CN-122024879-A - Prediction method of polyvinyl alcohol related chemical reaction based on fusion of deep learning framework and multi-scale features

Abstract

The invention provides a prediction method of a polyvinyl alcohol related chemical reaction based on deep learning framework and multi-scale feature fusion, which is characterized in that quantum chemical features (such as HOMO, LUMO, dipole moment and the like), molecular structure SMILES codes, reaction conditions and other multi-scale features calculated by a Density Functional Theory (DFT) are fused, and a end-to-end deep learning framework is constructed by combining a Transformer and ChemBERTa pre-training model and an LSTM network, so that high-precision prediction and reverse design of the reaction of the polyvinyl alcohol (PVA) and related polymers are realized. The method has the overall accuracy of more than 95%, has the strong generalization capability, can be extended to PE, PP, PVC and other polymer systems, and provides a general solution for intelligent reaction optimization in a low data scene.

Inventors

• ZHENG LANG
• WANG YUTUO
• LI YIJUN
• LIU XINGANG
• LI XINKANG
• WANG BIN

Assignees

• 天府永兴实验室

Dates

Publication Date

20260512

Application Date

20251201

Claims (6)

1. 1. A prediction method of polyvinyl alcohol related chemical reaction based on fusion of a deep learning framework and multi-scale features is characterized by mainly comprising the following steps: (1) Carrying out standardized coding on molecular structures of PVA and related polymers according to SMILES rules to obtain molecular structure SMILES codes; (2) Configuring multidimensional macroscopic reaction condition parameters comprising temperature, time and solvent selection, and carrying out normalization treatment to obtain normalized reaction condition parameters; (3) Using PySCF in a computing environment, sequentially computing HOMO, LUMO, dipole moment and total energy based on the molecular structure SMILES code obtained in the step (1), and configuring a basis group def2-SVP and functional B3LYP for optimization to obtain electronic structural characteristics; (4) Aligning and fusing the normalization reaction condition parameters obtained in the step (2) and the electronic structural features obtained in the step (3) to form a unified feature vector; (5) Inputting the feature vector obtained in the step (4) into a deep learning architecture fused with Transformer, chemBERTa a pre-training model and an LSTM network, carrying out model training by using an Adam optimizer and a Dropout regularization strategy, and finally optimizing model parameters by using a cross-validation and early-stop strategy to obtain an optimized PVA-ReAct prediction model.
2. 2. The method of claim 1, wherein the forming of the unified feature vector in step (4) further comprises normalizing the fused feature vector.
3. 3. The method of claim 1, wherein in step (5), the deep learning architecture of the LSTM network and the pre-training model are integrated Transformer, chemBERTa, specifically, a transducer encoder is used for extracting feature vectors from the bottom layer through a self-attention mechanism, the pre-training model is integrated ChemBERTa in the middle layer, and the LSTM network is deployed in the top layer.
4. 4. The method of claim 1, wherein in the step (5), the number of layers of the transducer is set to 6-12, the embedding dimension of the chemberta pre-training 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. 5. The PVA-ReAct prediction model obtained by the polyvinyl alcohol-related chemical reaction prediction method based on fusion of a deep learning framework and multi-scale features according to claim 1.
6. 6. The PVA-ReAct predictive model according to claim 5, which is used for predicting the structure of a product, the reaction conditions, and the reaction yield.

Description

Prediction method of polyvinyl alcohol related chemical reaction based on fusion of deep learning framework and multi-scale features Technical Field The invention belongs to the technical fields of chemical informatics and artificial intelligence, relates to a prediction method of a polyvinyl alcohol related chemical reaction based on deep learning framework and multi-scale feature fusion, and particularly relates to a method for realizing high-precision prediction and reverse design of a polyvinyl alcohol (PVA) and related polymer reaction by fusing quantum chemical features, molecular structures, reaction conditions and other multi-scale features to construct the deep learning framework. Background The demand for chemical reaction prediction is rapidly growing in the world as the material science and chemical industry rapidly evolve. Compared with the traditional experimental method, the machine learning has high prediction capability, and the model can simulate a complex reaction process with low calculation cost and can be optimized in a controllable manner, so that the machine learning has rapid application after being discovered. After the second battle, research and application of machine learning has become a hotspot. To date, machine learning has been widely used in the fields related to human production and life, such as drug discovery, material design, and response optimization. The computational characteristics of machine learning determine that the machine learning has great dependence on data, so that the problem of generalization in low data scenes is a core problem which needs to be solved in the development and utilization process of a model. Chemical reaction prediction is a new arduous task that accompanies the application of machine learning, especially in the case of small data sets. In the sixties of the last century, a few countries, represented by the united states, began systematic studies of chemical reaction prediction. Later, in the background that machine learning technology was popularized and applied worldwide, countries with more research and utilization of AI began to join in teams with research reaction prediction. Since the reactions to be predicted are various and the accuracy requirements of the prediction are different, a general prediction method is not available so far. Therefore, in order to cope with the complex situation, the development of prediction techniques also tends to be diversified. The prediction methods developed and mature at present are various, but can be divided into rule-based methods and data driving methods according to the action principle. The rule-based method can be divided into quantum mechanical simulation, empirical formulas and the like, and the data driving method comprises a graph neural network, a transducer model, a pre-training language model and the like. Although the prediction methods disclosed in the prior art are various and can solve various troublesome situations, the defects of a large amount of data and calculation resources are generally existed, and in addition, the selection of the prediction methods is restricted by the predicted reaction type, the data quality, the model architecture and the like. Disclosure of Invention In order to solve the problems in the background technology, the invention provides a polyvinyl alcohol related chemical reaction prediction method based on deep learning framework and multi-scale feature fusion, which is implemented by fusing multi-scale features such as quantum chemical features (such as HOMO, LUMO, dipole moment and the like), molecular structure SMILES codes and reaction conditions and the like calculated by Density Functional Theory (DFT), combining a transducer and ChemBERTa pre-training model and an LSTM network, constructing an end-to-end deep learning framework, and realizing high-precision prediction and reverse design of the reaction of polyvinyl alcohol (PVA) and related polymers. The method has the overall accuracy of more than 95%, has the strong generalization capability, can be extended to PE, PP, PVC and other polymer systems, and provides a general solution for intelligent reaction optimization in a low data scene. In order to achieve the above object, the present invention is realized by adopting a technical scheme comprising the following technical measures. In one aspect, the invention provides a prediction method of a polyvinyl alcohol-related chemical reaction based on fusion of a deep learning framework and multi-scale features, which mainly comprises the following steps: (1) According to SMILES (SIMPLIFIED MOLECULAR INPUT LINE ENTRY SYSTEM) rule, carrying out standardized coding on molecular structures of PVA and related polymers to obtain molecular structure SMILES codes; (2) Configuring multidimensional macroscopic reaction condition parameters comprising temperature, time and solvent selection, and carrying out normalization treatment to obtain normalized reac