Polymer multi-property prediction method based on self-supervised feature learning

Abstract

The invention discloses a polymer multi-property prediction method based on self-supervised feature learning, and relates to the technical field of high polymer material information detection, and the method comprises the following steps: constructing a polymer data set containing a plurality of property labels; carrying out monomer legality filtering on the data, and constructing a polymer virtual graph containing a real edge and a virtual edge based on a monomer structure; generating an electrical characteristic label of each atom through self-supervised learning, and training an improved 2D graph neural network and sequence prediction model based on the electrical characteristic label; respectively training the model for different target properties, and optimizing the model performance through hyper-parameter search; and finally, performing weighted fusion on the prediction results of the plurality of basic models by using an integrator, and outputting a final prediction value of the target property. According to the method, the characterization capability of the polymer is enhanced through the virtual graph structure, and the accuracy, generalization capability and extrapolation performance of polymer multi-property prediction are improved in combination with self-supervised learning and model integration.

Description

Technical Field

[0001] This invention relates to the field of polymer material information detection technology, specifically to a method for predicting multiple properties of polymers based on self-supervised feature learning. Background Technology

[0002] Key properties of polymers, such as atomization energy, bulk band gap, melting temperature, refractive index, and thermal decomposition temperature, directly affect the material's processing, mechanical properties, transport behavior, and service stability, making them core indicators for material design and selection. Obtaining these properties typically relies on three approaches: experimental testing, theoretical calculations / multi-scale simulations, and statistical / machine learning modeling based on molecular structure.

[0003] Experimental testing offers directness, such as differential scanning calorimetry, densitometer measurements, and small-angle scattering, but it often suffers from limitations: sensitivity to experimental conditions, measurement bias and repeatability issues between different laboratories, time and material consumption, and difficulty in achieving high-throughput coverage across a large pool of candidate materials. Theoretical calculations and multi-scale simulations can provide mechanistic explanations, but for polymer systems, the coupling across spatiotemporal scales and the complexity of chain conformations make high-precision predictions costly, parameterization difficult, and unsuitable for universal application across a large chemical space. Traditional quantitative structure-activity relationship (QSAR) methods often construct linear or weakly nonlinear models using manually selected molecular descriptors. While offering some interpretability, they lack universality and scalability when dealing with complex structural factors such as polymer end-group effects, multi-site polymerization of repeating units, crystallinity, and chain conformational constraints.

[0004] In recent years, machine learning models targeting small molecule properties have achieved good accuracy on several tasks. However, direct transfer to polymer prediction still faces challenges. For example, polymer structure representation needs to take into account the polymerization sites of repeating units and their connecting chemical environments; multi-source data, such as experimental and literature extraction, have differences in label conditions and missing values; the feature space spans semantic embedding, topological statistics, and physicochemical descriptors, which can easily lead to collinearity and data leakage risks; at the same time, the model's extrapolation ability (for new backbones / new repeating units) and reproducibility management are also key to engineering implementation.

[0005] Therefore, there is an urgent need for a polymer property prediction method that takes into account data standardization, feature engineering robustness, model integration and generalization, and full-process auditability, so as to improve prediction accuracy and applicability while ensuring scientific rigor. Summary of the Invention

[0006] The present invention aims to address the shortcomings of existing polymer characterization methods in encoding periodic monomer structures, capturing repeating patterns between monomers, and topological information.

[0007] To address the above problems, this invention provides a polymer multi-property prediction method based on self-supervised feature learning, comprising the following steps: Step 1: Construct a polymer multi-property dataset Data= { Y, M, A, P, R, N, S}; Y, M, A, P, R, N, S It is a set of different properties of polymers; Step 2, for Data Perform monomer-to-polymer characteristic calculations; Step 2.1, for Data Each SMILES expression in the code undergoes singleton validity filtering. Step 2.2, from M Construct a polymer virtual graph in the middle; Step 2.3: For each atom in each SMILES, generate a self-supervised tag to obtain the electrical properties of each polymer. GC i ; Step 3: Construct and train a self-supervised prediction model for polymer multi-property features, and perform inference predictions, including: Step 3.1: Construct a model set; Step 3.2, for each target property An improved 2D graph neural network model and sequence prediction model based on self-supervised label training; Step 3.3: Search and optimize the hyperparameter configuration of the model; Step 3.4: Use an integrator to perform weighted fusion of the prediction results from multiple base models to obtain the final predicted value of the target property; Step 3.5: Record and save all training and model parameters.

[0008] Further, in step 1, Y This is a polymer property tag matrix. M Let A be the set of polymer monomer structures, and let A be the set of data sources and collected metadata. P For the original polymer entity description set, R Representing a set by repeating elements, N This is a collection of samples linking the structure and properties of polymer monomers. S Strategies for partitioning the dataset.

[0009] Further, in step 2.2, from M The steps for constructing a polymer virtual graph include: Step 2.2.1: For each monomer, construct the real edge by the interaction of directly connected atom pairs. E 1i ={ e11 ,…,e 1q} i Virtual edges within a monomer are constructed by the interactions of non-directly connected atomic pairs. E 2i ={ e 21 ,…, e 2w} i p gauss ,in p gauss It is a Gaussian probability distribution; Step 2.2.2: Construct virtual edges between monomers through the periodic interactions of two atoms directly connected to the polymer reaction sites. E 3i ; Obtain the set of virtual edges E i ={ E 1 , E 2 , E 3} i ; Step 2.2.3: Treat each atom as a node. V i As a node feature, and E i Together they form a virtual graph g i ={ V , E} i .

[0010] Further, in step 3.1, the model set includes an improved 2D graph neural network and a sequence prediction model; wherein the improved 2D graph neural network model is based on a virtual graph. g i ,right E 1i , E 2i and E 3i Feature learning is performed using different message passing methods to output the electrical property sequence of the polymer. GC i The sequence prediction model uses GC i As input, the complete property output sequence is predicted through a decoder and attention mechanism.

[0011] Furthermore, the improved 2D graph neural network model outputs a sequence of the polymer's electrical properties.GC i The steps include: Step 3.1.1, based on the degree of the atom, for E 1i The neighboring atomic features of the connected atoms are then weighted and summed for updating. Step 3.1.2: Based on the degree of the virtual edge, for E 2i The neighboring atomic features of the connected atoms are then weighted and summed for updating. Step 3.1.3: Calculate learnable attention weights E 3i The edge weights between connected atoms are used to update the features of neighboring atoms through a weighted summation. Step 3.1.4, will... E 1i , E 2i and E 3i The sum of transmitted messages, plus the bias, yields the updated node features. These features are then processed through multiple convolutional and linear layers to output the electrical property sequence of the polymer. GC i .

[0012] Furthermore, the sequence prediction model employs a multi-head attention mechanism, the calculation method of which is as follows: ,in, All by Obtained by projection.

[0013] Furthermore, in step 3.2, the method for individual training is as follows: optimize network parameters through gradient descent, use MSE as training loss, and employ early stopping, learning rate scheduling, and regularization strategies.

[0014] Furthermore, in step 3.3, the method for searching and optimizing the hyperparameter configuration of the model is as follows: [The text abruptly shifts to a different topic] ...the constructed self-supervised labels... Y SSL and monomers m i A 2D graph neural network is used to predict the electrical properties of the corresponding polymer, and the results are used as input for subsequent sequence models. Then, for each self-supervised graph neural network model and sequence prediction model, a model configuration search is performed based on the adjustable range of its structure and training strategy. Model performance is evaluated using the validation set performance metric R0. 2 As an evaluation criterion, the configuration with the smallest overall error is selected as the model's target configuration. Y i The optimal solution.

[0015] Furthermore, in step 3.4, the step of using an integrator to weightedly fuse the prediction results of multiple base models to obtain the final predicted value of the target property includes: Step 3.4.1, for each target property Train multiple different base models and obtain the prediction results and evaluation metrics of each model on the validation set; Step 3.4.2: Sort and filter the base models according to the performance of the validation set, and assign fusion weights to the selected models based on performance metrics; Step 3.4.3: In the inference stage, the input features are passed through all selected base models, and the prediction results of each model are weighted and fused according to the obtained weights to output the final predicted value of the target property.

[0016] Furthermore, the feature is that, in step 3.5, when adding new samples or expanding features, the same data specifications, self-supervised feature construction process, and model training process are reused.

[0017] The polymer multi-property prediction method based on self-supervised feature learning of the present invention has, but is not limited to, the following beneficial effects compared with the prior art: (1) Refinement and global enhancement of polymer construction and feature representation: The virtual graph of “monomer + virtual edge connection” is adopted, which takes into account both polymer connection sites and intra-monomer connections, and realizes traceable mapping from monomer sequence representation to polymer electrical properties and graph features to polymer properties; enhances the integrity and globality of structural representation, improves the sensitivity to different internal connections and connection point connection modes, and significantly improves the extrapolation and generalization ability of the model in the absence of conformation and special cases.

[0018] (2) Enhanced extrapolation capability of self-supervised feature learning: Using high-dimensional polymer features and sequence representations generated by self-supervised training, the model learns the continuous mapping relationship from monomer sequence and virtual structure diagram to polymer properties; by capturing periodic monomer patterns and global structural information through sequence model, the model's extrapolation capability for unseen monomer combinations, special end groups and novel conformations is improved; by combining extrapolation prediction and regularization training strategies, the model effectively alleviates the problem of data sparsity and uneven distribution, and improves the robustness and generalization capability of the model in complex polymer structure space.

[0019] (3) Fine capture of structure and features of virtual two-dimensional graph: By constructing a two-dimensional virtual polymer graph of "monomer + virtual edge connection", local neighborhood information is aggregated and connection relationship is explicitly encoded to achieve unified modeling of main chain configuration, substituent effect and repeating pattern between monomers; virtual edge sampling and learnable feature projection ensure stable expression of information at different structural scales and avoid the loss and bias of manually designed features; under weak annotation and diverse polymer structure, it maintains sensitivity to internal connection pattern, connection point change and global topology, and provides high-fidelity and low-noise potential representation for downstream self-supervised feature learning and multi-objective property prediction.

[0020] (4) Inference speed and ease of deployment: The design of a streamlined feature generation process and the separation of core model components result in a low latency and low resource overhead throughout the entire process from molecular structure input to property output; the standardized service encapsulation supports heterogeneous hardware and multiple parallel execution frameworks, and can be flexibly scheduled in local data centers and cloud computing environments; at the same time, it supports uninterrupted hot updates and multi-version coexistence mechanisms to ensure that online tasks maintain stable response and low latency during model iteration.

[0021] (5) Reliable prediction and risk perception capabilities: By modeling and characterizing the uncertainty of the model output, prediction biases caused by data sparsity or novel structure can be identified, thereby outputting the corresponding confidence range when facing out-of-distribution samples or potentially high-risk decisions. Combining cross-validation mechanism and multi-model fusion method, the system can maintain stable prediction quality when the input changes or the task switches, providing interpretable, controllable decision support with safety boundaries for material research and development and screening. Attached Figure Description

[0022] The invention will now be further described with reference to the accompanying drawings.

[0023] Figure 1 This is a flowchart illustrating the polymer multi-property prediction method based on self-supervised feature learning according to an embodiment of the present invention. Figure 2 This is a test set prediction result diagram of the polymer multi-property prediction method based on self-supervised feature learning according to an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see Figure 1-2This invention provides a method for predicting multiple properties of polymers based on self-supervised feature learning, comprising the following steps: Step 1: Construct a polymer multi-property dataset Data= { Y, M, A, P, R, N, S}; Y, M, A, P, R, N, S It is a set of different properties of polymers.

[0026] in: Y This is a polymer property tag matrix. Y= { T g , Bandgap Chain, T m , Density, Hardness} ∈ R I×K , K This represents the number of types of polymer molecular properties. Y The properties of the polymers contained herein include, but are not limited to, glass transition temperature T. g (K), Bandgap Chain (eV), Melting Temperature T m (K), Density (g / ml) and Hardness (MPa); M It is a collection of polymer monomer structures. M = { m 1,…, m i ,…, m I}; m i Indicates the first i The repeating unit structure of a polymer, I The total number of samples; by using the active end-group identifier "[" on the polymer monomer structure. Identify the polymerization reaction sites and their adjacent chemical environment, and record the number of polymerization sites for each polymer monomer; A represents the set of data sources and collected metadata. A ={ a 1,…, a i ,…, a I}; a i Used to record the source category of samples during data acquisition, for example: samples generated from materials literature extraction (such as those based on MaterialsBERT) forming a set. As a supplement to the data; P For the original polymer entity description set,P = { p 1,…, p i ,…, p I}; p i It includes polymer-level information, such as formulation, end groups, and molecular weight range. P and M Correspondingly, it supports structural correlations traced back from the polymer level to the monomer level; R Representing a set by repeating elements, R = { r 1,…, r i ,…, r I}, R and M One-to-one correspondence, used to store the repeating unit representation of the model input; N This is a collection of samples linking the structure and properties of polymer monomers. N = {( m i , y i ), used to record the association between polymer monomer structures and corresponding property labels; S The dataset was partitioned into training, validation, and test sets in an 8:1:1 ratio during the experiment. S ={ N tr , N val , N te},satisfy N tr ∪ N val ∪ N te = N e ∪ N 0 and they do not overlap.

[0027] Step 2, for Data Calculations are performed on monomer-to-polymer related characteristics.

[0028] Step 2.1, for Data Each SMILES expression in the code undergoes singleton validity filtering.

[0029] In this embodiment, step 2.1 specifically involves: for each SMILES expression in the dataset, firstly performing singleton validity filtering, especially excluding structures containing R-based placeholders (such as [R], [R1], [R2], etc.); since such placeholders cannot be directly used to generate the sub-analyses required for subsequent analysis via RDKit. Sub-structure Therefore, it needs to be removed from the dataset. The specific steps are as follows: traverse all SMILES and check if they contain the above R-based identifiers. Once a match is found, collect them into the R-based SMILES set. N RG ={ N rg1 ,…, N rgb ,…, N rgB},in B For the number of identified, b Indicates the first b 1 record. Finally, by removing members from this set, the original sample set is updated to 1. N total =( N e ∪ N 0) - N RG .

[0030] Step 2.2, from M A polymer virtual graph is constructed in the middle.

[0031] We know the set of polymer monomer structures M ,from M Constructing a polymer virtual graph G ={ g 1,…, g i ,…, g I The process includes: Step 2.2.1: For each monomer, construct the real edge by the interaction of directly connected atom pairs. E 1i ={ e 11 ,…,e 1q} i Virtual edges within a monomer are constructed by the interactions of non-directly connected atomic pairs. E 2i ={ e 21 ,…, e 2w} i p gauss,in p gauss It is a Gaussian probability distribution.

[0032] In this embodiment, step 2.2.1 specifically refers to: [The text abruptly ends here, so the translation stops.] AP bond ={ ap 1,…, ap q} i In addition, there are also pairs of atoms that are not directly connected. AP vin ={ ap 1,…, ap w} i , AP vin Interactions may also occur under special usage or chemical conditions, so we randomly selected... AP vin Add virtual edges between E 2i ={ e 21 ,…, e 2w} i p gauss ,in p gauss It follows a Gaussian probability distribution, reflecting the randomness of this interaction; therefore, ordinary edges... E 1i ={ e 11 ,…,e 1q} i After constructing the virtual edge E 2i Then it is merged with it, at which point the virtual graph edge set becomes E i ={ E 1 , E 2} i .

[0033] Step 2.2.2, from the polymer reaction site (i.e. [ The periodic interaction of two directly connected atoms constructs virtual edges between monomers. E 3i ; Obtain the set of virtual edges E i ={ E 1 , E 2 ,E 3} i .

[0034] In this embodiment, step 2.2.2 specifically involves: to simulate the periodic interactions of polymers composed of numerous identical monomers, we constructed a second type of virtual edge. E 3i This is due to [ The virtual edge is constructed from two directly connected atoms; after constructing the virtual edge... E 3i After adding the virtual graph edge set, the edge set is now... E i ={ E 1 , E 2 , E 3} i .

[0035] Step 2.2.3: Treat each atom as a node. V i As a node feature, and E i Together they form a virtual graph g i ={ V , E} i .

[0036] Step 2.3: For each atom in each SMILES, generate a self-supervised tag to obtain the electrical properties of each polymer. GC i .

[0037] In this embodiment, step 2.3 specifically involves: for each atom in each SMILES, generating its Gasteiger Charge self-supervised label using the RDKit library; and obtaining the electrical properties of each polymer using a trained 2D graph neural network. GC i ={ gc 1 ,…, gc Z} i ,in Z It is the number of atoms in the monomer.

[0038] Step 3: Construct and train a self-supervised prediction model for polymer multi-property features, and perform inference predictions, including: Step 3.1: Construct a model set; The model set includes improved 2D graph neural networks and sequence prediction models, unifies the interfaces and evaluation metrics of each module, uses a fixed random seed, and supports early stopping / parallelization.

[0039] in: The improved 2D graph neural network model is based on a virtual graph. g i ,right E 1i , E 2i and E 3i Feature learning is performed using different message passing methods to output the electrical property sequence of the polymer. GC i The specific steps are as follows: Step 3.1.1, based on the degree of the atom, for E 1i The neighboring atomic features of the connected components are then weighted and summed for updating; specifically: Message passing of real edges within a single entity; for each single entity m i Its z The property of each atom is V i ={ v 1 ,…, v Z}, each atom f Real connected neighbor set c ∈( Neighbor bond (f) ∪ f ), calculate the degree of each atom DEG i ={ deg 1 ,…, deg Z} i .;for e hb ∈ E 1i Calculate the normalization coefficient of the edge. norm hb = ,in h and b These are the two atoms on this edge; the final atom f Upon receiving the message, update your corresponding characteristics. ; Step 3.1.2: Based on the degree of the virtual edge, for E 2iThe neighboring atomic features of the connected components are then weighted and summed for updating; specifically: The conditions are the same as in step 3.1.1; atoms f Upon receiving the message, update your corresponding characteristics. ,in deg vin Represents atoms vin The degree of the edge.

[0040] Step 3.1.3: Calculate learnable attention weights E 3i The edge weights between connected atoms are used to weight and sum the features of neighboring atoms accordingly for updating; specifically: The conditions are the same as in step 3.1.1; for virtual edges E 3i The connected atoms are respectively h and b Calculate their edge weights ,in It is the sigmoid activation function. These are learnable attention weights. scale It is the scaling factor; finally, the summation is performed over the corresponding neighbors. Through steps 3.1.1.1 and 3.1.1.2, the final node... f The sum of received messages plus the bias bias , .

[0041] Step 3.1.4, will... E 1i , E 2i and E 3i The sum of transmitted messages, plus the bias, yields the updated node features. These features are then processed through multiple convolutional and linear layers to output the electrical property sequence of the polymer. GC i Specifically: get L The total message after convolution is ={ ,…, } i Then, the polymer electrical property sequence is obtained through a linear layer MLP. .

[0042] The sequence prediction model GC i As input, the complete property output sequence is predicted through a decoder and attention mechanism; specifically: For the polymer electrical property sequence obtained in step 3.1.4 The input is fed into the sequence prediction model decoder for item-by-item prediction of the property sequence: The model models the property sequence as a parallel output form, and extracts the internal structural relationships of the sequence through a multi-head attention mechanism: ,in All by The projection allows the model to learn the dependencies, periodic trends, local correlations, and long-range relationships between different properties, thus obtaining a complete property prediction output. OUTPUT ={ output 1 ,…, output i ,…, output I}

[0043] Step 3.2, for each target property An improved 2D graph neural network model and sequence prediction model based on self-supervised label training.

[0044] In this embodiment, step 3.2 specifically includes: For the first i One goal Y i = { Atomization Energy , Bandgap Bulk , T m , Refractive Index , T d} i Self-supervised labels are constructed based on steps 1 and 2.3 and added to the corresponding training set. N tr Subsequently, a standard self-supervised graphical neural network model and a sequence prediction model are trained separately for each target. During model training, network parameters are optimized through gradient descent, MSE is used as the training loss, and early stopping, learning rate scheduling, and regularization strategies are adopted to prevent overfitting. After training, multiple evaluation metrics such as MAE and R² are calculated on the validation set to comprehensively evaluate the model performance. Candidate models corresponding to the target property are sorted from low to high according to the validation error. The optimal model configuration, training parameters, and evaluation performance are recorded to support subsequent model selection, ensemble decision-making, or unified training process.

[0045] Step 3.3: Search and optimize the hyperparameter configuration of the model; In this embodiment, step 3.3 specifically includes: Self-supervised labels for construction Y SSL and monomers m iA 2D graph neural network is used to predict the electrical properties of the corresponding polymer, and the results are used as input for subsequent sequence models. Then, for each self-supervised graph neural network model and sequence prediction model, a model configuration search is performed based on the adjustable range of its structure and training strategy (e.g., hidden layer dimension, number of layers, number of attention heads, learning rate, Dropout, pre-training weight freezing strategy, etc., which can be random search, grid search, or Bayesian optimization). During the search process, the same training-validation partitioning and fixed random seed settings as described above are used to ensure the reproducibility of experimental results. The model performance is evaluated primarily by the validation set performance index R², and the configuration with the smallest overall error is selected as the model's performance on the target set. Y i The optimal solution is determined by the search scope. For tasks with different physical properties, the search scope will be adjusted appropriately according to the task characteristics. At the same time, the use of any validation data to update model parameters will be strictly avoided throughout the process to maintain the independence and reliability of model evaluation.

[0046] Step 3.4: Use an integrator to perform weighted fusion of the prediction results from multiple base models to obtain the final predicted value of the target property; Step 3.4.1, for each target property Multiple different base models are trained, and the prediction results and evaluation metrics of each model are obtained on the validation set; specifically: For each target property Y i Using graph representations generated by a self-supervised graph neural network and property sequence representations generated by a sequence prediction model as inputs, multiple base models (including graph neural network models and sequence prediction models with different structures or hyperparameter configurations) are trained under a fixed, uniform random seed. In each training iteration, only the training subset is used to fit the corresponding model, and the model's performance on the validation subset is then evaluated. Y i The prediction results and evaluation indicators.

[0047] Step 3.4.2: Sort and filter the base models based on the validation set performance, and assign fusion weights to the selected models based on performance metrics; specifically: In the nature of the target Y i Based on the performance of the validation set, such as MAE or R... 2 Using error as the primary metric, candidate models are sorted from smallest to largest, and the best-performing model is selected as the ensemble member. Weights are generated using error inverse normalization or soft normalization. The outputs of the best-performing models are then weighted and fused based on these weights to form a ensemble member targeting the desired model. Y i An integrated predictor.

[0048] Step 3.4.3: In the inference phase, the input features are passed through all selected base models, and the prediction results of each model are weighted and fused according to the obtained weights to output the final predicted value of the target property; specifically: During the reasoning phase, for each target property Y i The standardized self-supervised graph representation and property sequence representation generated in step 3.1.1.4 are used as input, and are passed through all the best base models, and the prediction results of each model are collected; then, they are weighted and integrated according to the fixed fusion weights learned in step 3.4.2 to obtain the target. Y i The final predicted value is obtained; the process is repeated for all target properties, and the results are summarized to generate the final multi-property prediction results.

[0049] Step 3.5: Record and save all training and model parameters; specifically: After training, the system will fully record and save: all hyperparameter configurations, training logs and validation results of each supervised graph neural network model and sequence prediction model, final weights of the ensemble, input feature columns used, data standardization parameters, random seeds, model architecture versions (including graph encoder and sequence predictor), self-supervised task and pre-trained weight versions, and runtime dependency library information. When adding new samples or expanding features, simply reuse the same data specifications, self-supervised feature construction process, and model training process to achieve stable incremental updates, ensuring that training results at different stages remain comparable, traceable, and consistent.

[0050] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for predicting multiple properties of polymers based on self-supervised feature learning, characterized in that, Includes the following steps: Step 1: Construct a polymer multi-property dataset Data= { Y, M, A, P, R, N, S }; Y, M, A, P, R, N, S It is a collection of different properties of polymers; Step 2, for Data Perform characteristic calculations from monomers to polymers; Step 2.1, for Data Each SMILES expression in the code undergoes singleton validity filtering. Step 2.2, from M Construct a polymer virtual graph in the middle; Step 2.3: For each atom in each SMILES, generate a self-supervised tag to obtain the electrical properties of each polymer. GC i ; Step 3: Construct and train a self-supervised prediction model for polymer multi-property features, and perform inference predictions, including: Step 3.1: Construct a model set; Step 3.2, for each target property An improved 2D graph neural network model and sequence prediction model based on self-supervised label training; Step 3.3: Search and optimize the hyperparameter configuration of the model; Step 3.4: Use an integrator to perform weighted fusion of the prediction results from multiple base models to obtain the final predicted value of the target property; Step 3.5: Record and save all training and model parameters.

2. The polymer multi-property prediction method based on self-supervised feature learning according to claim 1, characterized in that, In step 1, Y This is a polymer property tag matrix. M Let A be the set of polymer monomer structures, and let A be the set of data sources and collected metadata. P For the original polymer entity description set, R Representing a set by repeating elements, N This is a collection of samples linking the structure and properties of polymer monomers. S Strategies for partitioning the dataset.

3. The polymer multi-property prediction method based on self-supervised feature learning according to claim 2, characterized in that, In step 2.2, from M The steps for constructing a polymer virtual graph include: Step 2.2.1: For each monomer, construct the real edge by the interaction of directly connected atom pairs. E 1i ={ e 11 ,…,e 1q } i Virtual edges within a monomer are constructed by the interactions of non-directly connected atomic pairs. E 2i ={ e 21 ,…, e 2w } i p gauss ,in p gauss It is a Gaussian probability distribution; Step 2.2.2: Construct virtual edges between monomers through the periodic interactions of two atoms directly connected to the polymer reaction sites. E 3i ; Obtain the set of virtual edges E i ={ E 1 , E 2 , E 3 } i ; Step 2.2.3: Treat each atom as a node. V i As a node feature, and E i Together they form a virtual graph g i ={ V , E } i .

4. The polymer multi-property prediction method based on self-supervised feature learning according to claim 3, characterized in that, In step 3.1, the model set includes an improved 2D graph neural network and a sequence prediction model; wherein the improved 2D graph neural network model is based on a virtual graph. g i ,right E 1i , E 2i and E 3i Feature learning is performed using different message passing methods to output the electrical property sequence of the polymer. GC i The sequence prediction model uses GC i As input, the complete property output sequence is predicted through a decoder and attention mechanism.

5. The polymer multi-property prediction method based on self-supervised feature learning according to claim 4, characterized in that, The improved 2D graph neural network model outputs a sequence of the polymer's electrical properties. GC i The steps include: Step 3.1.1, based on the degree of the atom, for E 1i The neighboring atomic features of the connected atoms are then weighted and summed for updating. Step 3.1.2: Based on the degree of the virtual edge, for E 2i The neighboring atomic features of the connected atoms are then weighted and summed for updating. Step 3.1.3: Calculate learnable attention weights E 3i The edge weights between connected atoms are used to update the features of neighboring atoms through a weighted summation. Step 3.1.4, will... E 1i , E 2i and E 3i The sum of transmitted messages, plus the bias, yields the updated node features. These features are then processed through multiple convolutional and linear layers to output the electrical property sequence of the polymer. GC i .

6. The polymer multi-property prediction method based on self-supervised feature learning according to claim 4, characterized in that, The sequence prediction model employs a multi-head attention mechanism, and its calculation method is as follows: ,in, All by Obtained by projection.

7. The polymer multi-property prediction method based on self-supervised feature learning according to claim 6, characterized in that, In step 3.2, the method for individual training is as follows: optimize network parameters through gradient descent, use MSE as training loss, and adopt early stopping, learning rate scheduling, and regularization strategies.

8. The polymer multi-property prediction method based on self-supervised feature learning according to claim 7, characterized in that, In step 3.3, the method for searching and optimizing the hyperparameter configuration of the model is as follows: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] Y SSL and monomers m i A 2D graph neural network is used to predict the electrical properties of the corresponding polymer, and the results are used as input for subsequent sequence models. Then, for each self-supervised graph neural network model and sequence prediction model, a model configuration search is performed based on the adjustable range of its structure and training strategy. Model performance is evaluated using the validation set performance index R. 2 As an evaluation criterion, the configuration with the smallest overall error is selected as the model's target configuration. Y i The optimal solution.

9. The polymer multi-property prediction method based on self-supervised feature learning according to claim 8, characterized in that, In step 3.4, the step of using an integrator to weight and fuse the prediction results of multiple base models to obtain the final predicted value of the target property includes: Step 3.4.1, for each target property Train multiple different base models and obtain the prediction results and evaluation metrics of each model on the validation set; Step 3.4.2: Sort and filter the base models according to the performance of the validation set, and assign fusion weights to the selected models based on performance metrics; Step 3.4.3: In the inference stage, the input features are passed through all selected base models, and the prediction results of each model are weighted and fused according to the obtained weights to output the final predicted value of the target property.

10. The polymer multi-property prediction method based on self-supervised feature learning according to claim 9, characterized in that, In step 3.5, when adding new samples or expanding features, the same data specifications, self-supervised feature construction process, and model training process are reused.

Images