Polyimide performance prediction method and system, polyimide structure design method and system, electronic equipment, computer readable storage medium and computer product

Abstract

The invention discloses a polyimide performance prediction method and system, a polyimide structure design method and system, electronic equipment, a computer readable storage medium and a computer product. The polyimide performance prediction method comprises the following steps: inputting a first molecular descriptor into an AFP model to obtain a to-be-measured physical property parameter; the first molecular descriptor is a molecular descriptor of polyimide to be detected; the AFP model is obtained through training of a training data set, and the training data set comprises a second molecular descriptor and a first physical property parameter; the second molecular descriptor and the first physical property parameter are a molecular descriptor and a physical property parameter of known polyimide respectively; each of the first molecular descriptor and the second molecular descriptor comprises a node feature and an edge feature, and the type of the first molecular descriptor is the same as that of the second molecular descriptor; the type of the to-be-measured physical property parameter is the same as that of the first physical property parameter. The method has high accuracy, and the effectiveness and reliability of de novo design are improved.

Description

Technical Field

[0001] The invention relates to a performance prediction method and system of polyimide, a structural design method and system, electronic equipment, a computer-readable storage medium and a computer product. Background Art

[0002] As a polymer material, polyimide (PI) is widely used in flexible display devices, optoelectronic device manufacturing, aerospace and automobile manufacturing due to its excellent mechanical properties, heat resistance and chemical stability. Polyimide films with high transparency and excellent comprehensive properties can improve the display quality of flexible display devices and enhance the efficiency and stability of optoelectronic devices.

[0003] However, finding such polyimides is a major challenge. Researchers have conducted a lot of work to further enhance the transparency and other properties of polyimide films. Methods such as introducing large groups, adding fluorinated groups, or introducing ring structures into polyimides have been applied to suppress the formation of charge transfer complexes (CTCs) and improve film transparency. Despite some progress, these methods also have drawbacks. Large groups lead to cumbersome and costly preparation processes, fluorinated groups reduce thermal stability and mechanical properties, and the introduction of ring structures may lead to poor temperature tolerance and insufficient rigidity, thus affecting the overall performance.

[0004] Therefore, researchers focus on rational design based directly on performance requirements. However, the chemical space of compounds is vast, reaching 10 60 (i.e., the number of compounds in the chemical space is nearly 10 to the power of 60) making traditional trial-and-error methods insufficient to provide material design solutions with excellent performance in a short period of time.

[0005] Currently, the design strategy based on forward design is widely used. However, this strategy is limited by the known material structure and properties, making it challenging to discover completely new material systems. In addition, forward design usually requires a lot of experimental and computational work, which is costly and time-consuming. The emergence of de novo design strategies provides a powerful means to address these challenges. The use of this strategy in material development starts from the desired material properties, enabling the discovery of new materials with target properties, providing a higher level of innovation and exploration.

[0006] In addition, considering the unique structure of polyimide (containing imide groups), it brings additional challenges to de novo design, that is, to design molecules with specific structures and specific properties at the same time. Summary of the invention

[0007] In order to overcome the defects in the prior art that it is difficult to quickly predict the performance parameters of polyimide, and the method of obtaining the polyimide structure is costly and time-consuming; the present invention provides a performance prediction method and system of polyimide (PI), a structural design method and system, an electronic device, a computer-readable storage medium and a computer product. The performance prediction method of the present invention uses specific deep learning to make the prediction results of performance have a higher accuracy rate; in addition, the present invention adopts a fragment-based design architecture to improve the efficiency of polyimide design, and combines the proximal policy optimization (PPO) in reinforcement learning for training, so that an intelligent agent with target performance polyimide can be designed from scratch; further improving the effectiveness and reliability of the de novo design.

[0008] The present invention solves the above technical problems through the following technical solutions.

[0009] The present invention provides a method for predicting the performance of polyimide, which comprises the following steps:

[0010] Inputting the first molecular descriptor into the AFP model to obtain the physical property parameter to be measured; the first molecular descriptor is the molecular descriptor of the polyimide to be measured;

[0011] The AFP model is obtained by training a training data set, wherein the training data set includes a second molecular descriptor and a first physical property parameter; the second molecular descriptor and the first physical property parameter are respectively the molecular descriptor and the physical property parameter of a known polyimide;

[0012] The first molecular descriptor and the second molecular descriptor both include node features and edge features, and the type of the first molecular descriptor is the same as the type of the second molecular descriptor;

[0013] The type of the physical property parameter to be measured is the same as the type of the first physical property parameter.

[0014] In the present invention, polyimide is a complex polymer with different types and structures. In the actual research process, since the experimental data of polyimide is relatively small and insufficient, the data set is relatively sparse, which makes model training more difficult. Therefore, for those skilled in the art, constructing a model that can accurately predict the performance of polyimide is a difficult technical problem to solve.

[0015] In the present invention, the full name of the AFP model is Attention Fingerprint. The Attention Fingerprint (AFP) graph neural network regards the chemical structure as a graph, in which atoms are represented as nodes and bonds are represented as edges, encoding direct and extended connections between atoms. In order to distinguish different nodes and edges (i.e., atoms and bonds), it is necessary to assign feature vectors (such as: atom type, valence, bond type, etc.) to these nodes and edges. Through the attention mechanism, the fingerprint obtained from the node-level and graph-level embedding is introduced, called the attention fingerprint (AFP), which can learn non-local molecular interactions. Due to its ability to learn the intrinsic connections in chemical structures, AFP exhibits excellent performance in building models.

[0016] In the present invention, "the type of the first molecular descriptor is the same as the type of the second molecular descriptor" means that the type of the second molecular descriptor corresponds to the type of the first molecular descriptor; for example, when the second molecular descriptor includes a chiral node feature, the first molecular descriptor also includes a chiral node feature; when the second molecular descriptor includes all feature vectors, the first molecular descriptor also includes the same feature vectors.

[0017] In the present invention, those skilled in the art should know that the first physical property parameter in the training data set is generally collected from public documents.

[0018] In some embodiments, the type of the physical property parameter to be measured and the type of the first physical property parameter are each independently a light transmittance (T 400 ), coefficient of linear thermal expansion (CLTE), tensile strength (σ b ) or glass transition temperature (T g ). These physical property parameters are extracted from publicly available literature. Generally, the amount of data used for deep learning is in the thousands or even tens of thousands, while the amount used for traditional machine learning is at least dozens.

[0019] Among them, the light transmittance performance T 400 Refers to the transmittance at 400nm, the linear thermal expansion coefficient CLTE to reflect dimensional stability, and the tensile strength σ b To reflect the mechanical properties, glass transition temperature T g To reflect the heat resistance.

[0020] In the present invention, light transmittance, dimensional stability, mechanical properties and heat resistance are all performance indicators of different dimensions, and the measurement method, unit and numerical range of each indicator are different. Data from different sources may have measurement errors, inconsistent data and other problems (the mean filling method is used to handle missing values ​​to ensure data integrity).

[0021] In the present invention, the molecular descriptor is obtained through RDKit or Mordred, which is used to represent the mathematical description of the chemical molecular structure; wherein, RDKit (Research Data Kit) is an open source chemical informatics toolkit; Mordred is a special descriptor calculation software.

[0022] In some embodiments, the node characteristics are one or more of atom type, number of bond connections, number of hydrogen atom connections, displayed valence, hybridization, aromaticity, chiral center, chiral type, and formal charge.

[0023] Among them, the atom type can be a unique hot encoding of "one or more of C, N, O, F, P, S, Cl, Br, I" as a node feature.

[0024] The number of key connections may be encoded as "0, 1, 2, 3 or 4" as a node feature.

[0025] The number of hydrogen atom connections may be encoded as "0, 1, 2, 3 or 4" as a node feature.

[0026] The displayed valence may be encoded as a node feature using a unique hot encoding of "0, 1, 2, 3, 4 or 5".

[0027] Wherein, the hybridization can be "sp, sp 2 、sp 3 、sp 3 d or sp 3 d 2 " is used as the node feature.

[0028] The aromaticity can be "0 or 1" as a node feature to indicate whether the atom is part of an aromatic system.

[0029] The chiral center can be "0 or 1" as a node feature to indicate whether the atom is a chiral center.

[0030] The chirality type may be "R or S" as a node feature to indicate the chirality type of the atom involved.

[0031] Wherein, when the formal charge is used as a node feature, it can be an integer to represent the charge assigned to a single atom in a molecule.

[0032] In some embodiments, the edge characteristics are one or more of bond type, conjugation, ring structure, and stereo bond.

[0033] The bond type may be a unique hot code of "single bond, double bond, triple bond or aromatic bond" as a node feature.

[0034] The conjugation condition can be "0 or 1" as a node feature to indicate whether the bond is conjugated.

[0035] The ring structure may use "0 or 1" as a node feature to indicate whether a key is part of the ring.

[0036] The stereo bond may be encoded as a node feature using a one-hot encoding of "none, any, Z / E or cis / trans".

[0037] In some embodiments, the model parameters of the AFP model include hidden layer dimension, initial learning rate, learning rate decay factor, weight decay, dropout rate, number of node-level embedding layers, or number of graph-level embedding layers.

[0038] Preferably, the hidden layer dimension is 16-128, for example, 87, 97, 111 or 116.

[0039] Preferably, the initial learning rate is 10 -3 -10 -1 , for example, 10 -3 , 10 -2.26 or 10 -1.92 .

[0040] Preferably, the learning rate attenuation factor is 0.4-0.9, for example, 0.6, 0.8 or 0.9.

[0041] Preferably, the weight decay is 0-0.4, for example, 0.

[0042] Preferably, the discard rate is 0-0.4, for example, 0.1, 0.2, 0.25 or 0.35.

[0043] Preferably, the number of node-level embedding layers is 1-6, such as 1, 2 or 3.

[0044] Preferably, the number of graph-level embedding layers is 1-4, such as 1, 2 or 4.

[0045] In one embodiment, when the physical property parameter is the transmittance T 400 When the hidden layer dimension is 116, the initial learning rate is 10 -3 , the learning rate decay factor is 0.90, the weight decay is 0, the dropout rate is 0.25, the number of node-level embedding layers is 3, and the number of graph-level embedding layers is 2.

[0046] In one embodiment, when the physical property parameter is the coefficient of linear thermal expansion CTLE, the hidden layer dimension is 87 and the initial learning rate is 10 -1.92, the learning rate decay factor is 0.60, the weight decay is 0, the dropout rate is 0.35, the number of node-level embedding layers is 2, and the number of graph-level embedding layers is 1.

[0047] In one embodiment, when the physical property parameter is tensile strength σ b When the hidden layer dimension is 97, the initial learning rate is 10 -2.26 , the learning rate decay factor is 0.80, the weight decay is 0, the dropout rate is 0.20, the number of node-level embedding layers is 3, and the number of graph-level embedding layers is 4.

[0048] In one embodiment, when the physical property parameter is the glass transition temperature T g When the hidden layer dimension is 111, the initial learning rate is 10 -3 , the learning rate decay factor is 0.90, the weight decay is 0, the dropout rate is 0.10, the number of node-level embedding layers is 1, and the number of graph-level embedding layers is 4.

[0049] In some embodiments, the first molecular descriptor and the second molecular descriptor are generated by RDKit or Mordred, respectively.

[0050] In some embodiments, the first molecular descriptor and the second molecular descriptor are generated in the same manner.

[0051] In some embodiments, the training data set is cleaned to exclude outliers.

[0052] In a specific embodiment, the cleaning process includes:

[0053] i. Excluding polyimide containing metal elements and / or ionic compounds; the polyimide is an organic substance;

[0054] ii. Exclude polyimide copolymers and blends and select only homopolymers; the homopolymer here refers to a polymer composed entirely of one polyimide monomer, that is, a dianhydride and a diamine are polymerized.

[0055] iii. When polyimide of the same structure has multiple data records of the same physical property parameter, the average value of each data record is taken.

[0056] In the present invention, the AFP model can extract the complex chemical structure characteristics of polyimide, thereby improving the accuracy of performance prediction; the AFP model is a graph neural network model of the graph attention mechanism framework, which belongs to a deep learning method; generally, under the premise that the data set is large enough, the prediction effect of the deep learning model is better.

[0057] Among them, by studying the influence of different models on the performance prediction results, it can be seen that compared with other deep learning methods (such as FraGAT model or BERT model), the AFP model has a better performance prediction effect on polyimide. Compared with traditional machine learning methods (such as multi-layer perceptron, which involves feature engineering processing), the AFP model has a better performance prediction effect on polyimide.

[0058] In the present invention, the structural information of the polyimide to be tested and the known polyimide can be obtained by converting the repeating units therein into corresponding SMILES. SMILES, the full name of which is Simplified molecular input line entry specification, is a specification that clearly describes the molecular structure using an ASCII string.

[0059] The present invention also provides a performance prediction system for polyimide, which includes a prediction module, which is used to input a first molecular descriptor into an AFP model to obtain a physical property parameter to be measured; the first molecular descriptor is a molecular descriptor of the polyimide to be measured; the AFP model is obtained by training a training data set, and the training data set includes a second molecular descriptor and a first physical property parameter; the second molecular descriptor and the first physical property parameter are respectively the molecular descriptor and the physical property parameter of a known polyimide; the first molecular descriptor and the second molecular descriptor both include node features and edge features, and the type of the first molecular descriptor is the same as the type of the second molecular descriptor; the type of the physical property parameter to be measured is the same as the type of the first physical property parameter.

[0060] In some embodiments, the prediction module is also used in the performance prediction method as described above.

[0061] The present invention also provides a structural design method of polyimide, which comprises: training and generating an intelligent agent of polyimide based on proximal policy optimization (PPO);

[0062] The proximal strategy optimization comprises the following steps:

[0063] S1, fragmenting the dianhydride molecules and the diamine molecules to obtain a dianhydride structure group containing a plurality of dianhydride fragments and a diamine structure group containing a plurality of diamine fragments;

[0064] S2, selecting a dianhydride starting fragment and a diamine starting fragment from the dianhydride structure group and the diamine structure group, respectively;

[0065] S3, according to the strategy of the neural network, selecting an action, the action comprising: obtaining a dianhydride selected fragment and a diamine selected fragment from the dianhydride structure group and the diamine structure group respectively through the dianhydride starting fragment and the diamine starting fragment in step S2, and obtaining a dianhydride synthesis fragment and a diamine synthesis fragment;

[0066] S4, judging whether the intelligent body of the polyimide can be generated by the dianhydride synthesis fragment and the diamine synthesis fragment; when the intelligent body of the polyimide can be generated, using the AFP model in the performance prediction method as described above or the AFP model in the performance prediction system as described above, calculating the physical property parameters of the intelligent body of the polyimide, and further calculating the reward value; when the intelligent body of the polyimide cannot be generated, returning to step S2 for loop iteration;

[0067] S5. Terminate the proximal strategy optimization step.

[0068] In the present invention, the above-mentioned structural design method can avoid constructing a candidate library of polyimide in advance; however, the disadvantage of constructing a candidate library of polyimide in advance is that the candidate library constructed artificially is inevitably limited and cannot include all structures in the chemical space.

[0069] In the present invention, polyimide that meets the conditions can be directly designed according to the target performance; this can be achieved by the proximal strategy optimization in reinforcement learning.

[0070] In the present invention, the designed target molecules can be decomposed into common chemical groups through RDKit as a candidate gene library for de novo design.

[0071] In some embodiments, in step S1, the fragmentation method includes a decomposition method for back-tracing the synthesis of interesting chemical substructures.

[0072] In some embodiments, in step S2, the selection method includes: extracting groups from the dianhydride molecules and the diamine molecules, and converting them into computer-recognizable 600-dimensional vectors through a polyBERT model.

[0073] In some embodiments, in step S2, the selection rule of the dianhydride starting fragment and the diamine starting fragment is: according to the occurrence frequency of the starting fragment, the dianhydride starting fragment and the diamine starting fragment with the highest occurrence frequency are selected as the only starting fragment.

[0074] In step S3 of the present invention, the method of "respectively obtaining dianhydride selected fragments and diamine selected fragments from the dianhydride structure group and the diamine structure group" may include: assigning different selection probabilities to all fragments in the dianhydride structure group and the diamine structure group based on the strategy of the neural network, and the strategy of the neural network will preferentially select fragments with large selection probabilities through the starting fragments of dianhydride and diamine in step S2. In the process of training the intelligent agent, the strategy of the neural network will continuously iteratively update the selection probability of each fragment.

[0075] In some implementations, in step S3, the model architecture of the strategy is a multi-layer perceptron.

[0076] In some implementations, in step S3, the number of hidden layers of the strategy is 1-5.

[0077] In some embodiments, in step S3, the number of hidden layer neurons of the strategy is 16-512, preferably 16, 32, 64, 128, 256 or 512.

[0078] In some implementations, in step S3, the hidden layer of the strategy is a fully connected layer.

[0079] In some implementations, in step S3, the activation function of the strategy is Relu, LeakyRelu or Tanh, preferably Relu.

[0080] In some embodiments, in step S3, the method for obtaining the dianhydride synthesis fragment includes: when the dianhydride starting fragment and the dianhydride selected fragment both contain a connector, splicing the dianhydride starting fragment and the dianhydride selected fragment to form the dianhydride synthesis fragment; when the dianhydride starting fragment contains a connector and the dianhydride selected fragment does not contain a connector, using the dianhydride selected fragment as the dianhydride synthesis fragment; when the dianhydride starting fragment does not contain a connector, using the dianhydride starting fragment as the dianhydride synthesis fragment.

[0081] In some embodiments, in step S3, the method for obtaining the diamine synthesis fragment includes: when the diamine starting fragment and the diamine selection fragment both contain a connector, splicing the diamine starting fragment and the diamine selection fragment to form the diamine synthesis fragment; when the diamine starting fragment contains a connector and the diamine selection fragment does not contain a connector, using the diamine selection fragment as the diamine synthesis fragment; when the diamine starting fragment does not contain a connector, using the diamine starting fragment as the diamine synthesis fragment.

[0082] In the present invention, after obtaining the predicted performance of the candidate polyimide agent, a comprehensive score can be obtained by weighted average, and the comprehensive score is used as a basis for customizing rewards in reinforcement learning.

[0083] In some embodiments, in step S4, the physical property parameters of the polyimide agent include light transmittance, linear thermal expansion coefficient, tensile strength and glass transition temperature; preferably, the expression of the reward value is R=R T400 ×(1+R CLTE +R σb +R Tg +R SAscore ) / 5; where R T400 , R CLTE , R σb , R Tg , R SAscore They represent the 400nm transmittance, linear thermal expansion coefficient, tensile strength, glass transition temperature, and normalized reward value of SAscore respectively.

[0084] The reward value is used to quantify the benefits obtained by the agent after performing a specific action or making a decision. In this field, the reward value is usually used to guide the learning process of the agent, so that it can identify which behaviors are better in a specific situation. In this application, the calculation formula of the reward value combines multiple physical performance parameters to evaluate the overall performance of the polyimide material. The influence or role of the reward value on the structural design results is reflected in:

[0085] (1) Guiding the search direction: The reward value provides an evaluation criterion. The agent can adjust its search strategy according to the size of the reward value to discover polyimide structures with better performance.

[0086] (2) Optimizing decisions: The agent will continuously try different combinations of material parameters during the structural design process. The feedback of reward values ​​can help the agent learn which combinations are more likely to lead to high-performance materials.

[0087] (3) Balanced performance: Since the reward value is a combination of multiple parameters, the intelligent agent will not only pursue the maximization of a single performance, but also seek to find the best balance between multiple performance indicators.

[0088] In some embodiments, in step S5, the termination condition of the proximal strategy optimization is any one of the following conditions: a. the intelligent agent of the polyimide is generated and the reward value is calculated; b. in step S3, the dianhydride synthesis fragment does not meet the generation condition of containing two amide groups; or, the diamine synthesis fragment does not meet the generation condition of containing two amine groups; c. the number of atoms of the intelligent agent of the polyimide exceeds 50, preferably exceeds 120; d. in step S4, the number of loop iterations exceeds 5, preferably exceeds 10.

[0089] In the present invention, the structural design method of polyimide belongs to de novo design, which means starting from the fragments of dianhydride and diamine respectively, so as to improve the effectiveness of polyimide design; the de novo design scheme based on fragments can ensure 100% design effectiveness. The difficulty of this design is greater than the forward design of the prior art, because the de novo design needs to directly design a structure that meets the conditions according to the target performance, and it is necessary to train an intelligent agent or model with sufficient strength; while the forward design of the prior art only needs to traverse and predict the candidate library based on the performance prediction model, and then screen it.

[0090] The present invention also provides a polyimide structural design system, which includes a structural design module for training and generating an intelligent agent of polyimide based on proximal strategy optimization;

[0091] The proximal strategy optimization comprises the following steps: S1, fragmenting dianhydride molecules and diamine molecules to obtain a dianhydride structure group containing a plurality of dianhydride fragments and a diamine structure group containing a plurality of diamine fragments; S2, selecting a dianhydride starting fragment and a diamine starting fragment from the dianhydride structure group and the diamine structure group, respectively; S3, selecting an action according to the strategy of the neural network, the action comprising obtaining a dianhydride selected fragment and a diamine selected fragment from the dianhydride structure group and the diamine structure group, respectively, through the dianhydride starting fragment and the diamine starting fragment in step S2, and obtaining a dianhydride synthesis fragment and a diamine synthesis fragment; S4, judging whether the intelligent agent of the polyimide can be generated through the dianhydride synthesis fragment and the diamine synthesis fragment; when the intelligent agent of the polyimide can be generated, using the AFP model in the performance prediction method as described above or the AFP model in the performance prediction system as described above to calculate the physical property parameters of the intelligent agent of the polyimide; when the intelligent agent of the polyimide cannot be generated, returning to step S2 for loop iteration; S5, terminating the step of the proximal strategy optimization.

[0092] In some embodiments, the structural design module is also used in the structural design method of polyimide as described above.

[0093] The present invention also provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned polyimide structural design method when executing the computer program.

[0094] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the structure design method of polyimide as described above is implemented.

[0095] The present invention also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the structure design method of polyimide as described above is implemented.

[0096] In the present invention, a polyimide de novo design strategy based on deep learning and reinforcement learning is used to efficiently discover polyimides with specific properties, such as higher transparency (or lower transparency), lower linear thermal expansion coefficient, higher tensile strength and higher glass transition temperature.

[0097] In an embodiment of the present invention, by comparing four machine learning models (a neural network (multi-layer perceptron), a natural language model (BERT model) and two graph neural network models (AFP model, FraGAT model)), it is found that the AFP graph neural network model of the present invention has better overall performance, and the most suitable performance estimator is determined.

[0098] On the basis of being in accordance with the common sense in the art, the above-mentioned preferred conditions can be arbitrarily combined to obtain the preferred embodiments of the present invention.

[0099] The reagents and raw materials used in the present invention are commercially available.

[0100] The positive and progressive effects of the present invention are:

[0101] (1) The prediction method of the present invention has a high prediction effect and prediction efficiency for the performance of polyimide; the prediction method is preferably suitable for predicting the light transmittance, linear thermal expansion coefficient, tensile strength and glass transition temperature of polyimide.

[0102] (2) The present invention is based on proximal strategy optimization in reinforcement learning, and trains a polyimide agent designed from scratch. The starting point of the present invention is preferably to design a polyimide with excellent comprehensive performance. Further, by using training, it is preferred to obtain a PI agent with high transparency (or low transparency), low linear thermal expansion coefficient, high tensile strength and high glass transition temperature.

[0103] (3) Compared with the forward design strategy, the method of the present invention avoids the prediction of a large number of candidate structures. It can automatically generate potential candidate structures with target performance in the vast PI chemical space, saving computing resources and computing time, and greatly improving the design efficiency of PI. BRIEF DESCRIPTION OF THE DRAWINGS

[0104] Figure 1 A flow chart showing the molecular diagram of the polyimide repeating unit structure and the model construction of different physical property parameters in Example 1;

[0105] Figure 2Flowchart for agent training of Example 1 and de novo design of polyimide with target performance by the agent.

[0106] Figure 3 It is a data distribution diagram of each physical property parameter data set in Example 1; Figure 3 (a) is the data distribution diagram of the light transmittance performance data set. Figure 3 (b) is the data distribution diagram of the linear thermal expansion coefficient data set. Figure 3 (c) is the data distribution diagram of the tensile strength data set. Figure 3 (d) is the data distribution diagram of the glass transition temperature data set.

[0107] Figure 4 The figure is a test effect diagram of the AFP model in Example 1 for different physical property parameters; Figure 4 (a) AFP model for light transmittance performance T 400 The test effect diagram, Figure 4 (b) is the test effect diagram of the AFP model for the linear thermal expansion coefficient CLTE. Figure 4 (c) AFP model for tensile strength σ b The test effect diagram, Figure 4 (d) AFP model for glass transition temperature T g Test effect diagram.

[0108] Figure 5 This is a comparison chart of the training results of different models under different performances in Example 1; Figure 5 (a) The light transmittance T of different models 400 Comparison chart of training results under Figure 5 (b) is a comparison chart of the training results of different models under the linear thermal expansion coefficient CLTE. Figure 5 (c) The tensile strength σ of different models b Comparison chart of training results under Figure 5 (d) The glass transition temperature T of different models g Comparison chart of the training results below.

[0109] Figure 6 Flowchart for training the reinforcement learning agent designed from scratch for polyimide in Example 2.

[0110] Figure 7 This is a design effect diagram of the reinforcement learning polyimide agent in effect embodiment 1; Figure 7 (a) is the reward dynamics (including maximum, minimum and average) of the agent in the entire training process of Example 1. Figure 7 (b) is the light transmittance performance T of the polyimide smart body 400 Design renderings, Figure 7(c) is the design effect diagram of the linear thermal expansion coefficient CLTE of the polyimide intelligent body. Figure 7 (d) is the tensile strength σ of the polyimide smart body b Design renderings, Figure 7 (e) is the glass transition temperature T of the polyimide smart body g Design renderings, Figure 7 (f) is a diagram showing the design effect of the synthetic accessibility SAscore of the polyimide agent.

[0111] Figure 8 This is the distribution of polyimide agents in the chemical space at different stages in Effect Example 1.

[0112] Fig. 9 The number of clusters of polyimide designed by the agent after 70 trainings is determined according to the silhouette score and inertia in Example 1.

[0113] Fig.10 This is the chemical spatial distribution of polyimide after being clustered into 6 clusters by K-means in Effect Example 1.

[0114] Fig.11 The polyimide with the highest reward value in clusters 1-6 in effect example 1 and its various properties.

[0115] Fig.12 is the chemical structural formula of the polyimide designed from scratch for the intelligent agent in Effect Example 1.

[0116] Fig.13 It is a polyimide film designed from scratch for the intelligent body in Effect Example 1.

[0117] Fig.14 The experimental values ​​and predicted values ​​of the thin film synthesized experimentally for the polyimide designed for the intelligent agent in Effect Example 1 are compared.

[0118] Fig.15 It is the fragment frequency obtained by decomposing dianhydride by the BRICS method in Effect Example 1.

[0119] Fig.16 This is the fragment frequency obtained by decomposing the diamine by the BRICS method in Effect Example 1.

[0120] Fig.17 Effect diagram of the design effect of the reinforcement learning polyimide agent in Example 2; Fig.17 (a) is the reward dynamics (including maximum, minimum and average values) of the agent in the entire training process of the effect embodiment 2. Fig.17 (b) is the light transmittance performance T of the polyimide smart body 400 Design renderings, Fig.17(c) is the design effect diagram of the linear thermal expansion coefficient CLTE of the polyimide intelligent body. Fig.17 (d) is the tensile strength σ of the polyimide smart body b Design renderings, Fig.17 (e) is the glass transition temperature T of the polyimide smart body g Design renderings, Fig.17 (f) is a diagram showing the design effect of the synthetic accessibility SAscore of the polyimide agent.

[0121] Fig.18 This is the distribution of polyimide agents in chemical space at different stages in effect example 2.

[0122] Fig.19 The number of clusters of polyimide designed by the agent after 70 trainings is determined according to the silhouette score and inertia in Example 2.

[0123] Fig. 20 This is the chemical spatial distribution of polyimide after being clustered into three clusters by K-means in Effect Example 2.

[0124] Fig.21 The polyimide with the highest reward value in clusters 1-3 in effect example 2 and its various properties. DETAILED DESCRIPTION

[0125] The present invention is further described below by way of examples, but the present invention is not limited to the scope of the examples. The experimental methods in the following examples without specifying specific conditions are carried out according to conventional methods and conditions, or selected according to the product specifications.

[0126] Example 1

[0127] This embodiment discloses a method for predicting the performance of polyimide. Figure 1 A flow chart showing the molecular diagram of the polyimide repeating unit structure and the model construction of different physical property parameters in Example 1; Figure 2 Flowchart for agent training of Example 1 and de novo design of polyimide with target performance by the agent.

[0128] The performance prediction method comprises the following steps:

[0129] ss1. Collect the light transmittance properties (T) of polyimide (PI) films from 137 public documents. 400 ), coefficient of linear thermal expansion (CLTE), tensile strength (σ b ) and glass transition temperature (T g ) are taken as independent data sets. Figure 3It is a data distribution diagram of each physical property parameter data set in this embodiment; Figure 3 (a) is the data distribution diagram of the light transmittance performance data set. Figure 3 (b) is the data distribution diagram of the linear thermal expansion coefficient data set. Figure 3 (c) is the data distribution diagram of the tensile strength data set. Figure 3 (d) is the data distribution diagram of the glass transition temperature data set; the data volume of the four physical property parameters collected is as follows Figure 3 As shown, they are: 389 items for light transmittance, 454 items for linear thermal expansion coefficient, 764 items for tensile strength, and 946 items for glass transition temperature. In each of the above data sets, the repeating unit structure of PI is stored through SMILES, and the performance data is recorded in floating point format. Due to the non-uniqueness of SMILES in representing the asymmetry of complex polymer structures (caused by the asymmetry of dianhydrides or diamines), for data enhancement, multiple isomers of some PIs are included in the data.

[0130] ss2. The descriptor feature generation of molecules is completed through RDKit, and some inappropriate descriptors are removed. For example, some descriptors may only be applicable to certain systems or specific situations and cannot generate effective features for the vast majority of samples. In addition, the molecular graph information containing features such as atomic types and bonding information in molecules is implemented through Python scripts written using conventional methods.

[0131] ss3. The molecular structure diagram of polyimide obtained by the above Python script is converted into a vector, and the node (atom) features and edge (bond) features in the process of using them as molecular descriptors are shown in Tables 1 and 2; Table 1 is the information of node (atom) features, and Table 2 is the information of edge (bond) features.

[0132] Table 1

[0133] feature describe size Atom Type Atom type (C,N,O,F,P,S,Cl,Br,I) (one-hot encoding) 9 Number of key connections The number of bonds to which the atom is connected (0,1,2,3,4) (one-hot encoding) 5 Number of hydrogen atoms connected The number of hydrogen atoms the atom is connected to (0,1,2,3,4) (one-hot encoding) 5 Display valence Display valence (0,1,2,3,4,5) (one-hot encoding) 6 Hybridization <![CDATA[Hybridization (sp, sp 2 , sp 3 , sp 3 d, sp 3 d 2 )(One - hot encoding)]]> 5 Aromatic Whether the atom is part of an aromatic system (0 or 1) 1 Chiral Center Whether the atom is a chiral center (0 or 1) 1 Chiral type The chirality of the atom (R or S) 2 Formal charge The charge assigned to a single atom in a molecule (integer) 1

[0134] Table 2

[0135] feature describe size Key Type Bond type (single bond, double bond, triple bond, aromatic bond) (one-hot encoding) 4 Conjugation Whether the bond is conjugated (0 or 1) 1 ring Whether the bond is part of a ring (0 or 1) 1 Stereo Key Bond stereochemistry (none, any, Z / E, cis / trans) (one-hot encoding) 6

[0136] ss4. A machine learning model of AFP was constructed on the light transmittance, linear thermal expansion coefficient, tensile strength, and glass transition temperature datasets. The model performance evaluation mainly focuses on the model's generalization ability to unseen PIs, the prediction ability for the entire PI dataset, and the training stability of the model. The dataset was randomly divided and kept consistent across all models to provide a common basis for comparing model performance. The ratio of training set, validation set, and test set is 80:10:10. The evaluation indicators mainly include mean absolute error (MAE), root mean square error (RMSE), and determination coefficient (R2).

[0137] Among them, the light transmittance performance (T 400 ), coefficient of linear thermal expansion (CLTE), tensile strength (σ b ) and glass transition temperature (T g The AFP model parameters constructed by the ) dataset are shown in Table 3 below:

[0138] Table 3

[0139] Model parameters <![CDATA[T 400 ]]> CLTE <![CDATA[σ b ]]> <![CDATA[T g ]]> Hidden layer dimensions 116 87 97 111 Initial learning rate <![CDATA[10 -3 ]]> <![CDATA[10 -1.92 ]]> <![CDATA[10 -2.26 ]]> <![CDATA[10 -3 ]]> Learning rate decay factor 0.90 0.60 0.80 0.90 Weight decay 0 0 0 0 Discard rate 0.25 0.35 0.20 0.10 Number of node-level embedding layers 3 2 3 1 Number of graph-level embedding layers 2 1 4 4

[0140] The descriptors to be tested are input into the AFP model to obtain the parameters of the physical properties to be tested. For the evaluation of the entire data set, no retraining is performed, and the evaluation is performed only on the complete data set. The training stability of the model is evaluated after repeated training 50 times under different initial parameter conditions.

[0141] This embodiment also discloses a polyimide performance prediction system, which includes a prediction module for the above-mentioned performance prediction method.

[0142] Comparative Example 1

[0143] This comparative example uses the FraGAT model, polyBERT model, and descriptor + multi-layer perceptron (Des+ANN) model to predict the performance of polyimide.

[0144] Steps ss1 and ss2 of the performance prediction method are the same as those in Example 1, the data set used is the same as that used in Example 1, and the division criteria of the training set, the validation set, and the test set are the same as those in Example 1; the method further comprises the following steps:

[0145] ss3. Feature engineering is used to select relevant features generated in ss2. First, correlation analysis is used to reduce the thousands of descriptors generated to 141, 158, 145, and 150, respectively. Second, Ridge regression is used to further reduce the feature dimension to achieve the best model performance. Ridge regression is a linear model that uses an L2 regularization term to reduce the complexity of the model, thereby compressing the number of features of the model. During the dimensionality reduction process, the performance of each attempt is evaluated based on the results of 5-fold cross-validation. After feature dimensionality reduction through Ridge regression, the descriptors corresponding to the light transmittance, linear thermal expansion coefficient, tensile strength, and glass transition temperature data sets are reduced to 52, 66, 57, and 63, respectively. The best feature combination obtained after feature engineering is used to build a multi-layer perceptron model.

[0146] ss4. Three machine learning models, FraGAT, polyBERT, and descriptor + multi-layer perceptron, were constructed on the light transmittance, linear thermal expansion coefficient, tensile strength, and glass transition temperature data sets respectively.

[0147] The model parameters in FraGAT, polyBERT, and Descriptor + Multilayer Perceptron (Des+ANN) were evaluated on each dataset using the default values.

[0148] Application Example 1

[0149] In this application example, the AFP model trained in the performance prediction method of Example 1 and the descriptor + neural network model (Des + ANN) trained in the performance prediction method of Comparative Example 1 are used to test the training set, validation set and test set of polyimide (PI) divided in Example 1, and the results of different models on different physical parameters (light transmittance T 400 , linear thermal expansion coefficient CLTE, tensile strength σ b , glass transition temperature T g )’s prediction results.

[0150] Figure 4 The figure is a test effect diagram of the AFP model in Example 1 for different physical property parameters; Figure 4 (a) AFP model for light transmittance performance T 400 The test effect diagram, Figure 4 (b) is the test effect diagram of the AFP model for the linear thermal expansion coefficient CLTE. Figure 4 (c) AFP model for tensile strength σ b The test effect diagram, Figure 4 (d) AFP model for glass transition temperature T g Test effect diagram.

[0151] For the training set, the transmittance performance T 400 , its MAE and RMSE are 2.41% and 3.52% respectively, R 2 The MAE and RMSE of the linear thermal expansion coefficient CLTE are 2.83 ppm·K. -1 and 4.19ppm·K -1 , R 2 is 0.95. For the tensile strength σ b , its MAE and RMSE are 4.34MPa and 6.13MPa respectively, R 2 is 0.97. For the glass transition temperature T g , its MAE and RMSE are 3.09℃ and 4.78℃ respectively, R 2 is 0.99.

[0152] For the validation set, the transmittance performance T 400 , its MAE and RMSE are 3.89% and 3.52% respectively, R 2The MAE and RMSE of the linear thermal expansion coefficient CLTE are 4.19 ppm·K. -1 and 5.97ppm·K -1 , R 2 is 0.94. For the tensile strength σ b , its MAE and RMSE are 10.88MPa and 14.03MPa respectively, R 2 is 0.86. For the glass transition temperature T g , its MAE and RMSE are 9.64℃ and 15.61℃ respectively, R 2 It is 0.94.

[0153] For the test set, the transmittance performance T 400 , its MAE and RMSE are 3.93% and 5.64% respectively, R 2 The MAE and RMSE of the linear thermal expansion coefficient CLTE are 5.39 ppm·K. -1 and 7.57ppm·K -1 , R 2 is 0.87. b , its MAE and RMSE are 12.75MPa and 20.37MPa respectively, R 2 is 0.87. For the glass transition temperature T g , its MAE and RMSE are 9.22℃ and 13.86℃ respectively, R 2 is 0.95.

[0154] Figure 5 This is a comparison chart of the training results of different models under different performances in Example 1; Figure 5 (a) The light transmittance T of different models 400 Comparison chart of training results under Figure 5 (b) is a comparison chart of the training results of different models under the linear thermal expansion coefficient CLTE. Figure 5 (c) The tensile strength σ of different models b Comparison chart of training results under Figure 5 (d) The glass transition temperature T of different models g Comparison of training results under . Figure 5 From (a) to (d), it can be seen that the AFP model has the best modeling effect.

[0155] The traditional descriptor + neural network model (Des + ANN), in its test set, targets the transmittance performance T 400, its MAE and RMSE are 4.56% and 8.86% respectively. For the coefficient of linear thermal expansion CLTE, its MAE and RMSE are 6.19 ppm·K -1 and 8.29ppm·K -1 For the tensile strength σ b , its MAE and RMSE are 15.11MPa and 22.14MPa respectively. For the glass transition temperature T g , and its MAE and RMSE are 12.27℃ and 16.89℃ respectively. In comparison, the generalization ability of the AFP model (the accuracy of predicting unknown polyimide) is significantly improved.

[0156] Application Example 2

[0157] The performance prediction method of polyimide in Example 1 is used to predict the light transmittance, linear thermal expansion coefficient, tensile strength and glass transition temperature of each polyimide in the following table, and the above values ​​are compared with their corresponding experimental values ​​to obtain Tables 4-7. Table 4 is a comparison between the experimental value and the predicted value of the light transmittance of polyimide (%); Table 5 is a comparison between the linear thermal expansion coefficient of polyimide (ppm·K -1 ) The experimental values ​​are compared with the predicted values; Table 6 is the comparison between the experimental values ​​and the predicted values ​​of the tensile strength of polyimide (MPa); Table 7 is the comparison between the experimental values ​​and the predicted values ​​of the glass transition temperature of polyimide (℃).

[0158] Among them, the experimental values ​​of light transmittance, linear thermal expansion coefficient, tensile strength and glass transition temperature in each table are all derived from publicly published SCI literature.

[0159] Table 4

[0160]

[0161]

[0162]

[0163] Table 5

[0164]

[0165]

[0166]

[0167]

[0168] Table 6

[0169]

[0170]

[0171]

[0172]

[0173] Table 7

[0174]

[0175]

[0176]

[0177]

[0178] It can be seen from Tables 4-7 that the performance prediction method of this embodiment has a high accuracy rate.

[0179] Example 2

[0180] This embodiment discloses a method for designing a structure of a polyimide. Figure 6 Flowchart for training a reinforcement learning agent designed from scratch for polyimide in this embodiment.

[0181] The design method includes: training an intelligent agent for generating polyimide based on proximal policy optimization (PPO); the proximal policy optimization includes the following steps:

[0182] S1, fragmenting the dianhydride molecules and the diamine molecules to obtain a dianhydride structure group containing a plurality of dianhydride fragments and a diamine structure group containing a plurality of diamine fragments;

[0183] S2, selecting a dianhydride starting fragment and a diamine starting fragment from the dianhydride structure group and the diamine structure group, respectively;

[0184] S3, according to the strategy of the neural network, selecting an action, the action comprising: obtaining a dianhydride selected fragment and a diamine selected fragment from the dianhydride structure group and the diamine structure group respectively through the dianhydride starting fragment and the diamine starting fragment in step S2, and obtaining a dianhydride synthesis fragment and a diamine synthesis fragment;

[0185] S4, judging whether an intelligent agent of polyimide can be generated by the dianhydride synthesis fragment and the diamine synthesis fragment; when the intelligent agent of polyimide can be generated, the AFP model in Example 1 is used to calculate the physical property parameters of the intelligent agent of polyimide, and further calculate the reward value; when the intelligent agent of polyimide cannot be generated, returning to step S2 for loop iteration;

[0186] S5, the step of terminating the proximal strategy optimization.

[0187] Specifically, in order to minimize the possibility of generating incorrect polyimide molecules, this embodiment uses the decomposition method of interesting chemical substructures in retrospective synthesis (BRICS) to decompose dianhydride and diamine molecules. The specific process is as follows: 856 dianhydrides and 1759 diamines were first collected from the public database. Then, the collected dianhydrides and diamines were decomposed into 576 fragment molecules and 1004 fragment molecules respectively by the BRICS method. Considering the frequency of occurrence of dianhydride and diamine fragments, the top-ranked fragments were selected as the starting fragments of dianhydride and diamine.

[0188] In step S3, the method of "obtaining dianhydride selected fragments and diamine selected fragments from the dianhydride structure group and the diamine structure group respectively" includes: assigning different selection probabilities to all fragments in the dianhydride structure group and the diamine structure group based on the strategy of the neural network, and through the starting fragments of dianhydride and diamine in step S2, the strategy of the neural network will preferentially select fragments with large selection probabilities. In the process of training the intelligent agent, the strategy of the neural network will continuously iteratively update the selection probability of each fragment.

[0189] In step S3, the model architecture of the strategy is a multi-layer perceptron; the number of hidden layers of the strategy is 3; the number of neurons in the hidden layer of the strategy is 256, 128, and 128; the hidden layer of the strategy is a fully connected layer; and the activation function of the strategy is Relu.

[0190] In step S3, the method for obtaining the dianhydride synthesis fragment includes: when the dianhydride starting fragment and the dianhydride selected fragment both contain a connector, splicing the dianhydride starting fragment and the dianhydride selected fragment to form a dianhydride synthesis fragment; when the dianhydride starting fragment contains a connector and the dianhydride selected fragment does not contain a connector, using the dianhydride selected fragment as the dianhydride synthesis fragment; when the dianhydride starting fragment does not contain a connector, using the dianhydride starting fragment as the dianhydride synthesis fragment.

[0191] In step S3, the method for obtaining the diamine synthesis fragment includes: when the diamine starting fragment and the diamine selection fragment both contain a connector, the diamine starting fragment and the diamine selection fragment are spliced ​​to form a diamine synthesis fragment; when the diamine starting fragment contains a connector and the diamine selection fragment does not contain a connector, the diamine selection fragment is used as the diamine synthesis fragment; when the diamine starting fragment does not contain a connector, the diamine starting fragment is used as the diamine synthesis fragment.

[0192] In step S4, the physical property parameters of the polyimide agent include light transmittance, linear thermal expansion coefficient, tensile strength and glass transition temperature; the expression of the reward value is R = R T400 ×(1+R CLTE +R σb +R Tg +R SAscore ) / 5; where R T400 , RCLTE , R σb , R Tg , R SAscore They represent the 400nm transmittance, linear thermal expansion coefficient, tensile strength, glass transition temperature, and normalized reward value of SAscore respectively.

[0193] In step S5, the termination conditions of the proximal strategy optimization are: i. an intelligent agent of polyimide is generated and a reward value is calculated; ii. in step S3, the dianhydride synthesis fragment does not meet the generation condition of containing two amide groups; or, the diamine synthesis fragment does not meet the generation condition of containing two amine groups; iii. the number of atoms of the intelligent agent of polyimide exceeds 120; iv. in step S4, the number of loop iterations exceeds 5.

[0194] To efficiently identify the best fragment combinations for polyimide design, we used Proximal Policy Optimization (PPO) designed for continuous control tasks within a reinforcement learning framework. In this case, the agent makes decisions based on a policy, which is usually represented as a neural network. The policy network computes the probabilities and state-value functions of different actions based on the current state. Essentially, at each time step, the action is probabilistically selected from the action space. The goal of the agent is to maximize the expected reward it receives during its interaction with the environment.

[0195] The present invention maps chemical fragments or molecules into a 600-dimensional vector as a state value through the polyBERT model in a specific state. The action space is defined as a discrete combination space of dianhydride and diamine fragments. Each action is sampled based on the state value function and the probability of expected reward. In order to design a polyimide with target performance, the higher the reward value of the fragment combination, the higher the probability of it being sampled. In addition, in order to enhance the effectiveness of polyimide design, the reward of the fragment combination that cannot generate a valid molecule is set to zero.

[0196] The polyimide de novo design method of this embodiment has high effectiveness and reliability.

[0197] Effect Example 1

[0198] This effect embodiment adopts the design method of embodiment 2 to design a polyimide with high light transmittance, low linear thermal expansion coefficient, high tensile strength and high glass transition temperature from scratch.

[0199] The performance prediction construction method model and reinforcement learning framework used in this effect embodiment are used to train an intelligent agent that can design polyimide with high light transmittance, low linear thermal expansion coefficient, high tensile strength and high glass transition temperature from scratch. The results show that the trained intelligent agent can design polyimide that meets the above requirements. The specific steps are:

[0200] S1. The AFP model trained in Example 1 is used as a performance prediction model for light transmittance, linear thermal expansion coefficient, tensile strength and glass transition temperature. Taking into account the synthetic accessibility, this effect example uses the synthetic accessibility score SAscore to evaluate the difficulty of synthesizing the designed polyimide. The smaller the SAscore value, the easier it is to synthesize. The goal of this effect example is to design a polyimide with high transparency, low linear thermal expansion coefficient, high tensile strength, high glass transition temperature and good synthetic accessibility. And achieving high transparency is the main goal. The better the comprehensive performance, the higher the reward value. The expression of the reward value is defined as: R T400 ,R CLTE ,R σb ,R Tg , and R SAscore They represent the normalized reward value of 400nm transmittance, linear thermal expansion coefficient, tensile strength, glass transition temperature and SAscore, respectively.

[0201] S2. Every 10 training cycles, 10,000 polyimide samples were randomly generated using the corresponding agent to evaluate the learning progress of the agent. For comparison, 10,000 random generation runs were performed before training the agent. Figure 7 This is a design effect diagram of the reinforcement learning polyimide agent in effect embodiment 1. Obviously, with the continuous reinforcement learning of the agent, the design strategy of the agent tends to a higher reward value.

[0202] in, Figure 7 (a) is the reward dynamics (including maximum, minimum and average) of the agent in the entire training process of this embodiment. Figure 7 (b) is the light transmittance performance T of the polyimide smart body 400 Design renderings, Figure 7 (b) shows T 400 The distribution of converges to about 80% after 70 training cycles; Figure 7 (c) is the design effect diagram of the linear thermal expansion coefficient CLTE of the polyimide intelligent body. Figure 7 (d) is the tensile strength σ of the polyimide smart body b Design renderings, Figure 7 (e) is the glass transition temperature T of the polyimide smart body g Design renderings, Figure 7 (f) is a diagram showing the design effect of the synthetic accessibility SAscore of the polyimide agent. Figure 7 (c) to Figure 7(f) shows that the other four performance indicators also converge in their respective optimization directions (i.e., lower linear thermal expansion coefficient, higher tensile strength, higher glass transition temperature, and lower SAscore).

[0203] The distribution of polyimide properties generated via de novo design changes significantly compared to a collected literature dataset, indicating that the agent learns a unique design strategy.

[0204] S3. UMAP dimensionality reduction was used to map it into a two-dimensional chemical space to explore the designed polyimide. Figure 8 The distribution of polyimide agents in chemical space at different stages of this effect embodiment. With continuous reinforcement learning, the chemical space range of the designed polyimide gradually decreases. The points in these spaces correspond to the potential polyimides that are expected to be discovered. After 70 training cycles, the polyimides designed by the agents are discretely distributed in multiple regions. In order to check the uniqueness of the polyimide in each region, the present invention uses unsupervised K-means clustering in this chemical space.

[0205] Fig. 9 In this embodiment, the number of clusters of polyimide designed by the agent after 70 trainings is determined according to the silhouette score and inertia, and the number of clusters K is set to 6. Fig.10 This is the chemical spatial distribution of polyimide after being clustered into 6 clusters by K-means in this embodiment. Fig.11 The polyimide with the highest reward value in clusters 1-6 and its various properties in this effect embodiment. Compared with the chemical space of polyimides in the literature, the intelligent agent designed new types of polyimides (such as cluster 2 and cluster 6). In addition, these representative polyimides showed excellent comprehensive properties (i.e., T 400 >80%, CLTE<30ppm·K -1 , σ b >100MPa,T g >350℃).

[0206] Fig.12 This is the chemical formula of the polyimide designed from scratch for the intelligent agent in this effect embodiment. Fig.13 A polyimide film designed from scratch for the intelligent body in this effect embodiment. Fig.14 The experimental values ​​and predicted values ​​of the thin film synthesized experimentally by the polyimide designed by the intelligent agent in this effect embodiment are compared.

[0207] Fig.15 This is the fragment frequency obtained by decomposing dianhydride by the BRICS method in this effect example. Fig.16 This is the fragment frequency obtained by decomposing diamine by the BRICS method in this effect example. Fig.15 , Fig.16The top 24 dianhydride fragments with the highest frequency of occurrence and the top 24 diamine fragments with the highest frequency of occurrence are shown respectively. The neural network-based strategy assigns different selection probabilities to all fragments in the dianhydride structure group and the diamine structure group, and the strategy gives priority to fragments with large selection probabilities. Although the starting fragments of dianhydride and diamine are fixed, since the selection probability of each fragment in the structure group is different, polyimides with different structures will be generated in the end.

[0208] Effect Example 2

[0209] This effect embodiment adopts the design method of embodiment 2 to design a polyimide with low light transmittance, low linear thermal expansion coefficient, high tensile strength and high glass transition temperature from scratch.

[0210] The performance prediction construction method model and reinforcement learning framework used in this effect embodiment are used to train an intelligent agent that can design polyimide with low light transmittance, low linear thermal expansion coefficient, high tensile strength and high glass transition temperature from scratch. The results show that the trained intelligent agent can design polyimide that meets the above requirements. The specific steps are:

[0211] S1. The AFP model trained in Example 1 is used as a performance prediction model for light transmittance, linear thermal expansion coefficient, tensile strength and glass transition temperature. The expression of the reward value is the same as that in Example 1.

[0212] S2. Every 10 training cycles, 10,000 polyimide samples were randomly generated using the corresponding agent to evaluate the learning progress of the agent. For comparison, 10,000 random generation runs were performed before training the agent. Fig.17 Effect diagram of the design effect of the reinforcement learning polyimide agent in Example 2. Obviously, with the continuous reinforcement learning of the agent, the design strategy of the agent tends to a higher reward value.

[0213] in, Fig.17 (a) is the reward dynamics (including maximum, minimum and average values) of the agent in the entire training process of the effect embodiment 2. Fig.17 (b) is the light transmittance performance T of the polyimide smart body 400 The design effect diagram converges to about 0% after 70 training cycles. Fig.17 (c) is the design effect diagram of the linear thermal expansion coefficient CLTE of the polyimide intelligent body. Fig.17 (d) is the tensile strength σ of the polyimide smart body b Design renderings, Fig.17 (e) is the glass transition temperature T of the polyimide smart body g Design renderings, Fig.17(f) is a diagram showing the design effect of the synthetic accessibility SAscore of the polyimide agent. Fig.17 (c)~ Fig.17 (f) shows that the other four performance indicators also converge in their respective optimization directions (i.e., lower linear thermal expansion coefficient, higher tensile strength, higher glass transition temperature, and lower SAscore).

[0214] The distribution of polyimide properties generated via de novo design changes significantly compared to a collected literature dataset, indicating that the agent learns a unique design strategy.

[0215] S3. The present invention uses UMAP dimensionality reduction to map it into a two-dimensional chemical space to explore the designed polyimide. Fig.18 The distribution of polyimide agents in chemical space at different stages in this effect embodiment. With continuous reinforcement learning, the chemical space range of the designed polyimide gradually decreases. The points in these spaces correspond to the potential polyimides that are expected to be discovered. After 70 training cycles, the polyimides designed by the agent are discretely distributed in multiple regions.

[0216] Fig.19 In this embodiment, the number of clusters of polyimide designed by the agent after training 70 times is determined based on the silhouette score and inertia. Fig. 20 This is the chemical spatial distribution of polyimide after being clustered into three clusters by K-means in this embodiment. Fig.21 The polyimide with the highest reward value in clusters 1-3 in this effect example and its various properties. The same clustering method as in effect example 1 is used, and the cluster number K is set to 3. These representative polyimides show excellent comprehensive properties (i.e., T 400 <10%, CLTE<20ppm·K -1 , σ b >150MPa, Tg>350℃).

Claims

1. A method for predicting the performance of polyimide, characterized in that: It includes the following steps: Inputting the first molecular descriptor into the AFP model to obtain the physical property parameter to be measured; the first molecular descriptor is the molecular descriptor of the polyimide to be measured; The AFP model is obtained by training a training data set, wherein the training data set includes a second molecular descriptor and a first physical property parameter; The second molecular descriptor and the first physical property parameter are respectively the molecular descriptor and the physical property parameter of the known polyimide; The first molecular descriptor and the second molecular descriptor both include node features and edge features, and the type of the first molecular descriptor is the same as the type of the second molecular descriptor; The type of the physical property parameter to be measured is the same as the type of the first physical property parameter.

2. The method for predicting the performance of polyimide according to claim 1, characterized in that: The performance prediction method meets one or more of the following conditions: ① The type of the physical property parameter to be measured and the type of the first physical property parameter are independently light transmittance, linear thermal expansion coefficient, tensile strength or glass transition temperature; ② The node characteristics are one or more of atom type, number of bond connections, number of hydrogen atom connections, displayed valence, hybridization, aromaticity, chiral center, chiral type and formal charge; ③ The edge characteristics are one or more of bond type, conjugation, ring structure and stereo bond; ④ The model parameters of the AFP model include hidden layer dimension, initial learning rate, learning rate decay factor, weight decay, dropout rate, number of node-level embedding layers or number of graph-level embedding layers; Preferably, the hidden layer dimension is 16-128, for example, 87, 97, 111 or 116; Preferably, the initial learning rate is 10 -3 -10 -1 , for example, 10 -3 , 10 -2.26 or 10 -1.92 ; Preferably, the learning rate attenuation factor is 0.4-0.9, for example, 0.6, 0.8 or 0.9; Preferably, the weight decay is 0-0.4, for example, 0; Preferably, the discard rate is 0-0.4, for example, 0.1, 0.2, 0.25 or 0.35; Preferably, the number of node-level embedding layers is 1-6, such as 1, 2 or 3; Preferably, the number of graph-level embedding layers is 1-4, such as 1, 2 or 4.

3. The method for predicting the performance of polyimide according to claim 1, characterized in that: The performance prediction method meets one or more of the following conditions: ① The first molecular descriptor and the second molecular descriptor are generated by RDKit or Mordred respectively; ② The first molecular descriptor and the second molecular descriptor are generated in the same manner; ③ The training data set is cleaned, and the cleaning method includes: i. excluding polyimides containing metal elements and / or ionic compounds; ii. excluding polyimide copolymers and blends; iii. when polyimide of the same structure has multiple data records of the same physical property parameters, taking the average value of each data record.

4. A polyimide performance prediction system, characterized in that: It includes a prediction module, which is used to input the first molecular descriptor into the AFP model to obtain the physical property parameter to be measured; the first molecular descriptor is the molecular descriptor of the polyimide to be measured; the AFP model is obtained by training a training data set, and the training data set includes the second molecular descriptor and the first physical property parameter; The second molecular descriptor and the first physical property parameter are respectively the molecular descriptor and the physical property parameter of the known polyimide; the first molecular descriptor and the second molecular descriptor both include node features and edge features, and the type of the first molecular descriptor is the same as the type of the second molecular descriptor; the type of the physical property parameter to be measured is the same as the type of the first physical property parameter; Preferably, the prediction module is also used in the performance prediction method as claimed in claim 2 or 3.

5. A method for structural design of polyimide, characterized in that: It includes: Based on proximal policy optimization, an agent is trained to generate polyimide; The proximal strategy optimization comprises the following steps: S1, fragmenting the dianhydride molecules and the diamine molecules to obtain a dianhydride structure group containing a plurality of dianhydride fragments and a diamine structure group containing a plurality of diamine fragments; S2, selecting a dianhydride starting fragment and a diamine starting fragment from the dianhydride structure group and the diamine structure group, respectively; S3, according to the strategy of the neural network, selecting an action, the action comprising: obtaining a dianhydride selected fragment and a diamine selected fragment from the dianhydride structure group and the diamine structure group respectively through the dianhydride starting fragment and the diamine starting fragment in step S2, and obtaining a dianhydride synthesis fragment and a diamine synthesis fragment; S4, judging whether the intelligent agent of the polyimide can be generated by the dianhydride synthesis fragment and the diamine synthesis fragment; when the intelligent agent of the polyimide can be generated, using the AFP model in the performance prediction method according to any one of claims 1 to 3 or the AFP model in the performance prediction system according to claim 4, calculating the physical property parameters of the intelligent agent of the polyimide, and further calculating the reward value; when the intelligent agent of the polyimide cannot be generated, returning to step S2 for loop iteration; S5. Terminate the proximal strategy optimization step.

6. The method for designing the structure of polyimide according to claim 5, characterized in that: The structural design method meets one or more of the following conditions: i. In step S1, the fragmentation method includes a decomposition method for synthesizing interesting chemical substructures in retrospect; ii. In step S2, the selection method includes: extracting groups from the dianhydride molecules and the diamine molecules, and converting them into a 600-dimensional vector recognizable by a computer through a polyBERT model; iii. In step S2, the selection rule of the dianhydride starting fragment and the diamine starting fragment is: according to the occurrence frequency of the starting fragment, the dianhydride starting fragment and the diamine starting fragment with the highest occurrence frequency are selected as the only starting fragment; iv. In step S3, the model architecture of the strategy is a multi-layer perceptron; v. In step S3, the number of hidden layers of the strategy is 1-5; vi. In step S3, the number of hidden layer neurons of the strategy is 16-512, preferably 16, 32, 64, 128, 256 or 512; vii. In step S3, the hidden layer of the strategy is a fully connected layer; viii. In step S3, the activation function of the strategy is Relu, LeakyRelu or Tanh, preferably Relu; iw. In step S3, the method for obtaining the dianhydride synthesis fragment includes: when the dianhydride starting fragment and the dianhydride selected fragment both contain a connector, splicing the dianhydride starting fragment and the dianhydride selected fragment to form the dianhydride synthesis fragment; when the dianhydride starting fragment contains a connector and the dianhydride selected fragment does not contain a connector, using the dianhydride selected fragment as the dianhydride synthesis fragment; when the dianhydride starting fragment does not contain a connector, using the dianhydride starting fragment as the dianhydride synthesis fragment; w. In step S3, the method for obtaining the diamine synthesis fragment includes: when the diamine starting fragment and the diamine selection fragment both contain a connector, the diamine starting fragment and the diamine selection fragment are spliced ​​to form the diamine synthesis fragment; when the diamine starting fragment contains a connector and the diamine selection fragment does not contain a connector, the diamine selection fragment is used as the diamine synthesis fragment; when the diamine starting fragment does not contain a connector, the diamine starting fragment is used as the diamine synthesis fragment; wi. In step S4, the physical property parameters of the polyimide intelligent body include light transmittance, linear thermal expansion coefficient, tensile strength and glass transition temperature; preferably, the expression of the reward value is R=R T400 ×(1+R CLTE +R σb +R Tg +R SAscore ) / 5; where R T400 , R CLTE , R σb , R Tg , R SAscore They represent the normalized reward value of 400nm transmittance, linear thermal expansion coefficient, tensile strength, glass transition temperature, and SAscore respectively; wii. The termination condition of the proximal strategy optimization is any one of the following conditions: a. The intelligent agent of the polyimide is generated and the reward value is calculated; b. In step S3, the dianhydride synthesis fragment does not meet the generation condition of containing two amide groups; or, the diamine synthesis fragment does not meet the generation condition of containing two amine groups; c. The number of atoms of the intelligent agent of the polyimide exceeds 50, preferably exceeds 120; d. In step S4, the number of loop iterations exceeds 5, preferably exceeds 10.

7. A structural design system for polyimide, characterized in that: It includes a structure design module for training an intelligent agent that generates polyimide based on proximal policy optimization; The proximal strategy optimization includes the following steps: S1, fragmenting dianhydride molecules and diamine molecules to obtain a dianhydride structure group containing a plurality of dianhydride fragments and a diamine structure group containing a plurality of diamine fragments; S2, selecting a dianhydride starting fragment and a diamine starting fragment from the dianhydride structure group and the diamine structure group, respectively; S3, selecting an action according to the strategy of the neural network, the action including obtaining a dianhydride selected fragment and a diamine selected fragment from the dianhydride structure group and the diamine structure group, respectively, through the dianhydride starting fragment and the diamine starting fragment in step S2. , and obtain dianhydride synthesis fragments and diamine synthesis fragments; S4, determine whether the intelligent body of the polyimide can be generated through the dianhydride synthesis fragments and the diamine synthesis fragments; when the intelligent body of the polyimide can be generated, use the AFP model in the performance prediction method according to any one of claims 1 to 3 or the AFP model in the performance prediction system according to claim 4 to calculate the physical property parameters of the intelligent body of the polyimide; when the intelligent body of the polyimide cannot be generated, return to step S2 for loop iteration; S5, terminate the step of proximal strategy optimization; Preferably, the structural design module is also used in the structural design method of polyimide as claimed in claim 6.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method for designing the structure of polyimide according to claim 5 or 6 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for designing the structure of polyimide according to claim 5 or 6 is implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method for designing the structure of polyimide according to claim 5 or 6 is implemented.

Images