Polyvinyl alcohol related chemical reaction prediction method based on deep learning framework and multi-scale feature fusion

Abstract

The invention provides a polyvinyl alcohol related chemical reaction prediction method based on a deep learning framework and multi-scale feature fusion, and the method comprises the steps: fusing quantum chemical features (such as HOMO, LUMO, dipole moment and the like) calculated by a density functional theory (DFT), molecular structure SMILES coding, reaction conditions and other multi-scale features, combining a Transform and ChemBERTa pre-training model and an LSTM network, carrying out the prediction of the chemical reaction related to polyvinyl alcohol, and carrying out the prediction of the chemical reaction related to polyvinyl alcohol. An end-to-end deep learning framework is constructed, and high-precision prediction and reverse design of polyvinyl alcohol (PVA) and related polymer reactions are realized. The overall accuracy of the method can exceed 95%, and the method has strong generalization ability, can be expanded to polymer systems such as PE, PP, PVC and the like, and provides a general solution for intelligent reaction optimization in a low-data scene.

Description

Technical Field

[0001] This invention belongs to the fields of cheminformatics and artificial intelligence, and relates to a method for predicting polyvinyl alcohol-related chemical reactions based on a deep learning framework and multi-scale feature fusion. Specifically, it involves constructing a deep learning framework by fusing multi-scale features such as quantum chemical features, molecular structure and reaction conditions to achieve high-precision prediction and reverse design of polyvinyl alcohol (PVA) and related polymer reactions. Background Technology

[0002] In today's world, with the rapid development of materials science and the chemical industry, the demand for chemical reaction prediction is growing rapidly. Compared with traditional experimental methods, machine learning has highly efficient predictive capabilities. Its models can simulate complex reaction processes with minimal computational cost and can be optimized in a controllable manner, thus leading to its rapid application since its discovery. After World War II, research and application of machine learning became a hot topic. To date, machine learning is widely used in fields related to human production and daily life, such as drug discovery, materials design, and reaction optimization. The computational characteristics of machine learning determine its significant dependence on data; therefore, avoiding generalization problems in low-data scenarios has become a core issue that must be addressed in the development and utilization of models. Chemical reaction prediction has become a new and challenging task arising from the application of machine learning, especially in the case of small datasets.

[0003] In the 1960s, a few countries, led by the United States, pioneered systematic research on chemical reaction prediction. Later, with the widespread application of machine learning technology globally, more countries researching and utilizing AI joined the ranks of reaction prediction researchers. Because the reactions to be predicted are diverse, and the required accuracy varies, there is currently no universally applicable prediction method.

[0004] Therefore, to cope with complex situations, the development of prediction technologies has become increasingly diversified. Currently, there are many mature prediction methods, but they can be categorized into rule-based methods and data-driven methods based on their operating principles. Rule-based methods include quantum mechanical simulations, empirical formulas, etc.; data-driven methods include graph neural networks, Transformer models, pre-trained language models, etc.

[0005] Although there are many publicly available prediction methods that can solve a variety of challenging situations, they all generally suffer from the lack of large amounts of data and computing resources. In addition, the type of reaction being predicted, the quality of the data, and the model architecture will all constrain the choice of prediction methods. Summary of the Invention

[0006] To address the problems in the background technology, this invention provides a method for predicting polyvinyl alcohol (PVA)-related chemical reactions based on a deep learning framework and multi-scale feature fusion. This method fuses multi-scale features, including quantum chemical features calculated by density functional theory (DFT) (such as HOMO, LUMO, dipole moment, etc.), SMILES molecular structure encoding, and reaction conditions. Combined with Transformer and ChemBERTa pre-trained models and an LSTM network, an end-to-end deep learning framework is constructed to achieve high-precision prediction and reverse design of PVA and related polymer reactions. The method achieves an overall accuracy exceeding 95%, possesses strong generalization ability, and can be extended to polymer systems such as PE, PP, and PVC, providing a general solution for intelligent reaction optimization in low-data scenarios.

[0007] To achieve the above objectives, the present invention is implemented by adopting a technical solution consisting of the following technical measures.

[0008] In one aspect, this invention provides a method for predicting polyvinyl alcohol-related chemical reactions based on a deep learning framework and multi-scale feature fusion, mainly including the following steps: (1) According to the SMILES (Simplified molecular input line entry system) rules, the molecular structure of PVA and related polymers is standardized and coded to obtain the molecular structure SMILES code; (2) Configure multi-dimensional macroscopic reaction condition parameters including temperature, time, and solvent selection, and normalize them to obtain normalized reaction condition parameters; (3) Using PySCF in the computing environment, based on the molecular structure SMILES encoding obtained in step (1), calculate HOMO, LUMO, dipole moment and total energy in sequence, and configure the basis set def2-SVP and functional B3LYP for optimization to obtain electronic structure features. (4) Align and merge the normalized reaction condition parameters obtained in step (2) with the electronic structure features obtained in step (3) to form a unified feature vector; (5) Input the feature vector obtained in step (4) into a deep learning architecture that integrates Transformer, ChemBERTa pre-trained model and LSTM network, train the model using Adam optimizer and Dropout regularization strategy, and finally optimize the model parameters through cross-validation and early stopping strategy to obtain the optimized PVA-ReAct prediction model.

[0009] The resulting PVA-ReAct prediction model requires no additional pre-training and can be used directly.

[0010] In this paper, step (1) describes the standardization and encoding of the molecular structure of PVA and related polymers according to the SMILES (Simplified molecular input line entry system) rules to obtain the molecular structure SMILES code. The SMILES (Simplified molecular input line entry system) rules are the publicly available specifications in this technical field that explicitly describe molecular structures using ASCII strings. Therefore, those skilled in the art can directly standardize and encode the molecular structure of PVA and related polymers using the above-mentioned SMILES rules.

[0011] In this paper, the configuration in step (2) includes multi-dimensional macroscopic reaction condition parameters such as temperature, time, and solvent selection. To facilitate normalization, the multi-dimensional macroscopic reaction condition parameters including temperature, time, and solvent selection are all standardized and coded. For example, in solvent selection, a coding system is constructed based on commonly used solvents and converted into numerical vectors.

[0012] In this paper, the normalization process described in step (2) includes, but is not limited to, using linear transformation, standard deviation standardization, and conversion of categorical variables into numerical vectors, so that the multidimensional macroscopic reaction condition parameters are converted into normalized reaction condition parameters.

[0013] In this paper, PySCF, mentioned in step (3), is an open-source quantum chemical computing framework based on Python. PySCF supports a variety of molecular input formats (such as XYZ coordinate files). Through tools such as RDKit, molecular structures SMILES can be encoded and converted into three-dimensional molecular structures. The computation time is 10~30 minutes per molecule, and preliminary geometric optimization is performed using the MMFF94 force field.

[0014] In this paper, the calculation of HOMO, LUMO, dipole moment and total energy in step (3) is performed by performing self-consistent field calculation through the scf.hf module of PySCF. The calculation is cyclically converged to the ground state wave function through SCF, and the electronic structure characteristics such as the highest occupied molecular orbital (HOMO), the lowest unoccupied molecular orbital (LUMO) energy level, dipole moment and total energy are output.

[0015] In this paper, the basis set def2-SVP and functional B3LYP mentioned in step (3) are integrated functions in PySCF. The purpose of configuring the basis set def2-SVP and functional B3LYP for optimization is to achieve a balance between computational accuracy and efficiency, while ensuring that the description of the complex polymer system is accurate enough, providing a reasonable initial structure for subsequent accurate calculations, and avoiding convergence difficulties caused by excessive consumption of computational resources.

[0016] In this paper, the normalized reaction condition parameters and electronic structure features described in step (4) are aligned and fused to form a unified feature vector. This usually involves a conventional mechanism for establishing a unified feature dimension. The multidimensional vector of electronic structure features and the multidimensional vector of normalized reaction condition parameters are mapped to the same dimension through a fully connected layer. An attention mechanism is used to perform weighted fusion of the two types of features. The attention weights are dynamically generated through learnable parameters. Then, feature cross-operations are introduced to calculate the cross-features of reaction condition parameters such as temperature, time, and solvent selection with HOMO, LUMO, dipole moment, and total energy.

[0017] In one of the technical solutions, the formation of a unified feature vector in step (4) also includes normalizing the fused feature vector to ensure numerical stability.

[0018] In this paper, the deep learning architecture that integrates Transformer, ChemBERTa pre-trained model and LSTM network described in step (5) is an open-source model architecture that can be downloaded from the network. It is integrated through a conventional open-source deep learning framework package (such as PyTorch) to build a hybrid deep learning architecture.

[0019] In one of the technical solutions, the deep learning architecture that integrates Transformer, ChemBERTa pre-trained model and LSTM network in step (5) is specifically: the bottom layer uses Transformer encoder to extract feature vectors through self-attention mechanism, the middle layer integrates ChemBERTa pre-trained model, and the top layer deploys LSTM network.

[0020] In this paper, the model training using the Adam optimizer and Dropout regularization strategy described in step (5) is a common open-source basic component of the entire model training framework, and it inserts Dropout layers between fully connected layers.

[0021] In this paper, step (5) involves optimizing the model parameters using cross-validation and an early stopping strategy. The cross-validation involves dividing the dataset into multiple subsets, using different subsets as validation sets in turn, and the rest as training sets, thereby evaluating the model performance multiple times and calculating the average metric. The early stopping strategy involves monitoring the validation set performance (such as accuracy or loss value) in real time during training. When the validation set metric stops improving or even begins to decline, training is terminated immediately. This effectively avoids overfitting and prevents the model from overfitting on the training set, ensuring generalization ability. At the same time, it saves computational resources, terminates ineffective training in advance, and improves efficiency.

[0022] In one preferred embodiment, the number of Transformer layers in step (5) is set to 6 to 12, the embedding dimension of the ChemBERTa pre-trained model is 768, the number of units in the LSTM network is 128 to 256, and the number of training rounds is set to 50 to 100.

[0023] Based on the PVA-ReAct prediction model provided by this invention, it can be applied in any of the following ways (I) to (III): (I) Forward prediction: Input reactants and conditions, and predict the product structure and reaction yield using the PVA-ReAct prediction model; (II) Reverse design: Input the target product or target yield, and use the gradient descent algorithm to search in the condition space and reverse-engineer the optimal reaction conditions; (III) The model outputs prediction results or design schemes, and can be generalized to other polymer systems such as PE, PP, and PVC.

[0024] The present invention has the following beneficial effects:

[0025] 1. This invention provides a method for predicting polyvinyl alcohol-related chemical reactions based on a deep learning framework and multi-scale feature fusion. It constructs a low-data-driven prediction framework that can be used for polyvinyl alcohol-related chemical reactions. For reaction prediction in low-data scenarios, after training on a small dataset, the framework can accurately predict the product and yield, and generalize to other systems, thereby achieving intelligent optimization of chemical reactions.

[0026] 2. This invention provides a method for predicting polyvinyl alcohol-related chemical reactions based on a deep learning framework and multi-scale feature fusion. This method has the advantages of high computational efficiency and scalability. It is a combination of quantum chemistry and deep learning: density functional theory (DFT) is used as a quantum component, features are extracted by PySCF, and Transformer-ChemBERTa-LSTM is used as a deep learning component. A deep learning architecture is constructed through multi-scale fusion.

[0027] 3. When the deep learning architecture of this invention is applied to PVA reaction, the results are accurately predicted, the accuracy improves with model optimization, and it is verified by evaluation metrics that it has advantages such as low data requirements and generalizability. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the steps in the polyvinyl alcohol-related chemical reaction prediction method based on deep learning framework and multi-scale feature fusion in Embodiment 1 of the present invention.

[0029] Figure 2This is a summary graph showing the accuracy of the prediction models obtained in Example 1 and Comparative Examples 1-3 of the present invention in predicting PVA reaction types on datasets with different proportions. In the graph, (A) corresponds to the overall accuracy; (B) corresponds to cross-linking reactions; (C) corresponds to addition reactions; (D) corresponds to condensation reactions; (E) corresponds to oxidation reactions; (F) corresponds to reduction reactions; (G) corresponds to hydrolysis reactions; (H) corresponds to esterification reactions; and (I) corresponds to photodegradation reactions. The proportions on the left of all graphs represent the dataset proportions: training set: validation set: test set. All results are expressed as the mean ± standard deviation of 10 replicates.

[0030] Figure 3 The absolute coefficients R for predicting PVA reaction types using the prediction models obtained in Example 1 and Comparative Examples 1-3 of this invention on datasets of different proportions. 2 Summary chart. Figure (A) corresponds to the overall absolute coefficient R. 2 Figure (B) corresponds to the cross-linking reaction; Figure (C) corresponds to the addition reaction; Figure (D) corresponds to the condensation reaction; Figure (E) corresponds to the oxidation reaction; Figure (F) corresponds to the reduction reaction; Figure (G) corresponds to the hydrolysis reaction; Figure (H) corresponds to the esterification reaction; Figure (I) corresponds to the photodegradation reaction; The scale on the left of all figures represents the dataset scale, training set: validation set: test set; All values ​​are expressed as the mean ± standard deviation of 10 replicates. Detailed Implementation

[0031] To further understand the present invention, preferred embodiments are described below with reference to examples. However, it should be understood that these descriptions are merely for illustrating the features and advantages of the invention and not for limiting the scope of the claims. Those skilled in the art can refer to the content herein to appropriately modify the parameters. It is particularly important to note that all similar substitutions and modifications are obvious to those skilled in the art and are considered to be included within the scope of the invention. The methods and applications of the present invention have been described through preferred embodiments, and those skilled in the art can obviously make modifications or appropriate changes and combinations to the methods and applications described herein without departing from the content, spirit, and scope of the invention to implement and apply the technology of the present invention. While it is believed that those skilled in the art will fully understand the following terms, the following definitions are set forth to help illustrate the subject matter disclosed in this invention.

[0032] In one aspect, this invention provides a method for predicting polyvinyl alcohol-related chemical reactions based on a deep learning framework and multi-scale feature fusion, mainly including the following steps: (1) According to the SMILES (Simplified molecular input line entry system) rules, the molecular structure of PVA and related polymers is standardized and coded to obtain the molecular structure SMILES code; (2) Configure multi-dimensional macroscopic reaction condition parameters including temperature, time, and solvent selection, and normalize them to obtain normalized reaction condition parameters; (3) Using PySCF in the computing environment, based on the molecular structure SMILES encoding obtained in step (1), calculate HOMO, LUMO, dipole moment and total energy in sequence, and configure the basis set def2-SVP and functional B3LYP for optimization to obtain electronic structure features. (4) Align and merge the normalized reaction condition parameters obtained in step (2) with the electronic structure features obtained in step (3) to form a unified feature vector; (5) Input the feature vector obtained in step (4) into a deep learning architecture that integrates Transformer, ChemBERTa pre-trained model and LSTM network, train the model using Adam optimizer and Dropout regularization strategy, and finally optimize the model parameters through cross-validation and early stopping strategy to obtain the optimized PVA-ReAct prediction model.

[0033] The resulting PVA-ReAct prediction model requires no additional pre-training and can be used directly.

[0034] In this paper, step (1) describes the standardization and encoding of the molecular structure of PVA and related polymers according to the SMILES (Simplified molecular input line entry system) rules to obtain the molecular structure SMILES code. The SMILES (Simplified molecular input line entry system) rules are the publicly available specifications in this technical field that explicitly describe molecular structures using ASCII strings. Therefore, those skilled in the art can directly standardize and encode the molecular structure of PVA and related polymers using the above-mentioned SMILES rules.

[0035] In this paper, the configuration in step (2) includes multi-dimensional macroscopic reaction condition parameters such as temperature, time, and solvent selection. To facilitate normalization, the multi-dimensional macroscopic reaction condition parameters including temperature, time, and solvent selection are all standardized and coded. For example, in solvent selection, a coding system is constructed based on commonly used solvents and converted into numerical vectors.

[0036] In this paper, the normalization process described in step (2) includes, but is not limited to, using linear transformation, standard deviation standardization, and conversion of categorical variables into numerical vectors, so that the multidimensional macroscopic reaction condition parameters are converted into normalized reaction condition parameters.

[0037] In this paper, PySCF, mentioned in step (3), is an open-source quantum chemical computing framework based on Python. PySCF supports a variety of molecular input formats (such as XYZ coordinate files). Through tools such as RDKit, molecular structures SMILES can be encoded and converted into three-dimensional molecular structures. The computation time is 10~30 minutes per molecule, and preliminary geometric optimization is performed using the MMFF94 force field.

[0038] In this paper, the calculation of HOMO, LUMO, dipole moment and total energy in step (3) is performed by performing self-consistent field calculation through the scf.hf module of PySCF. The calculation is cyclically converged to the ground state wave function through SCF, and the electronic structure characteristics such as the highest occupied molecular orbital (HOMO), the lowest unoccupied molecular orbital (LUMO) energy level, dipole moment and total energy are output.

[0039] In this paper, the basis set def2-SVP and functional B3LYP mentioned in step (3) are integrated functions in PySCF. The purpose of configuring the basis set def2-SVP and functional B3LYP for optimization is to achieve a balance between computational accuracy and efficiency, while ensuring that the description of the complex polymer system is accurate enough, providing a reasonable initial structure for subsequent accurate calculations, and avoiding convergence difficulties caused by excessive consumption of computational resources.

[0040] In this paper, the normalized reaction condition parameters and electronic structure features described in step (4) are aligned and fused to form a unified feature vector. This usually involves a conventional mechanism for establishing a unified feature dimension. The multidimensional vector of electronic structure features and the multidimensional vector of normalized reaction condition parameters are mapped to the same dimension through a fully connected layer. An attention mechanism is used to perform weighted fusion of the two types of features. The attention weights are dynamically generated through learnable parameters. Then, feature cross-operations are introduced to calculate the cross-features of reaction condition parameters such as temperature, time, and solvent selection with HOMO, LUMO, dipole moment, and total energy.

[0041] In one of the technical solutions, the formation of a unified feature vector in step (4) also includes normalizing the fused feature vector to ensure numerical stability.

[0042] In this paper, the deep learning architecture that integrates Transformer, ChemBERTa pre-trained model and LSTM network described in step (5) is an open-source model architecture that can be downloaded from the network. It is integrated through a conventional open-source deep learning framework package (such as PyTorch) to build a hybrid deep learning architecture.

[0043] In one implementation, the deep learning architecture that integrates Transformer, ChemBERTa pre-trained model and LSTM network in step (5) is specifically: the bottom layer uses Transformer encoder to extract feature vectors through self-attention mechanism, the middle layer integrates ChemBERTa pre-trained model, and the top layer deploys LSTM network.

[0044] In this paper, the model training using the Adam optimizer and Dropout regularization strategy described in step (5) is a common open-source basic component of the entire model training framework, and it inserts Dropout layers between fully connected layers.

[0045] In this paper, step (5) involves optimizing the model parameters using cross-validation and an early stopping strategy. The cross-validation involves dividing the dataset into multiple subsets, using different subsets as validation sets in turn, and the rest as training sets, thereby evaluating the model performance multiple times and calculating the average metric. The early stopping strategy involves monitoring the validation set performance (such as accuracy or loss value) in real time during training. When the validation set metric stops improving or even begins to decline, training is terminated immediately. This effectively avoids overfitting and prevents the model from overfitting on the training set, ensuring generalization ability. At the same time, it saves computational resources, terminates ineffective training in advance, and improves efficiency.

[0046] In one preferred embodiment, the number of Transformer layers in step (5) is set to 6 to 12, the embedding dimension of the ChemBERTa pre-trained model is 768, the number of units in the LSTM network is 128 to 256, and the number of training rounds is set to 50 to 100.

[0047] Based on the PVA-ReAct prediction model provided by this invention, it can be applied in any of the following ways (I) to (III): (I) Forward prediction: Input reactants and conditions, and predict the product structure and reaction yield using the PVA-ReAct prediction model; (II) Reverse design: Input the target product or target yield, and use the gradient descent algorithm to search in the condition space and reverse-engineer the optimal reaction conditions; (III) The model outputs prediction results or design schemes, and can be generalized to other polymer systems such as PE, PP, and PVC.

[0048] The present application will be further explained in detail below with reference to embodiments. However, those skilled in the art should understand that these embodiments are provided for illustrative purposes only and are not intended to limit the present application.

[0049] Example

[0050] The implementation schemes of this application will be described in detail below with reference to the embodiments. However, those skilled in the art will understand that the following embodiments are only used to illustrate this application and should not be regarded as limiting the scope of this application.

[0051] Example 1

[0052] This embodiment presents a method for predicting polyvinyl alcohol-related chemical reactions based on a deep learning framework and multi-scale feature fusion, which mainly includes the following steps: (1) According to the SMILES (Simplified molecular input line entry system) rule, the molecular structure of PVA is standardized and encoded to obtain the molecular structure SMILES code; (2) Obtain PVA reaction data from public databases, configure multi-dimensional macroscopic reaction condition parameters including temperature, time, and solvent selection, and normalize them to obtain normalized reaction condition parameters; (3) Using PySCF in the computing environment, based on the molecular structure SMILES encoding obtained in step (1), calculate HOMO, LUMO, dipole moment and total energy in sequence, and configure the basis set def2-SVP and functional B3LYP for optimization to obtain electronic structure features. (4) Align and merge the normalized reaction condition parameters obtained in step (2) with the electronic structure features obtained in step (3) to form a unified feature vector; (5) Input the feature vector obtained in step (4) into the deep learning architecture that integrates Transformer, ChemBERTa pre-trained model and LSTM network, train the model using Adam optimizer and Dropout regularization strategy, and finally optimize the model parameters through cross-validation and early stopping strategy to obtain the optimized PVA-ReAct prediction model. For easy comparison, it is called PVA-ReAct-LSTM. The deep learning architecture that integrates Transformer, ChemBERTa pre-trained model and LSTM network is specifically designed as follows: the bottom layer uses a Transformer encoder to extract feature vectors through a self-attention mechanism, the middle layer integrates a ChemBERTa pre-trained model, and the top layer deploys an LSTM network. The Transformer layer number is set to 6~12, the ChemBERTa pre-trained model embedding dimension is 768, the LSTM network has 128~256 units, and the number of training rounds is set to 50~100.

[0053] The resulting PVA-ReAct prediction model requires no additional pre-training and can be used directly.

[0054] like Figure 1 The diagram shows the method steps in Example 1.

[0055] To better illustrate the technical effects of the present invention and fully demonstrate the technical advantages of the deep learning architecture that integrates Transformer, ChemBERTa pre-trained model and LSTM network, in Comparative Examples 1 to 3, the LSTM network is replaced with a fully connected network (FCN), a convolutional neural network (CNN), and a reinforcement learning network (RL), respectively. The PVA-ReAct prediction model is trained according to the method and steps of Example 1 and is denoted as PVA-ReAct-FCN, PVA-ReAct-CNN, and PVA-ReAct-RL, respectively.

[0056] like Figure 2 , Figure 3 As shown, the accuracy and R-value of the prediction model constructed using this invention are... 2 It is optimal and has good predictive performance.

[0057] Embodiments of the present invention also provide an electronic device that can perform the polyvinyl alcohol-related chemical reaction prediction method based on deep learning framework and multi-scale feature fusion provided by the present invention.

[0058] This invention also provides a computer-readable storage medium for storing program code for executing the polyvinyl alcohol-related chemical reaction prediction method based on deep learning framework and multi-scale feature fusion described in the foregoing embodiments.

[0059] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to perform... Figure 1 The method shown.

[0060] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0061] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0062] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0063] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0064] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0065] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for predicting polyvinyl alcohol-related chemical reactions based on a deep learning framework and multi-scale feature fusion, characterized in that... The main steps include: (1) According to the SMILES rules, the molecular structures of PVA and related polymers are standardized and coded to obtain the molecular structure SMILES codes; (2) Configure multi-dimensional macroscopic reaction condition parameters including temperature, time, and solvent selection, and normalize them to obtain normalized reaction condition parameters; (3) Using PySCF in the computing environment, based on the molecular structure SMILES encoding obtained in step (1), calculate HOMO, LUMO, dipole moment and total energy in sequence, and configure the basis set def2-SVP and functional B3LYP for optimization to obtain electronic structure features. (4) Align and merge the normalized reaction condition parameters obtained in step (2) with the electronic structure features obtained in step (3) to form a unified feature vector; (5) Input the feature vector obtained in step (4) into a deep learning architecture that integrates Transformer, ChemBERTa pre-trained model and LSTM network, train the model using Adam optimizer and Dropout regularization strategy, and finally optimize the model parameters through cross-validation and early stopping strategy to obtain the optimized PVA-ReAct prediction model.

2. The method according to claim 1, characterized in that: The process of forming a unified feature vector in step (4) also includes normalizing the fused feature vector.

3. The method according to claim 1, characterized in that: The deep learning architecture that integrates Transformer, ChemBERTa pre-trained model and LSTM network in step (5) is specifically: the bottom layer uses Transformer encoder to extract feature vectors through self-attention mechanism, the middle layer integrates ChemBERTa pre-trained model, and the top layer deploys LSTM network.

4. The method according to claim 1, characterized in that: In step (5), the number of Transformer layers is set to 6~12, the embedding dimension of the ChemBERTa pre-trained model is 768, the number of units of the LSTM network is 128~256, and the number of training rounds is set to 50~100.

5. The PVA-ReAct prediction model obtained by the polyvinyl alcohol related chemical reaction prediction method based on deep learning framework and multi-scale feature fusion as described in claim 1.

6. The PVA-ReAct prediction model as described in claim 5 is applied to predict the structure of the product, reaction conditions, and reaction yield.

Images