Method for predicting NH3-TPD temperature based on machine learning

Abstract

The invention discloses a method for predicting NH3-TPD temperature based on machine learning. The method comprises the following steps: 1) collecting information of a known catalyst as data to establish a database; 2) screening and normalizing the data, and dividing the data in the database into a training set and a test set; 3) training the preprocessed data set by adopting a machine learning classification model; wherein the input features of the machine learning classification model comprise preparation information, composition information, pore structure information and physicochemical property information of the catalyst, and the peak number of each sample in each temperature interval is output; training the machine learning classification model through the training set, and verifying the trained machine learning classification model by using the test set to obtain a final machine learning classification model and a prediction accuracy rate thereof; 4) training the preprocessed data set by using a deep learning multi-output regression model; according to the peak number prediction result of each temperature interval by the machine learning classification model, determining the peak temperature number to be evaluated in each temperature interval, and only evaluating the regression peak temperature corresponding to the predicted peak number, thereby realizing interval peak temperature selective evaluation limited according to the classification result; and 5) predicting the temperature information of the NH3-TPD of the unknown catalyst by adopting the final deep learning multi-output regression model.

Description

Technical Field

[0001] This invention relates to a catalyst N - A method for predicting TPD temperature, specifically involving a method based on machine learning to predict N -TPD temperature method belongs to the intersection of artificial intelligence and materials science and technology. Background Technology

[0002] The ammonia desorption peak temperature on the surface of a metal catalyst is one of the key indicators characterizing its acidity and distribution characteristics. Traditional experimental methods, such as ammonia temperature-programmed desorption (N2-N2), are used for this purpose. While TPD and in-situ infrared spectroscopy can directly obtain peak temperature data, they are time-consuming, energy-intensive, and costly, making them unsuitable for high-throughput screening. First-principles simulations can reveal the energy barriers and kinetics of adsorption and desorption processes at the microscopic level, but for complex and structurally diverse catalytic systems, the computational load is enormous and inefficient, hindering their widespread application in large-scale material screening. In contrast, data-driven methods based on machine learning can utilize existing preparation conditions, structures, and physicochemical characteristics to establish catalyst characteristics and N2O... The nonlinear mapping between the desorption peak temperature and the TPD desorption peak temperature enables rapid, low-cost, and high-precision prediction of the desorption temperature, significantly improving screening efficiency and showing good prospects for widespread application. Summary of the Invention

[0003] The present invention aims to overcome the above-mentioned shortcomings of the prior art and provide a method for predicting the acid content on the catalyst surface based on machine learning technology, which can improve development efficiency.

[0004] This invention proposes a data-driven method for predicting the NH3-TPD temperature of catalysts. By establishing a quantitative structure-activity relationship between catalyst structural characteristics and ammonia desorption peak temperature, it provides reliable guidance for the rational design of novel catalysts, significantly improving R&D efficiency and reducing experimental screening costs.

[0005] The technical solution adopted by this invention to solve its technical problem is: A method for predicting NH3-TPD temperature based on machine learning includes the following steps: S1: Collect information on known catalysts to establish a database. The catalyst information includes preparation information, composition information, pore structure information, physicochemical property information, and temperature information. S2: Filter and normalize the data, and divide the data in the database into training set and test set; S3: Train a machine learning classification model on the preprocessed dataset; The machine learning classification model takes into account the catalyst's preparation, composition, pore structure, and physicochemical properties as inputs, and outputs the number of peaks for each sample in each temperature range. The machine learning classification model is trained using the training set and validated using the test set to obtain the final machine learning classification model and its prediction accuracy. S4: Train a deep learning multi-output regression model on the preprocessed dataset; The deep learning multi-output regression model uses catalyst preparation information, composition information, pore structure information and physicochemical property information as input features, and takes all peak temperatures of each sample in different temperature ranges as the multi-output regression target. The model is trained using the training set, and the predicted peak temperature is evaluated by partitioning based on the classification results of the machine learning classification model described in step S3. The trained model is then validated using the test set to obtain the final deep learning multi-output regression model. The partitioned evaluation process is as follows: based on the peak number prediction results of the machine learning classification model for each temperature interval, the number of peak temperatures to be evaluated in each temperature interval is determined, and only the regression peak temperatures corresponding to the predicted peak numbers are evaluated, thereby achieving selective evaluation of interval peak temperatures limited by the classification results. S5: The final deep learning multi-output regression model described in S4 is used to predict the temperature information of NH3-TPD for the unknown catalyst.

[0006] Preferably, in step S1, the catalyst preparation information includes the catalyst preparation method and preparation parameters, wherein the preparation parameters include one or more of calcination temperature and calcination time; Catalyst composition information includes: whether the catalyst contains a support, the constituent elements and their types and quantities, and one or more of the mass fraction and mole fraction of metal elements; Catalyst pore structure information includes one or more of the following: specific surface area, average pore size, and average pore volume; The physicochemical properties of the catalyst include one or more of the following parameters of the first metal, the second metal, and the support: molecular weight, melting point, volume, density, standard enthalpy of formation, decomposition energy, experimental formation energy, bulk modulus, coordination number of the corresponding metal element, conductivity, band gap, first to third ionization energies, number of valence electrons, Pauli electronegativity, ionic radius, and valence state. Temperature information includes N - The number of desorption peaks and their corresponding peak temperature positions obtained from the TPD test are recorded, and each peak temperature is divided into a low temperature region (0-250℃), a medium temperature region (200-400℃), and a high temperature region (400-800℃) for regional statistics.

[0007] Preferably, the data screening in step S2 includes determining numerical ranges, removing outliers, removing non-discriminatory features, and samples with no second metal component, allowing all of its corresponding second metal features to be collectively nullified without deletion.

[0008] Preferably, the data normalization in step S2 adopts the following formula: (1) wherein x is the original value, μ is the mean value of the feature, σ is the standard deviation of the feature, and x' is the normalized value.

[0009] Preferably, in step S2, the data in the database is divided into a training set and a test set in a random division method at a ratio of 8:2, with the training set accounting for 80% and the test set accounting for 20%.

[0010] Preferably, step S3 includes the following sub-steps: S3.1: For the peak number classification task of each temperature interval, select a suitable machine learning classification model; wherein the machine learning classification model includes but is not limited to a Random Forest Classifier, an Extremely Randomized Trees Classifier, an Artificial Neural Network, an eXtreme Gradient Boosting Classifier, a Support Vector Classifier, a Logistic Regression Classifier, a Naive Bayes Classifier, and a Gradient Boosting Decision Tree Classifier; S3.2: Use the Grid Search method to perform exhaustive optimization on the key parameters of the selected machine learning classification model to obtain the optimal parameter configuration of the machine learning classification model; The key parameters differ depending on the specific model type, mainly including but not limited to: the number of decision tree-based learners, the maximum depth, the minimum number of split samples, and the feature sampling ratio; the kernel function type, the regularization coefficient, and the kernel parameter of the support vector classifier; the hidden layer structure and learning rate of the artificial neural network classifier; the learning rate, the number of weak learners, and the maximum depth of the gradient boosting model; and the regularization form and penalty coefficient of the logistic regression model, etc. S3.3: Using the obtained optimal machine learning classification model parameter configuration, train the machine learning classification model on the training set and evaluate the classification performance on the independent evaluation validation set or test set.

[0011] Preferably, step S4 includes the following sub-steps: S4.1: According to the characteristics of the catalyst temperature information, select a suitable deep learning multi-output regression model; The deep learning multi-output regression model includes but is not limited to: Multi-Output Regression Network; Convolutional Neural Network; Residual Neural Network, ResNet; Recurrent Neural Network, Self-Attention Network and its regression variants; S4.2: Use the grid search method to perform exhaustive optimization on the key parameters of the selected deep learning multi-output regression model to determine the optimal parameter combination; The key parameters differ depending on the specific model type, mainly including but not limited to: the number of decision tree-based learners, the maximum depth, the minimum number of split samples, and the feature sampling ratio; the kernel function type, the regularization coefficient, and the kernel parameter of the support vector classifier; the hidden layer structure and learning rate of the artificial neural network classifier; the learning rate, the number of weak learners, and the maximum depth of the gradient boosting model; and the regularization form and penalty coefficient of the logistic regression model, etc. S4.3: Using the obtained optimal deep learning multi-output regression model parameter configuration, train the deep learning multi-output regression model on the training set and evaluate the regression performance on the independent evaluation validation set or test set.

[0012] Preferably, in step S3 and step S4, when the model accuracy of any step does not meet the preset threshold, by replacing the machine learning classification model, the deep learning multi-output regression model, or adjusting the parameters of the machine learning classification model and the deep learning multi-output regression model, step S3 and step S4 are repeatedly executed until the model accuracy meets the requirements.

[0013] Preferably, in step S3, the evaluation index is accuracy.

[0014] Preferably, in step S4, the evaluation index is one or more of the coefficient of determination (R2), the mean absolute error (MAE), and the root mean square error (RMSE).

[0015] Preferably, the calculation formula of the accuracy is: (2) The calculation formula of the coefficient of determination (R2) is:

[0016] The calculation formula of the root mean square error (RMSE) is:

[0017] The calculation formula of the mean absolute error (MAE) is: (5) In the formula, Ncorrect is the number of samples whose predicted value is equal to the actual value, Ntotal is the total number of samples for the output, n is the total number of samples, And respectively represent the actual value and the predicted value of the i th sample, represent the average value of all target variable actual values.

[0018] The working principle of the present application is: based on the multi-dimensional characteristics of the preparation information, composition information, pore structure information, and physical and chemical property information of the catalyst, and the N Based on the statistical correlation between the number of peaks and peak temperature distribution across temperature ranges in the TPD characterization, a data-driven approximation of this complex mapping relationship is achieved by constructing a dedicated database and introducing machine learning and deep learning models. First, a machine learning classification model is used to learn the relationship between input features and the number of peaks in each temperature range, enabling the prediction of the existence and quantity of peaks in each temperature zone. Then, a deep learning multi-output regression model is employed to simultaneously predict all potential peak temperatures within each temperature zone based on shared feature representations. By incorporating the peak number prediction results from the classification model into the partitioning evaluation process of the regression model, selective evaluation is performed only on the peak temperature output corresponding to the predicted peak number, effectively suppressing the interference of invalid peak positions, thus obtaining a peak temperature prediction model that combines physical rationality and high accuracy. Finally, for unknown catalysts, only their preparation process, composition, pore structure, and physicochemical parameters need to be input to obtain their N values ​​through the above cascaded model. - The number of peaks and peak temperatures in each temperature zone of the TPD enable a forward-looking quantitative characterization of the acidity distribution of the catalyst.

[0019] The innovation of this invention is: 1. Existing technologies mostly use N -TPD, as a post-hoc characterization method, only characterizes catalyst acidity by performing peak fitting, kinetic analysis, or simply extracting a few indicators such as total acidity and main peak temperature from measured spectra. It is insufficient to fundamentally reveal the systematic relationship between multi-temperature ranges, multi-peak distributions, and catalyst preparation conditions and physicochemical properties, nor can it address the N content of untested samples. -TPD behavior is prospectively predicted. To address the above issues, this invention proposes a cascaded modeling framework of "peak number classification + peak temperature multi-output regression": First, a machine learning classification model is used to learn the correspondence between catalyst features and the number of peaks in each temperature range, achieving quantitative prediction of the number of peaks in each temperature range; then, a deep learning multi-output regression model is used to simultaneously predict all peak temperatures in different temperature ranges within a unified feature space. By organically linking peak number prediction and peak temperature regression, this invention can fully explore N... - The inherent correlation between multiple peaks and intervals in TPD enables a more refined and predictable characterization of the acidity distribution of catalysts. 2. A partitioned selective evaluation mechanism based on classification results is introduced into peak temperature prediction. Based on the number of peaks in each temperature zone given by the classification model, loss calculation and performance evaluation are performed only on the regression peak temperatures corresponding to that number of peaks. This effectively avoids interference from invalid outputs on model training and evaluation, significantly improving the accuracy and stability of peak temperature prediction. 3. The final model obtained by this invention can be used to predict the N of unknown catalysts. -TPD temperature information enables forward-looking quantitative characterization of catalyst acid strength and acid site distribution, reduces the number of experimental measurements, lowers R&D costs, and provides a powerful tool for rapid catalyst screening and rational design.

[0020] The application provides a catalyst NH3-TPD desorption temperature prediction method combining machine learning and deep learning. The temperature data obtained through the NH3-TPD test are used as the basis to construct a mixed prediction model, so that the desorption peak temperature of unknown catalysts can be quickly and accurately predicted. Compared with the traditional experimental characterization method, the method can significantly reduce the screening cost, shorten the research and development cycle, and has high throughput, low energy consumption and excellent prediction reliability. In addition, the method has strong universality and can be popularized to the intelligent design and rapid development of various adsorption materials and catalyst materials. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, below the drawings needed to be used in the embodiment or prior art description will be briefly introduced. Obviously, the drawings in the following description are only embodiments of the application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.

[0022] Figure 1 is a method flowchart of the application; Figures 2a-2c is a prediction result graph of the machine learning classification model of the application in the training set, Figure 2a is a classification accuracy rate graph of the low-temperature zone-training set, Figure 2b is a classification accuracy rate graph of the medium-temperature zone-training set, Figure 2c is a classification accuracy rate graph of the high-temperature zone-training set; Figures 3a-3c is a prediction result graph of the machine learning classification model of the application in the test set, Figure 3a is a classification accuracy rate graph of the low-temperature zone-test set, Figure 3b is a classification accuracy rate graph of the medium-temperature zone-test set, Figure 3c is a classification accuracy rate graph of the high-temperature zone-test set; Figures 4a-4c is a prediction result graph of the deep learning multi-output regression model of the application in the training set, Figure 4a is a regression prediction effect graph of the low-temperature zone-training set, Figure 4b is a regression prediction effect graph of the medium-temperature zone-training set, Figure 4c is a regression prediction effect graph of the high-temperature zone-training set; Figures 5a-5c is a prediction result graph of the deep learning multi-output regression model of the application in the test set, Figure 5a is a regression prediction effect graph of the low-temperature zone-test set, Figure 5b is a regression prediction effect graph of the medium-temperature zone-test set, Figure 5c is a regression prediction effect graph of the high-temperature zone-test set. Detailed Implementation

[0023] The technical solution of the present invention will be further described in detail below through specific embodiments and in conjunction with the accompanying drawings. It should be understood that the implementation of the present invention is not limited to the following embodiments, and any modifications and / or alterations made to the present invention will fall within the protection scope of the present invention.

[0024] In this invention, unless otherwise specified, all parts and percentages are by weight, and the equipment and raw materials used are commercially available or commonly used in the art. Unless otherwise specified, the methods in the following embodiments are conventional methods in the art. Unless otherwise specified, the components or equipment in the following embodiments are general standard parts or components known to those skilled in the art, and their structures and principles can be learned by those skilled in the art through technical manuals or conventional experimental methods.

[0025] This invention proposes a method for predicting NH3-TPD temperature based on machine learning. By collecting experimental data and relevant catalyst information from existing literature, a large database containing catalyst acidity data is constructed. Using machine learning methods, a nonlinear mapping relationship is established between catalyst preparation conditions, compositional characteristics, pore structure, physicochemical properties, and desorption temperature, thereby achieving intelligent prediction of the strength and distribution characteristics of surface acidity on metal oxide catalysts. This method can serve as an effective auxiliary tool for developing catalysts with controllable acidity and optimized reaction performance, possessing advantages such as high efficiency, intelligence, and scalability.

[0026] The concept of the present application is that, in order to solve the problems of long time consumption and low efficiency in the process of obtaining catalyst ammonia desorption temperature information, a prediction method based on the combination of machine learning and deep learning is proposed. The method collects the preparation information, composition information, pore structure information, physical and chemical property information, and temperature information of the catalyst, and constructs an acid amount database. The collected data is preprocessed and normalized, and the key descriptors closely related to the temperature are selected through feature engineering. Then the data is divided into training set and test set according to a certain proportion, and a machine learning algorithm is used to construct a catalyst temperature prediction model. During the model training process, the cross-validation method is used to evaluate the generalization ability of the model to prevent overfitting; the performance of the model is verified by the test set, and the accuracy, determination coefficient (R²), mean absolute error (MAE) and root mean square error (RMSE) are used to quantitatively evaluate the accuracy of the model, so as to obtain the prediction model with the best performance. The model can realize the rapid prediction of the catalyst temperature (including different intensity of temperature) in the target system, provide effective support for the high-throughput screening, directional regulation and optimization design of the catalyst, help to improve the research and development efficiency and shorten the development cycle. The data source of the constructed database is not limited, which can come from the published literature, report, patent or open database, or can be obtained by experimental synthesis and characterization, provided that the data has reliability and accuracy.

[0027] The technical solutions of the present application are described in detail below in combination with embodiments.

[0028] Embodiment 1: As shown in a method for predicting NH3-TPD temperature based on machine learning, the method comprises the following steps: Figure 1 S1: Collect information of known catalysts as data to establish a database, wherein the catalyst information includes preparation information, composition information, pore structure information, physical and chemical property information, and temperature information; S2: Screen and normalize the data, and divide the data in the database into training set and test set; S3: Train the preprocessed data set using a machine learning classification model; Wherein the input features of the machine learning classification model include the preparation information, composition information, pore structure information and physical and chemical property information of the catalyst, and the output is the peak number of each sample in each temperature interval; The machine learning classification model is trained by the training set, and the trained machine learning classification model is verified by the test set to obtain the final machine learning classification model and its prediction accuracy; S4: Train the preprocessed data set using a deep learning multi-output regression model; ​The input features of the deep learning multi-output regression model include preparation information, composition information, pore structure information and physical and chemical property information of the catalyst as input features, and all peak temperatures of each sample in different temperature intervals as multi-output regression targets. The model is trained by the training set, the predicted peak temperatures are evaluated based on the classification results of the machine learning classification model in step S3, and the completed model is verified by the test set to obtain the final deep learning multi-output regression model. The evaluation process is as follows: according to the peak number prediction results of each temperature interval by the machine learning classification model, the number of peak temperatures to be evaluated in each temperature interval is determined, and only the regression peak temperatures corresponding to the predicted peak number are evaluated, so as to realize the selective evaluation of the interval peak temperature according to the classification result. S5: Using the final deep learning multi-output regression model in S4 to predict the temperature information of unknown catalysts in NH3-TPD.

[0029] The catalyst preparation information includes the preparation method of the catalyst and the preparation parameters, and the preparation parameters include one or more of calcination temperature and calcination time; The catalyst composition information includes one or more of whether the catalyst contains a carrier, the types and amounts of constituent elements, and the mass fraction and molar fraction of metal elements; The catalyst pore structure information includes one or more of specific surface area, average pore size and average pore volume; The catalyst physical and chemical property information includes one or more of the following parameters of the first metal, the second metal and the carrier: molecular weight, melting point, volume, density, standard enthalpy of formation, decomposition energy, experimental formation energy, bulk modulus, coordination number of corresponding metal element, electrical conductivity, band gap, first to third ionization energy, valence electron number, Pauli electronegativity, ionic radius and valence state; The temperature information includes the number of desorption peaks and the corresponding peak temperature positions obtained by N TPD test, and each peak temperature is divided into a low temperature zone (0-250℃), a medium temperature zone (200-400℃) and a high temperature zone (400-800℃) for zoned statistics.

[0030] The following specific information is selected for the machine learning classification model and the deep learning multi-output regression model respectively in this embodiment: Catalyst preparation information: calcination temperature and calcination time of the catalyst; Catalyst composition information: whether the catalyst contains a carrier, the elements of the catalyst composition and the mass fraction of metal elements; Catalyst pore structure information: specific surface area of the catalyst; Catalyst physical and chemical information: volume, density, bulk modulus, molecular weight, band gap, melting point of metal oxides, coordination number, charge, first ionization energy and ionic radius of metal elements; Catalyst temperature information: temperature of three different intervals.

[0031] In the step S2, the collected data is preliminarily screened, including but not limited to: screening and normalizing the data, dividing the data in the database into a training set and a test set. Data screening includes determining the numerical range, deleting outliers, deleting non-discriminatory features, and samples with empty second metal components, allowing all corresponding second metal features to be collectively empty without deletion; using a random division method to divide the data into two parts according to an 8:2 ratio, with 80% as the training set and 20% as the test set; the training set is used for machine learning classification model training, and the test set is used to test and verify the accuracy of the machine learning classification model after the model training is completed.

[0032] Wherein the data normalization adopts the following formula: (1) Wherein x is the original value, μ is the mean of the feature, σ is the standard deviation of the feature, and x' is the normalized value.

[0033] The step S3 includes the following sub-steps: S3.1: For the peak number classification task of each temperature interval, select a suitable machine learning classification model; Wherein the machine learning classification model includes but is not limited to: random forest classifier (Random Forest Classifier); extremely randomized tree classifier (Extremely Randomized Trees Classifier); artificial neural network classifier (Artificial Neural Network); eXtreme Gradient Boosting Classifier; support vector classifier (Support Vector Classifier); logistic regression classifier (Logistic Regression Classifier); Naive Bayes Classifier; Gradient Boosting Decision Tree Classifier; S3.2: Use the grid search (Grid Search) method to traverse the key parameters of the selected machine learning classification model to obtain the optimal parameter configuration of the machine learning classification model. The key parameters differ depending on the specific model type, mainly including but not limited to: the number of decision tree-based learners, the maximum depth, the minimum number of split samples, and the feature sampling ratio; the kernel function type, the regularization coefficient, and the kernel parameter of the support vector classifier; the hidden layer structure and learning rate of the artificial neural network classifier; the learning rate, the number of weak learners, and the maximum depth of the gradient boosting model; and the regularization form and penalty coefficient of the logistic regression model, etc. S3.3: Using the obtained optimal machine learning classification model parameter configuration, train the machine learning classification model on the training set and evaluate the classification performance on the independent evaluation validation set or test set; The step S4 includes the following sub-steps: S4.1: According to the characteristics of the catalyst temperature information, select a suitable deep learning multi-output regression model; The deep learning multi-output regression model includes but is not limited to: Multi-Output Regression Network, Convolutional Neural Network, Residual Neural Network (ResNet), Recurrent Neural Network, Self-Attention Network, and its regression variants. S4.2: Adopting grid search method to perform traversal optimization on the key parameters of the selected deep learning multi-output regression model to determine the optimal parameter combination; The key parameters differ depending on the specific model type, mainly including but not limited to: the number of decision tree-based learners, the maximum depth, the minimum number of split samples, and the feature sampling ratio; the kernel function type, the regularization coefficient, and the kernel parameter of the support vector classifier; the hidden layer structure and learning rate of the artificial neural network classifier; the learning rate, the number of weak learners, and the maximum depth of the gradient boosting model; and the regularization form and penalty coefficient of the logistic regression model, etc. S4.3: Using the obtained optimal deep learning multi-output regression model parameter configuration, train the deep learning multi-output regression model on the training set and evaluate the regression performance on the independent evaluation validation set or test set; Deep learning classification algorithms include, but are not limited to: Multilayer Perceptron (MLP) multi-output regression networks; Convolutional Neural Networks (CNN); Residual Neural Networks (ResNet) multi-output regression models; Recurrent Neural Networks (RNN) with self-attention mechanisms (Transformer) and their regression variants; and joint regression networks based on multi-task learning.

[0034] In steps S3 and S4, if the model accuracy in any step does not meet the preset threshold, steps S3 and S4 are repeated until the model accuracy meets the requirements by changing the machine learning classification model or the deep learning multi-output regression model, or by adjusting the parameters of the machine learning classification model and the deep learning multi-output regression model.

[0035] In step S3, the evaluation metric is accuracy.

[0036] In step S4, the evaluation index is one or more of the following: coefficient of determination (R²), mean absolute error (MAE), and root mean square error (RMSE).

[0037] In this embodiment, regarding catalyst N -The TPD desorption temperature peaks in each interval were predicted using a random forest classifier; the evaluation metric was classification accuracy, calculated as follows:

[0038] In the formula, Ncorrect is the number of samples whose predicted value is equal to the actual value, and Ntotal is the total number of samples for this output.

[0039] In this embodiment, regarding catalyst N -The total number of TPD desorption peaks and peak temperatures in each temperature range are predicted by regression analysis using a residual neural network multi-output regression model; the evaluation index is the coefficient of determination (R²), calculated as follows:

[0040] In the formula, n is the total number of samples. and These represent the actual value and the predicted value of the i-th element, respectively. This represents the average of the actual values ​​of all target variables.

[0041] The hyperparameters of the model are optimized using a grid search method. On the constructed training set, the machine learning classification model of the application is used to classify N -TPD desorption total peak number and peak number in each temperature interval are classified and predicted respectively. As shown in Table 6, the results show that the peak number classification accuracy in the low temperature zone, the medium temperature zone and the high temperature zone is 1.000, 1.000 and 1.000 respectively. The above data show that the model has excellent fitting effect and high prediction accuracy on the training set. Figures 2a-2c

[0042] On the constructed training set, the deep learning multi-output regression model of the application is used to predict N -TPD peak temperature in each temperature interval is predicted. As shown in Table 7, the determination coefficient (R²) of the prediction result in the low temperature zone, the medium temperature zone and the high temperature zone is 0.932, 0.906 and 0.739 respectively. The above data show that the model has excellent fitting effect, high prediction accuracy and generalization ability on the training set. Figures 4a-4c

[0043] It should be pointed out that if mean absolute error (MAE) and root mean square error (RMSE) are used as evaluation indexes, the calculation formula is as follows: (4) (5) In the formula, n is the total number of samples, and represent the actual value and the predicted value of the i-th, represent the average value of all target variables actual values.

[0044] On the constructed training set, the machine learning classification model of the application is used to classify N -TPD desorption total peak number and peak number in each temperature interval are classified and predicted respectively. As shown in Table 8, the results show that the peak number classification accuracy in the low temperature zone, the medium temperature zone and the high temperature zone is 0.9333, 0.9333 and 0.9000 respectively. The above data show that the model has excellent fitting effect and high prediction accuracy on the test set. Figures 3a-3c

[0045] On the constructed training set, the deep learning multi-output regression model of the application is used to predict N -TPD peak temperature in each temperature interval is predicted. As shown in Table 9, the determination coefficient (R²) of the prediction result in the low temperature zone, the medium temperature zone and the high temperature zone is 0.845, 0.736 and 0.786 respectively. The above data show that the model has excellent fitting effect, high prediction accuracy and generalization ability on the test set. Figures 5a-5c

[0046] ​​​​The embodiments of the present specification are merely a list of implementation forms of the inventive concept, and the protection scope of the present application should not be regarded as being limited to the specific forms stated in the embodiments, and the protection scope of the present application also encompasses equivalent technical means that can be thought of by those skilled in the art according to the inventive concept.

Claims

1. A method for predicting NH3-TPD temperature based on machine learning, characterized in that, Includes the following steps: S1: Collect information on known catalysts to establish a database. The catalyst information includes preparation information, composition information, pore structure information, physicochemical property information, and temperature information. S2: Filter and normalize the data, and divide the data in the database into training set and test set; S3: Train a machine learning classification model on the preprocessed dataset; The machine learning classification model takes into account the catalyst's preparation, composition, pore structure, and physicochemical properties as inputs, and outputs the number of peaks for each sample in each temperature range. The machine learning classification model is trained using the training set and validated using the test set to obtain the final machine learning classification model and its prediction accuracy. S4: Train a deep learning multi-output regression model on the preprocessed dataset; The deep learning multi-output regression model uses catalyst preparation information, composition information, pore structure information and physicochemical property information as input features, and takes all peak temperatures of each sample in different temperature ranges as the multi-output regression target. The model is trained using the training set, and the predicted peak temperature is evaluated by partitioning based on the classification results of the machine learning classification model described in step S3. The trained model is then validated using the test set to obtain the final deep learning multi-output regression model. The partitioned evaluation process is as follows: based on the peak number prediction results of the machine learning classification model for each temperature interval, the number of peak temperatures to be evaluated in each temperature interval is determined, and only the regression peak temperatures corresponding to the predicted peak numbers are evaluated, thereby achieving selective evaluation of interval peak temperatures limited by the classification results. S5: The final deep learning multi-output regression model described in S4 is used to predict the temperature information of NH3-TPD for the unknown catalyst.

2. The prediction method according to claim 1, characterized in that... In step S1, Catalyst preparation information includes the catalyst preparation method and preparation parameters, wherein the preparation parameters include one or more of calcination temperature and calcination time; Catalyst composition information includes: whether the catalyst contains a support, the constituent elements and their types and quantities, and one or more of the mass fraction and mole fraction of metal elements; Catalyst pore structure information includes one or more of the following: specific surface area, average pore size, and average pore volume; The physicochemical properties of the catalyst include one or more of the following parameters of the first metal, the second metal, and the support: molecular weight, melting point, volume, density, standard enthalpy of formation, decomposition energy, experimental formation energy, bulk modulus, coordination number of the corresponding metal element, conductivity, band gap, first to third ionization energies, number of valence electrons, Pauli electronegativity, ionic radius, and valence state. Temperature information includes N - The number of desorption peaks and their corresponding peak temperature positions obtained from the TPD test are recorded, and each peak temperature is divided into a low temperature region (0-250℃), a medium temperature region (200-400℃), and a high temperature region (400-800℃) for regional statistics.

3. The method according to claim 1, characterized in that: The data filtering in step S2 includes determining the numerical range, deleting outliers, deleting indifferent features, and for samples with an empty second metal component, allowing all corresponding second metal features to be left empty without deletion.

4. The method according to claim 1, characterized in that: The data normalization process in step S2 uses the following formula: （1） In the formula, x is the original value, μ is the mean of the feature, σ is the standard deviation of the feature, and x′ is the normalized value.

5. The method according to claim 1, characterized in that: In step S2, the data in the database is randomly divided into a training set and a test set in a ratio of 8:2, with the training set accounting for 80% and the test set accounting for 20%.

6. The method according to claim 1, characterized in that: Step S3 includes the following sub-steps: S3.1: Select an appropriate machine learning classification model for the peak number classification task in each temperature range; The machine learning classification models include, but are not limited to, Random Forest Classifier; Extremely Randomized Trees Classifier; Artificial Neural Network Classifier; eXtreme Gradient Boosting Classifier; Support Vector Classifier; Logistic Regression Classifier; Naive Bayes Classifier; and Gradient Boosting Decision Tree Classifier. S3.2: Use the grid search method to perform traversal optimization on the key parameters of the selected machine learning classification model to obtain the optimal parameter configuration of the machine learning classification model; The key parameters vary depending on the specific model type, and mainly include, but are not limited to: the number of decision tree base learners, maximum depth, minimum number of split samples, and feature sampling ratio; the kernel function type, regularization coefficient, and kernel parameters of the support vector classifier; the hidden layer structure and learning rate of the artificial neural network classifier; the learning rate, number of weak learners, and maximum depth of the gradient boosting model; and the regularization form and penalty coefficient of the logistic regression model, etc. S3.3: Using the obtained optimal machine learning classification model parameter configuration, train the machine learning classification model on the training set and evaluate the classification performance on an independent evaluation validation set or test set.

7. The method according to claim 1, characterized in that: Step S4 includes the following sub-steps: S4.1: Select a suitable deep learning multi-output regression model based on the characteristics of catalyst temperature information; Deep learning multi-output regression models include, but are not limited to: multi-output regression networks with multi-layer perceptrons; convolutional neural networks; residual neural networks (ResNet); recurrent neural networks; self-attention networks and their regression variants; S4.2: Use the grid search method to perform traversal optimization of the key parameters of the selected deep learning multi-output regression model in order to determine the optimal parameter combination; The key parameters vary depending on the specific model type, and mainly include, but are not limited to: activation function type (such as ReLU, LeakyReLU, GELU, etc.); loss function type and its penalty weights (including weighted mean squared error, negative penalty, range penalty, order penalty and corresponding coefficients); network structure hyperparameters (such as hidden layer dimension, number of feature channels, number of attention heads, recurrent unit dimension, Dropout ratio, etc.); and optimizer type and learning rate scheduling strategy (such as Adam, SGD and their learning rate and decay settings). S4.3: Using the obtained optimal deep learning multi-output regression model parameter configuration, train the deep learning multi-output regression model on the training set and evaluate the regression performance on the independent evaluation validation set or test set.

8. The method according to claims 6 and 7, characterized in that: In steps S3 and S4, if the model accuracy in any step does not meet the preset threshold, steps S3 and S4 are repeated until the model accuracy meets the requirements by changing the machine learning classification model or the deep learning multi-output regression model, or by adjusting the parameters of the machine learning classification model and the deep learning multi-output regression model.

9. The method according to claim 8, characterized in that: In step S4, the evaluation index is one or more of the following: coefficient of determination (R²), mean absolute error (MAE), and root mean square error (RMSE).

10. The method according to claim 9, characterized in that: In step S3, the evaluation metric is accuracy, and the formula for calculating accuracy is as follows: The determination coefficient R 2 The calculation formula is: The formula for calculating the root mean square error (RMSE) is as follows: The formula for calculating the mean absolute error (MSE) is as follows: （5） In the formula, Ncorrect is the number of samples where the predicted value equals the actual value, Ntotal is the total number of samples for this output, and n is the total number of samples. and These represent the actual value and the predicted value of the i-th element, respectively. This represents the average of the actual values ​​of all target variables.

Images