Methods and systems for predicting the performance of polyimide, methods and systems for structural design, electronic devices, computer-readable storage media, and computer products.

Abstract

This invention discloses a method and system for predicting the performance of polyimides, a method and system for structural design, an electronic device, a computer-readable storage medium, and a computer product. The method for predicting the performance of polyimides includes the following steps: inputting a first molecular descriptor into an AFP model to obtain the physical property parameter to be tested; the first molecular descriptor is the molecular descriptor of the polyimide to be tested; the AFP model is trained using a training dataset, which includes a 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 physical property parameter of a known polyimide; both the first and second molecular descriptors 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 tested is the same as the type of the first physical property parameter. This invention has high accuracy and improves the effectiveness and reliability of de novo design.

Description

Technical Field

[0001] This invention relates to a method and system for predicting the performance of polyimide, a method and system for structural design, an electronic device, a computer-readable storage medium, and a computer product. Background Technology

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

[0003] However, finding such polyimides is a significant challenge. Researchers have done extensive work to further enhance the transparency and other properties of polyimide films. Methods such as introducing large groups, adding fluorinated groups, or introducing cyclic 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 cyclic structures may result in poor temperature tolerance and insufficient rigidity, thus affecting overall performance.

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

[0005] Currently, the widely adopted design strategy is based on forward design. However, this strategy is limited by known material structures and properties, making the discovery of entirely new material systems challenging. Furthermore, forward design typically requires extensive experimental and computational work, resulting in high costs and time consumption. The emergence of de novo design strategy offers a powerful means to address these challenges. Adopting this strategy in materials development, starting with the desired material properties, enables the discovery of new materials with the target properties, providing a higher level of innovation and exploration.

[0006] Furthermore, the unique structure of polyimide (containing imide groups) presents additional challenges for de novo design, namely, designing molecules with both specific structures and specific properties. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, such as the difficulty in rapidly predicting the performance parameters of polyimides and the high cost and time consumption of polyimide structure acquisition, this invention provides a method and system for predicting the performance of polyimides (PI), a method and system for structural design, an electronic device, a computer-readable storage medium, and a computer product. The performance prediction method of this invention achieves high accuracy in performance prediction through specific deep learning. Furthermore, this invention employs a fragment-based design architecture to improve the efficiency of polyimide design, and combines it with proximal policy optimization (PPO) in reinforcement learning for training, enabling the de novo design of agents with target polyimide performance; further improving the effectiveness and reliability of de novo design.

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

[0009] This invention provides a method for predicting the performance of polyimide, comprising the following steps:

[0010] The first molecular descriptor is input into the AFP model to obtain the physical property parameters to be tested; the first molecular descriptor is the molecular descriptor of the polyimide to be tested;

[0011] The AFP model is trained using a training dataset, which includes a second molecular descriptor and a first physical property parameter; the second molecular descriptor and the first physical property parameter are the molecular descriptor and physical property parameter of a known polyimide, respectively.

[0012] Both the first molecular descriptor and the second molecular descriptor include node features and edge features, and the types of the first molecular descriptor and the second molecular descriptor are the same.

[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 this invention, polyimide is a complex polymer with different types and structures. In actual research, due to the limited and insufficient experimental data on polyimide and the sparse dataset, model training is 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 this invention, the full name of the AFP model is Attention Fingerprint. The Attention Fingerprint (AFP) graph neural network treats chemical structures as graphs, where atoms represent nodes and bonds represent edges, encoding direct and extended connections between atoms. To distinguish different nodes and edges (i.e., atoms and bonds), feature vectors (e.g., atom type, valence, bond type, etc.) need to be assigned to these nodes and edges. Through an attention mechanism, fingerprints obtained from node-level and graph-level embeddings are introduced, called Attention Fingerprint (AFP), which can learn nonlocal molecular interactions. Because it can learn the intrinsic connections in chemical structures, AFP exhibits excellent performance in model construction.

[0016] In this 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 contains chiral node features, the first molecular descriptor also contains chiral node features; when the second molecular descriptor contains all feature vectors, the first molecular descriptor also contains the same feature vectors.

[0017] In this invention, those skilled in the art should know that the first physical property parameters in the training dataset are generally collected from publicly available literature.

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

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

[0020] In this invention, light transmittance, dimensional stability, mechanical properties, and heat resistance are all performance indicators in different dimensions, and the measurement methods, units, and numerical ranges for each indicator are different. Data from different sources may have problems such as measurement errors and data inconsistencies (the mean-filling method is used to handle missing values ​​and ensure data integrity).

[0021] In this invention, molecular descriptors are obtained through RDKit or Mordred, which are used to represent mathematical descriptions of chemical molecular structures; RDKit (Research Data Kit) is an open-source cheminformatics toolkit; Mordred is a specialized descriptor computation software.

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

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

[0024] The number of key connections can be one-hot encoded as a node feature using "0, 1, 2, 3 or 4".

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

[0026] The displayed valence can be a one-hot encoded form of "0, 1, 2, 3, 4 or 5" as a node feature.

[0027] The hybridization can be "sp, sp". 2 sp 3 sp 3 d or sp 3 d 2 The one-hot encoding of “” is used as the node feature.

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

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

[0030] The chirality type can be "R or S" as a node feature to indicate the chirality type involving atoms.

[0031] 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 the molecule.

[0032] In some implementations, the edge features are one or more of bond type, conjugation, loop structure, and 3D bond.

[0033] The bond type can be a one-hot encoding 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 can use "0 or 1" as a node feature to indicate whether the key is part of the ring.

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

[0037] In some implementations, 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 dimension of the hidden layer 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 decay 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, for example, 1, 2 or 3.

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

[0045] In one embodiment, when the physical property parameter is transmittance T 400 At that time, the hidden layer dimension was 116, and the initial learning rate was 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 linear thermal expansion coefficient CTLE, the hidden layer dimension is 87 and the initial learning rate is 10. -1.92The 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 At that time, the hidden layer dimension was 97, and the initial learning rate was 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 At that time, the hidden layer dimension was 111, and the initial learning rate was 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 implementations, the first molecular descriptor and the second molecular descriptor are generated using RDKit or Mordred, respectively.

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

[0051] In some implementations, the training dataset is cleaned to remove outliers.

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

[0053] i. Exclude polyimides containing metallic elements and / or ionic compounds; the polyimide must be an organic compound;

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

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

[0056] In this invention, the AFP model can extract the complex chemical structural features of polyimide, thereby improving the accuracy of performance prediction; the AFP model is a graph neural network model with a graph attention mechanism framework, which belongs to deep learning methods; generally, given a sufficiently large dataset, deep learning models have better prediction performance.

[0057] The study of the impact of different models on performance prediction results shows that the AFP model performs better in predicting the performance of polyimide compared to other deep learning methods (such as the FraGAT model or the BERT model). Furthermore, the AFP model performs better in predicting the performance of polyimide compared to traditional machine learning methods (such as multilayer perceptrons, which involve feature engineering).

[0058] In this invention, the structural information of both the polyimide to be tested and the known polyimide can be obtained by converting the repeating units therein into corresponding SMILES. SMILES, short for Simplified Molecular Inputline Entry Specification, is a specification that explicitly describes the molecular structure using ASCII strings.

[0059] This invention also provides a performance prediction system for polyimide, comprising a prediction module for inputting a first molecular descriptor into an AFP model to obtain the physical property parameter to be tested; the first molecular descriptor is a molecular descriptor of the polyimide to be tested; the AFP model is trained using a training dataset, the training dataset including a second molecular descriptor and the first physical property parameter; the second molecular descriptor and the first physical property parameter are respectively a molecular descriptor and a physical property parameter of a known polyimide; both the first molecular descriptor and the second molecular descriptor include node features and edge features, 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 tested is the same as the type of the first physical property parameter.

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

[0061] The present invention also provides a method for structural design of polyimide, comprising: training an agent to generate polyimide based on proximal policy optimization (PPO);

[0062] The near-end strategy optimization includes the following steps:

[0063] S1. The dianhydride molecule and the diamine molecule are fragmented to obtain a dianhydride structure group containing several dianhydride fragments and a diamine structure group containing several diamine fragments.

[0064] S2. Select dianhydride initiating fragments and diamine initiating fragments from the dianhydride structure group and the diamine structure group, respectively;

[0065] S3. According to the strategy of the neural network, select an action, the action including: obtaining a dianhydride selection fragment and a diamine selection fragment from the dianhydride structure group and the diamine structure group respectively through the dianhydride initiation fragment and the diamine initiation fragment in step S2, and obtaining a dianhydride synthesis fragment and a diamine synthesis fragment;

[0066] S4. Determine whether the polyimide agent can be generated through the dianhydride synthesis fragment and the diamine synthesis fragment; if the polyimide agent can be generated, use the AFP model in the performance prediction method described above or the AFP model in the performance prediction system described above to calculate the physical property parameters of the polyimide agent and further calculate the reward value; if the polyimide agent cannot be generated, return to step S2 for iterative iteration.

[0067] S5. Terminate the near-end strategy optimization step.

[0068] In this invention, the above-described structural design method avoids the need to construct a candidate library of polyimides in advance. The disadvantage of constructing a candidate library of polyimides in advance is that the candidate library constructed artificially is inevitably limited and cannot include all structures in the chemical space.

[0069] In this invention, polyimides that meet the requirements can be directly designed based on the target performance; this can be achieved in the proximal policy optimization in reinforcement learning.

[0070] In this invention, the target molecule can be broken down into common chemical groups using RDKit, which can then be used as a candidate gene library for de novo design.

[0071] In some embodiments, step S1, the fragmentation process includes a method for retrospectively synthesizing interesting chemical substructures.

[0072] In some embodiments, step S2 includes selecting groups from the dianhydride molecule and the diamine molecule and converting them into a computer-recognizable 600-dimensional vector using a polyBERT model.

[0073] In some embodiments, in step S2, the selection rule for the dianhydride initiation fragment and the diamine initiation fragment is as follows: based on the occurrence frequency of the initiation fragment, the dianhydride initiation fragment and the diamine initiation fragment with the highest occurrence frequency are selected as the unique initiation fragment.

[0074] In step S3 of this invention, the method of "obtaining dianhydride selection fragments and diamine selection fragments from the dianhydride structure group and the diamine structure group respectively" 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; using the initial fragments of dianhydride and diamine in step S2, the strategy of the neural network will preferentially select fragments with higher selection probabilities. During the training of the 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 multilayer perceptron.

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

[0077] In some implementations, in step S3, the number of hidden layer neurons in 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, step S3 involves obtaining the dianhydride synthesis fragment in the following ways: when both the dianhydride starting fragment and the dianhydride selected fragment contain connectors, the dianhydride starting fragment and the dianhydride selected fragment are spliced ​​together to form the dianhydride synthesis fragment; when the dianhydride starting fragment contains a connector and the dianhydride selected fragment does not contain a connector, the dianhydride selected fragment is used as the dianhydride synthesis fragment; when the dianhydride starting fragment does not contain a connector, the dianhydride starting fragment is used as the dianhydride synthesis fragment.

[0081] In some embodiments, step S3 involves obtaining the diamine synthesis fragment in the following ways: when both the diamine initiation fragment and the diamine selection fragment contain connectors, the diamine initiation fragment and the diamine selection fragment are spliced ​​together to form the diamine synthesis fragment; when the diamine initiation 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 initiation fragment does not contain a connector, the diamine initiation fragment is used as the diamine synthesis fragment.

[0082] In this invention, after obtaining the predictive performance of the candidate polyimide agent, a comprehensive score can be obtained by weighted averaging, which serves as the basis for customizing rewards in reinforcement learning.

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

[0084] The reward value is used to quantify the benefit gained by the agent after performing a specific action or making a decision. In this field, the reward value is typically used to guide the learning process of an agent, enabling it to identify which behaviors are superior in a specific context. In this application, the formula for calculating the reward value integrates multiple physical performance parameters to evaluate the overall performance of the polyimide material. The impact or effect of the reward value on the structural design results is reflected in:

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

[0086] (2) Optimization decision: During the structural design process, the agent will continuously try different combinations of material parameters. The feedback of reward value 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 agent will not only pursue the maximization of a single performance, but also seek to find the best balance among multiple performance indicators.

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

[0089] In this invention, the polyimide structural design method is a de novo design, which means designing from fragments of dianhydride and diamine respectively, thereby improving the effectiveness of polyimide design; adopting a fragment-based de novo design scheme can ensure 100% design effectiveness. This design is more difficult than the forward design of existing technologies because de novo design requires directly designing a structure that meets the conditions based on the target performance, and requires training a sufficiently strong agent or model; while the forward design of existing technologies only requires traversing and predicting the candidate library based on the performance prediction model to perform screening.

[0090] The present invention also provides a polyimide structure design system, which includes a structure design module for training an agent that generates polyimide based on proximal strategy optimization;

[0091] The proximal strategy optimization includes the following steps: S1, fragmenting dianhydride and diamine molecules to obtain a dianhydride structure group containing several dianhydride fragments and a diamine structure group containing several diamine fragments; S2, selecting dianhydride initiating fragments and diamine initiating fragments from the dianhydride structure group and diamine structure group, respectively; S3, selecting actions according to the neural network strategy, the actions including obtaining dianhydride selection fragments and diamine selection fragments from the dianhydride structure group and diamine structure group, respectively, through the dianhydride initiating fragments and diamine initiating fragments in step S2, and obtaining dianhydride synthesis fragments and diamine synthesis fragments; S4, determining whether the polyimide agent can be generated through the dianhydride synthesis fragments and diamine synthesis fragments; when the polyimide agent can be generated, calculating the physical property parameters of the polyimide agent using the AFP model in the performance prediction method or the AFP model in the performance prediction system as described above; when the polyimide agent cannot be generated, returning to step S2 for iterative iteration; S5, terminating the proximal strategy optimization steps.

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

[0093] The present invention also provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the polyimide structural design method described above.

[0094] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the polyimide structural design method as described above.

[0095] The present invention also provides a computer program product comprising a computer program that, when executed by a processor, implements the polyimide structural design method as described above.

[0096] In this invention, a de novo design strategy for polyimides 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 coefficient of thermal expansion, higher tensile strength, and higher glass transition temperature.

[0097] In an embodiment of the present invention, by comparing four machine learning models (one neural network (multilayer perceptron), one natural language model (BERT model), and two graph neural network models (AFP model and 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] Based on common knowledge in the field, the above-mentioned preferred conditions can be combined arbitrarily to obtain various preferred embodiments of the present invention.

[0099] The reagents and raw materials used in this invention are all commercially available.

[0100] The positive and progressive effects of this invention are as follows:

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

[0102] (2) This invention is based on proximal policy optimization in reinforcement learning to train a de novo polyimide agent. The preferred starting point of this invention is to design a polyimide with excellent overall performance. Furthermore, through training, it is preferable to obtain a PI agent that combines high transparency (or low transparency), low linear coefficient of thermal expansion, 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 a broad PI chemical space, saving computational resources and computation time, and greatly improving the design efficiency of PI. Attached Figure Description

[0104] Figure 1 The flowchart shows the molecular diagram representation of the polyimide repeating unit structure in Example 1 and the model construction of different physical property parameters;

[0105] Figure 2The flowchart illustrates the agent training in Example 1 and the de novo design of the agent with target performance using polyimide.

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

[0107] Figure 4 The image shows the test results of the AFP model in Example 1 for different physical property parameters; Figure 4 (a) is the AFP model for light transmittance T 400 The test results are shown in the image. Figure 4 (b) is a graph showing the test results of the AFP model for the linear thermal expansion coefficient CLTE. Figure 4 (c) is the AFP model for tensile strength σ b The test results are shown in the image. Figure 4 (d) is the AFP model for the glass transition temperature T. g The test results are shown in the image.

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

[0109] Figure 6 This is a flowchart of the training process for the de novo polyimide reinforcement learning agent in Example 2.

[0110] Figure 7 The diagram shows the design effect of the reinforcement learning polyimide agent in Example 1. Figure 7 (a) The reward dynamics (including maximum, minimum, and average values) of the agent throughout the entire training process in Example 1. Figure 7 (b) Regarding the light transmittance T of the polyimide smart body 400 Design renderings, Figure 7(c) is a design rendering of CLTE, which relates to the linear thermal expansion coefficient of polyimide smart cells. Figure 7 (d) represents the tensile strength σ of the polyimide smart body. b Design renderings Figure 7 (e) represents the glass transition temperature T of the polyimide smart device. g Design renderings, Figure 7 (f) is a design rendering of SAscore, which relates to the synthetic accessibility of polyimide smart agents.

[0111] Figure 8 The distribution of polyimide smart bodies in chemical space at different stages in Example 1 is shown.

[0112] Figure 9 In Example 1, the number of polyimide clusters designed for the agent after 70 training cycles was determined based on contour scores and inertia.

[0113] Figure 10 The chemical spatial distribution of polyimide after K-means clustering into 6 clusters in Example 1 is shown.

[0114] Figure 11 The polyimide with the highest reward value among clusters 1-6 in Example 1 and its various properties are used to illustrate the effects.

[0115] Figure 12 The chemical structure of the polyimide used in the de novo design of the agent in Example 1 is shown.

[0116] Figure 13 The polyimide film for de novo design of the agent in Example 1 is used for the effect.

[0117] Figure 14 The experimental and predicted values ​​of the polyimide film synthesized from the smart agent designed in Example 1 are compared.

[0118] Figure 15 The frequency of the fragment obtained by decomposing dianhydride using the BRICS method in Example 1 is shown.

[0119] Figure 16 The frequency of the fragment obtained by decomposing diamine using the BRICS method in Example 1 is shown.

[0120] Figure 17 The design effect diagram of the reinforcement learning polyimide agent in Example 2; Figure 17 (a) The reward dynamics (including maximum, minimum, and average values) of the agent throughout the entire training process in Example 2. Figure 17 (b) Regarding the light transmittance T of the polyimide smart body 400 Design renderings, Figure 17(c) is a design rendering of CLTE, which relates to the linear thermal expansion coefficient of polyimide smart cells. Figure 17 (d) represents the tensile strength σ of the polyimide smart body. b Design renderings Figure 17 (e) represents the glass transition temperature T of the polyimide smart device. g Design renderings, Figure 17 (f) is a design rendering of SAscore, which relates to the synthetic accessibility of polyimide smart agents.

[0121] Figure 18 The distribution of polyimide smart bodies in chemical space at different stages in Example 2 is shown.

[0122] Figure 19 In Example 2, the number of polyimide clusters designed for the agent after 70 training cycles was determined based on contour scores and inertia.

[0123] Figure 20 The chemical spatial distribution of polyimide after K-means clustering into 3 clusters is shown in Example 2.

[0124] Figure 21 The polyimide with the highest reward value in clusters 1-3 of Example 2 and its various properties are used to illustrate the effects. Detailed Implementation

[0125] The present invention is further illustrated below by way of embodiments, but the invention is not limited to the scope of the embodiments described herein. Experimental methods in the following embodiments that do not specify specific conditions were performed according to conventional methods and conditions, or as selected according to the product instructions.

[0126] Example 1

[0127] This embodiment discloses a method for predicting the performance of polyimide. Figure 1 The flowchart shows the molecular diagram representation of the polyimide repeating unit structure in Example 1 and the model construction of different physical property parameters; Figure 2 The flowchart illustrates the agent training in Example 1 and the de novo design of the agent with target performance using polyimide.

[0128] This performance prediction method includes the following steps:

[0129] ss1. Collected data on light transmittance (T) properties related to polyimide (PI) films from 137 published articles. 400 ), linear thermal expansion coefficient (CLTE), tensile strength (σ) b ) and glass transition temperature (T g The data are treated as separate datasets. Figure 3This is a data distribution diagram of the datasets for each physical property parameter in this embodiment; Figure 3 (a) is a data distribution diagram of the light transmittance performance dataset. Figure 3 (b) is a data distribution diagram of the linear thermal expansion coefficient dataset. Figure 3 (c) is a data distribution diagram of the tensile strength dataset. Figure 3 (d) is a data distribution diagram of the glass transition temperature dataset; the amount of data collected for the four physical property parameters is as follows: Figure 3 As shown, the data are as follows: light transmittance (389 data points), linear thermal expansion coefficient (454 data points), tensile strength (764 data points), and glass transition temperature (946 data points). In each of the above datasets, the repeating unit structures of PI are stored using SMILES, and the performance data are 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), several isomers of PI are included in the data for data augmentation.

[0130] ss2. Molecular descriptor feature generation was accomplished using RDKit, with some inappropriate descriptors 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. Furthermore, molecular graph information, including features such as atom types and bonding information within the molecule, was generated using a Python script written using conventional methods.

[0131] ss3. The molecular structure diagram of polyimide obtained by the above Python script is converted into a vector and used as a molecular descriptor. The node (atom) features and edge (bond) features are shown in Table 1 and Table 2. Table 1 shows the information of node (atom) features and Table 2 shows 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) (unique thermal encoding) 9 Number of key connections The number of bonds connected to an atom (0, 1, 2, 3, 4) (one-heat encoding) 5 Number of hydrogen atoms The number of hydrogen atoms bonded to the atom (0, 1, 2, 3, 4) (unique thermal encoding) 5 Display valence Display valence (0, 1, 2, 3, 4, 5) (One-hot encoding) 6 Hybrid <![CDATA[Hybridization (sp, sp 2 , sp 3 , sp 3 d, sp 3 d 2 )(One - hot encoding)]]> 5 Aromatic Is the atom part of an aromatic system (0 or 1)? 1 Chiral center Is the atom a chiral center (0 or 1)? 1 Chiral type Chiral type involving atoms (R or S) 2 formal charge The charge (integer) assigned to a single atom in a molecule. 1

[0134] Table 2

[0135] feature describe size key type Bond type (single bond, double bond, triple bond, aromatic bond) (one-hot encoding) 4 Conjugate Whether the bonds are conjugated (0 or 1) 1 ring Is the bond part of a ring (0 or 1)? 1 Stereo keys Bond stereochemistry (none, any, Z / E, cis / trans) (unique thermal encoding) 6

[0136] 4. Machine learning models for AFP were built on datasets for light transmittance, linear thermal expansion coefficient, tensile strength, and glass transition temperature. Model performance evaluation focused on the model's generalization ability to unseen PI datasets, its predictive ability across the entire PI dataset, and the model's training stability. The datasets were randomly split and kept consistent across all models to provide a common basis for comparing model performance. The ratio of training, validation, and test sets was 80:10:10. Evaluation metrics primarily included mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²).

[0137] Among them, light transmittance (T) 400 ), linear thermal expansion coefficient (CLTE), tensile strength (σ) b ) and glass transition temperature (T g The parameters of the AFP model constructed from 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 Drop 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 descriptor to be tested is input into the AFP model to obtain the physical property parameters to be tested. For evaluation on the entire dataset, no retraining is performed; evaluation is only conducted on the complete dataset. The training stability of the model is evaluated after 50 iterations under different initial parameter conditions.

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

[0142] Comparative Example 1

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

[0144] The performance prediction method uses the same steps ss1 and ss2 as in Example 1, employs the same dataset as in Example 1, and uses the same criteria for dividing the training set, validation set, and test set as in Example 1. The method also includes the following steps:

[0145] SS3. Feature engineering was used to select relevant features generated in SS2. First, correlation analysis was used to reduce the number of descriptors from thousands to 141, 158, 145, and 150, respectively. Second, Ridge regression was used for further feature dimensionality reduction to optimize model performance. Ridge regression is a linear model that uses L2 regularization to reduce model complexity, thereby compressing the number of features. During dimensionality reduction, the performance of each attempt was evaluated based on the results of 5x cross-validation. After feature dimensionality reduction using Ridge regression, the descriptors corresponding to the data sets of light transmittance, linear thermal expansion coefficient, tensile strength, and glass transition temperature were reduced to 52, 66, 57, and 63, respectively. The optimal feature combination obtained after feature engineering was used to construct the multilayer perceptron model.

[0146] ss4. Three machine learning models, FraGAT, polyBERT, and descriptor + multilayer perceptron, were constructed on datasets for light transmittance, linear thermal expansion coefficient, tensile strength, and glass transition temperature, respectively.

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

[0148] Application Example 1

[0149] This application example uses 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 to test the training set, validation set, and test set of polyimide (PI) divided in Example 1, and obtains the results of different models for different physical property parameters (transmittance T). 400 Linear thermal expansion coefficient CLTE, tensile strength σ b Glass transition temperature T g The prediction results.

[0150] Figure 4 The image shows the test results of the AFP model in Example 1 for different physical property parameters; Figure 4 (a) is the AFP model for light transmittance T 400 The test results are shown in the image. Figure 4 (b) is a graph showing the test results of the AFP model for the linear thermal expansion coefficient CLTE. Figure 4 (c) is the AFP model for tensile strength σ b The test results are shown in the image. Figure 4 (d) is the AFP model for the glass transition temperature T. g The test results are shown in the image.

[0151] For the training set, regarding the light transmittance T 400 Its MAE and RMSE were 2.41% and 3.52% respectively, R 2 The value is 0.98. For the linear coefficient of thermal expansion CLTE, its MAE and RMSE are 2.83 ppm·K. -1 and 4.19ppm·K -1 R 2 It is 0.95. For tensile strength σ... b Its MAE and RMSE were 4.34 MPa and 6.13 MPa, respectively. 2 It is 0.97. For the glass transition temperature T... g Its MAE and RMSE were 3.09℃ and 4.78℃, respectively, R 2 It is 0.99.

[0152] For the validation set, regarding the light transmittance T 400 Its MAE and RMSE were 3.89% and 3.52% respectively, R 2The value is 0.95. For the linear coefficient of thermal expansion CLTE, its MAE and RMSE are 4.19 ppm·K. -1 and 5.97ppm·K -1 R 2 It is 0.94. For tensile strength σ... b Its MAE and RMSE were 10.88 MPa and 14.03 MPa, respectively. 2 It is 0.86. For the glass transition temperature T... g Its MAE and RMSE were 9.64℃ and 15.61℃, respectively, R 2 It is 0.94.

[0153] For the test set, regarding the light transmittance T 400 Its MAE and RMSE were 3.93% and 5.64% respectively, R 2 The value is 0.93. For the linear coefficient of thermal expansion CLTE, its MAE and RMSE are 5.39 ppm·K. -1 and 7.57ppm·K -1 R 2 It is 0.87. For tensile strength σ... b Its MAE and RMSE were 12.75 MPa and 20.37 MPa, respectively. 2 It is 0.87. For the glass transition temperature T... g Its MAE and RMSE were 9.22℃ and 13.86℃, respectively, R 2 It is 0.95.

[0154] Figure 5 This is a comparison chart of the training results of different models under different performance conditions in Example 1; Figure 5 (a) The light transmittance T of different models 400 The following is a comparison chart of the training results. Figure 5 (b) is a comparison graph of the training results of different models under the linear thermal expansion coefficient CLTE. Figure 5 (c) For different models, the tensile strength σ b The following is a comparison chart of the training results. Figure 5 (d) shows the glass transition temperature T for different models. g The following is a comparison chart of the training results. Figure 5 As can be seen from (a) to (d), the AFP model has the best modeling effect.

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

[0156] Application Example 2

[0157] Using the performance prediction method for polyimides described in Example 1, the light transmittance, linear thermal expansion coefficient, tensile strength, and glass transition temperature of each polyimide listed in the table below were predicted. These values ​​were then compared with their corresponding experimental values, resulting in Tables 4-7. Table 4 compares the experimental and predicted light transmittance (%) values ​​of polyimides; Table 5 shows the linear thermal expansion coefficient (ppm·K) of polyimides. -1 Table 6 compares the experimental and predicted values ​​of polyimide tensile strength (MPa); Table 7 compares the experimental and predicted values ​​of polyimide glass transition temperature (°C).

[0158] The experimental values ​​for 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] As shown in Table 4-7, the performance prediction method in this embodiment has a high accuracy.

[0179] Example 2

[0180] This embodiment discloses a structural design method for polyimide. Figure 6 This is a flowchart of the training process for a reinforcement learning agent designed from scratch using polyimide in this embodiment.

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

[0182] S1. The dianhydride molecule and the diamine molecule are fragmented to obtain a dianhydride structure group containing several dianhydride fragments and a diamine structure group containing several diamine fragments.

[0183] S2. Select dianhydride initiating fragments and diamine initiating fragments from the dianhydride structure group and the diamine structure group, respectively;

[0184] S3. According to the strategy of the neural network, select actions, including: obtaining dianhydride selection fragments and diamine selection fragments from the dianhydride structure group and the diamine structure group respectively through the dianhydride initiation fragment and diamine initiation fragment in step S2, and obtaining dianhydride synthesis fragments and diamine synthesis fragments.

[0185] S4. Determine whether a polyimide agent can be generated through the dianhydride synthesis fragment and the diamine synthesis fragment; if a polyimide agent can be generated, use the AFP model as in Example 1 to calculate the physical property parameters of the polyimide agent and further calculate the reward value; if a polyimide agent cannot be generated, return to step S2 for iterative iteration.

[0186] S5, the steps to terminate near-end strategy optimization.

[0187] Specifically, to minimize the possibility of generating incorrect polyimide molecules, this embodiment employs the Backward Synthesis of Interesting Chemical Substructures (BRICS) method to decompose dianhydride and diamine molecules. The specific process is as follows: 856 dianhydrides and 1759 diamines were first collected from publicly available databases. Then, the collected dianhydrides and diamines were decomposed into 576 and 1004 fragment molecules, respectively, using the BRICS method. Considering the frequency of occurrence of dianhydride and diamine fragments, the highest-ranking fragments were selected as the starting fragments for dianhydride and diamine.

[0188] In step S3, the method of "obtaining dianhydride selection fragments and diamine selection 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 a neural network strategy. Using the initial fragments of dianhydride and diamine in step S2, the neural network strategy will prioritize selecting fragments with higher selection probabilities. During the training of the agent, the neural network strategy will continuously iteratively update the selection probability of each fragment.

[0189] In step S3, the model architecture of the policy is a multilayer perceptron; the number of hidden layers of the policy is 3; the number of neurons in the hidden layers of the policy is 256, 128, and 128; the hidden layers of the policy are fully connected layers; and the activation function of the policy is ReLU.

[0190] In step S3, the dianhydride synthesis fragment is obtained in the following ways: when both the dianhydride starting fragment and the dianhydride selected fragment contain connectors, the dianhydride starting fragment and the dianhydride selected fragment are spliced ​​together to form the dianhydride synthesis fragment; when the dianhydride starting fragment contains a connector and the dianhydride selected fragment does not contain a connector, the dianhydride selected fragment is used as the dianhydride synthesis fragment; when the dianhydride starting fragment does not contain a connector, the dianhydride starting fragment is used as the dianhydride synthesis fragment.

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

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

[0193] In step S5, the termination conditions for proximal strategy optimization are: i. A polyimide agent 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 in the polyimide agent exceeds 120; iv. In step S4, the number of iterations exceeds 5.

[0194] To efficiently identify the optimal combination of fragments for polyimide design, we employ proximal policy optimization (PPO), specifically designed for continuous control tasks, within a reinforcement learning framework. In this context, the agent makes decisions based on a policy, typically represented as a neural network. The policy network computes the probabilities of different actions and a state-value function based on the current state. Essentially, at each time step, an action is probabilistically selected from the action space. The agent's goal is to maximize the expected reward gained through its interaction with the environment.

[0195] This invention maps chemical fragments or molecules in specific states to a 600-dimensional vector as state values ​​using a polyBERT model. The action space is defined as a discrete combination space of dianhydride and diamine fragments. Each action is sampled based on a state-value function and the probability of the expected reward. To design polyimides with target performance, the higher the reward value of a fragment combination, the higher its probability of being sampled. Furthermore, to enhance the effectiveness of polyimide design, the reward for fragment combinations that cannot generate effective molecules is set to zero.

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

[0197] Example 1

[0198] This 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 model and reinforcement learning framework used in this embodiment were employed to train an agent capable of de novo design of polyimides with high light transmittance, low linear coefficient of thermal expansion, high tensile strength, and high glass transition temperature. Results show that the trained agent can design polyimides that meet these requirements. The specific steps are as follows:

[0200] S1. The AFP model trained in Example 1 is used as the performance prediction model for light transmittance, linear coefficient of thermal expansion, tensile strength, and glass transition temperature. Considering synthetic accessibility, this effect example uses the synthetic accessibility score (SAscore) to evaluate the ease of synthesis of 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 coefficient of thermal expansion, high tensile strength, high glass transition temperature, and good synthetic accessibility. Achieving high transparency is the primary objective, and the better the overall performance, the higher the reward value. The expression for the reward value is defined as: R T400 ,R CLTE ,R σb ,R Tg , and R SAscore These represent the transmittance at 400 nm, linear coefficient of thermal expansion, tensile strength, glass transition temperature, and normalized reward value of SAscore, respectively.

[0201] S2. Every 10 training cycles, 10,000 polyimide samples are randomly generated for the corresponding agent to evaluate the agent's learning progress. For comparison, 10,000 random generation cycles are performed before training the agent. Figure 7 The diagram shows the design effect of the reinforcement learning polyimide agent in Example 1. Clearly, as the agent continues reinforcement learning, its design strategy tends towards higher reward values.

[0202] in, Figure 7 (a) shows the reward dynamics (including maximum, minimum, and average values) of the agent throughout the entire training process in this embodiment. Figure 7 (b) Regarding the light transmittance T of the polyimide smart body 400 Design renderings, Figure 7 (b) Showing T 400 The distribution converges to approximately 80% after 70 training epochs; Figure 7 (c) is a design rendering of CLTE, which relates to the linear thermal expansion coefficient of polyimide smart cells. Figure 7 (d) represents the tensile strength σ of the polyimide smart body. b Design renderings Figure 7 (e) represents the glass transition temperature T of the polyimide smart device. g Design renderings, Figure 7 (f) is a design rendering of SAscore, which relates to the synthetic accessibility of polyimide smart agents. Figure 7 (c) to Figure 7(f) shows that the other four performance metrics 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] Compared to the collected literature dataset, the performance distribution of polyimides generated through de novo design showed a significant change, indicating that the agent learned a unique design strategy.

[0204] S3. UMAP was used to reduce the dimension and map the polyimide to a two-dimensional chemical space to explore the designed polyimide. Figure 8 This diagram illustrates the distribution of polyimide agents in the chemical space at different stages in this embodiment. With continuous reinforcement learning, the chemical space of the designed polyimides gradually decreases. Points in these spaces correspond to potential polyimides to be discovered. After 70 training epochs, the polyimides designed by the agents are discretely distributed across multiple regions. To examine the uniqueness of the polyimides within each region, this invention uses unsupervised K-means clustering in this chemical space.

[0205] Figure 9 In this effect embodiment, the number of polyimide clusters designed by the agent after 70 training cycles is determined based on contour scores and inertia, and the number of clusters K is set to 6. Figure 10 This is the chemical spatial distribution of polyimide after K-means clustering into 6 clusters in this effect embodiment. Figure 11 The polyimides with the highest reward values ​​among clusters 1-6 in this effect embodiment and their individual properties are presented. Compared to the chemical space of polyimides in the literature, the agent designed novel types of polyimides (such as clusters 2 and 6). Furthermore, these representative polyimides exhibit excellent overall performance (i.e., Ti). 400 >80%, CLTE <30ppm·K -1 , σ b >100MPa, T g >350℃).

[0206] Figure 12 This is the chemical structure of the polyimide used in the de novo design of the agent in this effect embodiment. Figure 13 This is a polyimide film for de novo design of the agent in this effect embodiment. Figure 14 This example compares the experimental and predicted values ​​of the polyimide film synthesized from the designed agent using experimental methods.

[0207] Figure 15 The frequencies of the fragments obtained by decomposing dianhydrides using the BRICS method in this effect embodiment are shown. Figure 16 The frequency of the fragment obtained by decomposing diamine using the BRICS method in this effect embodiment is shown. Figure 15 , Figure 16The top 24 most frequent dianhydride fragments and the top 24 most frequent diamine fragments are shown respectively. A neural network-based strategy assigns different selection probabilities to all fragments in the dianhydride and diamine structure groups, prioritizing fragments with higher selection probabilities. Although the initial fragments for dianhydrides and diamines are fixed, the different selection probabilities of each fragment in the structure group result in polyimides with distinct structures.

[0208] Example 2

[0209] This embodiment adopts the design method of Example 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 model and reinforcement learning framework used in this embodiment were used to train an agent capable of de novo design of polyimides with low light transmittance, low linear coefficient of thermal expansion, high tensile strength, and high glass transition temperature. Results show that the trained agent can design polyimides that meet these requirements. The specific steps are as follows:

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

[0212] S2. Every 10 training cycles, 10,000 polyimide samples are randomly generated for the corresponding agent to evaluate the agent's learning progress. For comparison, 10,000 random generation cycles are performed before training the agent. Figure 17 The diagram shows the design effect of the reinforcement learning polyimide agent in Example 2. Clearly, as the agent continues reinforcement learning, its design strategy tends towards higher reward values.

[0213] in, Figure 17 (a) The reward dynamics (including maximum, minimum, and average values) of the agent throughout the entire training process in Example 2. Figure 17 (b) Regarding the light transmittance T of the polyimide smart body 400 The design renderings converged to approximately 0% after 70 training cycles. Figure 17 (c) is a design rendering of CLTE, which relates to the linear thermal expansion coefficient of polyimide smart cells. Figure 17 (d) represents the tensile strength σ of the polyimide smart body. b Design renderings Figure 17 (e) represents the glass transition temperature T of the polyimide smart device. g Design renderings, Figure 17(f) is a design rendering of SAscore, which relates to the synthetic accessibility of polyimide smart agents. Figure 17 (c)~ Figure 17 (f) shows that the other four performance metrics 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] Compared to the collected literature dataset, the performance distribution of polyimides generated through de novo design showed a significant change, indicating that the agent learned a unique design strategy.

[0215] S3. This invention uses UMAP dimensionality reduction to map it to a two-dimensional chemical space to explore the designed polyimide. Figure 18 This diagram shows 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 polyimides gradually decreases. Points in these spaces correspond to potential polyimides to be discovered. After 70 training epochs, the polyimides designed by the agents are discretely distributed across multiple regions.

[0216] Figure 19 In this effect embodiment, the number of polyimide clusters designed for the agent after 70 training cycles is determined based on contour scores and inertia. Figure 20 This is the chemical spatial distribution of polyimide after being clustered into 3 clusters using K-means in this effect embodiment. Figure 21 The polyimides with the highest reward values ​​in clusters 1-3 of this effect example and their individual properties are identified. The same clustering method as in effect example 1 was used, with the cluster number K set to 3. These representative polyimides exhibit excellent overall performance (i.e., Ti). 400 <10%, CLTE <20ppm·K -1 , σ b >150MPa, Tg>350℃).

Claims

1. A method for structural design of polyimide, characterized in that, It includes: Based on proximal policy optimization, an agent generating polyimide is trained; the proximal policy optimization includes the following steps: S1. The dianhydride molecule and the diamine molecule are fragmented to obtain a dianhydride structure group containing several dianhydride fragments and a diamine structure group containing several diamine fragments; the fragmentation method includes a decomposition method for retrospectively synthesizing interesting chemical substructures. S2. Select dianhydride initiating fragments and diamine initiating fragments from the dianhydride structure group and the diamine structure group, respectively. The selection method includes: extracting groups from the dianhydride molecule and the diamine molecule, and converting them into a computer-recognizable 600-dimensional vector using the polyBERT model. The selection rule for the dianhydride initiating fragment and the diamine initiating fragment is: select the dianhydride initiating fragment and the diamine initiating fragment with the highest occurrence frequency as the unique initiating fragment. S3. According to the neural network strategy, select an action, the action including: obtaining a dianhydride selection fragment and a diamine selection fragment from the dianhydride structure group and the diamine structure group respectively, using the dianhydride initiation fragment and the diamine initiation fragment from step S2, and obtaining a dianhydride synthesis fragment and a diamine synthesis fragment; the method of obtaining the dianhydride synthesis fragment includes: when both the dianhydride initiation fragment and the dianhydride selection fragment contain connectors, concatenating the dianhydride initiation fragment and the dianhydride selection fragment to form the dianhydride synthesis fragment; when the dianhydride initiation fragment contains a connector and the dianhydride selection fragment does not contain a connector, ... The dianhydride selected fragment is used as the dianhydride synthesis fragment; when the dianhydride initiating fragment does not contain a connector, the dianhydride initiating fragment is used as the dianhydride synthesis fragment; the diamine synthesis fragment is obtained by: when both the diamine initiating fragment and the diamine selected fragment contain connectors, splicing the diamine initiating fragment and the diamine selected fragment to form the diamine synthesis fragment; when the diamine initiating fragment contains a connector and the diamine selected fragment does not contain a connector, the diamine selected fragment is used as the diamine synthesis fragment; when the diamine initiating fragment does not contain a connector, the diamine initiating fragment is used as the diamine synthesis fragment. S4. Determine whether the polyimide smart agent can be generated through the dianhydride synthesis fragment and the diamine synthesis fragment; if the polyimide smart agent can be generated, use the AFP model in the performance prediction method to calculate the physical property parameters of the polyimide smart agent, and further calculate the reward value; if the polyimide smart agent cannot be generated, return to step S2 for iterative iteration; the physical property parameters of the polyimide smart agent include light transmittance, linear thermal expansion coefficient, tensile strength, and glass transition temperature; The performance prediction method includes the following steps: The first molecular descriptor is input into the AFP model to obtain the physical property parameters to be tested; the first molecular descriptor is the molecular descriptor of the polyimide to be tested; The AFP model is trained using a training dataset, which includes a second molecular descriptor and a first physical property parameter; the second molecular descriptor and the first physical property parameter are the molecular descriptor and physical property parameter of a known polyimide, respectively. Both the first molecular descriptor and the second molecular descriptor include node features and edge features, and the types of the first molecular descriptor and the second molecular descriptor are the same. The type of the physical property parameter to be measured is the same as the type of the first physical property parameter; S5. Terminate the near-end strategy optimization step.

2. The polyimide structural design method as described in claim 1, characterized in that, The performance prediction method satisfies one or more of the following conditions: ①The type of the physical property parameter to be measured and the type of the first physical property parameter are each independently light transmittance, linear thermal expansion coefficient, tensile strength or glass transition temperature; ②The node features are one or more of the following: atom type, number of bond connections, number of hydrogen atom connections, displayed valence, hybridization, aromaticity, chiral center, chiral type, and formal charge; ③The edge features are one or more of the following: bond type, conjugation, loop structure, and 3D 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.

3. The polyimide structural design method as described in claim 2, characterized in that, The dimensions of the hidden layer are 16-128.

4. The polyimide structural design method as described in claim 3, characterized in that, The dimensions of the hidden layer are 87, 97, 111, or 116.

5. The polyimide structural design method as described in claim 2, characterized in that, The initial learning rate is 10 -3 -10 -1 .

6. The polyimide structural design method as described in claim 5, characterized in that, The initial learning rate is 10 -3 , 10 -2.26 , or 10 -1.92 .

7. The polyimide structural design method as described in claim 2, characterized in that, The learning rate decay factor is 0.4-0.

9.

8. The method for structural design of polyimide as described in claim 7, characterized in that, The learning rate decay factor is 0.6, 0.8, or 0.

9.

9. The method for structural design of polyimide as described in claim 2, characterized in that, The weight decays to 0-0.

4.

10. The method for structural design of polyimide as described in claim 9, characterized in that, The weight decay is 0.

11. The polyimide structural design method as described in claim 2, characterized in that, The discard rate is 0-0.4%.

12. The method for designing the structure of polyimide as described in claim 11, characterized in that, The discard rate is 0.1, 0.2, 0.25, or 0.

35.

13. The polyimide structural design method as described in claim 2, characterized in that, The number of node-level embedding layers is 1-6.

14. The polyimide structural design method as described in claim 13, characterized in that, The number of node-level embedding layers is 1, 2, or 3.

15. The method for structural design of polyimide as described in claim 2, characterized in that, The number of graph-level embedding layers is 1-4.

16. The method for structural design of polyimide as described in claim 15, characterized in that, The number of graph-level embedding layers is 1, 2, or 4.

17. The method for structural design of polyimide as described in claim 1, characterized in that, The performance prediction method satisfies one or more of the following conditions: ① The first molecule descriptor and the second molecule descriptor are generated using RDKit or Mordred, respectively; ②The first molecule descriptor and the second molecule descriptor are generated in the same way; ③ The training dataset has been cleaned, and the cleaning process includes: i. excluding polyimides containing metal elements and / or ionic compounds; ii. excluding polyimide copolymers and blends; iii. when polyimides with the same structure have multiple data records with the same physical property parameters, taking the average value of each data record.

18. The method for structural design of polyimide as described in claim 1, characterized in that, The structural design method satisfies one or more of the following conditions: i. In step S3, the model architecture of the strategy is a multilayer perceptron; ii. In step S3, the number of hidden layers of the strategy is 1-5; iii. In step S3, the number of hidden layer neurons in the strategy is 16-512; iv. In step S3, the hidden layer of the strategy is a fully connected layer; v. In step S3, the activation function of the strategy is ReLU, LeakyReLU, or Tanh; vi. The termination condition for the near-end strategy optimization is any one of the following: a. The 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 agent of the polyimide exceeds 50; d. In step S4, the number of iterations exceeds 5.

19. The method for structural design of polyimide as described in claim 18, characterized in that, The structural design method satisfies one or more of the following conditions: i. In step S3, the number of hidden layer neurons in the strategy is 16, 32, 64, 128, 256 or 512. ii. In step S3, the activation function of the strategy is ReLU; iii. The termination condition for the near-end strategy optimization is any one of the following: a. The number of atoms of the polyimide smart agent exceeds 120; b. In step S4, the number of loop iterations exceeds 10.

20. The method for structural design of polyimide as described in claim 1, characterized in that, The expression of the reward value is R = R T400 × (1 + R CLTE + R σb + R Tg + R SAscore ) / 5; wherein R T400 , R CLTE , R σb , R Tg , R SAscore respectively represent the normalized reward values of 400 nm transmittance, linear thermal expansion coefficient, tensile strength, glass transition temperature, and synthetic accessibility SAscore.

21. A structural design system for polyimide, characterized in that, It includes a structural design module for training and generating polyimide-based agents based on proximal policy optimization; The proximal strategy optimization includes the following steps: S1, fragmenting dianhydride and diamine molecules to obtain a dianhydride structure group containing several dianhydride fragments and a diamine structure group containing several diamine fragments; S2, selecting dianhydride initiating fragments and diamine initiating fragments from the dianhydride structure group and diamine structure group, respectively; S3, selecting actions according to the neural network strategy, the actions including obtaining dianhydride selection fragments and diamine selection fragments from the dianhydride structure group and diamine structure group, respectively, using the dianhydride initiating fragments and diamine initiating fragments from step S2, and obtaining dianhydride synthesis fragments and diamine synthesis fragments; S4, determining whether the polyimide agent can be generated using the dianhydride synthesis fragments and diamine synthesis fragments; when the polyimide agent can be generated, using the AFP model in the performance prediction system, calculating the physical property parameters of the polyimide agent, and further calculating the reward value; when the polyimide agent cannot be generated, returning to step S2 for iterative iteration; S5, terminating the proximal strategy optimization steps; In step S1, the fragmentation process includes a decomposition method for retrospectively synthesizing interesting chemical substructures; In step S2, the selection method includes: extracting groups from the dianhydride molecule and the diamine molecule, and converting them into a computer-recognizable 600-dimensional vector using the polyBERT model; the selection rule for the dianhydride initiation fragment and the diamine initiation fragment is: based on the occurrence frequency of the initiation fragment, the dianhydride initiation fragment and the diamine initiation fragment with the highest occurrence frequency are selected as the unique initiation fragment; In step S3, the method for obtaining the dianhydride synthesis fragment includes: when both the dianhydride starting fragment and the dianhydride selected fragment contain connectors, splicing the dianhydride starting fragment and the dianhydride selected fragment together 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. The method for obtaining the diamine synthesis fragment includes: when both the diamine starting fragment and the diamine selected fragment contain connectors, splicing the diamine starting fragment and the diamine selected fragment together to form the diamine synthesis fragment; when the diamine starting fragment contains a connector and the diamine selected fragment does not contain a connector, using the diamine selected 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. In step S4, the physical properties of the polyimide smart body include light transmittance, linear thermal expansion coefficient, tensile strength, and glass transition temperature. The polyimide performance prediction system includes a prediction module for inputting a first molecular descriptor into an AFP model to obtain the physical property parameter to be tested. The first molecular descriptor is a molecular descriptor of the polyimide to be tested. The AFP model is trained using a training dataset, which includes a 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 physical property parameter of a known polyimide. Both the first molecular descriptor and the second molecular descriptor include node features and edge features. 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 tested is the same as the type of the first physical property parameter.

22. The polyimide structural design system as described in claim 21, characterized in that, The expression for 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 These represent the normalized reward values ​​of SAscore for transmittance at 400 nm, linear coefficient of thermal expansion, tensile strength, glass transition temperature, and synthetic accessibility, respectively.

23. The polyimide structural design system according to claim 21, characterized in that, The structural design module is used in the structural design method of the polyimide as described in any one of claims 1-20.

24. 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, it implements the structural design method for polyimide as described in any one of claims 1-20.

25. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the structural design method for polyimide as described in any one of claims 1-20.

26. A computer program product comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the structural design method for polyimide as described in any one of claims 1-20.

Images