Integrated Intelligent System for Novel Structure Generation and Performance Prediction of Alkaline Polymer Electrolytes

Abstract

This invention belongs to the field of materials design computational methods and discloses an integrated intelligent system for generating novel structures and predicting the properties of alkaline polymer electrolytes. This integrated intelligent system iteratively generates a large number of potential repeating units for alkaline polymer electrolytes through machine learning algorithms, significantly expanding the existing chemical space. Utilizing a self-built high-quality database, a high-precision property prediction module is trained, capable of accurately predicting the water absorption rate, swelling degree, hydroxide ion conductivity, and alkaline stability of the generated electrolytes, and further screening them. This integrated intelligent system effectively overcomes the problems of long development cycles and low efficiency associated with traditional experimental trial-and-error methods, improving the R&D efficiency of high-performance alkaline polymer electrolytes.

Description

Technical Field

[0001] This invention belongs to the field of materials design calculation methods and discloses an integrated intelligent system for generating novel structures and predicting the performance of alkaline polymer electrolytes based on machine learning. Background Technology

[0002] Alkaline polymer electrolytes play a crucial role in electrochemical technologies such as water electrolyzers and fuel cells. As ionomers in anion exchange membranes or catalyst layers, they require high hydroxide ion (OH-) conductivity and excellent alkaline stability. To improve overall performance, researchers have developed various chemical structures, such as block copolymers and multi-ionic side chains, to construct ordered ion channels. However, current development processes still rely heavily on trial-and-error experiments, resulting in long development cycles, high costs, and low efficiency.

[0003] Machine learning techniques, by revealing the complex relationship between structure and performance, can accelerate material screening and optimization, and can also significantly reduce the number of experiments and R&D costs through performance prediction and key variable identification. Currently, most research focuses on performance prediction models for polymer electrolytes, employing algorithms such as decision trees, gradient boosting, and deep neural networks to estimate key properties such as OH- conductivity and alkali stability. However, these models typically decompose polymer electrolytes into basic chemical units (e.g., ether aryl, methyl) and use traditional chemical fingerprints (e.g., ECFP, MACCS, Mordred) to extract features, making it difficult to capture the sequence and configuration information of complex polymers.

[0004] Machine learning can not only be used for performance prediction, but also for directly designing novel polymer electrolyte structures through generative models. Combined with predictive models, it can screen for high-performance structures, achieving comprehensive integration and intelligent design. This combination of generation and prediction is more promising than relying solely on predictive models. The key to achieving this goal lies in accurately identifying the complex molecular structures of basic polymer electrolytes. The Transformer architecture, with its self-attention mechanism and parallel design, possesses powerful sequence data processing capabilities, making it particularly suitable for analyzing long-chain polymer structures. Originally developed for Natural Language Processing (NLP), the Transformer has been widely applied in molecular generation, property prediction, and drug screening. In the task of identifying and generating basic polymer electrolytes, the Transformer demonstrates great potential, providing a new approach for the intelligent design of related materials. Summary of the Invention

[0005] This invention proposes an integrated intelligent system for the generation and performance prediction of novel alkaline polymer electrolyte structures. The system applies a self-attention mechanism to the generation and prediction models, directly processing complex SMILES codes to extract high-dimensional molecular features and reveal contextual relationships within the chemical structure. Under controlled conditions, the system generates a large number of novel and reliable repeating units; and using a high-precision prediction model trained with a self-built database, it accurately predicts the water absorption rate, swelling degree, OH- conductivity, and alkaline stability of the generated electrolyte. Finally, high-throughput screening yields superior alkaline polymer electrolyte structures.

[0006] The technical solution of the present invention:

[0007] An integrated intelligent system for generating novel alkaline polymer electrolyte structures and predicting their performance is proposed. This system utilizes experimental data of alkaline polymer electrolytes to train a machine learning model that integrates generation and prediction functions. The integrated intelligent system can automatically generate novel and reliable alkaline polymer electrolyte structures and accurately predict their water absorption rate, swelling degree, OH- conductivity and alkaline stability. At the same time, it can efficiently screen out high-quality structures with excellent performance.

[0008] The alkaline polymer electrolyte is of the quaternary ammonium type, which achieves OH- conduction through quaternary ammonium groups. This electrolyte is widely used in electrochemical devices such as fuel cells and water electrolysis, and can be used as an ionomer in anion exchange membranes or catalyst layers.

[0009] The integrated intelligent system mainly consists of four parts: a new database, a structure generation module, a performance prediction module, and a high-throughput screening module.

[0010] The newly created database was constructed by extracting the structures and corresponding properties of 1133 alkaline polymer electrolytes from 333 relevant academic papers, and building a high-quality database. According to the needs of different tasks, it was further divided into three database subsets, and the data was preprocessed.

[0011] The three database subsets are: ConDs for predicting water absorption, swelling degree, and OH- conductivity; AsDs for predicting alkali stability; and StrDs for training the generative model and predicting module parameters. StrDs contains only the chemical structure information of all repeating units of the basic polymer electrolyte, represented using Simplified Molecular Linear Input Specification (SMILES) encoding. Both ConDs and AsDs contain the SMILES encoding of the repeating units of the basic polymer electrolyte, the polymer type of the basic polymer electrolyte, and polymer composition data. Furthermore, ConDs also includes data on water absorption, swelling degree, and OH- conductivity at different temperatures; while AsDs further includes the residual conductivity or IEC residual rate in alkali stability testing and its corresponding test conditions.

[0012] The preprocessing includes: dividing the long-chain structure of the polyelectrolyte into repeating units and representing them using SMILES encoding; for cross-linked or branched polymers, combining the cross-linked or branched structure with one of the repeating units and treating it as a whole unit; supplementing missing values ​​with average values ​​and filtering out outliers in the database.

[0013] The database is mainly used for training the encoder of the generative model, fine-tuning the encoder of the predictive model, and training the predictor of the predictive model.

[0014] The aforementioned training and generation model refers to training a structure generation module using SMILES encoded data from the StrDs subset, enabling the model to learn the chemical structure representation of existing alkaline polymer electrolytes. During training, SMILES codes are input character by character into the model through the input layer. The model generates the probability distribution of the next character based on the current input character, selects the character with the highest probability value, compares it with the actual character at the corresponding position in the currently input SMILES code, and calculates the cross-entropy. The model parameters are continuously adjusted through backpropagation until the loss function value of the generation model converges, thus achieving the learning of the alkaline polymer electrolyte structure.

[0015] The fine-tuning prediction model encoder refers to: calling a Transformer encoding model that has been pre-trained on a large-scale polymer dataset, embedding the SMILES encoding from the StrDs subset into the model input layer, randomly masking the input sequence, training the model to predict the accuracy of characters at the mask positions, adjusting the internal parameters through backpropagation, and freezing the underlying parameters of the model until the model's loss function converges.

[0016] The predictor in the training prediction model refers to the complete feature vector formed by combining the input features from the ConDs and AsDs subsets with the encoder output, and then inputting this vector into the predictor for training. The predictor is a specially designed deep neural network containing five hidden layers with 1024, 512, 256, 256, and 256 neurons respectively. Residual connections exist between layers 1 and 3, and between layers 3 and 5. The predictor fits target values ​​for different prediction tasks based on the input feature vectors, and supervised training enables the system to predict the performance of new structures. During training, the ConDs and AsDs datasets are divided into training, test, and validation sets in a ratio of 0.75:0.15:0.15.

[0017] The input features include experimental test conditions, polymer type of alkaline polymer electrolyte, and polymer composition data. The experimental conditions include test temperature, immersion time and concentration of alkaline solution; the polymer type of alkaline polymer electrolyte includes random polymer, block polymer and cross-linked polymer; the polymer composition data includes the proportion of different repeating units of alkaline polymer electrolyte and ion exchange capacity.

[0018] The structure generation module refers to the iterative generation of potential basic polymer electrolyte repeating units using machine learning algorithms. This module is a generative model based on the Transformer architecture, comprising an input layer, a Transformer layer, and an output layer. During operation, a single carbon atom ("C") is input into the model via the input layer, and structure generation is performed under the control of total charge (TC) and hydrophobicity (represented by the logarithm of the octanol / water partition coefficient, log P). TC and log P are selectable control attributes; inputting TC and log P individually constitutes a single-attribute control generation task, while inputting both simultaneously constitutes a dual-attribute control generation task. By setting the generative model parameters, new basic polymer electrolyte repeating units can be iteratively generated step-by-step. The generated repeating units are classified into hydrophilic and hydrophobic repeating units. Finally, the generated hydrophilic and hydrophobic repeating units are randomly combined to form complete novel potential basic polymer electrolyte chemical structures, and the newly generated chemical structures are compiled into a GenDs database.

[0019] The generation model parameters are "gen_size" used to control the number of iterations, "batch_size" used to divide the generation task into multiple small batches, and "vocab_size" and "block_size" used to adjust the number of character types and the number of characters in SMILES encoding.

[0020] The performance prediction module is used to predict the key properties of the new alkaline polymer electrolyte generated by the structure generation module, including OH- conductivity, alkaline stability, water absorption, and swelling degree. This module combines a Transformer-based encoder with a predictor for regression prediction. During operation, the encoder encodes the chemical structure of the alkaline polymer electrolyte input from GenDs into a 600-dimensional vector, and constructs a complete feature vector based on the target performance requirements and auxiliary features. The predictor uses this feature vector as input to perform high-precision regression prediction of the target performance.

[0021] The auxiliary features mentioned refer to specific prediction conditions, namely the prediction temperature (20-90℃) for water absorption rate, swelling degree and OH- conductivity, and the prediction conditions for alkali stability include temperature (20-90℃), test duration (0-5000h) and alkali concentration (0.1-10M).

[0022] The high-throughput screening refers to the process of screening for single or multiple attributes based on the evaluation results of GenDs by the performance prediction module, according to preset thresholds for water absorption, swelling degree, OH- conductivity, or alkali stability, to extract the top-ranked structures. Combined with human experience evaluation, this further screens out high-quality novel basic polyelectrolyte structures.

[0023] The aforementioned human experience evaluation refers to assessing the synthetic feasibility and the rationality of the property prediction results of the generated structure in actual experiments based on past experience in synthesis and testing.

[0024] The beneficial effects of this invention are as follows: It constructs an integrated intelligent system capable of generating novel structures, predicting performance with high accuracy, and screening high-performance structures for alkaline polymer electrolytes. This system, based on the Transformer framework with a multi-head self-attention mechanism, directly analyzes the SMILES encoding of repeating units in alkaline polymer electrolytes, accurately capturing high-dimensional molecular features and their contextual relationships. This method significantly improves the efficiency and reliability of structure generation, while greatly enhancing the accuracy of performance prediction. This invention effectively overcomes the limitations of traditional experimental trial-and-error methods, which suffer from long development cycles and low efficiency, significantly accelerating the research and development process of high-performance alkaline polymer electrolytes. Attached Figure Description

[0025] Figure 1 The technical workflow of this invention is summarized. This system includes a new database for model training, a structure generation module, a property prediction module, and a high-throughput screening module.

[0026] Figure 2 This is a t-SNE two-dimensional projection of the structure generation module in Example 1 of the present invention, under the dual control of total charge and hydrophobicity. The results show that the generated structure distribution range is significantly expanded, effectively filling the sparse regions in the chemical space and extending the boundary of the original chemical space to a certain extent.

[0027] Figure 3 This refers to the fitting accuracy of the prediction module in Example 1 of this invention, tested on the validation set after training. Figure 3 (a) is a task for predicting water absorption rate; Figure 3 (b) is the task of predicting the degree of swelling; Figure 3 (c) is OH - Electrical conductivity prediction task; Figure 3(d) is the prediction task for the residual conductivity after alkali treatment. It can be seen that the prediction model demonstrates excellent fitting ability in all four regression tasks, and the coefficient of determination R on the validation set is [value missing]. 2 All values ​​are above 0.9, which fully demonstrates that the prediction module can reliably and accurately predict the key performance of the generated structure. Detailed Implementation

[0028] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0029] Example 1:

[0030] First, given the generative model parameters: gen_size = 100000, batch_size = 200, vocab_size = 53, block_size = 313, and using TC and log P dual-attribute control with control values ​​ranging from (0,0) to (10,10), structure generation was performed. The validity of the generated results was mainly around 0.5, the specificity was 0.374, and the novelty was 1. The generated structures achieved complete novelty relative to the training set while maintaining high validity and low repetition rate, effectively expanding the chemical space. This example generated a total of 1,150,686 valid structures, which, after screening and combination, resulted in a GenDs set of 10,536,000 structures. The GenDs were then input into the prediction module, with prediction conditions set as follows: water absorption rate, swelling degree, and OH group. - The predicted temperature for conductivity is 80℃, and the predicted conditions for alkali stability are 1M alkali solution, 1000h test duration, and 80℃. The model outputs property labels for all structures, based on OH... - Through conductivity screening, a novel polyarylpiperidine structure was finally obtained. As shown in the figure below, this structure is a cross-linked structure with an IEC of 2.36 mmol / g and a high conductivity of 255.06 mS / cm. Furthermore, after treatment in 1M NaOH solution at 80℃ for 1000 h, it still retains 86.78% of its conductivity.

[0031]

[0032] Example 2:

[0033] First, given the generative model parameters: gen_size = 500, batch_size = 50, vocab_size = 53, block_size = 313, and using TC and log P dual-attribute control with control values ​​ranging from (0,0) to (4,4), structure generation was performed. The validity of this generation batch was 0.404, specificity was 0.485, and novelty was 1. The generated structures achieved complete novelty relative to the training set while maintaining high validity and low repetition rate. Subsequently, the generated alkali polyelectrolyte database (GenDs) was constructed, and key properties were predicted. Prediction conditions were set as follows: water absorption rate, swelling degree, OH-. - The predicted temperature for conductivity was 80℃, and the predicted conditions for alkali stability were 1M alkali solution, 1000h test duration, and 80℃. The prediction model output attribute labels for all structures, and performed two rounds of screening based on conductivity and alkali stability, ultimately yielding a novel hyperbranched cross-linked polymer electrolyte. As shown in the figure below, this structure is a cross-linked structure with an IEC of 2.38 mmol / g and a high conductivity of 225.17 mS / cm. Furthermore, after 1000h treatment in 1M NaOH solution at 80℃, it still retains 91.5% of its conductivity, demonstrating high chemical stability.

[0034]

[0035] Example 3:

[0036] First, given the generative model parameters: gen_size = 1000, batch_size = 100, vocab_size = 53, block_size = 313, and using dual-attribute control of TC and log P, with control values ​​ranging from (0,0) to (6,6), structure generation was performed. The validity of this generation batch was 0.696, specificity was 0.362, and novelty was 1. The generated structures achieved complete novelty relative to the training set while maintaining high validity and low repetition rate. Subsequently, the generated alkali polyelectrolyte database (GenDs) was constructed, and key properties were predicted. Prediction conditions were set as follows: water absorption rate, swelling degree, and OH-. - The predicted temperature for conductivity was 80℃, and the predicted conditions for alkali stability were 1M alkali solution, 1000h test duration, and 80℃. The prediction model output attribute labels for all structures, and the results were filtered based on alkali stability, ultimately yielding a novel structure with fluorinated side chains. As shown in the figure below, this structure is a random polymer with an IEC of 2.13 mmol / g. After treatment in 1M NaOH solution at 80℃ for 1000h, it still retains 94.73% of its conductivity, exhibiting high chemical stability, while also achieving a conductivity of 186.72 mS / cm.

[0037]

[0038] 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. An integrated intelligent system for generating novel structures and predicting the performance of alkaline polymer electrolytes, wherein the alkaline polymer electrolyte is a quaternary ammonium type; characterized in that, The integrated intelligent system mainly consists of a new database, a structure generation module, a performance prediction module, and a high-throughput screening module. The structure and properties of alkaline polymer electrolytes were extracted by consulting academic literature, and a high-quality new database was constructed. The new database is mainly used to train the encoder of the structure generation model, fine-tune the performance prediction model, and train the predictor of the performance prediction model. The structure generation module iteratively generates potential basic polymer electrolyte repeating units using machine learning algorithms. The module is a generative model based on the Transformer architecture, comprising an input layer, a Transformer layer, and an output layer. During the operation of the generative model, a single carbon atom serves as the starting point, input into the model via the input layer. Structure generation is controlled by total charge and hydrophobicity, which are selectable control attributes. Inputting total charge and hydrophobicity individually constitutes a single-attribute control generation task, while inputting both simultaneously constitutes a dual-attribute control generation task. The generative model parameters are set, and new basic polymer electrolyte repeating units are iteratively generated. These new repeating units are classified as hydrophilic or hydrophobic. Finally, the hydrophilic and hydrophobic repeating units are randomly combined to form complete new potential basic polymer electrolyte chemical structures, which are then compiled into a generated structure database. The performance prediction module is used to predict the key properties of the new potential alkaline polymer electrolytes generated by the structure generation module, including OH- conductivity, alkaline stability, water absorption, and swelling degree; the performance prediction module combines an encoder based on the Transformer architecture with a predictor for regression prediction; During operation, the encoder encodes the chemical structures of alkaline polymer electrolytes in the structure database into 600-dimensional vectors, and constructs a complete feature vector based on the target performance requirements and auxiliary features; the predictor uses this feature vector as input to complete a high-precision regression prediction of the target performance. The high-throughput screening module is based on the evaluation results of the generated structure database by the performance prediction module. According to the preset thresholds for water absorption rate, swelling degree, OH- conductivity and alkali stability, it performs screening for single or multiple attributes and extracts the top-ranked structures as novel alkaline polymer electrolyte structures.

2. The integrated intelligent system for generating novel structures and predicting the performance of alkaline polymer electrolytes according to claim 1, characterized in that, Based on the needs of different tasks, the newly created database is further divided into three database subsets, and the data is preprocessed. The three database subsets are: ConDs for predicting water absorption, swelling degree, and OH- conductivity; AsDs for predicting alkali stability; and StrDs for training the parameters of the structure generation module and performance prediction module. Among them, the StrDs subset contains only the chemical structure information of all repeating units of alkaline polymer electrolytes, and the chemical structure information is represented using SMILES encoding. Both the ConDs and AsDs subsets contain the SMILES encoding of the repeating units of alkaline polymer electrolytes, the polymer type of the alkaline polymer electrolyte, and polymer composition data. The ConDs subset also includes water absorption, swelling degree, and OH- conductivity data at different temperatures. The AsDs subset also includes the conductivity residual rate or IEC residual rate in alkali stability testing and its corresponding test conditions. Preprocessing includes dividing the long-chain structure of the alkaline polymer electrolyte into repeating units and representing them using SMILES encoding; for cross-linked or branched alkaline polymer electrolytes, the cross-linked or branched structure is combined with one of the repeating units and treated as a whole unit; missing values ​​are supplemented using the average value and outliers in the newly created database are screened out.

3. The integrated intelligent system for generating novel structures and predicting the performance of alkaline polymer electrolytes according to claim 1, characterized in that, The training structure generation model described above uses SMILES encoded data in the StrDs subset to train the structure generation module, enabling the generative model to learn the chemical structure expression of existing alkaline polymer electrolytes. During training, SMILES codes are input character by character into the generative model through the input layer. The generative model generates the probability distribution of the next character based on the current input character, selects the character with the highest probability value, compares it with the actual character at the corresponding position of the current input SMILES code, and calculates the cross-entropy. The parameters of the generative model are continuously adjusted through backpropagation until the loss function value of the generative model converges, thereby achieving the learning of the structure of alkaline polymer electrolyte.

4. The integrated intelligent system for generating novel structures and predicting the performance of alkaline polymer electrolytes according to claim 1, characterized in that, The fine-tuned performance prediction model encoder calls a Transformer encoding model that has been pre-trained on a large-scale polymer dataset. It embeds the SMILES encoding from the StrDs subset into the input layer of the Transformer encoding model, randomly masks the input sequence, trains the Transformer encoding model to predict the accuracy of the characters at the mask positions, and adjusts the internal parameters through backpropagation. At the same time, it freezes the underlying parameters of the Transformer encoding model until the loss function value of the Transformer encoding model converges.

5. The integrated intelligent system for generating novel structures and predicting the performance of alkaline polymer electrolytes according to claim 1, characterized in that, The predictor of the training performance prediction model is formed by combining the input features from the ConDs subset and the AsDs subset with the encoder output to create a complete feature vector, which is then input into the predictor for training. The predictor is a deep neural network containing five hidden layers with 1024, 512, 256, 256, and 256 neurons respectively. There are residual connections between layers 1 and 3, and between layers 3 and 5. The predictor fits the target values ​​of different prediction tasks according to the input feature vector. Through supervised training, the system is equipped with the ability to predict the performance of new structures. During training, the ConDs subset and the AsDs subset are divided into training set, test set and validation set in a ratio of 0.75:0.15:0.

15.

6. The integrated intelligent system for generating novel structures and predicting the performance of alkaline polymer electrolytes according to claim 5, characterized in that, The input features include experimental test conditions, polymer type of alkaline polymer electrolyte, and polymer composition data. The experimental test conditions include test temperature, immersion time and concentration of alkaline solution; the polymer type of alkaline polymer electrolyte includes random polymer, block polymer and cross-linked polymer; the polymer composition data includes the proportion of different repeating units of alkaline polymer electrolyte and ion exchange capacity.

7. The integrated intelligent system for generating novel structures and predicting the performance of alkaline polymer electrolytes according to claim 1, characterized in that, The parameters of the generative model are "gen_size" for controlling the number of iterations, "batch_size" for dividing the generation task into multiple small batches, and "vocab_size" and "block_size" for adjusting the number of character types and the number of characters in SMILES encoding.

8. The integrated intelligent system for generating novel structures and predicting the performance of alkaline polymer electrolytes according to claim 1, characterized in that, The auxiliary features mentioned refer to specific prediction conditions, namely, the prediction temperature for water absorption, swelling degree and OH- conductivity is 20-90℃, and the prediction conditions for alkali stability include temperature 20-90℃, test duration 0-5000h and alkali concentration 0.1-10M.

Images