Method and device for training catalyst activity prediction model, and storage medium

Abstract

This invention provides a training method for a catalyst activity prediction model, belonging to the field of catalytic chemistry technology. The training method includes: acquiring multiple molecular structure images of multiple catalysts and the catalytic activity of each catalyst at multiple polymerization reaction temperatures; performing feature extraction processing on the multiple catalyst molecular structure images to obtain multiple catalyst molecular structure feature vectors; fusing the multiple catalyst molecular structure feature vectors with temperature features corresponding to the multiple polymerization reaction temperatures to obtain multiple fused input feature vectors; obtaining a model training set based on the multiple fused input feature vectors and their corresponding catalytic activities; and training a pre-constructed initial catalyst activity prediction model using the model training set to obtain the final catalyst activity prediction model. This application can improve the efficiency of determining the catalytic activity of catalysts.

Description

Technical Field

[0001] This invention relates to the field of catalytic chemistry technology, and more specifically to a training method, apparatus, and storage medium for a catalyst activity prediction model. Background Technology

[0002] In existing technologies, catalyst activity assessment mainly relies on experimental testing or an experience-driven "trial and error" approach, requiring repeated polymerization experiments under different catalyst structures and reaction temperatures. This method suffers from long experimental cycles and high costs in determining the quantitative relationship between catalyst molecular structure, reaction temperature, and catalytic activity. Therefore, existing technologies have the problem of relatively low efficiency in determining catalyst catalytic activity. Summary of the Invention

[0003] The purpose of this application is to provide a training method, apparatus, storage medium, and computer program product for a catalyst activity prediction model, in order to solve the problem of low efficiency in determining the catalytic activity of catalysts in the prior art.

[0004] To achieve the above objectives, a first aspect of this application provides a method for training a catalyst activity prediction model, the method comprising:

[0005] To obtain multiple molecular structure images of multiple catalysts and the catalytic activity of each catalyst at multiple polymerization reaction temperatures; Feature extraction was performed on multiple catalyst molecular structure images to obtain multiple catalyst molecular structure feature vectors; Multiple catalyst molecular structure feature vectors and temperature features corresponding to multiple polymerization reaction temperatures are fused to obtain multiple fused input feature vectors; The model training set is obtained based on multiple fused input feature vectors and their corresponding catalytic activities; The pre-constructed initial catalyst activity prediction model is trained using the model training set to obtain the final catalyst activity prediction model.

[0006] In this embodiment of the application, fusing multiple catalyst molecular structure feature vectors and multiple temperature features corresponding to polymerization reaction temperatures to obtain multiple fused input feature vectors includes: converting multiple polymerization reaction temperatures into multiple Kelvin temperatures and determining the reciprocals of multiple Kelvin temperatures to obtain multiple temperature features; and fusing multiple catalyst molecular structure feature vectors and multiple temperature features to obtain multiple fused input feature vectors.

[0007] In this embodiment of the application, feature extraction processing is performed on multiple catalyst molecular structure images to obtain multiple catalyst molecular structure feature vectors, including: based on a preset correspondence between catalyst molecular structure images and catalyst molecular structure strings, determining the target molecular structure strings corresponding to the multiple catalyst molecular structure images respectively; and based on a pre-determined string vectorization encoding model, vectorizing the target molecular structure strings corresponding to the multiple catalyst molecular structure images to obtain multiple catalyst molecular structure feature vectors.

[0008] In this embodiment of the application, obtaining the model training set based on multiple fused input feature vectors and corresponding catalytic activities includes: converting the catalytic activity into logarithmic form to obtain the converted logarithmic catalytic activity; and obtaining the model training set based on multiple fused input feature vectors and corresponding converted logarithmic catalytic activities.

[0009] A second aspect of this application provides a method for determining the catalytic activity of a catalyst, the method comprising: Obtain molecular structure images of the catalyst under test at the polymerization temperature under test; Feature extraction processing is performed on the molecular structure image of the catalyst to be tested to obtain the feature vector of the molecular structure of the catalyst to be tested; The molecular structure feature vector of the catalyst to be tested and the temperature feature corresponding to the polymerization reaction temperature to be tested are fused to obtain the target fused input feature vector; Based on the pre-trained catalyst activity prediction model, the catalytic activity of the catalyst to be tested is determined according to the target fusion input feature vector. The catalyst activity prediction model is trained using the training method described above.

[0010] In this embodiment of the application, the feature extraction processing of the image of the catalyst molecular structure to be tested to obtain the feature vector of the catalyst molecular structure to be tested includes: based on the preset correspondence between the catalyst molecular structure image and the catalyst molecular structure string, determining the string of the molecular structure to be tested corresponding to the image of the catalyst molecular structure to be tested; and based on a pre-determined string vectorization encoding model, vectorizing the string of the molecular structure to be tested to obtain the feature vector of the catalyst molecular structure to be tested.

[0011] In this embodiment of the application, fusing the molecular structure feature vector of the catalyst to be tested and the temperature feature corresponding to the polymerization reaction temperature to be tested to obtain the target fusion input feature vector includes: converting the polymerization reaction temperature to be tested into the Kelvin temperature to be tested, and determining the reciprocal of the Kelvin temperature to be tested to obtain the temperature feature; fusing the molecular structure feature vector of the catalyst to be tested and the temperature feature to be tested to obtain the target fusion input feature vector.

[0012] A third aspect of this application provides a training apparatus for a catalyst activity prediction model, comprising: The memory is configured to store instructions; and the processor is configured to retrieve instructions from the memory and, when executing the instructions, to implement the training method for the catalyst activity prediction model described above.

[0013] A fourth aspect of this application provides a machine-readable storage medium storing instructions for causing a machine to execute the above-described training method for a catalyst activity prediction model.

[0014] The fifth aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method for training a catalyst activity prediction model.

[0015] The above technical solutions (the beneficial effects of the embodiments of the present invention are described in conjunction with the technical solutions, working principles, etc.)

[0016] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 The schematic diagram illustrates a flowchart of a method for training a catalyst activity prediction model according to an embodiment of this application; Figure 2 A diagram illustrating the comparison between predicted and actual catalytic activity results according to embodiments of this application is provided. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0019] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with the relevant provisions of national laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0020] It should be noted that if the embodiments of this application involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0021] Furthermore, if the embodiments of this application involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0022] Figure 1 The illustration shows a flowchart of a training method for a catalyst activity prediction model according to an embodiment of this application. Figure 1 As shown in the figure, this application provides a method for training a catalyst activity prediction model. Taking the application of this method to a processor as an example, the method may include the following steps: Step S101: Obtain multiple molecular structure images of multiple catalysts and the catalytic activity of each catalyst at multiple polymerization reaction temperatures; Step S102: Perform feature extraction processing on multiple catalyst molecular structure images to obtain multiple catalyst molecular structure feature vectors; Step S103: Fuse multiple catalyst molecular structure feature vectors and multiple temperature features corresponding to polymerization reaction temperatures to obtain multiple fused input feature vectors. Step S104: Obtain the model training set based on multiple fused input feature vectors and their corresponding catalytic activities; Step S105: Train the pre-constructed initial catalyst activity prediction model according to the model training set to obtain the final catalyst activity prediction model.

[0023] The catalyst molecular structure image is an image that reflects the molecular structure of the catalyst. The polymerization reaction temperature is the temperature applied during the catalytic reaction process. The catalyst molecular structure feature vector is a feature vector used to express the molecular structure of the catalyst. The temperature feature is the feature related to temperature after processing the polymerization reaction temperature. The fused input feature vector is the feature vector obtained by fusing the catalyst molecular structure feature vector and the temperature feature. The model training set is the training set used to train the model. The catalyst activity prediction model is the model used to predict the activity of the catalyst.

[0024] Specifically, first, the processor collects images of the molecular structures of various catalysts, along with their catalytic activity data at different polymerization temperatures. Next, each catalyst molecular structure image is analyzed to extract key structural information, forming a corresponding catalyst molecular structure feature vector. Then, the extracted feature vectors are combined with the corresponding polymerization temperature information (temperature features) to obtain a fused input feature vector. These fused input feature vectors are then paired with their corresponding catalytic activity data to construct a complete dataset for model training. Finally, this constructed training set is used to train and optimize a pre-designed initial catalyst activity prediction model, ultimately resulting in a reliable catalyst activity prediction model.

[0025] The aforementioned technical solution collects images of the molecular structures of various catalysts, along with catalytic activity data of these catalysts at different polymerization temperatures. This overcomes the limitations of traditional methods that rely solely on single molecular structures or single temperature conditions. The inclusion of multiple catalysts and different polymerization temperatures allows the training data to cover a wider range of substances and process conditions, solving the problem that traditional models can only adapt to specific catalysts or temperatures. Next, each catalyst molecular structure image is analyzed to extract key structural information, forming a corresponding catalyst molecular structure feature vector. This transforms unstructured catalyst molecular structure images into structured feature vectors, addressing the technical challenge of directly inputting image data into machine learning models. Then, the extracted catalyst molecular structure feature vectors are combined with the corresponding polymerization temperature information (temperature features) to obtain a fused input feature vector. This precisely matches the characteristic that catalytic activity during the catalytic reaction is determined by both the catalyst structure and the reaction temperature. These fused input feature vectors are then paired with their corresponding catalytic activity data to construct a complete dataset for model training. Finally, using this constructed training set, the pre-designed initial catalyst activity prediction model is trained and optimized, ultimately resulting in a reliable catalyst activity prediction model. By using a high-quality model training set to train the initial catalyst activity prediction model, the model can learn the complex nonlinear relationship between molecular structure, temperature, and catalytic activity, gradually correcting the model parameters and significantly improving the accuracy and precision of the prediction.

[0026] In one embodiment, fusing multiple catalyst molecular structure feature vectors and multiple temperature features corresponding to polymerization reaction temperatures to obtain multiple fused input feature vectors includes: converting multiple polymerization reaction temperatures into multiple Kelvin temperatures and determining the reciprocals of the multiple Kelvin temperatures to obtain multiple temperature features; and fusing multiple catalyst molecular structure feature vectors and multiple temperature features to obtain multiple fused input feature vectors.

[0027] It can be understood that Kelvin temperature is temperature expressed in Kelvin. Temperature characteristics are the reciprocal of Kelvin temperature.

[0028] Specifically, the processor first converts multiple polymerization reaction temperatures into multiple temperatures in Kelvin and determines the reciprocals of these temperatures in Kelvin to obtain multiple temperature features. Then, it fuses multiple catalyst molecular structure feature vectors with the corresponding temperature features to obtain multiple fused input feature vectors.

[0029] In the embodiments of this application, the reaction temperature is uniformly converted to the Kelvin scale and used as a continuous numerical feature input model in the form of the reciprocal. This makes the catalyst activity prediction model conform to the physicochemical laws of the Arrhenius equation in reaction kinetics.

[0030] In one embodiment, feature extraction processing is performed on multiple catalyst molecular structure images to obtain multiple catalyst molecular structure feature vectors, including: based on a preset correspondence between catalyst molecular structure images and catalyst molecular structure strings, determining the target molecular structure strings corresponding to the multiple catalyst molecular structure images respectively; and based on a pre-determined string vectorization encoding model, vectorizing the target molecular structure strings corresponding to the multiple catalyst molecular structure images to obtain multiple catalyst molecular structure feature vectors.

[0031] It can be understood that the pre-defined correspondence between the catalyst molecular structure image and the catalyst molecular structure string is the pre-defined correspondence between the catalyst molecular structure image and the catalyst molecular structure string. The target molecular structure string is the finally determined string used to express the molecular structure of the catalyst. The string vectorization encoding model is the model used to vectorize and encode the catalyst molecular structure string.

[0032] Specifically, based on a preset correspondence between catalyst molecular structure images and catalyst molecular structure strings, the processor converts multiple catalyst molecular structure images into target molecular structure strings corresponding to each catalyst molecular structure image. Based on a pre-determined string vectorization encoding model, the processor performs vectorization encoding on the target molecular structure strings corresponding to the multiple catalyst molecular structure images to obtain multiple catalyst molecular structure feature vectors.

[0033] In this embodiment, a string vectorization encoding model (Transformer model) is used to perform end-to-end semantic encoding on multiple catalyst molecular structure strings. This automatically learns the local and global semantic features of the molecular structure, eliminating the need for manual construction of molecular descriptors or graph structure rules. This avoids the reliance on experience in selecting atomic features and setting adjacency relationships during graph construction, and can more completely preserve the sequence, ring structure, and long-range association information in the molecular structure. It also improves the model's ability to express and generalize the structural differences of complex metallocene catalysts.

[0034] In one embodiment, obtaining the model training set based on multiple fused input feature vectors and their corresponding catalytic activities includes: converting the catalytic activity into logarithmic form to obtain the converted logarithmic catalytic activity; and obtaining the model training set based on multiple fused input feature vectors and their corresponding converted logarithmic catalytic activities.

[0035] It is understandable that the logarithmic catalytic activity is the same as the logarithmic catalytic activity.

[0036] Specifically, the processor converts the catalytic activity into logarithmic form to obtain the converted logarithmic catalytic activity, and obtains the model training set based on multiple fused input feature vectors and the corresponding converted logarithmic catalytic activities.

[0037] In the embodiments of this application, the natural logarithm of catalytic activity is taken so that the model learning process conforms to the physicochemical laws of the Arrhenius equation. This allows the model to more naturally capture the intrinsic effects of temperature changes on reaction rate and catalytic activity, and significantly improves the stability and consistency of prediction results in different temperature ranges, while reducing the prediction bias of the model when extrapolating temperature conditions.

[0038] In one embodiment, this application provides a method for determining the catalytic activity of a catalyst, the method comprising: Obtain molecular structure images of the catalyst under test at the polymerization temperature under test; Feature extraction processing is performed on the molecular structure image of the catalyst to be tested to obtain the feature vector of the molecular structure of the catalyst to be tested; The molecular structure feature vector of the catalyst to be tested and the temperature feature corresponding to the polymerization reaction temperature to be tested are fused to obtain the target fused input feature vector; Based on the pre-trained catalyst activity prediction model, the catalytic activity of the catalyst to be tested is determined according to the target fusion input feature vector. The catalyst activity prediction model is trained using the training method described above.

[0039] It can be understood that the catalyst under test is the catalyst with the desired activity. The polymerization reaction temperature to be tested is the reaction temperature of the catalyst under test during the catalytic reaction process. The molecular structure image of the catalyst under test is the molecular structure image of the catalyst under test. The molecular structure feature vector of the catalyst under test is the molecular structure feature vector of the catalyst under test. The temperature feature to be tested is the temperature feature corresponding to the polymerization reaction temperature to be tested after processing. The target fusion input feature vector is the fused input feature vector corresponding to the fusion of the molecular structure feature vector of the catalyst under test and the temperature feature to be tested.

[0040] Specifically, firstly, the processor acquires an image of the molecular structure of the catalyst under test and the target polymerization temperature. Next, the image of the molecular structure is analyzed to extract key structural information, resulting in a feature vector of the catalyst's molecular structure. Then, the extracted feature vector is fused with the corresponding temperature feature to obtain a target fused input feature vector. Finally, based on a pre-trained catalyst activity prediction model, the catalytic activity of the catalyst is determined according to the target fused input feature vector. This catalyst activity prediction model is trained using the aforementioned training method.

[0041] The aforementioned technical solution collects images of the molecular structure of the catalyst under test and the polymerization reaction temperature of the catalyst under test, overcoming the limitations of traditional methods that rely solely on single molecular structures or single temperature conditions. Next, the molecular structure images of the catalyst under test are analyzed to extract key structural information, forming corresponding feature vectors of the catalyst's molecular structure. This transforms unstructured images of the catalyst's molecular structure into structured feature vectors, addressing the technical challenge of directly inputting image data into machine learning models for computation. Then, the extracted feature vectors of the catalyst's molecular structure are combined with the corresponding temperature features to obtain a fused target input feature vector, accurately matching the characteristic that catalytic activity during the catalytic reaction is determined by both the catalyst's structure and the reaction temperature. Finally, the target fused input feature vector is input into a pre-trained initial catalyst activity prediction model, thereby outputting accurate catalytic activity.

[0042] In one embodiment, feature extraction processing of the image of the catalyst molecular structure to be tested to obtain the feature vector of the catalyst molecular structure to be tested includes: based on a preset correspondence between the catalyst molecular structure image and the catalyst molecular structure string, determining the string of the molecular structure to be tested corresponding to the image of the catalyst molecular structure to be tested; and based on a pre-determined string vectorization encoding model, vectorizing the string of the molecular structure to be tested to obtain the feature vector of the catalyst molecular structure to be tested.

[0043] It can be understood that the molecular structure string to be tested is the molecular structure string corresponding to the molecular structure image of the catalyst to be tested.

[0044] Specifically, the processor converts the image of the catalyst molecular structure to be tested into a string of molecular structures based on a preset correspondence between the catalyst molecular structure image and the catalyst molecular structure string. Based on a pre-determined string vectorization encoding model, the processor performs vectorization encoding on the string of molecular structures to be tested to obtain the feature vector of the catalyst molecular structure to be tested.

[0045] In this embodiment, unstructured molecular structure images are transformed into standardized molecular structure strings (such as SMILES) through a predefined correspondence, eliminating differences caused by image format, resolution, and drawing style, and achieving a unified expression of molecular structure information. Using the catalyst molecular structure string directly as input, a predefined string vectorization encoding model (Transformer model) is used to perform end-to-end semantic encoding of the molecular structure. This automatically learns the local and global semantic features of the molecular structure, eliminating the need for manually constructing molecular descriptors or graph structure rules. This avoids the reliance on experience in selecting atomic features and setting adjacency relationships during graph construction, and more completely preserves sequence, ring structures, and long-range association information in the molecular structure, improving the model's ability to express and generalize complex catalyst structural differences.

[0046] In one embodiment, fusing the molecular structure feature vector of the catalyst to be tested and the temperature feature corresponding to the polymerization reaction temperature to be tested to obtain the target fusion input feature vector includes: converting the polymerization reaction temperature to be tested into the Kelvin temperature to be tested, and determining the reciprocal of the Kelvin temperature to be tested to obtain the temperature feature; fusing the molecular structure feature vector of the catalyst to be tested and the temperature feature to be tested to obtain the target fusion input feature vector.

[0047] It can be understood that the Kelvin temperature to be measured is the polymerization reaction temperature in Kelvin. The characteristic of the temperature to be measured is the reciprocal of the Kelvin temperature to be measured.

[0048] Specifically, the processor converts the polymerization reaction temperature to be measured into the Kelvin temperature to be measured, and determines the reciprocal of the Kelvin temperature to obtain the temperature characteristics. Then, it fuses the catalyst molecular structure feature vector and the temperature characteristics to obtain the target fused input feature vector.

[0049] In this embodiment, based on the principle of chemical reaction kinetics, the reaction temperature is uniformly converted to the Kelvin temperature scale, and the reciprocal of the temperature (1 / T) is used as the model input feature. This makes the model learning process conform to the physicochemical laws of the Arrhenius equation, enabling the model to capture the effect of temperature changes on the reaction rate more naturally, significantly improving the stability and consistency of the prediction results in different temperature ranges, and reducing the prediction bias of the model when extrapolating temperature conditions.

[0050] A specific embodiment of this application provides a training method for a catalyst activity prediction model and a method for determining the catalytic activity of a catalyst. The specific steps are as follows: (1) Collect literature data A large number of literatures on the synthesis of polyethylene by metallocene catalysis were collected, and the temperature, catalyst images, and catalytic activity were listed in a table.

[0051] (2) Data processing of temperature and activity In metallocene-catalyzed polymerization of ethylene or ethylene / α-olefins, catalytic activity is influenced by both reaction temperature and catalyst structure, with reaction temperature having a significant effect on the polymerization rate. To improve the ability of catalytic activity prediction models to characterize the effect of temperature and to ensure that the prediction results conform to the physicochemical nature of catalytic reactions, this invention performs unified data processing and standardized transformation on temperature parameters and catalytic activity data.

[0052] First, regarding temperature data processing, the experimentally recorded reaction temperatures were uniformly converted from Celsius to Kelvin temperatures on the thermodynamic temperature scale. Furthermore, the reciprocal form of the temperature was used as the model input feature, specifically 1 / T, where T is the absolute temperature expressed in Kelvin. This approach stems from the Arrhenius relation in chemical reaction kinetics, which indicates an exponential relationship between the reaction rate constant and the reciprocal of temperature. By converting the temperature to its reciprocal form, the model can more easily capture the influence of temperature changes on the catalytic reaction rate and catalytic activity, thereby improving the stability and physical plausibility of the predictions.

[0053] Secondly, regarding the processing of catalytic activity data, since the units of catalytic activity reported under different literature and experimental conditions vary, this invention first standardizes the units of the original catalytic activity data to kg / (mol). h), so that they can be compared and modeled under the same dimensional system. After unifying the units, the natural logarithm of the catalytic activity values ​​is taken to obtain the logarithmic form of the catalytic activity characterization. By taking the logarithm of the catalytic activity, on the one hand, the numerical range of the original activity data can be compressed, reducing the impact of extreme values ​​on the model training process; on the other hand, this method is consistent with the characteristic that the logarithm of the reaction rate constant in the Arrhenius equation has a linear relationship with the reciprocal of temperature, which helps the model learn the intrinsic relationship between temperature, catalyst structure and catalytic activity.

[0054] Through the above data processing methods, this invention transforms the reaction temperature characteristics and catalytic activity characterization into a form that is more in line with the laws of reaction kinetics. This allows the AI-based prediction model to simultaneously take into account both data-driven characteristics and physicochemical mechanism constraints during training and prediction, thereby improving the accuracy, stability, and interpretability of the catalytic activity prediction results.

[0055] (3) Convert the catalyst image from the uniform extraction rule (Canonical SMILES) to SMILES. The rules used for image transformation here are as follows: ① Atomic Representation: Basic nonmetallic atoms are represented by their respective uppercase symbols, such as carbon atom "C", sulfur atom "S", and oxygen atom "O". Atoms in aromatic rings are represented by their corresponding lowercase letters, such as benzene ring "c1ccccc1" and benzoic acid "O=C(O)c1ccccc1". Ions and metal atoms are usually represented by square brackets and their charge values, such as ferrous ion "[Fe+2]", hydroxide ion "[OH-]", and gold "[Au]". Symbols that are adjacent in arrangement are also connected in the molecule, such as in benzoic acid "O=C(O)c1ccccc1", where oxygen atom "O" is connected to carbon atom "C", and in benzene ring, carbon atom "c1" is connected to subsequent carbon atoms "c". Hydrogen atoms are usually omitted and are not represented in the SMILES string. Furthermore, isotopes with special masses require the atomic mass to be added before the atomic symbol, and in this case, hydrogen atoms must be explicitly indicated, for example, carbon-13 methane is represented as "13CH4".

[0056] ② Representation of chemical bonds: Single bonds between atoms are represented by "-", but are usually omitted. Double bonds, triple bonds, quadruple bonds, and aromatic bonds are represented by "=", "#", "$", and ":", respectively. Double, triple, and quadruple bonds cannot be omitted, but aromatic bonds are similar to single bonds and are generally omitted. However, when aromatic compounds are linked to form a new compound, the specific position and symbol of the bond must be indicated. For example, the chemical formula of ethane is "CH3CH3", so its smils are represented by "CC". The chemical formula of acetylene is "C2H2", so its smils are represented by "C#C". The smils of benzoic acid containing a benzene ring are represented by "O=C(O)c1ccccc1".

[0057] ③ Special Structure Representation: Branched structures use "()" and ")" to indicate the beginning and end of the branch, respectively. For example, in benzoic acid "O=C(O)c1ccccc1", "(O)" indicates that there is one oxygen atom linking the carbon backbone. It is generally preferred to write shorter, easier-to-identify, and simpler branched structures within parentheses to represent the branch, while treating relatively complex branches as the backbone, in order to simplify the representation of branched structures. Ring structures need to be opened and represented as linear structures when generating SMILES representations. The atoms forming the ring are identified by numbers. For example, in benzoic acid "O=C(O)c1ccccc1", "c1ccccc1" indicates the benzene ring is opened, "c1" indicates the starting point of the benzene ring opening, and also indicates that the two "c1" carbon atoms were originally linked. Free structures without interacting chemical bonds are separated by ".". For example, the SMILES representation of copper sulfate pentahydrate is "OOOOO[O-]S(=O)(=O)[O-].[Cu+2]". The stereostructure of a compound is indicated by “ / ” and “\” to represent cis and trans, and the chiral structure is indicated by “@” and “@@”, for example, “[C@H]” and “[C@@H]”.

[0058] (4) Organize and output the processed data as a .csv file. (5) Machine learning (the molecular vector obtained through Transformer encoding - RandomForest model) Step 1: Data Reading and Sample Construction The above .csv file was read using Python. The sample data included: catalyst molecular structure information; polymerization reaction temperature parameters; and corresponding catalytic activity data. The catalyst molecular structure was represented as a string, the reaction temperature data was preprocessed to reflect the reaction kinetics, and the catalytic activity data were activity characterization quantities after unit standardization and logarithmic transformation.

[0059] Step 2: Deep semantic encoding of catalyst molecular structure The catalyst molecular structure string is input into a pre-trained molecular structure encoding model to vectorize the molecular structure, obtaining the corresponding high-dimensional molecular structure feature vector. The molecular structure encoding model is based on a deep learning architecture and can automatically extract semantic information reflecting atomic connections and local and global structural features from the molecular structure sequence. During the encoding process, features at various positions in the molecular sequence can be aggregated to obtain a fixed-dimensional molecular vector representation.

[0060] Step 3: Temperature Feature Acquisition and Construction The reaction temperature parameters corresponding to each catalytic reaction sample are obtained and input into the model as independent numerical features. These temperature features are constructed based on thermodynamic principles to enhance the model's ability to learn the influence of temperature changes on catalytic activity.

[0061] Step 4: Feature Fusion The molecular structure feature vector obtained in step 2 is concatenated or combined with the temperature feature vector obtained in step 3 to construct a joint input feature vector for predicting catalytic activity. Through feature fusion, the model can simultaneously utilize catalyst structure information and reaction temperature information to model their synergistic effect.

[0062] Step 5: Dataset Partitioning The joint input feature vector and the corresponding catalytic activity label data are divided into training dataset, validation dataset and test dataset for model training, parameter optimization and performance evaluation.

[0063] Step 6: Training the Catalytic Activity Prediction Model The training dataset is input into the regression prediction model for training, establishing a mapping relationship between catalyst structure, reaction temperature, and catalytic activity. During training, the model parameters are evaluated and adjusted using the validation dataset to improve the model's prediction accuracy and generalization ability. The regression prediction model can be a regression model based on ensemble learning or gradient boosting, capable of handling high-dimensional feature inputs and characterizing complex nonlinear relationships.

[0064] Step 7: Model Validation and Testing The trained prediction model was used to predict catalytic activity on validation and test datasets, and the prediction results were evaluated using error and correlation metrics to verify the model's prediction performance and stability on different datasets.

[0065] Step 8: Model Saving and Retrieval The trained catalytic activity prediction model is persistently stored so that it can be directly called upon in subsequent catalyst screening or process optimization, enabling rapid prediction of catalytic activity under new catalyst structure and reaction temperature conditions.

[0066] In practical applications, users can input the reaction temperature parameter to be evaluated and the string representation of the catalyst molecular structure into the system. After receiving the input information, the system automatically calls the pre-trained catalytic activity prediction model to perform joint calculations on the input catalyst structure and reaction temperature, and outputs the corresponding catalytic activity prediction results.

[0067] The prediction results include the logarithmic form of the catalytic activity and the converted actual catalytic activity value, which can intuitively reflect the catalytic performance level under given reaction temperature and catalyst structure conditions. Through this method, the present invention automates and visualizes the catalytic activity prediction process, allowing users to quickly obtain catalytic activity assessment results without conducting actual polymerization experiments.

[0068] This system can be used for rapid comparative analysis of different catalyst structures and under different reaction temperature conditions, providing auxiliary decision support for catalyst screening and process optimization.

[0069] like Figure 2 As shown, Figure 2A schematic diagram illustrating the comparison between predicted and actual catalytic activity results according to embodiments of this application is provided. The horizontal axis represents the actual logarithmic value of catalytic activity, and the vertical axis represents the predicted logarithmic value of catalytic activity obtained by the method of this invention. The dashed line in the figure represents the reference line where the predicted value and the actual value are equal under ideal prediction conditions. As can be seen from the figure, most sample points are distributed near the ideal prediction line, and the predicted value and the actual value show a good linear correlation, indicating that the prediction model constructed by this invention can accurately predict catalytic activity over a wide range of catalytic activity values. Quantitative evaluation of the prediction results on the test set shows that the model's prediction error on the test set is at a low level and has a high correlation index, indicating that this method can effectively capture the combined influence of catalyst molecular structure characteristics and reaction temperature on catalytic activity. The above results demonstrate that the method of this invention still has good prediction accuracy and stability under unknown sample conditions, and is suitable for applications such as catalyst screening and process parameter optimization.

[0070] The following describes the training method for a catalyst activity prediction model and the technical effects of a method for determining the catalytic activity of a catalyst, as described in the embodiments of this application: I. This invention achieves high-precision prediction of catalytic activity by innovatively combining the Transformer model and the RandomForest regression model. This innovation is mainly reflected in two aspects: deep encoding of catalyst structural characteristics and accurate modeling of temperature effects, significantly improving the accuracy, stability, and generalization ability of the prediction model.

[0071] First, regarding the feature extraction of catalyst molecular structure, this invention employs the Transformer model to perform deep semantic encoding on the SMILES representation of the catalyst. As a powerful deep learning architecture, the Transformer model can automatically learn the atomic connections and local and global structural features in the molecular structure, extracting a more detailed and accurate molecular representation through a self-attention mechanism. This encoding method can capture the complexity of the catalyst structure at the molecular level, thus providing rich feature information for catalytic activity prediction. Compared with traditional methods of manually designing molecular descriptors, the Transformer model can more effectively handle the complexity and high-dimensionality of molecular structures, providing more expressive features for subsequent prediction models.

[0072] Secondly, to further improve the model's predictive performance, this invention combines the molecular features obtained from Transformer encoding with the RandomForest regression model for catalytic activity prediction. RandomForest is an ensemble learning method based on bagging, possessing powerful nonlinear modeling capabilities and an efficient training algorithm. In this invention, the RandomForest model is used to learn the complex relationship between the reciprocal of temperature and the catalyst molecular structure features and catalytic activity. By jointly modeling Transformer features and temperature information, RandomForest can effectively capture the synergistic effect of both, enhancing the model's predictive ability for changes in catalytic activity under the combined influence of temperature and catalyst structure.

[0073] This innovative approach, combining the Transformer and RandomForest models, enables the invention to simultaneously consider the structural characteristics of the catalyst and the influence of reaction temperature, thereby effectively improving the accuracy and reliability of catalytic activity prediction. Through this method, the model can not only handle high-dimensional structural data but also capture the complex nonlinear relationship between the catalyst and temperature, thus providing a powerful tool to assist in catalyst screening and process optimization.

[0074] Through this innovation, the present invention introduces the powerful capabilities of deep learning and ensemble learning into the prediction of catalytic activity, enabling the prediction model to have higher accuracy, stability and wider applicability, and can provide effective technical support for the research and development of high-end polyolefin materials, catalyst design and reaction condition optimization.

[0075] Second, this invention, through an innovative method for predicting catalytic activity based on joint modeling of the reciprocal of temperature and catalyst structure, combined with unified data processing and standardized conversion technology, effectively solves the problems of low accuracy in predicting catalytic activity and insufficient characterization of the effect of reaction temperature on activity in traditional technologies.

[0076] First, this invention uses the reaction temperature as its reciprocal form (1 / T) as a model input feature and integrates it with the deep semantic features of the catalyst molecular structure. The reciprocal temperature treatment, derived from the physicochemical basis of the Arrhenius equation, helps to more accurately capture the nonlinear relationship between reaction temperature and catalytic activity. Simultaneously, by combining catalyst structural information, it effectively characterizes the synergistic effect of the catalyst and reaction temperature on the catalytic reaction rate and catalytic activity.

[0077] Secondly, to address the issue of inconsistent units in catalytic activity data across different literature sources, this invention standardizes the units of the original catalytic activity data and employs a natural logarithmic transformation technique to convert the catalytic activity into logarithmic form. This compresses extreme fluctuations in the data and reduces the impact of extreme values ​​on model training. This approach not only improves data consistency and stability but also allows the model to better conform to reaction kinetics during training, thereby enhancing prediction accuracy and stability.

[0078] Through the above innovations, this invention provides a highly efficient, stable, and physicochemically sound method and system for predicting catalytic activity, enabling rapid prediction and accurate evaluation during catalyst screening and process optimization, and has broad application prospects.

[0079] III. Catalytic Activity Prediction Panel System Based on Direct Input-Output The third innovation of this invention lies in the automation and visualization of the catalytic activity prediction process by establishing a panel system that directly outputs from inputs. This system innovation not only optimizes the operational process of catalytic activity prediction but also greatly improves user convenience and efficiency, playing a crucial role, especially in catalyst screening and reaction condition optimization.

[0080] First, the system features an interactive panel interface. Users simply input the reaction temperature (in degrees Celsius) and the SMILES string representing the catalyst molecular structure, and the system calculates and provides real-time predictions of catalytic activity. Specifically, users input the temperature value and catalyst structure (in normalized SMILES notation) into the interface, and the system automatically invokes a pre-trained catalytic activity prediction model to process the input data and output the corresponding prediction results. By combining a deep learning model (Transformer encoding) and a RandomForest regression model, this system can provide catalytic activity predictions under given temperature and catalyst structure conditions in a short time, greatly simplifying the traditional experimental trial-and-error process.

[0081] Secondly, the visualization of prediction results is another significant innovation of this panel system. The system helps users quickly understand the reliability of the predictions. Furthermore, the system displays both the logarithmic form of the catalytic activity and the converted actual catalytic activity value, allowing users to clearly and intuitively see the catalytic activity levels of different catalysts at different reaction temperatures. This combination of real-time feedback and visualization not only enhances the user experience but also helps researchers make faster and more accurate decisions in catalyst development and optimization.

[0082] In one embodiment, this application provides a training apparatus for a catalyst activity prediction model, comprising: The memory is configured to store instructions; and the processor is configured to retrieve instructions from the memory and, when executing the instructions, to implement the training method for the catalyst activity prediction model described above.

[0083] In one embodiment, this application provides a machine-readable storage medium storing instructions for causing a machine to execute the above-described training method for a catalyst activity prediction model.

[0084] In one embodiment, this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method for training a catalyst activity prediction model.

[0085] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0086] The above are merely embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application. It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

Claims

1. A training method for a catalyst activity prediction model, characterized in that, The training method includes: Obtain multiple molecular structure images of multiple catalysts and the catalytic activity of each catalyst at multiple polymerization reaction temperatures; Feature extraction processing is performed on multiple catalyst molecular structure images to obtain multiple catalyst molecular structure feature vectors; Multiple catalyst molecular structure feature vectors and multiple temperature features corresponding to polymerization reaction temperatures are fused to obtain multiple fused input feature vectors; The model training set is obtained based on the multiple fused input feature vectors and the corresponding catalytic activities; The pre-constructed initial catalyst activity prediction model is trained based on the model training set to obtain the final catalyst activity prediction model.

2. The method according to claim 1, characterized in that, The process of fusing multiple catalyst molecular structure feature vectors and multiple temperature features corresponding to polymerization reaction temperatures to obtain multiple fused input feature vectors includes: The multiple polymerization reaction temperatures are converted into multiple Kelvin temperatures, and the reciprocals of the multiple Kelvin temperatures are determined to obtain multiple temperature characteristics; Multiple catalyst molecular structure feature vectors and multiple temperature features are fused to obtain multiple fused input feature vectors.

3. The method according to claim 1, characterized in that, The step of performing feature extraction processing on the multiple catalyst molecular structure images to obtain multiple catalyst molecular structure feature vectors includes: Based on the preset correspondence between catalyst molecular structure images and catalyst molecular structure strings, the target molecular structure strings corresponding to the multiple catalyst molecular structure images are determined according to the multiple catalyst molecular structure images respectively; Based on a pre-determined string vectorization encoding model, the target molecular structure strings corresponding to the multiple catalyst molecular structure images are vectorized and encoded to obtain multiple catalyst molecular structure feature vectors.

4. The method according to claim 1, characterized in that, The process of obtaining the model training set based on multiple fused input feature vectors and their corresponding catalytic activities includes: The catalytic activity is converted into logarithmic form to obtain the converted logarithmic catalytic activity; The model training set is obtained based on the multiple fused input feature vectors and the corresponding transformed logarithmic catalytic activities.

5. A method for determining the catalytic activity of a catalyst, characterized in that, The method includes: Obtain molecular structure images of the catalyst under test at the polymerization temperature under test; The molecular structure image of the catalyst to be tested is subjected to feature extraction processing to obtain the feature vector of the molecular structure of the catalyst to be tested; The molecular structure feature vector of the catalyst to be tested and the temperature feature corresponding to the polymerization reaction temperature to be tested are fused to obtain the target fused input feature vector; Based on the pre-trained catalyst activity prediction model, the catalytic activity of the catalyst to be pre-treated is determined according to the target fusion input feature vector, wherein the catalyst activity prediction model is trained by the training method of the catalyst activity prediction model according to any one of claims 1 to 4.

6. The method according to claim 5, characterized in that, The step of performing feature extraction processing on the molecular structure image of the catalyst to be tested to obtain the feature vector of the molecular structure of the catalyst to be tested includes: Based on the preset correspondence between catalyst molecular structure images and catalyst molecular structure strings, the test molecular structure string corresponding to the test catalyst molecular structure image is determined according to the test catalyst molecular structure image. Based on a pre-determined string vectorization encoding model, the string of the molecular structure to be tested is vectorized to obtain the feature vector of the molecular structure of the catalyst to be tested.

7. The method according to claim 5, characterized in that, The process of fusing the molecular structure feature vector of the catalyst to be tested and the temperature feature corresponding to the polymerization reaction temperature to be tested to obtain the target fused input feature vector includes: The polymerization reaction temperature to be measured is converted into the Kelvin temperature to be measured, and the reciprocal of the Kelvin temperature to be measured is determined to obtain the characteristics of the temperature to be measured. The catalyst molecular structure feature vector and the temperature to be measured are fused to obtain the target fused input feature vector.

8. A training device for a catalyst activity prediction model, characterized in that, include: The memory is configured to store instructions; as well as The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the training method for the catalyst activity prediction model according to any one of claims 1 to 4.

9. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores instructions for causing the machine to execute the training method for the catalyst activity prediction model according to any one of claims 1 to 4.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the training method for the catalyst activity prediction model according to any one of claims 1 to 4.

Images