global sparsity fitting; the two types of interface descriptors from the validation set are input into the candidate prediction models to complete model validation and select the optimal candidate model; then, the two types of interface descriptors from the test set are input into the optimal candidate model to complete independent testing. After confirming that the fitting results meet the preset requirements, the constant parameters and analytical expression of the model are fixed, and finally, the interfacial adhesion energy prediction model is obtained through fitting; the SISSO interpretable symbolic regression algorithm... The proposed algorithm is a symbolic regression algorithm based on deterministic feature selection and sparsification operators. It comprises three core functional modules: feature independence selection, global sparsification fitting, and model quantization optimization. It is adapted to the construction of explicit analytical models between low-dimensional physically interpretable features and target variables. In this invention, the algorithm uses two types of element-level interface descriptors as input feature dimensions and interface adhesion energy as a single target variable for model construction. The algorithm operates solely based on the numerical correlation between the input features and the target variable for fitting operations, without introducing additional feature mapping, dimensionality enhancement, or nonlinear transformation operations. The feature independence selection module employs a feature-by-feature correlation test to verify the correlation between the two types of input interface descriptors and the interface adhesion energy target variable. The global sparsification fitting module uses a preset feature dimension as a constraint to perform basic symbolic operations and numerical fitting on the verified features. The model quantization optimization module quantifies and selects candidate models generated by the fitting based on numerical error evaluation indicators.

[0013] Furthermore, the mathematical expression for the adhesion energy-contact angle correlation algorithm is: ; ;in The absolute value of the predicted interfacial adhesion energy is... The surface energy of the loaded metal M in the collected metal surface energy data is the surface energy value of a smooth and dense metal surface under standard conditions of 298 K and 101.325 kPa. is the wetting contact angle of the metal on the surface of the metal oxide carrier, and arccos is the inverse cosine function, with a calculated value range of 0° to 180°.

[0014] Furthermore, when ranking the heat resistance sintering performance of candidate metal-oxide support systems, the absolute value of the difference between the wetting contact angle and 90° for each candidate system is first calculated. All candidate systems are then ranked in ascending order of absolute value. If multiple candidate systems have the same absolute value, they are then ranked a second time in descending order of absolute value of interfacial adhesion energy. After ranking, the candidate material combinations are output in list form. The list sequentially labels the candidate system with the type of loaded metal, the type of metal oxide support, the predicted value of interfacial adhesion energy, the calculated value of the wetting contact angle, and the absolute value of the difference between the wetting contact angle and 90°.

[0015] Beneficial effects Compared with existing technologies, this machine learning-based screening method for thermally sintered supported metal catalysts has the following advantages: I. This invention establishes evaluation criteria that align with the physical essence of anti-sintering performance, constructs a standardized dataset around the metal-oxide support interface system, and innovatively designs two types of element-level interface descriptors as input features for the prediction model. This achieves precise and scientific construction of the fundamental steps in catalyst screening. Targeted screening is conducted on experimental data, and appropriate normalization, outlier removal, and missing value completion methods are adopted for different types of data to ensure the integrity and reliability of the dataset. This effectively avoids interference from invalid data on subsequent model training. The calculation of the interface descriptors relies solely on the inherent thermodynamic properties of the loaded metal and the support metal cations, without introducing additional test or simulation parameters. Dimensional uniformity processing follows thermodynamic conversion rules, allowing the descriptors to accurately characterize the intensity of interface interactions, aligning with the physical essence of interface adhesion energy, and providing highly adaptable input features for adhesion energy prediction models.

[0016] II. This invention employs an algorithm adapted to low-dimensional physically interpretable features for model training, combines an adhesion energy-contact angle correlation algorithm to achieve accurate conversion of key parameters, and designs multi-dimensional scientific ranking rules. This enables efficient screening and performance quantification of anti-sintering catalyst systems. Model training relies on the core functional modules of the algorithm to complete feature screening and fitting without additional feature mapping or nonlinear transformation, preserving the physical interpretability of the model and making the predicted results of interfacial adhesion energy more credible and referential. The conversion from adhesion energy to wetting contact angle is achieved through a clear correlation algorithm. During ranking, the core criteria are used as the primary basis, followed by auxiliary indicators for secondary ranking, making the performance ranking of candidate systems more in line with practical application needs. The overall method deeply integrates machine learning and interface science theory, replacing the traditional screening method that relies on a large number of experiments, significantly improving screening efficiency and providing reliable technical support for the development of anti-sintering supported metal catalysts.

[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 This is a flowchart of a machine learning-based method for screening thermally sintered supported metal catalysts. Figure 2 A schematic diagram illustrating the construction of the interface descriptor and adhesion energy prediction model. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] Example 1: Machine learning screening of heat-resistant sintering supported metal catalyst systems (Pt / Al2O3 system as the core test system).

[0022] This embodiment uses a binary interface system composed of Pt single-component transition metal nanoparticles and Al2O3 single-component metal oxide support as the core test system. Other Pt group metal-Al2O3 systems, such as Pd / Al2O3 and Rh / Al2O3, are also included as comparative candidate systems. The machine learning screening method of this invention is used to screen for resistance to thermal sintering. Specific steps are as follows: Figure 1 As shown: S1, setting target criteria: The binary interface system composed of Pt nanoparticles and Al2O3 support, as well as other comparative candidate metal-oxide binary interface systems, are selected as screening objects. The interfacial adhesion energy is determined as the model prediction target. The wetting contact angle of the metal on the surface of Al2O3 support is calculated through the interfacial adhesion energy. It is determined that a wetting contact angle close to 90° is the core criterion for evaluating the heat resistance sintering performance of all candidate systems.

[0023] S2, Constructing a Standardized Dataset: First, collect experimental interfacial adhesion energy data, surface energy data of metals such as Pt, oxide structure information of Al2O3, and fundamental thermodynamic parameters of Pt and Al2O3 for all candidate systems. The experimental interfacial adhesion energy data was collected from SCI core journals in the fields of catalysis and interface science, proceedings of authoritative international conferences, and peer-reviewed public experimental databases. During collection, the corresponding test methods, ambient temperature, vacuum level, sample preparation process, and surface cleanliness control methods were recorded simultaneously. The oxide structure information of Al2O3 was extracted from the MaterialsProject and ICSD databases, including the stable crystal form, crystal structure parameters, and atomic occupancy information of Al2O3. The fundamental thermodynamic parameters of Pt and Al2O3 were collected from the FactSage and JANAF databases, including the formation enthalpy, sublimation enthalpy, and mixing formation enthalpy of Pt, as well as the thermodynamic stability parameters of Al2O3. The collected data underwent a series of processing steps. Only the experimental interface adhesion energy data were screened, retaining data with consistent measurement methods, clearly controlled or thoroughly discussed surface cleanliness interference factors, and at least three repeated tests. Metal surface energy data, Al2O3 structural information, and fundamental thermodynamic parameters of metals and oxides retained their original measured and extracted results. Z-score normalization was used to normalize all collected data, ensuring a mean of 0 and a variance of 1. Grubbs' criterion with a significance level of 0.05 was used to remove outliers, directly removing data deemed outliers from the dataset. Linear interpolation based on the periodic law was used to complete missing values ​​for fundamental thermodynamic parameters of metals, and the Miedema semi-empirical thermodynamic model was used to complete missing values ​​for the enthalpy of metal mixing. The completed results retained significant figures and dimensions according to dataset requirements and were directly incorporated into the dataset, ultimately forming a complete standardized dataset.

[0024] S3, Constructing the Interface Descriptor: Based on the elemental-level thermodynamic parameters in the standardized dataset, this involves targeting the metal cation Al in the supported metal Pt and Al2O3 support. 3+ First, extract Pt and Al from the standardized dataset. 3+ All corresponding elemental thermodynamic fundamental parameters were extracted, preserving the original measurement accuracy and standard state annotation information. Then, based on the metal-oxygen interfacial interaction mechanism, an oxygen affinity descriptor algorithm was used to calculate the interaction strength between Pt and oxygen using elemental thermodynamic fundamental parameters under standard states of 298 K and 101.325 kPa from a standardized dataset. The calculation process relied solely on the interaction between Pt and Al. 3+ The inherent thermodynamic properties of oxygen, the mathematical expression for the oxygen affinity descriptor algorithm is: ;in This is a descriptor for metal oxyphilicity. The molar enthalpy of formation for the supported metal M to form its thermodynamically most stable oxide is given by values ​​at standard conditions of 298 K and 101.325 kPa. The molar enthalpy of sublimation of the metal cation M' in the metal oxide support is taken as the value under standard conditions of 298K and 101.325kPa. No additional experimental test parameters or simulation calculation parameters are introduced. After the calculation is completed, a unique metal oxyphilicity descriptor characterizing the strength of the interaction between Pt and oxygen is obtained. Furthermore, based on the direct affinity mechanism between metals and metal cations across interfaces, the metal interaction descriptor algorithm is used to calculate the interaction between Pt and Al using elemental-level thermodynamic parameters under standard conditions of 298 K and 101.325 kPa. 3+ The interaction strength between them was calculated solely based on the interaction between Pt and Al. 3+ The intrinsic thermodynamic parameters of the metal interaction descriptor algorithm are expressed mathematically as follows: ;in This is a descriptor for metal-carrier metal interactions. Let be the molar enthalpy of formation of metal cation M' in an infinitely diluted state supported by metal M, and take the value under standard conditions of 298 K and 101.325 kPa. The molar enthalpy of sublimation of the metal cation M' The molar enthalpy of sublimation of the loaded metal M is taken as the value under standard conditions of 298 K and 101.325 kPa. No additional crystal structure simulation or interface configuration calculation is performed. After the calculation, the unique corresponding characterizations of Pt and Al are obtained. 3+ A descriptor for the metal-carrier metal interaction strength between the two.

[0025] After calculating the above two types of descriptors for all candidate systems, the dimensions of all calculated descriptor values ​​are uniformized to the general dimensions of the thermodynamic fundamental parameters. The dimension processing follows the conversion rules of thermodynamic parameters and does not change the original size and relative difference of the descriptor values. After processing, two types of element-level interface descriptors are formed, which serve as input features for the interface adhesion energy prediction model.

[0026] S4, Training and Predicting Adhesion Energy: Using the interface adhesion energy in the standardized dataset as the target variable and the two types of interface descriptors obtained in step S3 as input features, the SISSO interpretable symbolic regression algorithm is used for model training. First, the standardized dataset is divided into training, validation, and test sets using stratified random sampling. Then, the interface adhesion energy in the training set is used as the target variable, and the two types of interface descriptors are used as input features, imported into the SISSO algorithm training framework. The algorithm's maximum feature dimension is set to 2, the sparsity parameter to 1e-3 to 1e-2, and the number of iterations to be no less than 1000. This is then processed by the algorithm's built-in... The feature independence screening and global sparsification fitting are used to generate candidate prediction models. Then, the two types of interface descriptors from the validation set are input into the candidate prediction models to complete model validation and select the optimal candidate model. Then, the two types of interface descriptors from the test set are input into the optimal candidate model to complete independent testing. After confirming that the fitting results meet the preset requirements, the constant parameters and analytical expression of the model are fixed, and finally the interface adhesion energy prediction model is obtained. The interface adhesion energy prediction model is used to predict the interface adhesion energy of the Pt / Al2O3 core test system and other comparative candidate systems, and the predicted interface adhesion energy values ​​of each system are obtained.

[0027] S5, Calculate and sort the contact angles: Based on the predicted interfacial adhesion energy values ​​of Pt / Al2O3 and other candidate systems obtained in step S4, and combined with the surface energy data of the corresponding metals in each system, the adhesion energy-contact angle correlation algorithm is used to calculate the wetting contact angle of the metal on the oxide support surface in each system. The mathematical expression of the adhesion energy-contact angle correlation algorithm is: ; ;in The absolute value of the predicted interfacial adhesion energy is... The surface energy of the loaded metal M in the collected metal surface energy data is the surface energy value of a smooth and dense metal surface under standard conditions of 298 K and 101.325 kPa. The wetting contact angle of the metal on the metal oxide support surface is denoted as , and arccos is the inverse cosine function, with a calculated value ranging from 0° to 180°. Based on the criterion that the wetting contact angle is close to 90°, the heat resistance sintering performance of all candidate metal-oxide support systems is ranked. First, the absolute value of the difference between the wetting contact angle and 90° is calculated for each candidate system. The candidate systems are initially ranked in ascending order of this absolute value. If multiple candidate systems have the same absolute value, they are then ranked in descending order of the absolute value of the interfacial adhesion energy. After ranking, the candidate material combinations are output in list form. The list sequentially labels the loaded metal type, metal oxide support type, predicted interfacial adhesion energy, calculated wetting contact angle, and absolute value of the difference between the wetting contact angle and 90° for each candidate system. This completes the screening of the heat resistance sintering performance of the Pt / Al2O3 system and the comparative candidate systems.

[0028] In summary, this embodiment uses Pt / Al2O3 as the core test system, and compares it with other Pt group metal-Al2O3 systems, strictly following the machine learning screening method of this invention to complete the entire process of heat resistance sintering performance screening. Starting from the criterion of a wetting contact angle close to 90°, multiple types of data are collected and processed to form a standardized dataset, and Pt and Al2O3 are specifically generated for comparison. 3+ The two types of interface descriptors were trained using the SISSO interpretable symbolic regression algorithm to obtain a reliable adhesion energy prediction model. Finally, the contact angle was calculated using the adhesion energy-contact angle correlation algorithm and sorted according to rules to output a fully labeled list of candidate materials. This fully verified the applicability and operability of the method in the screening of Pt group metal-Al2O3 binary interface systems.

[0029] Example 2: Machine learning screening of heat-resistant sintering supported metal catalyst systems (Pd / ZrO2 system as the core test system).

[0030] This embodiment uses a binary interface system composed of Pd single-component transition metal nanoparticles and ZrO2 single-component metal oxide support as the core test system, while incorporating other transition metal-ZrO2 systems such as Ni / ZrO2 and Co / ZrO2 as comparative candidate systems. The machine learning screening method of this invention is used to screen for resistance to thermal sintering. Specific steps are as follows: Figure 2 As shown: S1, setting target criteria: The binary interface system composed of Pd nanoparticles and ZrO2 support, as well as other comparative candidate metal-oxide binary interface systems, are selected as screening objects. The interfacial adhesion energy is determined as the model prediction target. The wetting contact angle of the metal on the ZrO2 support surface is calculated through the interfacial adhesion energy. It is determined that a wetting contact angle close to 90° is the core criterion for evaluating the heat resistance sintering performance of all candidate systems.

[0031] S2, Constructing a Standardized Dataset: First, collect experimental interfacial adhesion energy data, surface energy data of metals such as Pd, oxide structure information of ZrO2, and fundamental thermodynamic parameters of Pd and ZrO2 for all candidate systems. The experimental interfacial adhesion energy data was collected from SCI core journals in the fields of catalysis and interface science, proceedings of authoritative international conferences, and peer-reviewed public experimental databases. During collection, the corresponding test methods, ambient temperature, vacuum level, sample preparation process, and surface cleanliness control methods were recorded simultaneously. The oxide structure information of ZrO2 was extracted from the MaterialsProject and ICSD databases, including the stable crystal form, crystal structure parameters, and atomic occupancy information of ZrO2. The fundamental thermodynamic parameters of Pd and ZrO2 were collected from the FactSage and JANAF databases, including the formation enthalpy, sublimation enthalpy, and mixing formation enthalpy of Pd, as well as the thermodynamic stability parameters of ZrO2. The collected data underwent a series of processing steps. Only the experimental interface adhesion energy data were screened, retaining data with consistent measurement methods, clearly controlled or thoroughly discussed surface cleanliness interference factors, and at least three repeated tests. Metal surface energy data, ZrO2 structural information, and fundamental thermodynamic parameters of metals and oxides retained their original measured and extracted results. Z-score normalization was used to normalize all collected data, ensuring a mean of 0 and a variance of 1. Grubbs' criterion with a significance level of 0.05 was used to remove outliers directly from the dataset. Linear interpolation based on the periodic law was used to complete missing values ​​for fundamental thermodynamic parameters of metals, and the Miedema semi-empirical thermodynamic model was used to complete missing values ​​for the enthalpy of metal mixing. The completed results retained significant figures and dimensions according to dataset requirements and were directly incorporated into the dataset, ultimately forming a complete standardized dataset.

[0032] S3, Constructing the Interface Descriptor: Based on the elemental-level thermodynamic parameters in the standardized dataset, this involves targeting the metal cation Zr in the loaded metal Pd and ZrO2 support. 4+ First, extract Pd and Zr from the standardized dataset. 4+ All corresponding elemental thermodynamic fundamental parameters were extracted, preserving the original measurement accuracy and standard state annotation information. Then, based on the metal-oxygen interfacial interaction mechanism, an oxygen affinity descriptor algorithm was used to calculate the interaction strength between Pd and oxygen using elemental thermodynamic fundamental parameters under standard states of 298 K and 101.325 kPa from a standardized dataset. The calculation process relied solely on the interaction between Pd and Zr. 4+ The inherent thermodynamic properties of oxygen, the mathematical expression for the oxygen affinity descriptor algorithm is: ;in This is a descriptor for metal oxyphilicity. The molar enthalpy of formation for the supported metal M to form its thermodynamically most stable oxide is given by values ​​at standard conditions of 298 K and 101.325 kPa. The molar enthalpy of sublimation of the metal cation M' in the metal oxide support is taken as the value under standard conditions of 298K and 101.325kPa. No additional experimental test parameters or simulation calculation parameters are introduced. After the calculation is completed, a unique metal oxyphilicity descriptor characterizing the strength of the interaction between Pd and oxygen is obtained. Furthermore, based on the direct affinity mechanism between metals and metal cations across interfaces, the metal interaction descriptor algorithm is used to calculate the Pd and Zr interactions using elemental-level thermodynamic parameters under standard conditions of 298 K and 101.325 kPa. 4+ The interaction strength between them is calculated solely based on the interaction between Pd and Zr. 4+ The intrinsic thermodynamic parameters of the metal interaction descriptor algorithm are expressed mathematically as follows: ;in This is a descriptor for metal-carrier metal interactions. Let be the molar enthalpy of formation of metal cation M' in an infinitely diluted state supported by metal M, and take the value under standard conditions of 298 K and 101.325 kPa. The molar enthalpy of sublimation of the metal cation M' The molar enthalpy of sublimation of the loaded metal M is taken as the value under standard conditions of 298 K and 101.325 kPa. No additional crystal structure simulation or interface configuration calculation is performed. After the calculation, the unique corresponding characterizations of Pd and Zr are obtained. 4+ A descriptor for the metal-carrier metal interaction strength between the two.

[0033] After calculating the above two types of descriptors for all candidate systems, the dimensions of all calculated descriptor values ​​are uniformized to the general dimensions of the thermodynamic fundamental parameters. The dimension processing follows the conversion rules of thermodynamic parameters and does not change the original size and relative difference of the descriptor values. After processing, two types of element-level interface descriptors are formed, which serve as input features for the interface adhesion energy prediction model.

[0034] S4, Training and Predicting Adhesion Energy: Using the interface adhesion energy in the standardized dataset as the target variable and the two types of interface descriptors obtained in step S3 as input features, the SISSO interpretable symbolic regression algorithm is used for model training. First, the standardized dataset is divided into training, validation, and test sets using stratified random sampling. Then, the interface adhesion energy in the training set is used as the target variable, and the two types of interface descriptors are used as input features, imported into the SISSO algorithm training framework. The algorithm's maximum feature dimension is set to 2, the sparsity parameter to 1e-3 to 1e-2, and the number of iterations to be no less than 1000. The feature independence screening and global sparsification fitting are used to generate candidate prediction models. Then, the two types of interface descriptors from the validation set are input into the candidate prediction models to complete model validation and select the optimal candidate model. Then, the two types of interface descriptors from the test set are input into the optimal candidate model to complete independent testing. After confirming that the fitting results meet the preset requirements, the constant parameters and analytical expression of the model are fixed, and finally the interface adhesion energy prediction model is obtained. The interface adhesion energy prediction model is used to predict the interface adhesion energy of the Pd / ZrO2 core test system and other comparative candidate systems, and the predicted interface adhesion energy values ​​of each system are obtained.

[0035] S5, Calculate and sort the contact angles: Based on the predicted interfacial adhesion energy values ​​of Pd / ZrO2 and other candidate systems obtained in step S4, and combined with the surface energy data of the corresponding metals in each system, the adhesion energy-contact angle correlation algorithm is used to calculate the wetting contact angle of the metal on the oxide support surface in each system. The mathematical expression of the adhesion energy-contact angle correlation algorithm is: ; ;in The absolute value of the predicted interfacial adhesion energy is... The surface energy of the loaded metal M in the collected metal surface energy data is the surface energy value of a smooth and dense metal surface under standard conditions of 298 K and 101.325 kPa. The wetting contact angle of the metal on the metal oxide support surface is denoted as arccosine, with a calculated value ranging from 0° to 180°. Based on the criterion that the wetting contact angle is close to 90°, the heat resistance sintering performance of all candidate metal-oxide support systems is ranked. First, the absolute value of the difference between the wetting contact angle and 90° is calculated for each candidate system. The candidate systems are initially ranked in ascending order of this absolute value. If multiple candidate systems have the same absolute value, they are then ranked in descending order of the absolute value of the interfacial adhesion energy. After ranking, the candidate material combinations are output in list form. The list sequentially marks the loaded metal type, metal oxide support type, predicted interfacial adhesion energy, calculated wetting contact angle, and absolute value of the difference between the wetting contact angle and 90° for each candidate system. This completes the screening of the heat resistance sintering performance of the Pd / ZrO2 system and the comparative candidate systems.

[0036] In summary, this embodiment uses Pd / ZrO2 as the core test system and Ni / ZrO2, Co / ZrO2, and other systems as comparative candidates, and conducts heat resistance sintering performance screening according to the screening method of this invention. The entire process includes setting criteria, acquiring multi-source data, and standardizing the data as required, providing a basis for Pd and ZrO2 testing. 4+ The specific metal oxygen affinity and metal-carrier metal interaction descriptors were calculated. After training, verification and testing with the SISSO algorithm, an accurate adhesion energy prediction model was obtained. Then, the contact angle was calculated by the adhesion energy-contact angle correlation algorithm and multi-level sorting was completed. The candidate material combination with complete information was output, which verified the effectiveness and practicality of the method in the screening of transition metal-ZrO2 binary interface system.

[0037] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A machine learning-based method for screening thermally sintering-resistant supported metal catalysts, characterized in that, The method includes: S1, set target criteria: take the interface system composed of metal nanoparticles and metal oxide carrier as the screening object, take the interface adhesion energy as the prediction target, calculate the wetting contact angle of metal on the carrier surface through the interface adhesion energy, and take the wetting contact angle close to 90° as the criterion for evaluating the heat resistance of the system. S2, Construct a standardized dataset: Collect experimental interface adhesion energy data, metal surface energy data, oxide structural information, and basic thermodynamic parameters of metals and oxides at the metal-oxide carrier interface. Filter, normalize, remove outliers, and fill in missing values ​​on the collected data to form a standardized dataset. S3, Constructing Interface Descriptors: Based on the elemental thermodynamic parameters in the standardized dataset, for the loaded metal M and the metal cation M′ in the metal oxide support, the interaction strength between the metal and oxygen is calculated using the oxyphilicity descriptor algorithm, and the interaction strength between the metal and the metal cation is calculated using the metal interaction descriptor algorithm. Unique metal oxyphilicity descriptors and metal-support metal interaction descriptors are generated respectively, which serve as input features for the interface adhesion energy prediction model. S4, Training and Prediction of Adhesion Energy: Using the interfacial adhesion energy in the standardized dataset as the target variable and two types of interface descriptors as input features, the SISSO interpretable symbolic regression algorithm is used to train the model and fit it to obtain the interfacial adhesion energy prediction model. The interfacial adhesion energy prediction model is then used to predict the interfacial adhesion energy of the metal-oxide support system under test. S5, Calculate and sort the contact angle: Based on the interfacial adhesion energy predicted in step S4, combined with the metal surface energy data, the wetting contact angle is calculated using the adhesion energy-contact angle correlation algorithm. Based on the criterion that the wetting contact angle is close to 90°, the candidate metal-oxide support systems are sorted according to their resistance to thermal sintering, and the sorted candidate material combinations are output.

2. The method for screening heat-resistant sintering supported metal catalysts based on machine learning according to claim 1, characterized in that, In step S1, the metal nanoparticles are single-component transition metal nanoparticles selected from Pt, Pd, Rh, Au, Ag, Ni, Co, and Cu. The metal oxide support is a single-component metal oxide selected from Al2O3, ZrO2, CeO2, SiO2, MgO, TiO2, and ZnO. The metal nanoparticles and the metal oxide support form a binary interface system.

3. The method for screening heat-resistant sintering supported metal catalysts based on machine learning according to claim 1, characterized in that, In step S2, the experimental interface adhesion energy data are collected from SCI core journals in the fields of catalysis and interface science, proceedings of international authoritative conferences, and peer-reviewed public experimental databases. During the collection, the test methods, test environment temperature, test vacuum degree, sample preparation process, and surface cleanliness control methods of the corresponding data are recorded simultaneously. The oxide structure information is extracted from the MaterialsProject database and ICSD database. The extracted content includes the stable crystal form of the oxide, crystal structure parameters, and atomic occupancy information. The thermodynamic fundamental parameters of the metal and oxide are collected from the FactSage database and JANAF database. The collected content includes the metal formation enthalpy, metal sublimation enthalpy, metal mixing formation enthalpy, and oxide thermodynamic stability parameters.

4. The method for screening heat-resistant sintering supported metal catalysts based on machine learning according to claim 1, characterized in that, In step S2, the screening of collected data is only carried out on experimental interface adhesion energy data. Experimental interface adhesion energy data with consistent measurement methods, clearly controlled or discussed in detail surface cleanliness interference factors, and no less than 3 test repetitions are retained. Metal surface energy data, oxide structure information, and thermodynamic basic parameters of metals and oxides are all retained as original measured and extracted results without screening.

5. The method for screening heat-resistant sintering supported metal catalysts based on machine learning according to claim 1, characterized in that, In step S2, the Z-score standardization method is used to normalize all collected data. The mean of the processed data is 0 and the variance is 1. The Grubbs criterion is used to remove outliers from the collected data. The significance level is set to 0.

05. Data that are identified as outliers are directly removed from the dataset. The missing values ​​of the basic thermodynamic parameters of the metal are filled in using the linear interpolation method based on the periodic law. The missing values ​​of the metal mixing enthalpy are filled in using the Miedema semi-empirical thermodynamic model.

6. The method for screening heat-resistant sintering supported metal catalysts based on machine learning according to claim 1, characterized in that, In step S3, the mathematical expression of the oxyphilicity descriptor algorithm is: ;in This is a descriptor for metal oxyphilicity. The molar enthalpy of formation for the supported metal M to form its thermodynamically most stable oxide is given by values ​​at 298 K and 101.325 kPa under standard conditions. The molar enthalpy of sublimation of the metal cation M' in the metal oxide support is taken as the value under standard conditions of 298 K and 101.325 kPa.

7. The method for screening heat-resistant sintering supported metal catalysts based on machine learning according to claim 1, characterized in that, In step S3, the mathematical expression of the metal interaction descriptor algorithm is: ;in This is a descriptor for metal-carrier metal interactions. Let be the molar enthalpy of formation of metal cation M' in an infinitely diluted state supported by metal M, and take the value under standard conditions of 298 K and 101.325 kPa. The molar enthalpy of sublimation of the metal cation M' Let M be the molar enthalpy of sublimation of the loaded metal M, and both are taken as values ​​under standard conditions of 298 K and 101.325 kPa.

8. The method for screening heat-resistant sintering supported metal catalysts based on machine learning according to claim 1, characterized in that, In step S4, the specific process of using the SISSO interpretable symbolic regression algorithm to train and fit the interfacial adhesion energy prediction model is as follows: The standardized dataset is divided into a training set, a validation set, and a test set using stratified random sampling. The interfacial adhesion energy in the training set is used as the target variable, and two types of interface descriptors are used as input features. The algorithm's maximum feature dimension is set to 2, the sparsity parameter to be between 1e-3 and 1e-2, and the number of iterations to be no less than 1000. Candidate prediction models are generated through the algorithm's built-in feature independence screening and global sparsity fitting. The two types of interface descriptors from the validation set are input into the candidate prediction models to complete model validation and select the optimal candidate model. Then, the two types of interface descriptors from the test set are input into the optimal candidate model to complete independent testing. After confirming that the fitting results meet the preset requirements, the constant parameters and analytical expression of the model are fixed, and finally, the interfacial adhesion energy prediction model is obtained.

9. The method for screening heat-resistant sintering supported metal catalysts based on machine learning according to claim 1, characterized in that, In step S5, the mathematical expression for the adhesion energy-contact angle correlation algorithm is: ; ;in The absolute value of the predicted interfacial adhesion energy is... The surface energy of the loaded metal M in the collected metal surface energy data is the surface energy value of a smooth and dense metal surface under standard conditions of 298 K and 101.325 kPa. is the wetting contact angle of the metal on the surface of the metal oxide carrier, and arccos is the inverse cosine function, with a calculated value range of 0° to 180°.

10. The method for screening heat-resistant sintering supported metal catalysts based on machine learning according to claim 1, characterized in that, In step S5, when ranking the anti-sintering performance of candidate metal-oxide support systems, the absolute value of the difference between the wetting contact angle and 90° for each candidate system is first calculated. All candidate systems are then ranked in ascending order of absolute value. If multiple candidate systems have the same absolute value, they are then ranked a second time in descending order of absolute value of interfacial adhesion energy. After ranking, the candidate material combinations are output in list form. The list sequentially marks the candidate system's loading metal type, metal oxide support type, predicted interfacial adhesion energy value, calculated wetting contact angle value, and absolute value of the difference between the wetting contact angle and 90°.

Images