A dual-channel polymer performance prediction system and synthesis scheme recommendation system

Abstract

This disclosure belongs to the field of machine learning technology. This disclosure provides a dual-channel polymer performance prediction system and a synthesis scheme recommendation system. The polymer performance prediction system includes: a parsing module for parsing the SMILES string of monomers to obtain a first feature vector and a second feature vector; a first encoder for inputting the first feature vector and outputting a first embedding vector; a second encoder for inputting the second feature vector and outputting a second embedding vector; a feature concatenation module for concatenating the first embedding vector and the second embedding vector into a monomer embedding vector; a feature fusion module for performing a weighted summation of the monomer embedding vectors according to the molar ratio coefficient of each monomer in the polymer to obtain a polymer embedding vector; and a first decoder for inputting the polymer embedding vector and outputting a first performance prediction vector.

Description

Technical Field

[0001] This disclosure belongs to the field of machine learning technology, and in particular relates to a polymer performance prediction system and a synthesis scheme recommendation system based on a dual-channel architecture. Background Technology

[0002] Currently, the acquisition of polymer performance parameters mainly relies on two methods: experimental testing and computational simulation. However, both have significant shortcomings. Experimental testing methods require a series of complex steps, including monomer screening, polymerization reaction, film formation treatment, and performance characterization. This process is not only time-consuming (usually taking several months) but also costly (a single synthesis test costs over ten thousand yuan). Furthermore, due to the vast chemical space of monomer combinations, experimental methods cannot comprehensively cover all possible polymer structures, resulting in low research and development efficiency. Computational simulation methods, taking traditional machine learning (ML) as an example, while able to establish a mapping relationship between polymer structure and performance, face three major bottlenecks in practical applications: 1. Data characterization bottleneck: The complexity of polymer structures and the lack of a unified and computable descriptor system make data characterization difficult; 2. Model generalization bottleneck: Most existing machine learning models focus on homopolymers, leaving gaps in the prediction of copolymers and key thermal properties (such as glass transition temperature and half-decomposition temperature); 3. Interdisciplinary barriers: Materials scientists often lack programming and algorithm knowledge, while computer experts are unfamiliar with material mechanisms, limiting the practicality of the models.

[0003] In recent years, artificial intelligence (AI) technology has provided a new paradigm for materials research and development. In the field of small molecules, linear symbolic languages ​​such as SMILES (Simplified Molecular Input Line Entry System) can encode chemical structures into strings, and combined with natural language processing (NLP) techniques (such as Transformer models), significant breakthroughs have been achieved in drug molecule generation and property prediction. However, in the field of polymers, the application of AI technology is still in its early stages. Since 2022, teams from Georgia Tech, Carnegie Mellon University, and others have attempted to apply SMILES to representation learning of homopolymers, but currently, they only use the Transformer encoder for feature extraction and cannot achieve structure generation or performance prediction. While large language models (LLMs) have performed well in the fields of proteins and small molecules, mature models have not yet been established in the field of polymers. Summary of the Invention

[0004] This disclosure provides a dual-channel polymer performance prediction system and a synthesis scheme recommendation system, which can effectively solve the above-mentioned problems.

[0005] This disclosure is implemented as follows:

[0006] In a first aspect, this disclosure provides a polymer performance prediction system based on a dual-channel architecture, the polymer performance prediction system comprising:

[0007] The parsing module is used to parse the SMILES string of the monomer to obtain a first feature vector and a second feature vector, wherein the first feature vector is used to characterize the atomic and bond features of the monomer, and the second feature vector is used to characterize the global molecular features of the monomer.

[0008] A first encoder, which employs a graph neural network, is used to take the first feature vector as input and output a first embedding vector.

[0009] The second encoder, which employs a fully connected neural network, is used as input to the second feature vector and outputs a second embedding vector.

[0010] The feature concatenation module is used to concatenate the first embedding vector and the second embedding vector into a single embedding vector.

[0011] The feature fusion module is used to perform a weighted summation on the monomer embedding vector according to the molar ratio coefficient of each monomer in the polymer to obtain the polymer embedding vector;

[0012] The first decoder, which employs a Transformer Decoder, takes a polymer embedding vector as input and outputs a first performance prediction vector, which is used to predict the first performance of the polymer.

[0013] Secondly, this disclosure provides a polymer synthesis scheme recommendation system based on a large language model. The polymer synthesis scheme recommendation system is used to predict the synthesis scheme of a second polymer. The polymer synthesis scheme recommendation system includes:

[0014] The text recognition module is used to identify the first performance data to be queried based on the query text of the first performance.

[0015] The combination recommendation module is used to obtain a recommended monomer combination based on the first performance data to be queried, wherein the first predicted data of the first performance of the second polymer matches the first performance data to be queried, and the recommended monomer combination includes each of the second monomers in the second polymer and its molar ratio, and the mapping between the recommended monomer combination and the first predicted data is determined by the polymer performance prediction system described in the first aspect.

[0016] The data acquisition module is used to call a vector database to obtain first synthesis scheme data based on the recommended combination of entities, and to call a graph database to obtain second synthesis scheme data based on the recommended combination of entities. Each vector in the vector database is a vector representation of a synthesis experimental data fragment in the literature; the graph database is used to record entities in the literature and the relationships between those entities.

[0017] The scheme recommendation module is used to obtain a recommended synthesis scheme based on the first synthesis scheme data and the second synthesis scheme data.

[0018] Compared with the prior art, the beneficial effects of this disclosure are:

[0019] 1. This disclosure provides a polymer performance prediction system that uses a dual-channel end-to-end machine learning model of GNN (Graph Neural Network) and FCNN (Fully Connected Neural Network) to achieve high-precision prediction from monomer combination to key polymer performance. It can screen monomer combinations that meet the target performance before the experiment, avoid the traditional "synthesis-test" cycle, break through the bottleneck of high-performance polymer design, and screen monomer combinations for specific polymer performance.

[0020] 2. This disclosure provides a polymer synthesis scheme recommendation system, which adopts a dual-engine RAG (Retrieval-Augmented Generation) system, combining vector retrieval and graph reasoning to generate synthesis schemes that are both data-supported and scientifically explained, helping researchers understand the performance optimization mechanism. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this disclosure and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1This is a schematic diagram of the structure of the polymer performance prediction system 1000 provided in the embodiments of this disclosure.

[0023] Figure 2 This is a schematic diagram of the model framework used in the polymer performance prediction system 1000 provided in this embodiment of the disclosure.

[0024] Figure 3 This is an example diagram of the QM9 (Quantum Machines 9) small molecule dataset used in the pre-training phase of the polymer performance prediction system 1000 provided in this embodiment of the disclosure.

[0025] Figure 4 This is an example diagram of the polyimide dataset used in the fine-tuning stage of the polymer performance prediction system 1000 provided in this embodiment of the disclosure.

[0026] Figure 5 This is a flowchart of the structured experimental data extraction process provided in this embodiment.

[0027] Figure 6 This is a distribution diagram of various first performance data in the polyimide dataset provided in this embodiment of the disclosure.

[0028] Figure 7 This is a graph showing the results of the polymer performance prediction system 1000 provided in this embodiment predicting different first properties.

[0029] Figure 8 This is a visualization of the attention weights of the GNN (Graph Neural Network) model provided in this embodiment.

[0030] Figure 9 This is a t-SNE (t-distributed Stochastic Neighbor Embedding) visualization analysis diagram of the polymer performance prediction system 1000 provided in this embodiment of the present disclosure during monomer screening.

[0031] Figure 10 This is a flowchart of the bimodal knowledge base construction process provided in this embodiment.

[0032] Figure 11 This is a schematic diagram of the polymer synthesis scheme recommendation system 2000 provided in the embodiments of this disclosure. Detailed Implementation

[0033] The present disclosure will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the disclosure. Furthermore, it should be noted that, for ease of description, only the parts relevant to the present disclosure are shown in the accompanying drawings.

[0034] It should be noted that, where there is no conflict, the embodiments and features described in this disclosure can be combined with each other. The technical solutions of this disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0035] Unless otherwise stated, the exemplary implementations / embodiments shown are to be understood as providing exemplary features of various details that provide ways in which the technical concepts of this disclosure can be implemented in practice. Therefore, unless otherwise stated, the features of various implementations / embodiments may be additionally combined, separated, interchanged and / or rearranged without departing from the technical concepts of this disclosure.

[0036] Example 1

[0037] This disclosure provides a dual-channel polymer performance prediction system 1000.

[0038] like Figure 1 As shown, the polymer performance prediction system 1000 includes: a parsing module 1002, a first encoder 1004, a second encoder 1006, a feature splicing module 1008, a feature fusion module 1010, and a first decoder 1012.

[0039] Specifically, the parsing module 1002 is used to parse the SMILES string of the monomer to obtain a first feature vector and a second feature vector, wherein the first feature vector is used to characterize the atomic and bond features of the monomer, and the second feature vector is used to characterize the global molecular features of the monomer.

[0040] The first encoder 1004 employs a graph neural network, which takes the first feature vector as input and outputs a first embedding vector.

[0041] The second encoder 1006 employs a fully connected neural network, which is used as input to the second feature vector and outputs a second embedding vector.

[0042] The feature concatenation module 1008 is used to concatenate the first embedding vector and the second embedding vector into a single embedding vector.

[0043] The feature fusion module 1010 is used to perform a weighted summation on the monomer embedding vector according to the molar ratio coefficient of each monomer in the polymer to obtain the polymer embedding vector;

[0044] The first decoder 1012 employs a Transformer Decoder, which takes a polymer embedding vector as input and outputs a first performance prediction vector, which is used to predict the first performance of the polymer.

[0045] In some implementations, the parsing module 1002 uses the cheminformatics toolkit RDKit to parse the SMILES string of the monomer, and obtains a first feature vector and a second feature vector, respectively.

[0046] The analytical module 1002 involves multi-scale feature extraction to comprehensively characterize the chemical structure of monomers.

[0047] 1. Atomic / Bond-Level Features: RDKit is used to analyze SMILES, generating molecular diagram structures (nodes = atoms, edges = chemical bonds), and extracting local features such as atomic hybridization type, valence state, bond order, and aromaticity. Intrinsic properties (atomic physical properties) such as atomic radius, electronegativity, and electron affinity are supplemented by the Mendeleev periodic table.

[0048] 2. Global molecular characteristics: Calculate macroscopic descriptors such as molecular weight, surface energy, and functional group ratio.

[0049] In some implementations, the first feature vector contains (26N+5M) dimensions, where N is the number of atoms in the monomer, i.e., each atom is converted into a 26-dimensional atomic feature, and M is the number of chemical bonds, i.e., each chemical bond is converted into a 5-dimensional bond feature.

[0050] Specifically, based on the SMILES string of the monomer, the molecule is converted into an undirected graph. .in, The set of nodes representing atoms E The edge set representing a chemical bond.

[0051] In some implementations, the second feature vector contains 399-dimensional global molecular features.

[0052] like Figure 2 As shown, the polymer performance prediction system 1000 employs a dual-channel encoder. The first encoder 1004 uses a graph neural network (GNN) to process molecular topology. The second encoder 1006 uses a fully connected neural network (FCNN) to process global molecular features. The dual-channel encoder is designed to process molecular information through two distinct paths. The GNN branch focuses on extracting atomic-level and bond properties from the molecular graph, transforming them into meaningful embeddings to capture the molecular topology. These embeddings are then aggregated through average pooling to form a unified molecular representation. Simultaneously, the FCNN branch processes global molecular descriptors, covering a wide range of physicochemical properties. The outputs of the two branches are concatenated to form a comprehensive feature vector, which is then input into a Transformer-based output layer for final property prediction. This architecture enables the model to capture both local structural features and global properties of molecules, providing a solid foundation for generative design tasks.

[0053] In some implementations, the graph neural network is a graph neural network based on GATv2 (Graph Attention Networks v2) layers. The GNN branches are constructed using GATv2 layers. Shaked Brody et al. proposed GATv2 in the paper "How Attentive are Graph Attention Networks?", an improved variant of GAT (Graph Attention Networks) capable of implementing a dynamic attention mechanism. The main improvement of GATv2 lies in addressing the limitations of static attention in the original GAT. In the original GAT, the ranking of attention scores is fixed and independent of the query node; this mechanism is called "static attention." GATv2, by adjusting the order of operations, introduces a "dynamic attention" mechanism, allowing the attention scores to be dynamically adjusted based on the characteristics of the query node.

[0054] Specifically, the 26-dimensional original features of atoms are projected to a higher dimension through the fully connected layers of the GNN. For example, using a 64-dimensional hidden layer design, the features of each atom are mapped to a 64-dimensional vector. The key features are used for information transfer between neighboring atoms during graph convolution (such as attention weight calculation), but do not change the dimension of the atom features themselves.

[0055] In some implementations, the final outputs of both GNN and FCNN are 32-dimensional data. The atomic vectors output by GNN are aggregated into 32-dimensional molecular vectors through average pooling, and then concatenated with the 32-dimensional output of FCNN by the feature concatenation module 1008 to form a 64-dimensional final embedding vector, i.e., the single-unit embedding vector.

[0056] GNN captures local topological structure, while FCNN integrates global molecular properties; the two complement each other. The concatenated monomer embedding vector contains both structural and global information, which enhances the output performance of the first decoder 1012.

[0057] Specifically, the model framework adopted by the polymer performance prediction system 1000 is shown in Table 1.

[0058]

[0059] The feature fusion module 1010 uses a molar ratio coefficient. ( ), for single-entity embedding vectors Perform a weighted summation to obtain the polymer embedding vector. ,in, i Indicates the order of monomer types in a polymer. , nThis indicates the number of monomer types in the polymer. The polymer is formed by polymerizing at least one monomer. That is, the polymer can be a homopolymer or a copolymer.

[0060] The polymer embedding vector serves as the input data for the first decoder 1012. The Transformer Decoder of the first decoder 1012 captures the complex dependencies between local topology (atomic / bond level), global molecular properties, and prediction performance through a self-attention mechanism.

[0061] The predicted first property could be a polymer-specific key property, such as the half-decomposition temperature. T d Carbon dioxide permeability, glass transition temperature T g Dielectric constant ε Tensile strength R m At least one of them. These predictions serve as key inputs to guide the design process of polymer synthesis schemes and to set the target performance range of polymer molecules.

[0062] Example 1

[0063] This disclosure provides a training method for the polymer performance prediction system 1000, specifically including:

[0064] I. Pre-training phase:

[0065] The pre-training phase involves training the first encoder 1004 and the second encoder 1006 to enable the encoders to capture universal molecular features.

[0066] The polymer performance prediction system further includes a second decoder, which employs a TransformerDecoder to take as input the monomer embedding vector of the small molecule monomer and output a second performance prediction vector, which is used to predict the quantum chemical properties of the small molecule monomer.

[0067] The training methods for the first encoder and the second encoder include:

[0068] The small molecule monomer dataset is used as the first training set, and each sample in the first training set contains the SMILES string of the small molecule monomer and its quantum chemical properties data.

[0069] The SMILES string of the small molecule monomer is input into the parsing module to obtain the first feature vector and the second feature vector of the small molecule monomer, respectively.

[0070] The first feature vector and the second feature vector of the small molecule monomer are respectively input into the first encoder and the second encoder to be trained to obtain the first embedding vector and the second embedding vector of the small molecule monomer.

[0071] The feature splicing module splices the first embedding vector and the second embedding vector of the small molecule monomer into the monomer embedding vector of the small molecule monomer.

[0072] The monomer embedding vector of the small molecule monomer is input into the second decoder to obtain the second performance prediction vector;

[0073] The parameters of the first encoder and the second encoder are optimized based on the quantum chemical property data of the small molecule monomer and the second performance prediction vector.

[0074] like Figure 3 As shown, the pre-training phase uses the open-source QM9 dataset, which includes 133,000 small molecular quantum mechanical records.

[0075] The SMILES of a single molecule in the QM9 dataset can be parsed by RDKit to obtain a total of 26N+5M+399 features (26-dimensional atomic features, 5-dimensional bond features, and 399-dimensional global molecular features).

[0076] As shown in Table 1, the input to the GNN branch consists of 26N-dimensional atomic features (hybridization type, valence state, etc.) and 5M-dimensional bond features (bond order, aromaticity, etc.). The GNN has 6 hidden layers, and in the 2nd to 6th hidden layers, the input and output of each layer are 64N-dimensional. The final output of the GNN is 32-dimensional data.

[0077] The input to the FCNN branch is 399-dimensional global molecular features (molecular weight, surface energy, functional group ratio, etc.). FCNN has two hidden layers. The first hidden layer has a 399-dimensional input and a 128-dimensional output. The second hidden layer has a 128-dimensional input and a 128-dimensional output. The final output of FCNN is 32-dimensional data.

[0078] The 32-dimensional data output from the GNN branch and the 32-dimensional data output from the FCNN branch are concatenated to form a total of 64-dimensional data, which serves as the input data for the second decoder. The second decoder's Transformer Decoder has two hidden layers, each with a dimension of 32. The final output of the second decoder, i.e., the second performance prediction vector, is 12-dimensional, representing 12 quantum mechanical properties of the small molecule, such as the HOMO-LUMO band gap and internal energy.

[0079] In the pre-training phase, a second decoder is introduced to predict various quantum properties of small molecule monomers, enabling the first encoder 1004 and the second encoder 1006 to learn the general molecular representation.

[0080] The pre-training phase configuration is shown in Table 2.

[0081] Table 2 Pre-training Phase Configuration

[0082]

[0083] II. Fine-tuning stage

[0084] After pre-training, the parameters of GNN and FCNN are fixed, and the first decoder 1012 is trained.

[0085] The training method for the first decoder 1012 includes:

[0086] The first polymer dataset is used as the second training set. Each sample in the second training set contains the SMILES string of each first monomer in the first polymer and its molar ratio coefficient, as well as the first performance data of the first polymer. The first polymer dataset is derived from publicly available literature in the field of synthesizing the polymer.

[0087] The SMILES string of the first monomer is input into the parsing module to obtain the first feature vector and the second feature vector of the first monomer, respectively.

[0088] The first feature vector and the second feature vector of the first single entity are respectively input into the trained first encoder and the second encoder to obtain the first embedding vector and the second embedding vector of the first single entity.

[0089] The feature splicing module splices the first embedding vector of the first monomer with the second embedding vector to form the monomer embedding vector of the first monomer.

[0090] The feature fusion module performs a weighted summation on the monomer embedding vector of the first monomer to obtain the polymer embedding vector of the first polymer.

[0091] The polymer embedding vector of the first polymer is input into the first decoder to be trained to obtain the first performance prediction vector of the first polymer.

[0092] The parameters of the first decoder are optimized based on the first performance data and the first performance prediction vector of the first polymer.

[0093] like Figure 4As shown, the fine-tuning phase utilizes a polyimide dataset. This dataset records the SMILES string for each monomer, along with experimental datasets of synthesized polyimides extracted from polyimide literature, including structured data such as monomer combinations, synthesis conditions (e.g., 60°C, 16 hours), and performance characterization data. Figure 5 As shown, the structured data extraction process is as follows: Public literature related to synthetic polymers (taking polyimide PI as an example) is retrieved from major databases (Web of Science, Semantic Scholar, PubMed, etc.) using domain keywords. Specifically, this embodiment uses over 6000 top-tier (Q1) polyimide literature articles. Based on Natural Language Processing (NLP) technology, structured data is systematically extracted from the literature, including: 1. The product is a homopolymer / polymer. 2. Types of diamine monomers, dianhydride monomers, and their molar ratios. 3. Solvent type, such as DMF (Dimethylformamide) / NMP (Methylpyrrolidone) / DMAC (Dimethylacetamide), etc. 4. Solid content calculated before the chemical reaction. 5. Chemical reaction temperature. 6. Chemical reaction time. 7. Post-treatment temperature. 8. Post-treatment time. 9. Characterization data, such as molecular weight, thermal properties, rheological / mechanical properties, electronic properties, optical properties, and structural features.

[0094] The distribution of various first performance data in the polyimide dataset is as follows: Figure 6 As shown. The polyimide dataset includes the half-decomposition temperature of polyimide. T d Glass transition temperature T g band gap E g, dielectric constant ε Various performance characterization data.

[0095] Furthermore, RDKit was used to generate multiple randomized SMILES from the standard SMILES for each monomer to perform data augmentation. These randomized sequences represent different textual representations of the same molecule while maintaining chemical structural equivalence. The augmented polyimide fine-tuning dataset contains approximately 2.5k strings. In other words, data augmentation was performed using a SMILES randomization strategy to generate multiple equivalent SMILES strings for the same monomer.

[0096] During the fine-tuning process, hyperparameters are adjusted through Bayesian optimization to address the problem of uneven data distribution.

[0097] The fine-tuning phase configuration is shown in Table 3.

[0098] Table 3 Fine-tuning Phase Configuration

[0099]

[0100] During fine-tuning, pre-trained weights from GNN and FCNN are transferred to the polyimide dataset. Simultaneously, SMILES of two monomers are input (because the synthesis of polyimide requires dianhydride and diamine monomers respectively). Each monomer's SMILES is encoded separately through pre-trained GNN and FCNN branches, concatenated to generate a 64-dimensional monomer feature vector, then weighted by the molar ratio coefficient of the monomer combination, and processed by the first decoder 1012 to be trained. The Transformer Decoder of the first decoder 1012 has two hidden layers, each with 32 dimensions. The final output of the first decoder 1012, i.e., the first performance prediction vector, is 1-dimensional data used to characterize a specific performance of polyimide. The prediction results of the trained polymer performance prediction system 1000 for different first performances are as follows: Figure 7 As shown. Figure 7 The scatter plot shows a comparison between the predicted and actual values ​​of various first properties, indicating that the polymer performance prediction system 1000, after fine-tuning, can accurately predict the specific properties of polyimide.

[0101] Example 2

[0102] I. Model Interpretability

[0103] Attention-based graph neural network models can effectively calculate the attention weights between adjacent atoms in a molecular structure and map these weights to the corresponding chemical bonds. Through visualization techniques, we present the attention weights of nodes as color gradients, where color depth is positively correlated with the attention weight value. This characterization method provides an intuitive basis for quantitatively analyzing the contribution of different chemical fragments to the prediction of molecular properties.

[0104] like Figure 8 As shown, the trained graph neural network model of the polymer performance prediction system 1000 is analyzed. Based on different monomer combinations, the model attention weights can be extracted to draw a heatmap matrix of different atoms and bonds. The different shades of grid color represent the degree of contribution of different atoms and bonds in the monomer to the output; the darker the color, the higher the degree of attention the GNN pays to that group. This embodiment of the disclosure reveals the influence mechanism of specific functional groups on the first performance of polyimide molecules by systematically analyzing the distribution law of attention weights. Figure 8The visualization results correspond to a graph neural network model of the polymer property prediction system 1000 used for predicting glass transition temperature. Corresponding to the prediction of glass transition temperature, the visualization results show:

[0105] 1. The high degree of rotational freedom of ether bonds in the molecular structure leads to enhanced chain segment mobility, negatively impacting the glass transition temperature (GVT). 2. The presence of symmetrical methyl groups significantly reduces the effective inertia of the system, also resulting in a lower GVT. 3. The effect of introducing a benzene ring structure on increasing the GVT is insufficient to offset the lowering effect of ether bonds. 4. When the dianhydride structure contains a macrocyclic system, molecular rigidity is enhanced, resulting in a higher GVT. 5. Introducing functional groups such as diphenylenediamine, chlorine, nitrate, and hydroxyl groups into the diamine structure significantly increases the GVT of polyimides. Specifically, diphenylenediamine increases molecular rigidity through steric hindrance; chlorine atoms promote the formation of intermolecular cross-linking networks at high temperatures, thus improving the material's thermal stability; and nitrate and hydroxyl groups enhance intermolecular interactions to increase the GVT.

[0106] II. Molecular Design / Formulation Screening

[0107] Based on the establishment of a high-performance prediction model, this disclosure systematically explores potential high-performance polyimide materials using high-throughput virtual screening technology. Based on the known reaction system of diamine / diisocyanate and dianhydride, a theoretical molecular library containing 60,000 polyimides is constructed.

[0108] First, corresponding to the prediction of the glass transition temperature, this embodiment of the disclosure uses model prediction to initially screen 2,000 polyimide molecular structures with a glass transition temperature greater than 400℃. The model extracts the t-SNE clustering visualization distribution of different polyimide monomer combination characteristics as shown below. Figure 9 As shown, the part circled in red represents the monomer combination range with a higher predicted glass transition temperature.

[0109] Considering the feasibility of actual synthesis, this disclosure further introduces a synthesis accessibility score (SAscore) and dielectric constant as important screening indicators. The SAscore scoring system uses a 1-10 scale; a lower score indicates easier synthesis. The dielectric constant is a physical quantity that measures a material's ability to store electrostatic energy in an electric field; a higher value indicates a stronger response to the electric field. For polyimide, its dielectric constant is typically between 2 and 4. This characteristic makes polyimide an ideal choice for high-frequency circuits and insulating materials. The dielectric constant value of the polyimide molecule is predicted using another polymer performance prediction system 1000 for predicting dielectric constant. This polymer performance prediction system 1000 also uses the aforementioned polyimide dataset (corresponding to the dielectric constant). ε The data was trained using characterization data. Glass transition temperature was selected based on the three screening indicators mentioned above. T g Based on the synthesis accessibility score (SAscore) and dielectric constant, 20 molecular structures with an SAscore < 3.8 and a dielectric constant > 3 were selected from the polyimide molecules initially screened above.

[0110] Furthermore, to address the specific requirements of flexible electronic devices and aerospace materials, this disclosure establishes a multi-objective optimization screening process. By comprehensively evaluating key parameters such as thermal properties, mechanical properties, dielectric properties, and synthesis feasibility, several polyimide materials with excellent overall performance were ultimately identified. The molecular structures of the selected materials were compared with those of commercial polyimides, and they were superior in all key parameters. Based on the synthesis accessibility score (SAscore) analysis, these selected materials all have good synthesis feasibility, and complete polymerization reaction routes have been designed.

[0111] The polymer performance prediction system 1000 disclosed herein can overcome the bottleneck of high-performance material design, screen monomer combinations for specific polymer properties, and achieve high-precision prediction from monomer combinations to key properties (such as glass transition temperature, dielectric constant, and tensile strength) through an end-to-end machine learning model (GNN+FCNN dual-channel). The system can screen monomer combinations that meet target properties before experimentation, avoiding the traditional "synthesis-testing" cycle. Furthermore, while traditional performance prediction models only support homopolymer prediction, this system, through transfer learning (QM9 dataset pre-training + polyimide dataset fine-tuning) and the SMILES data augmentation strategy, can predict the thermal, mechanical, and dielectric properties of copolymer systems containing multiple monomers.

[0112] This system integrates materials expertise with advanced AI technology. By establishing a structural coding standard applicable to complex polymers (especially copolymers), it overcomes the ambiguity of SMILES in representing long polymer chains. Its end-to-end intelligent framework enables high-precision prediction from molecular structure input to key properties (e.g., glass transition temperature). It lowers the barrier to interdisciplinary applications, allowing materials researchers to accelerate polymer molecular design using AI tools without requiring programming expertise. This system has significant application value in efficiently predicting specific polymer properties. Predicting relevant characteristics before laboratory synthesis and testing will significantly shorten the R&D cycle and promote the industrial application of polymer materials.

[0113] Example 3

[0114] I. Bimodal Knowledge Extraction and Storage

[0115] The process of acquiring bimodal knowledge is performed in parallel with the extraction of the structured data described above. Specifically, as mentioned above, this disclosure uses over 6000 polyimide-related documents.

[0116] Vector databases: such as Figure 10 As shown, text and table data from polyimide literature are segmented and then vectorized using the Gemini Embedding-001 embedding model. These vectors are then stored as high-dimensional vectors in the Pinecone vector database to enable semantic retrieval. Pinecone is a high-performance vector database designed for processing and storing large amounts of vector data. Its core functions include adding, deleting, and searching vectors for similarity, and it supports use as a backend for large language model applications.

[0117] Graph databases: such as Figure 10 As shown, the indexing phase used Deepseek-R1 to extract all entity relationships related to chemical synthesis from polyimide literature texts, constructing a knowledge graph, which was then stored using Neo4j (e.g., "BPDA combined with ODA →"). T g >300℃ → Suitable for aerospace coatings).

[0118] Both the vector and graph databases are built upon the same foundation of over 6,000 professional polyimide literature articles. By integrating the vector and graph databases, a bimodal knowledge base for efficient retrieval is constructed, enabling deep semantic search and logical reasoning. The vector database utilizes the high-performance vector database Pinecone to store and retrieve high-dimensional semantic vectors, facilitating rapid semantic search and similarity matching. The graph database uses Neo4j to store chemical knowledge graphs, a setup that supports complex logical reasoning and multi-hop relationship analysis.

[0119] II. Dual-engine collaborative retrieval mechanism:

[0120] This disclosure breaks through the limitations of traditional single-retrieval modes and innovatively constructs a dual-engine retrieval enhanced generation (RAG) system. It employs a "dual-engine driven" knowledge acquisition mechanism, deeply collaborating high-precision vector semantic retrieval (factual text fragments) with deep knowledge graph reasoning (e.g., the causal chain of "fluorine-containing monomer → reduced polarizability → decreased dielectric constant"). This integrated design not only focuses on the surface semantic matching of user queries but also strives to uncover the implicit logical connections (such as causality, attributes, hierarchy, etc.) between knowledge fragments, thereby fundamentally improving the depth, logical rigor, and scientific interpretability of the generated answers. This ensures that the conclusions are supported by extensive data and clearly demonstrate their inherent scientific reasoning paths.

[0121] The system deeply mines user query intent and performs multimodal knowledge retrieval. When a user submits a query (e.g., "How to reduce the dielectric constant of polyimide?"), two complementary and powerful search engines are launched in parallel:

[0122] The vector retrieval engine utilizes the advanced Gemini Embedding-001 embedding model to accurately map users' natural language queries into high-dimensional semantic vectors. This retrieval process performs an approximate nearest neighbor search on the vast Pinecone vector database, efficiently recalling the top 50 relevant literature fragments most similar to the query vector (e.g., paragraphs specifically discussing methods for synthesizing fluorinated monomers, experimental reports on dielectric constant optimization formulations, conclusions of studies on specific material modification, etc.), providing a massive amount of directly relevant and detailed factual fragments. These fragments typically contain specific experimental data, material formulations, process parameters, etc., serving as the fundamental source of facts for answering the question. Its advantages lie in its broad coverage, rapid recall, and ability to capture subtle semantic changes.

[0123] The knowledge graph reasoning engine utilizes the powerful Deepseek-R1 as a semantic parser to intelligently identify and extract core entities (e.g., "polyimide", "dielectric constant", "reduce") and implicit relationships (e.g., resolving "how to reduce" as "impact" or "optimization method") from user queries. The retrieval process uses the parsed entities and relationships as query clues to perform graph traversal in the structured Neo4j graph database. The engine not only finds directly related nodes but also explores multi-hop associations along relational edges (e.g., "has attributes", "affected by", "can be reduced by", "is a subclass of"), retrieving and matching up to 50 layers of adjacent entities and relationships around each entity. This constructs a knowledge subgraph of associations around the core entity (e.g., the mechanism of "specific crosslinking agent → increase free volume → reduce dielectric constant"). By providing deep logical networks and conceptual associations, it reveals the scientific principles, influencing factors, and potential solution paths behind the problem. Its advantages lie in understanding complex conceptual relationships, supporting reasoning, and discovering non-obvious knowledge connections.

[0124] The retrieved knowledge is intelligently integrated and structured. Deepseek-R1, acting as the brain of the retrieval system, receives and understands heterogeneous information from two retrieval channels: specific document fragments (rich in detailed data) from vector retrieval and structured relational networks (containing logical relationships) from the knowledge graph. Deepseek-R1 does not simply piece together information; instead, it performs deep semantic fusion, identifying evidence in document fragments that supports graph relationships, while utilizing the graph's logical framework to organize and interpret scattered facts within the documents. Finally, Deepseek-R1 synthesizes all the information to generate a clearly structured, logically rigorous, and fluently written answer. This answer not only directly addresses the user's question (e.g., listing specific methods to reduce the dielectric constant of polyimide) but also naturally incorporates the underlying scientific explanation (e.g., "Because fluorinated groups reduce molecular polarizability, therefore..." or "Studies have shown that increasing free volume can effectively reduce the dielectric constant, for example, by using a certain crosslinking agent...").

[0125] Traceability and credibility assurance mechanisms:

[0126] Literature citation (data credibility): For any specific performance data, experimental results, or key conclusions cited in the response, the retrieval system will clearly indicate their source. These sources are directly linked to the original literature fragments and their metadata (such as precise DOI number, paper title, author, journal / conference information) recalled during the vector retrieval phase.

[0127] Logical Verification (Reliable Reasoning): The retrieval system explicitly presents the key logical chains supporting the final conclusion, directly derived from the results of the knowledge graph retrieval. For example, in answering the question "Fluorine-containing monomers help reduce the dielectric constant of polyimide," the retrieval system can provide a verifiable causal path such as "Introducing fluorine-containing monomers → reducing molecular polarizability (increasing electron cloud displacement resistance) → leading to a decrease in the macroscopic dielectric constant." This provides transparent reasoning support for the conclusion. Users can examine this logical chain to determine whether each step of reasoning is reasonable and whether the evidence is sufficient, significantly enhancing the scientific rigor and verifiability of the answer.

[0128] Example 2

[0129] This disclosure provides a polymer synthesis scheme recommendation system 2000 based on a large language model. The polymer synthesis scheme recommendation system 2000 is used to predict the synthesis scheme of a second polymer.

[0130] Specifically, such as Figure 11 As shown, the polymer synthesis scheme recommendation system 2000 includes:

[0131] The text recognition module 2002 is used to recognize the first performance data to be queried based on the query text of the first performance.

[0132] The combination recommendation module 2004 is used to obtain a recommended monomer combination based on the first performance data to be queried, wherein the first predicted data of the first performance of the second polymer matches the first performance data to be queried, and the recommended monomer combination includes each of the second monomers in the second polymer and its molar ratio, and the mapping between the recommended monomer combination and the first predicted data is determined by the polymer performance prediction system 1000.

[0133] The data acquisition module 2006 is used to call a vector database to obtain first synthesis scheme data based on the recommended combination of entities, and to call a graph database to obtain second synthesis scheme data based on the recommended combination of entities. Each vector in the vector database is a vector representation of a synthesis experimental data fragment in the literature; the graph database is used to record entities in the literature and the relationships between those entities.

[0134] The scheme recommendation module 2008 is used to obtain a recommended synthesis scheme based on the first synthesis scheme data and the second synthesis scheme data.

[0135] Based on the aforementioned multimodal knowledge base and dual-engine RAG retrieval mechanism, as well as the polymer performance prediction system 1000, materials researchers can obtain targeted synthesis schemes through natural language queries using the polymer synthesis scheme recommendation system 2000, without any programming background, thus breaking down the technical barriers between computers and materials science.

[0136] To improve the usability of the polymer synthesis scheme recommendation system 2000, a locally running, visual web interface was developed, which can be used without relying on a cloud server. This interface allows users to efficiently input queries related to the target performance of polymers to filter recommended monomer combinations, and combines these with search results from vector databases and knowledge graphs to provide recommended synthesis schemes.

[0137] In some implementations, the target performance proposed by the user is a variety of primary performance characteristics. For each primary performance characteristic, the combination recommendation module 2004 makes a prediction using a pre-trained and fine-tuned polymer performance prediction system 1000. The recommended monomer combination is then derived by combining the prediction results from each system.

[0138] In some implementations, the combined recommendation module 2004 simultaneously predicts multiple first properties using a pre-trained and fine-tuned polymer performance prediction system 1000. This involves adjusting the output dimensions of the first decoder 1012 accordingly and training it.

[0139] The combined recommendation module 2004, through the pre-trained and fine-tuned polymer performance prediction system 1000, recommends performance based on the user-suggested target performance (e.g., "..."). T g >350℃, tensile strength >200MPa”, and other recommended monomer combinations were selected for polymerization synthesis of polymers with the target properties.

[0140] The target performance is the first performance predicted by the polymer performance prediction system 1000. The text recognition module 2002 identifies the first performance data to be queried based on the user's query.

[0141] The first predicted data of the first property of the second polymer matches the first property data to be queried, that is, the predicted data of the polymer performance prediction system 1000 falls into the first property data to be queried (e.g., T g Within the range of >350℃.

[0142] In some implementations, the combination recommendation module 2004 pre-constructs a mapping database between commercially available monomer combinations and first prediction data using the pre-trained and fine-tuned polymer performance prediction system 1000. Then, it retrieves recommended monomer combinations from this mapping database based on the first performance data required for query.

[0143] In some implementations, the data acquisition module 2006 is further configured to call a vector database to obtain first candidate combination data based on the first performance data to be queried, and to call a graph database to obtain second candidate combination data based on the first performance data to be queried.

[0144] The polymer synthesis scheme recommendation system 2000 also includes:

[0145] The combination filtering module is used to obtain the candidate single-unit combination based on the first candidate combination data and the second candidate combination data.

[0146] The combination recommendation module is used to obtain a recommended combination of individual entities based on the candidate combinations of the individual entities and the first performance data to be queried.

[0147] That is, the dual-engine RAG system, namely the data acquisition module 2006, filters out candidate monomer combinations. For example, as can be seen from the above analysis, the introduction of fluorine-containing monomers can reduce the dielectric constant, thereby narrowing the scope of the combination recommendation module 2004. Then, the combination recommendation module 2004 further filters and verifies the performance of these candidate monomer combinations.

[0148] Both candidate monomer combinations and recommended monomer combinations can be multiple.

[0149] In some embodiments, the polymer synthesis scheme recommendation system 2000 further includes:

[0150] A polymer property calculation engine is used to predict second predicted data for the first property of the second polymer, and the second predicted data corresponding to the recommended monomer combination matches the first property data to be queried.

[0151] The polymer property calculation engine used in this embodiment is the IBM polymer property calculation engine described by Giro et al. in the paper "npj Computational Materials" (2023).

[0152] As a complement to the internal model, the polymer property calculation engine is used for polymer performance prediction. The polymer property calculation engine employs the quantitative structure-property relationship (QSPR) method, utilizing topological indices, geometric variables, and other structural descriptors to predict polymer properties. The engine inputs the polymer corresponding to each recommended monomer combination for performance prediction, evaluating the prediction results of the polymer performance prediction system 1000. If the prediction results match, for example, if the deviation of the predicted values ​​is within a predetermined range, then the monomer combination is used to retrieve a synthesis scheme.

[0153] The data acquisition module 2006 retrieves relevant experimental scheme data from the vector database and the graph database according to the recommended monomer combination. The experimental scheme data includes reaction temperature, time, catalyst dosage, etc.

[0154] The synthesis scheme corresponds to the recommended monomer combination; that is, when there are multiple monomer combinations, the synthesis scheme corresponds to the number of monomer combinations.

[0155] Furthermore, the polymer synthesis scheme recommendation system 2000 also includes a hardware component, which includes a hardware instruction generation agent, an automated synthesis experimental device, and performance characterization equipment.

[0156] In some embodiments, the polymer synthesis scheme recommendation system further includes:

[0157] A synthesis module for synthesizing the second polymer according to a recommended synthesis scheme;

[0158] The testing module is used to test and obtain test data of the first property of the second polymer synthesized by the synthesis module.

[0159] The parameter optimization module is used to optimize the parameters of the polymer performance prediction system based on the test data of the second polymer and its first performance.

[0160] Once the complete synthesis path is planned, the hardware instruction generator, combined with the hardware instruction document, converts the complete synthesis scheme into hardware device instructions, which are then transmitted to the automated synthesis experimental device for experimental operation. The performance characterization device acquires feedback data in real time, compares it with the predicted value, and marks the deviation.

[0161] The polymer synthesis scheme recommendation system 2000 can achieve full-process optimization of polymer synthesis from planning to experiment through a closed loop of "AI planning - hardware execution - feedback optimization".

[0162] Closed-loop feedback and iterative optimization:

[0163] Data-driven retraining: Experimental data is accumulated and fed back to the polymer performance prediction system 1000, and the model weights are updated through incremental learning to correct performance prediction errors.

[0164] Bayesian optimization parameter tuning: Experimental parameters are dynamically adjusted through a Gaussian process to maximize the convergence efficiency of the target performance. The new experimental scheme can drive the hardware to conduct new experiments, and the synthesis efficiency and quality are improved in a cyclical manner.

[0165] The hardware architecture of the polymer synthesis scheme recommendation system 2000, implemented using a processor-based hardware approach, can be implemented using a bus architecture. The bus architecture can include any number of interconnect buses and bridges, depending on the specific application and overall design constraints of the hardware. Bus 2100 connects various circuits including one or more processors 2200, memory 2300, and / or hardware modules. Bus 2100 can also connect various other circuits 2400 such as peripheral devices, voltage regulators, power management circuits, external antennas, etc.

[0166] Bus 2100 can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Component (EISA) bus, etc. Bus 2100 can be divided into address bus, data bus, control bus, etc. For ease of representation, only one connection line is used in this diagram, but this does not indicate that there is only one bus or one type of bus.

[0167] Any process or method description in the flowcharts or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of preferred embodiments of this disclosure includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of this disclosure pertain. The processor performs the various methods and processes described above. For example, the method embodiments of this disclosure may be implemented as software programs tangibly contained in a machine-readable medium, such as memory. In some embodiments, part or all of the software program may be loaded and / or installed via memory and / or a communication interface. When the software program is loaded into memory and executed by the processor, one or more steps of the methods described above may be performed. Alternatively, in other embodiments, the processor may be configured to perform one of the methods described above by any other suitable means (e.g., by means of firmware).

[0168] The logic and / or steps represented in the flowchart or otherwise described herein may be specifically implemented in any readable storage medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-based system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0169] For the purposes of this specification, a "readable storage medium" can be any means capable of containing, storing, communicating, propagating, or transmitting a program for use in or in conjunction with an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of readable storage media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable read-only memory (CDROM). Furthermore, a readable storage medium can even be paper or other suitable media on which a program can be printed, since a program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in memory.

[0170] It should be understood that various parts of this disclosure can be implemented in hardware, software, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0171] Those skilled in the art will understand that all or part of the steps of the methods described above can be implemented by a program instructing related hardware. The program can be stored in a readable storage medium, and when executed, the program includes one or a combination of the steps of the method implementation.

[0172] Furthermore, the functional modules in the various embodiments of this disclosure can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a readable storage medium. The storage medium can be a read-only memory, a disk, or an optical disk, etc.

[0173] Those skilled in the art should understand that the above embodiments are merely for illustrating the present disclosure and are not intended to limit the scope of the disclosure. Those skilled in the art can make other changes or modifications based on the above disclosure, and these changes or modifications still fall within the scope of the present disclosure.

Claims

1. A dual-channel based polymer performance prediction system, characterized in that, The polymer property prediction system comprises: a parsing module configured to parse a SMILES string of a monomer to obtain a first feature vector and a second feature vector, wherein the first feature vector is configured to represent atomic features and bond features of the monomer, and the second feature vector is configured to represent global molecular features of the monomer; a first encoder configured to input the first feature vector and output a first embedding vector, wherein the first encoder adopts a graph neural network; a second encoder configured to input the second feature vector and output a second embedding vector, wherein the second encoder adopts a fully connected neural network; a feature splicing module configured to splice the first embedding vector and the second embedding vector into a monomer embedding vector; a feature fusion module configured to perform weighted summation on the monomer embedding vector according to a molar proportion coefficient of each monomer in the polymer to obtain a polymer embedding vector; a first decoder configured to input the polymer embedding vector and output a first property prediction vector, wherein the first decoder adopts a Transformer Decoder, and the first property prediction vector is configured to predict a first property of the polymer; the polymer property prediction system further comprises a second decoder configured to input a monomer embedding vector of a small molecule monomer and output a second property prediction vector, wherein the second decoder adopts a Transformer Decoder, and the second property prediction vector is configured to predict a quantum chemical property of the small molecule monomer; a training method of the first encoder and the second encoder, comprising: using a small molecule monomer dataset as a first training set, wherein each sample in the first training set comprises a SMILES string of a small molecule monomer and quantum chemical property data of the small molecule monomer; inputting the SMILES string of the small molecule monomer into the parsing module to obtain the first feature vector and the second feature vector of the small molecule monomer; inputting the first feature vector and the second feature vector of the small molecule monomer into the first encoder and the second encoder to be trained to obtain the first embedding vector and the second embedding vector of the small molecule monomer; splicing the first embedding vector and the second embedding vector of the small molecule monomer into a monomer embedding vector of the small molecule monomer through the feature splicing module; inputting the monomer embedding vector of the small molecule monomer into the second decoder to obtain a second property prediction vector; optimizing parameters of the first encoder and the second encoder according to the quantum chemical property data of the small molecule monomer and the second property prediction vector.

2. The system of claim 1, wherein, a training method of the first decoder, comprising: using a first polymer dataset as a second training set, wherein each sample in the second training set comprises a SMILES string of each first monomer in a first polymer, a molar proportion coefficient of the first monomer, and first property data of the first polymer, and the first polymer dataset is derived from public literature in a field related to synthesis of the polymer; inputting the SMILES string of the first monomer into the parsing module to obtain the first feature vector and the second feature vector of the first monomer respectively; inputting the first feature vector and the second feature vector of the first monomer into the trained first encoder and the trained second encoder respectively to obtain the first embedding vector and the second embedding vector of the first monomer; splicing the first embedding vector and the second embedding vector of the first monomer into a monomer embedding vector of the first monomer through the feature splicing module; performing weighted summation on the monomer embedding vector of the first monomer through the feature fusion module to obtain the polymer embedding vector of the first polymer; inputting the polymer embedding vector of the first polymer into the first decoder to be trained to obtain the first performance prediction vector of the first polymer; optimizing the parameters of the first decoder according to the first performance data and the first performance prediction vector of the first polymer.

3. The system of claim 1, wherein, Data augmentation is performed through a SMILES randomization strategy to generate multiple equivalent representation SMILES strings for the same monomer.

4. The system of claim 1, wherein, The graph neural network is a graph neural network based on a GATv2 layer.

5. The system of claim 1, wherein, The polymer is a polyimide.

6. The system of claim 1, wherein, The first performance is glass transition temperature and / or dielectric constant.

7. A large language model-based polymer synthesis protocol recommendation system, characterized by, The polymer synthesis scheme recommendation system is used to predict a synthesis scheme of a second polymer, and the polymer synthesis scheme recommendation system comprises: a text recognition module configured to recognize the first performance data to be queried according to query text of the first performance; a combination recommendation module configured to obtain a recommended monomer combination according to the first performance data to be queried, wherein first prediction data of the first performance of the second polymer matches the first performance data to be queried, the recommended monomer combination contains each second monomer in the second polymer and a molar ratio thereof, and the mapping between the recommended monomer combination and the first prediction data is determined by the polymer performance prediction system of any one of claims 1-6; a data acquisition module configured to call a vector database to obtain first synthesis scheme data according to the recommended monomer combination, and call a graph database to obtain second synthesis scheme data according to the recommended monomer combination, wherein each vector in the vector database is a vector representation of a synthesis experiment data segment in a literature; and the graph database is used to record entities in the literature and relationships between the entities; a scheme recommendation module configured to obtain a recommended synthesis scheme according to the first synthesis scheme data and the second synthesis scheme data.

8. The system of claim 7, wherein, The polymer synthesis scheme recommendation system further comprises: a polymer property calculation engine configured to predict second prediction data of the first performance of the second polymer, and the second prediction data corresponding to the recommended monomer combination matches the first performance data to be queried.

9. The system of claim 7, wherein, The polymer synthesis scheme recommendation system further comprises: a synthesis module configured to synthesize the second polymer according to the recommended synthesis scheme. a test module for testing test data of the first property of the second polymer resulting from the synthesis by the synthesis module, a parameter optimization module for optimizing parameters of the polymer property prediction system based on the second polymer and the test data of the first property thereof.

Images