An anion exchange membrane molecular design method based on a chemical knowledge enhanced dual-channel graph attention network

Abstract

The application belongs to the field of material design calculation method, and discloses a kind of anion exchange membrane molecular design method based on chemical knowledge enhanced double-channel graph attention network.The method embeds chemical priori knowledge into molecular graph representation, and learns hydrophilic ion segment and hydrophobic non-ion segment features through double-channel graph attention network respectively to explicitly represent hydrophilic and hydrophobic microphase separation;Meanwhile, introduce feature linear modulation to dynamically integrate experimental conditions into molecular features, realize multi-scale fusion of molecular structure information and external conditions.The method can realize the level prediction and evaluation of OH⁻ conductivity and alkali stability of anion exchange membrane, and through attention weight visualization, analyze the structure-activity relationship between key structural units and performance, provide interpretable basis for molecular structure optimization and new structure construction, so as to establish the iterative process of "prediction-analysis-structure modification / new structure construction-re-prediction", realize molecular design closed loop.

Description

Technical Field

[0001] This invention belongs to the field of materials design calculation methods and discloses a molecular design method for anion exchange membranes based on chemical knowledge-enhanced dual-channel graph attention networks. Background Technology

[0002] Alkaline anion exchange membranes are core components in emerging clean energy technologies such as water electrolysis for hydrogen production, carbon dioxide electroreduction, and fuel cells. Their performance directly affects the system's energy efficiency and lifespan. An ideal anion exchange membrane should simultaneously possess high OH content... - It exhibits high ionic conductivity and excellent alkaline stability. However, membrane performance is influenced by a complex array of factors across multiple scales, including molecular structure, hydrophilic / hydrophobic microphase separation morphology, and external operating conditions. This complex structure-property relationship makes the traditional experience- and trial-and-error-based material development model time-consuming, costly, and inefficient, severely limiting the development and commercial application of high-performance materials.

[0003] In recent years, the rapid development of artificial intelligence and machine learning technologies has provided a new data-driven approach for modeling the structure-property relationships of materials, which is expected to accelerate the performance prediction and molecular design of anion exchange membranes. Currently, machine learning research in the field of anion exchange membranes mainly focuses on the analysis of OH groups. - Regression predictions of continuous values ​​for ionic conductivity or alkaline stability are used. However, these models are susceptible to factors such as high noise in experimental data and uneven sample distribution, limiting their generalization ability and predictive stability. Discretizing performance indicators into several levels and performing classification predictions can, to some extent, weaken the influence of noise and improve prediction robustness. However, dedicated hierarchical prediction and design models for anion exchange membranes are still relatively lacking.

[0004] Furthermore, existing machine learning models suffer from significant shortcomings in feature extraction and information fusion strategies when processing anion exchange membrane systems. Traditional feature engineering or general neural networks struggle to effectively capture the factors that determine OH groups within the membrane. - The key physical mechanisms of ion transport, especially the separation of hydrophilic and hydrophobic microphases and the characteristics related to ion channels, lack systematic embedding of prior chemical knowledge. Furthermore, existing methods often employ simple splicing when fusing information such as molecular structure, physicochemical parameters, and experimental conditions, making it difficult to handle the differences and correlations between variables of different scales and sources. This can lead to models becoming overly sensitive to conditional variables, weakening the learning of intrinsic molecular structures. Therefore, there is an urgent need to develop a novel machine learning framework capable of explicitly modeling microphase separation, embedding chemical knowledge, and achieving conditionally adaptive fusion to support the stable hierarchical prediction and interpretable molecular design of anion exchange membranes. Summary of the Invention

[0005] The purpose of this invention is to design a molecular design method for anion exchange membranes based on a chemical knowledge-enhanced dual-channel graph attention network. This method uses a specially designed chemical knowledge-enhanced dual-channel graph attention network model to predict the performance grading of anion exchange membranes. It also incorporates an attention parsing module to extract and standardize the attention weight parameters of the hidden layers in the trained chemical knowledge-enhanced dual-channel graph attention network model, thereby visualizing the atomic attention weights of the molecular structure. The chemical knowledge-enhanced dual-channel graph attention network model is implemented based on a graph attention network (GAT) framework. First, prior chemical knowledge is introduced into the molecular graph embedding layer of the GAT model to form a knowledge-embedded molecular graph representation. Then, by decoupling the hydrophilic and hydrophobic unit structures of the anion exchange membrane, a dual-channel GAT layer is designed to learn the features of hydrophilic ionic fragments and hydrophobic nonionic fragments respectively, explicitly representing microphase separation. Simultaneously, a feature linear modulation (FiLM) module is introduced to dynamically integrate experimental operating conditions into the model training and inference process, achieving multi-scale information fusion of anion exchange membrane molecular structure information and experimental operating conditions. Finally, the model uses a multi-layer fully connected neural network as a performance classifier to achieve performance classification of the core performance of the anion exchange membrane, namely OH... - Hierarchical prediction of conductivity and base stability. The attention parsing module, based on node attention extraction and normalization strategies, extracts the influence weight of each atom in the hidden layer of the computational model on model inference. A chemical knowledge-enhanced dual-channel graph attention network model and the attention parsing module are linked to form a closed-loop design method for anion exchange membrane molecules: candidate molecules are ranked and ordered; favorable / unfavorable structural features are extracted from the attention heatmap; structural modifications or new structures are constructed based on these features and then input into the model for pre-evaluation; high-potential molecular schemes are obtained through iterative screening. This molecular design method can achieve [the following is incomplete and requires further context: OH] - The conductivity and base stability are used to predict performance levels, and the key structural contributions are revealed by visualization based on attention weights, providing an interpretable basis for molecular structure optimization and new structure design. By combining model pre-evaluation and iterative screening, interpretable design and rapid optimization of closed-loop molecules before synthesis can be achieved, reducing trial and error costs and improving R&D efficiency.

[0006] The technical solution of this invention: A molecular design method for anion exchange membranes based on a chemical knowledge-enhanced dual-channel graph attention network (DBN) comprises a DBN model and an attention weight analysis module, as well as a closed-loop molecular design process implemented by the DBN model and the attention weight analysis module. The first part of the process involves the DBN model used to perform hierarchical prediction of key performance characteristics of the anion exchange membrane. The second part involves the attention weight analysis module extracting the attention weight parameters of the hidden layers in the trained DBN model and performing standardization and visualization of the molecular attention weights. By connecting these two parts and iteratively following a "prediction-analysis-structural modification or new structure construction-re-prediction" process, a closed-loop molecular design for the anion exchange membrane is achieved. The chemical knowledge-enhanced dual-channel graph attention network model is built upon the Graph Attention Network (GAT) framework. GAT's architecture is designed for efficient and expressive representation learning of graph-structured data. Its core lies in utilizing a self-attention mechanism to dynamically calculate the importance weights of each node's neighbors in the graph. The model incorporates four core design features based on GAT: chemical knowledge embedding, a dual-channel graph attention encoder, experimental feature linear modulation fusion, and a performance classifier. Chemical knowledge embedding refers to an improved design of the molecular graph embedding layer in GAT. Chemical molecules need to be converted into differentiable graph-structured data to serve as input data for GAT computation. This graph-structured data includes vectors representing node information, vectors representing edge information, and matrices representing node-edge connection information. For chemical molecules, nodes represent atoms, edges represent chemical bonds, and the connection information between atoms and bonds is represented by an index matrix, which records which atoms are connected by chemical bonds. Chemical knowledge embedding involves assigning prior chemical knowledge as initial features to nodes and edges during molecular graph construction, forming a knowledge-enhanced molecular graph representation. This prior chemical knowledge includes: Atomic characteristics: atomic number, connectivity, formal charge, chiral marking, number of hydrogen atoms bonded, hybridization type, aromaticity, whether it is in a ring structure, atomic mass, electronegativity, van der Waals radius, ionization energy, number of valence electrons, acidity / basicity, electrophilicity, and thioether sulfur structure, aluminum hydroxyl oxygen structure; Bond characteristics: bond type, stereochemistry, conjugation information, whether it is a cyclic bond.

[0007] The dual-channel graphical attention encoder leverages the structural differences between hydrophilic and hydrophobic units in ion exchange membranes. It designs a dual-channel Gaussian Attention (GAT) layer to learn the features of hydrophilic ionic segments and hydrophobic nonionic segments respectively, thus explicitly characterizing microphase separation. The dual-channel GAT encoder employs a parallel dual-channel structure: the first channel encodes hydrophilic ionic segments in the molecule; the second channel encodes hydrophobic nonionic segments. Each channel is based on a pre-trained GAT, learning interatomic interactions through an attention mechanism to obtain the intermediate feature representation for that channel.

[0008] Experimental feature linear modulation fusion dynamically integrates external experimental conditions into the molecular characterization learning process. These external experimental conditions include: OH - Conductivity testing temperature (20–90 °C); alkali stability prediction conditions, including temperature (20–90 °C), testing duration (0–5000 h), and alkali concentration (0.1–10 M). This design uses these experimental conditions as input to generate scaling factors and bias parameters, and linearly modulates the intermediate feature representation output by the dual-channel encoder dimension by feature, thereby learning the coupling mapping relationship between "molecular structure-experimental conditions-performance level".

[0009] The performance classifier takes the data fused by linear modulation of experimental features as input and outputs OH through a multi-layer fully connected neural network. - Predicted performance grades for conductivity and alkali stability. Performance grade refers to: the performance grade of OH... - The continuous values ​​of conductivity and alkalinity stability are discretely divided into 2 to 10 levels; among which, alkalinity stability is represented by the OH content in the alkalinity stability test. - The conductivity margin or ion exchange capacity (IEC) margin is quantified.

[0010] The dual-channel graph attention encoder in the chemical knowledge-enhanced dual-channel graph attention network model requires pre-training, and the entire model needs full fine-tuning. Both pre-training and fine-tuning rely on a comprehensive anion exchange membrane database. This database is constructed by systematically collecting and organizing experimental data from publicly available literature. The anion exchange membranes in the database are quaternary ammonium type anion exchange membranes, which achieve OH- ion exchange through quaternary ammonium groups. -This technology is applicable to electrochemical devices such as water electrolysis for hydrogen production, carbon dioxide electroreduction, and fuel cells. The detailed database construction process is as follows: using "Anion Exchange Membrane" as the keyword, all research papers published between 2010 and June 2025 were retrieved from Web of Science, totaling over 2000 papers. Based on impact factors, 367 high-quality, non-composite membrane papers were extracted, from which 1320 sets of anion exchange membrane structures and experimental parameters were extracted. The database entries include: anion exchange membrane molecular structures (represented in simplified linear input notation SMILES), key performance data (OH... - The data included conductivity and alkali stability, along with corresponding test conditions (including temperature, test duration, and alkali concentration). The collected data underwent further preprocessing, which included: segmenting the long polymer chain structure of the anion exchange membrane into fragment structures and encoding them using SMILES; for cross-linked or branched polymers, combining the cross-linked or branched structure with one of the repeating units to treat it as a single unit; and completing missing values ​​and removing outliers.

[0011] Pre-training of the dual-channel graph attention encoder refers to self-supervised training based on the molecular structures of anion exchange membranes in a comprehensive database of anion exchange membranes, with the goal of reconstructing mask node information. This allows the dual-channel graph attention encoder to pre-learn the basic chemical representation and structural characteristics of anion exchange membrane molecular structures. The weight parameters in the pre-trained dual-channel graph attention encoder will be saved and loaded during the overall full-scale fine-tuning of the model, achieving parameter sharing in the dual-channel graph attention encoder.

[0012] The full-scale fine-tuning of the chemical knowledge-enhanced dual-channel graph attention network model refers to supervised end-to-end hierarchical prediction training of the complete model after loading the pre-trained weight parameters of the dual-channel graph attention encoder, using molecular graphs and experimental conditions as inputs and performance levels as labels. During full-scale fine-tuning, the parameters in the dual-channel graph attention encoder, experimental feature linear modulation fusion, and performance classifier are updated through training. After training, the model parameters are saved and loaded when performing inference tasks.

[0013] The attention parsing module, based on node attention extraction and standardization strategies, extracts the influence weights of each atom in the hidden layer of the computational model on model inference and generates a molecular graph visualizing node attention. The detailed process of the attention parsing module is as follows: First, it extracts the multi-head attention weight matrix of the last graph attention layer from the GAT hidden layer of the fully trained chemical knowledge-enhanced dual-channel graph attention network model. This matrix contains the attention weight coefficients calculated by multiple independent attention heads for each pair of atomic nodes in the molecular graph when the model makes predictions. Then, to fuse heterogeneous information from different attention heads and obtain a global, unified measure of node importance, the extracted attention weight coefficients are standardized. Specifically, max pooling is applied to the attention weights of each atomic node across all attention heads. This operation extracts the maximum value of the attention weights on all associated edges when each atom is considered a "focused node," thereby capturing the most significant local interaction strength of that atom in any attention head. Simultaneously, the original attention values ​​of each node obtained after pooling are standardized and scaled to the [0,1] interval to generate the relative attention weight parameters of each atom during model inference. The standardization process ensures the comparability of weights between different molecules or between different nodes within the same molecule. Finally, based on the molecule's two-dimensional structure, these calculated attention weight values, corresponding one-to-one with each atom, are mapped to color gradients. Higher weight values ​​are represented by darker colors, and lower weight values ​​by lighter colors, thus generating an intuitive molecular heatmap. This visualization clearly and qualitatively displays the molecular substructures or key atoms that the model is "focused on." The attention parsing module can identify high OH... - Atoms, chemical bonds, or substructures that contribute significantly to electrical conductivity or high base stability can generate actionable structural optimization suggestions. The closed-loop process of molecular design is as follows: predicting and ranking candidate molecules; analyzing attention heatmaps to extract favorable / unfavorable structural features; modifying structures or constructing new structures accordingly; inputting the data into the model for pre-evaluation; and obtaining high-potential molecular schemes through iterative screening to reduce trial-and-error costs and improve R&D efficiency.

[0014] The beneficial effects of this invention are as follows: It proposes a data-driven intelligent molecular design method for anion exchange membranes. This method, through chemical knowledge embedding and a dual-channel architecture, deeply integrates domain knowledge into the model and explicitly models the core physical mechanisms, thereby achieving high-precision and robust predictions of conductivity and stability. Furthermore, this invention utilizes an attention weighting analysis module to transform the prediction results into interpretable structural design guidelines, establishing a direct bridge from performance prediction to molecular design and forming a verifiable intelligent design closed loop. This effectively overcomes the limitations of traditional trial-and-error methods and traditional machine learning models that "predict without interpreting," providing a powerful new tool for systematically overcoming the conductivity-stability trade-off problem of anion exchange membranes. Attached Figure Description

[0015] Figure 1 The technical workflow of this invention is summarized. This method combines anion exchange membrane performance level prediction model and attention weight analysis module to realize closed-loop design of anion exchange membrane molecular structure. Figure 2 The diagrams show the attention weight analysis before and after the structural design of this invention. (a) represents Example 1; (b) represents Example 2; and (c) represents Example 3. Darker colors indicate higher attention weights in that region and a greater impact on the prediction results. Figure 3 This is the confusion matrix obtained by testing the level prediction model of Example 1 of this invention on the test set after training. (a) is the OH... - (a) is the conductivity level prediction task; (b) is the conductivity residual rate level prediction task after alkaline treatment; it can be seen that the main diagonal elements of the two confusion matrices are significantly dominant, indicating that the model accurately predicts the level of most samples on the test set and the overall prediction performance is good. Detailed Implementation

[0016] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0017] Example 1: This embodiment targets anion exchange membranes with a terphenyl structure to enhance their OH- - Electrical conductivity is used as the design target for structural optimization design, as shown in the attached figure. Figure 2 As shown in (a), the SMILES coding, repeating unit ratio data (1:1), and test temperature (80°C) of the original terphenyl anion exchange membrane are presented. o C) Input the performance level prediction model to predict its OH -The conductivity level is 3. Further attention weight analysis of the model revealed a significant concentration of attention in the quaternized piperidine region. Based on this, a design strategy synergistic between the fluorinated side chain and the amphiphilic piperidine cation was adopted to enhance the formation ability of continuous ion channels. Two new structural fragments were obtained after optimization, and these were combined with the original fragments in a ratio of 0.7:0.1:0.2 to form the optimized molecule. The model was then input again for prediction. Under the same testing conditions, the OH of the optimized structure... - The conductivity level was improved to level 1. Attention weight analysis showed that both the fluorinated segment and the piperidine cation unit were assigned high weights, indicating that they make key contributions to microphase separation and ion transport, confirming the effective improvement in conductivity.

[0018] Example 2: This embodiment focuses on optimizing a terphenyl-structured anion exchange membrane to improve its alkaline stability, as shown in the attached figure. Figure 2 (b) shows the molecular SMILES encoding, repeating unit ratio data (1:1), and test conditions (80). o The model, inputting C (1 M alkali solution, 1000 hours), predicted an alkali stability level of 2. Attention weight analysis showed that attention was highly concentrated in the main chain region. Based on this, an octylcarbazole structure was introduced into the main chain to construct a dual-cation main chain strategy, optimizing the electrostatic potential distribution and dispersing the cation charge density, thereby mitigating hydroxyl attack. The optimized molecule consisted of two fragments in a 0.8:0.2 ratio. Under the same testing conditions, the predicted alkali stability level increased to 1. Further attention weight analysis showed that the attention weight of the optimized structure was concentrated in the octylcarbazole main chain region, confirming that this structure played a key role in improving stability and successfully enhancing alkali stability.

[0019] Example 3: This embodiment focuses on anion exchange membranes with a meta-terphenylene structure, aiming to synergistically enhance their OH- ion exchange capacity. - Electrical conductivity and alkaline stability, as shown in the attached figure. Figure 2 As shown in (c). The molecular SMILES encoding, repeating unit ratio data (1:1), and OH... - Conductivity test temperature (80) o C) and alkali stability test conditions (80) oThe model was input with C (1 M alkali solution, 1000 hours), predicting a conductivity level of 3 and a stability level of 2. Analysis of the attention distribution in the conductivity and stability prediction tasks revealed that both were concentrated in the main chain region, with lower attention paid to the quaternized side chains. Based on this, the quaternary ammonium group was moved from the flexible aliphatic chain to the rigid terphenyl main chain, achieving localization and control of the side chains. The optimized structure's SMILES encoding and the same testing conditions were input into the model, predicting that both conductivity and stability levels improved to level 1. Further attention weight analysis showed that in conductivity prediction, cationic side chains received significant attention, indicating a more ordered arrangement of ion transport sites; in stability prediction, both cationic units and rigid terphenyl fragments were assigned high weights, indicating that the latter can protect ionic groups through π-π stacking and shielding effects. This design successfully achieved a comprehensive improvement in both conductivity and stability.

[0020] The above embodiments are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited to the specific implementations described in the above embodiments. All improvements and modifications made based on the technical solutions, algorithm architectures, or model designs of the present invention fall within the scope of protection of the present invention. Adjustments, optimizations, or extensions made by those skilled in the art without departing from the core ideas and technical solutions of the present invention should also be considered part of the scope of protection of the present invention.

Claims

1. A method for designing anion exchange membrane molecules based on chemical knowledge-enhanced dual-channel graph attention networks, characterized in that, The molecular design method for anion exchange membranes consists of a chemical knowledge-enhanced dual-channel graph attention network model and an attention weight analysis module, as well as a molecular closed-loop design process implemented by the chemical knowledge-enhanced dual-channel graph attention network model and the attention weight analysis module. The first part of the process is that the chemical knowledge-enhanced dual-channel graph attention network model is used to achieve hierarchical prediction of the key performance of anion exchange membranes. The second part of the process: The attention weight parsing module is used to extract the attention weight parameters of the hidden layers in the trained chemical knowledge-enhanced dual-channel graph attention network model, and to perform standardization and molecular attention weight visualization processing; the two parts of the process are connected in series, and the closed loop of anion exchange membrane molecular design is realized through the iterative process of "prediction-analysis-structural modification or new structure construction-re-prediction".

2. The anion exchange membrane molecular design method according to claim 1, characterized in that, The chemical knowledge-enhanced dual-channel graph attention network model is built on the graph attention network framework. Based on the graph attention network framework, four core designs are implemented: chemical knowledge embedding, dual-channel graph attention encoder, experimental feature linear modulation fusion, and performance classifier. Chemical knowledge embedding is an improved design of the molecular graph embedding layer in the graph attention network framework. Chemical molecules must be converted into differentiable graph structure data before they can be used as input data for computation in the graph attention network framework. The graph structure data includes vectors representing node information, vectors representing edge information, and matrices representing the connection information between nodes and edges. For chemical molecules, nodes represent atoms, edges represent chemical bonds, and the connection information between atoms and bonds is represented by an index matrix, which records which atoms are connected by chemical bonds. Chemical knowledge embedding involves assigning prior chemical knowledge as initial features to nodes and edges when constructing the molecular graph, forming a knowledge-enhanced molecular graph representation. The dual-channel graph attention encoder is based on the structural differences between hydrophilic and hydrophobic units in anion exchange membranes. It designs a dual-channel GAT layer to learn the features of hydrophilic ionic fragments and hydrophobic nonionic fragments respectively, so as to explicitly characterize microphase separation. The dual-channel graph attention encoder adopts a parallel dual-channel GAT structure: the first channel is used to encode the hydrophilic ionic chain segments in the molecule. The second channel is used to encode hydrophobic nonionic segments in the molecule; each channel is based on a pre-trained graph attention network framework, which learns interatomic interactions through the attention mechanism to obtain the intermediate feature representation of the corresponding channel; Experimental feature linear modulation fusion dynamically integrates external experimental conditions into the molecular characterization learning process. Using external experimental conditions as input, scaling factors and bias parameters are generated, and the intermediate feature representation output by the dual-channel graph attention encoder is linearly modulated along the feature dimension, thereby learning the coupling mapping relationship between "molecular structure-experimental conditions-performance level". The performance classifier takes the data fused by linear modulation of experimental features as input and outputs OH through a multi-layer fully connected neural network. - Predicted performance levels for electrical conductivity and alkali stability.

3. The anion exchange membrane molecular design method according to claim 2, characterized in that, The aforementioned prior chemical knowledge includes: Atomic characteristics: atomic number, connectivity, formal charge, chiral marking, number of hydrogen atoms bonded, hybridization type, aromaticity, whether it is in a ring structure, atomic mass, electronegativity, van der Waals radius, ionization energy, number of valence electrons, acidity / basicity, electrophilicity, and thioether sulfur structure, aluminum hydroxyl oxygen structure; Bond characteristics: bond type, stereochemistry, conjugation information, whether it is a cyclic bond.

4. The anion exchange membrane molecular design method according to claim 2, characterized in that, The external experimental conditions include: OH - Conductivity testing temperature: 20–90 °C; Alkali stability prediction conditions: temperature: 20–90 °C; testing duration: 0–5000 h; alkali concentration: 0.1–10 M.

5. The anion exchange membrane molecular design method according to claim 2, characterized in that, The performance level results: OH - The continuous values ​​of conductivity and alkalinity stability are discretely divided into 2 to 10 levels; among which, alkalinity stability is represented by the OH content in the alkalinity stability test. - The remaining conductivity or ion exchange capacity is quantified.

6. The anion exchange membrane molecular design method according to claim 2, characterized in that, The dual-channel graph attention encoder in the chemical knowledge-enhanced dual-channel graph attention network model needs to be pre-trained, and the entire chemical knowledge-enhanced dual-channel graph attention network model needs to be fully fine-tuned; both pre-training and full fine-tuning rely on the anion exchange membrane comprehensive database.

7. The anion exchange membrane molecular design method according to claim 6, characterized in that, The aforementioned anion exchange membrane comprehensive database refers to a database constructed by systematically collecting and organizing experimental report data from publicly available literature. The anion exchange membranes in the database are quaternary ammonium type anion exchange membranes, which achieve OH- ion exchange through quaternary ammonium groups. - Conductive properties are suitable for hydrogen production via water electrolysis, carbon dioxide electroreduction, and fuel cells. The detailed construction process of the comprehensive anion exchange membrane database is as follows: Using "Anion Exchange Membrane" as the keyword, all research papers published between 2010 and June 2025 were retrieved from Web of Science, totaling over 2000 papers. Based on the impact factor, 367 high-quality, non-composite membrane papers were extracted, from which 1320 sets of anion exchange membrane structures and experimental parameters were extracted. The data entries in the comprehensive anion exchange membrane database include the molecular structure of the anion exchange membrane, key performance data, and corresponding test condition parameters. The molecular structure of the anion exchange membrane is represented using the simplified linear input canonical notation SMILES, and the key performance data is OH... - Electrical conductivity and alkaline stability were tested under the following parameters: temperature, test duration, and alkaline concentration. The collected data underwent further preprocessing, which included: dividing the long polymer chain structure of the anion exchange membrane into fragment structures and encoding them using SMILES; for cross-linked or branched polymers, the cross-linked or branched structure was combined with one of the repeating units and treated as a whole unit; missing values ​​were filled in and outliers were removed.

8. The anion exchange membrane molecular design method according to claim 7, characterized in that, The pre-training of the dual-channel graph attention encoder is based on the molecular structures of anion exchange membranes in the comprehensive database of anion exchange membranes. Self-supervised training is carried out with the goal of reconstructing mask node information, so that the dual-channel graph attention encoder can learn the basic chemical representation and structural characteristics of anion exchange membrane molecular structures in advance. The weight parameters in the pre-trained dual-channel graph attention encoder will be saved and loaded during the full fine-tuning of the chemical knowledge-enhanced dual-channel graph attention network model, so as to realize parameter sharing of the dual-channel graph attention encoder.

9. The anion exchange membrane molecular design method according to claim 8, characterized in that, The overall full-scale fine-tuning of the chemical knowledge-enhanced dual-channel graph attention network model involves loading the pre-trained weight parameters of the dual-channel graph attention encoder, and then using molecular graphs and experimental conditions as inputs and performance levels as labels to perform supervised end-to-end hierarchical prediction training on the chemical knowledge-enhanced dual-channel graph attention network model. During full fine-tuning, the parameters in the dual-channel graph attention encoder, experimental feature linear modulation fusion, and performance classifier are trained and updated. After training, the parameters of the chemical knowledge-enhanced dual-channel graph attention network model are saved and loaded when performing inference tasks.

10. The anion exchange membrane molecular design method according to claim 8, characterized in that, The attention parsing module, based on node attention extraction and standardization strategies, extracts the influence weights of each atom in the GAT hidden layer of the computational chemistry knowledge-enhanced dual-channel graph attention network model on the inference of the model and generates a molecular graph for node attention visualization. The detailed process of the attention parsing module is as follows: First, the multi-head attention weight matrix of the last graph attention layer is extracted from the GAT hidden layer of the fully trained computational chemistry knowledge-enhanced dual-channel graph attention network model. This multi-head attention weight matrix contains the attention weight coefficients calculated by multiple independent attention heads for each pair of atomic nodes in the molecular graph when the computational chemistry knowledge-enhanced dual-channel graph attention network model makes predictions. Subsequently, in order to integrate heterogeneous information from different attention heads and obtain a global, unified node importance metric, the extracted attention weight coefficients are standardized. Specifically, max pooling is applied to the attention weights of each atomic node across all attention heads. This operation extracts the maximum value of the attention weights on all associated edges of each atom when it is a "focused node", thereby capturing the most significant local interaction strength of the atom in any attention head. At the same time, the original attention values ​​of each node obtained after pooling are standardized and scaled to the [0,1] interval to generate the relative attention weight parameters of each atom in the reasoning process of the chemical knowledge-enhanced dual-channel graph attention network model. The standardization process ensures the comparability of weights between different molecules or between different nodes within the same molecule. Finally, based on the two-dimensional structure of the molecule, the calculated attention weight values, corresponding one-to-one with each atom, are mapped to color gradients; higher weight values ​​are represented by darker colors, and lower weight values ​​by lighter colors, thus generating an intuitive molecular heatmap. The attention parsing module can identify high OH levels. - Atoms, chemical bonds, or substructures that contribute significantly to electrical conductivity or high base stability can generate actionable structural optimization suggestions.

Images