A deep learning-based specific glass transition temperature polyimide repeat unit generation method

Abstract

This invention discloses a deep learning-based method for generating polyimide repeating units with a specific glass transition temperature. By constructing a polyimide repeating unit dataset, SMILES are converted to SELFIES representations and random equivalent enumeration enhancement is performed to calculate the statistical parameters of the glass transition temperature and the boundary values ​​of the bucket intervals. A conditional autoregressive generation model is designed, employing a dual-path conditional representation method that parallels continuous scalar encoding paths and discrete bucket encoding paths. The target glass transition temperature is simultaneously mapped to a standardized continuous vector and a discrete interval embedding vector. The two vectors are fused to form a conditional memory representation, driving a Transformer-based decoder to generate candidate repeating unit SELFIES sequences bit by bit in an autoregressive manner. During inference, the target temperature is input to generate candidate structures that are decoded into SMILES, and candidate polyimide repeating units that satisfy the constraints are selected.

Description

Technical Field

[0001] This invention relates to the fields of intelligent design of polymer materials, generative artificial intelligence-assisted material discovery, and performance-oriented molecular generation technology of polyimide, and particularly to a method for generating repeating polyimide units at specific glass transition temperatures based on deep learning. Background Technology

[0002] Polyimide materials are widely used in aerospace, electronic packaging, flexible devices, and high-temperature separation membranes due to their excellent heat resistance, dielectric properties, mechanical properties, and chemical stability. In the development of polyimides, the glass transition temperature (Tg) is one of the crucial performance indicators determining the material's service temperature window and structural stability. In actual research and development, it is necessary to conduct long-term trial and error, empirical enumeration, and experimental screening of repeating unit structures around the target Tg, resulting in long design cycles, high experimental costs, and limitations imposed by existing experience.

[0003] In recent years, generative artificial intelligence methods have been gradually introduced into the fields of molecular design and materials discovery, enabling the automatic exploration of unknown structural spaces through statistical learning of existing structural samples. However, for target material systems such as polyimide, existing general molecular generation methods typically suffer from the following problems: First, the generated objects are mostly general small molecules, lacking structural constraints specific to polyimide repeating units; second, target performance conditions are often input only in a simple scalar manner, making it difficult to simultaneously express fine continuous variations and performance range semantics; third, although the generated results may be valid at the string level, they may not meet the requirements of the polyimide framework, engineering dimensions, or basic stability.

[0004] Therefore, how to establish a reverse design method that can both guide the conditions around the target glass transition temperature, ensure that the generated result has the structural characteristics of polyimide repeating units, and output reasonable candidate structures through subsequent rule screening remains a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a method for generating repeating units of polyimide with a specific glass transition temperature based on deep learning, in order to solve the problems of excessive reliance on human experience, insufficient expression of performance conditions, poor matching between the generated results and the target material system, and low efficiency in candidate structure screening in the existing polyimide design process.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for generating polyimide repeating units with a specific glass transition temperature based on deep learning. This method preprocesses the original data of high-performance polymer repeating units, constructs a conditional autoregressive generation model with SELFIES sequences as the generation object, maps the target glass transition temperature into a continuous conditional representation and a discrete interval conditional representation, and generates candidate polymer repeating unit structures bit by bit based on the conditional memory vector formed by fusing the continuous and discrete interval conditional representations. The generated candidate polymer repeating unit structures are then subjected to chemical resolvability verification, target skeleton structure rule verification, and molecular property stability screening, ultimately outputting candidate polymer repeating units that satisfy the rule constraints.

[0007] Raw data of high-performance polymer repeating units containing repeating unit structures and glass transition temperatures were collected, and missing values ​​were removed, structures were normalized, invalid structures were removed, and duplicate samples were removed. The normalized SMILES representation was converted into SELFIES representation, and the training set and validation set were divided according to a preset ratio. Random equivalent enumeration enhancement was performed only on the training set. A SELFIES vocabulary was constructed and the mean, standard deviation, and binning interval boundary values ​​of the glass transition temperatures were statistically analyzed.

[0008] The mapping constructs a dual-path conditional representation model of the target glass transition temperature, simultaneously mapping the target glass transition temperature into a standardized continuous scalar and a discrete bucket number. The standardized continuous scalar is input into a multilayer perceptron through a continuous scalar encoding path to obtain a continuous conditional vector, and the discrete bucket number is input into an embedding layer through a discrete bucket encoding path to obtain a discrete conditional vector. The two conditional vectors are then fused to form a conditional memory representation for subsequent generation model calls. Standardized continuous scalar Calculate using the following formula: ; T g Indicates the target glass transition temperature. This represents the average glass transition temperature of the training set. This represents the standard deviation of the glass transition temperature of the training set.

[0009] The conditional autoregressive generative model employs a conditional encoding method that combines continuous and discrete paths. Standardized continuous scalars are input into a multilayer perceptron, and discrete bucket numbers are input into the embedding layer. The two conditional vectors are fused to form a conditional memory representation. Then, a Transformer-based sequence decoding unit, guided by this conditional memory representation, generates a high-performance polymeric repeating unit SELFIES sequence. The fusion of the conditional vectors from the continuous and discrete paths is achieved through the following formula: Where z represents a normalized continuous scalar, c represents the discrete barrel number corresponding to the target glass transition temperature, MLP() represents a continuous conditional coding network, and Emb() represents a discrete interval embedding layer. The fused condition vector is further extended into a condition memory sequence of length 1 and input into the condition memory encoding unit.

[0010] The sequence decoding unit of the conditional autoregressive generative model adopts a Transformer-based decoding architecture. It uses a representation that adds token embeddings and learnable position embeddings to the input SELFIES token sequence. In each generation step, it simultaneously performs mask self-attention calculation on the generated SELFIES prefix sequence and cross-attention calculation on the conditional memory sequence to predict the probability distribution of the next position token.

[0011] The output of the conditional autoregressive generative model includes a layer normalization layer and an output projection layer. The weights of the output projection layer are shared with the weights of the token embedding layer. The model training adopts a teacher-forced approach, using the real structure prefix sequence and the target glass transition temperature condition as inputs, and the next position token as the supervision signal. The model parameters are optimized by ignoring the cross-entropy loss of the padding bit.

[0012] From the initial label during the reasoning phase <bos>The process begins by generating SELFIES tokens one by one. First, the vocabulary probabilities output from the current generation step are temperature-scaled. Then, the tokens are sorted from highest to lowest probability. The set of tokens with the smallest cumulative probability reaching a preset threshold is retained. Within this retained set, tokens are re-normalized and then randomly sampled until a termination token is generated. <eos>Or it may reach a preset maximum length; the temperature scaling is calculated according to the following formula: ; in, This represents the output logit of the j-th token in the current generation step. Indicates the current generation step word in the word list. The output logit for each token This is the index of each candidate token in the vocabulary; Indicates the temperature coefficient. This represents the probability of the corresponding token after temperature scaling; subsequently, random sampling is performed only within a subset of tokens that meet the cumulative probability threshold condition.

[0013] After decoding the generated SELFIES sequences into SMILES, a multi-level structure screening mechanism is constructed, including: using cheminformatics tools to determine whether the SMILES can be successfully resolved into molecular objects; judging whether the candidate structures meet the structural requirements of the set high-performance polymer repeating units based on the target skeleton pattern; and performing heuristic filtering of candidate structures based on molecular weight range, heavy atom number range, and predefined unstable group patterns.

[0014] In the multi-level structure screening mechanism, the structure rule verification can be configured with one or a combination of the following skeleton modes for different polymer systems: (1) Polyimide: imide skeleton -C(=O)-NC(=O)- and five-membered imide ring; (2) Polybenzoxazole: benzoxazole ring and corresponding precursor hydroxyamide structure; (3) Polybenzimidazole: benzimidazole ring skeleton; (4) Polyamide imide: amide bond and imide ring coexisting structure; The molecular property stability screening includes at least the following constraints: Where MW represents molecular weight, with units of g / mol. The range indicates the number of heavy atoms; the above range is determined based on the typical structural characteristics of polyimide repeating units, with the lower limit of molecular weight corresponding to the smallest effective imide unit and the upper limit covering complex repeating units containing multiple aromatic rings and bridging groups.

[0015] The trained conditional autoregressive generative model generates multiple candidate polymer structure repeating units for the input target glass transition temperature, and outputs the selected candidate structures and their corresponding structural identifiers, SMILES and SELFIES information.

[0016] Compared with the prior art, the present invention has at least the following beneficial effects: This invention improves the ability to express target performance conditions by simultaneously introducing continuous scalar encoding paths and discrete bucket encoding paths, enabling the generative model to be more structurally guided around the target glass transition temperature. This invention generates polyimide repeating units in the SELFIES sequence space, reducing the proportion of invalid structures and improving the resolvability of candidate structures; This invention introduces a multi-level screening process based on imide structure rules and fundamental stability rules after generation, which significantly improves the consistency between candidate structures and the target material system. This invention forms an integrated reverse design process of "target performance input - condition generation - structure screening - candidate output", which can be used for polyimide lead structure mining and candidate output. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall process of this invention. Figure 2 A schematic diagram of the dual-path conditional autoregressive generative model structure for target Tg; Figure 3 Flowchart for multi-level structured filtering of results; Figure 4 A schematic diagram showing the candidate structure pass rate under the criteria at 200℃, 300℃, and 400℃. Figure 5 (a) shows a schematic diagram of the generated polyimide candidate structure; (a)-(f) show schematic diagrams of different candidate structures. Detailed Implementation

[0018] A method for generating repeating polyimide units at specific glass transition temperatures based on deep learning includes the following steps: Step S1: Construct a high-performance polymer repeating unit original dataset, extract the repeating unit structure and corresponding glass transition temperature from the original data, delete missing samples, normalize the structure and remove invalid or duplicate samples to obtain the cleaned basic dataset, and divide it into training set, test set and validation set according to 8:1:1. Step S2 involves converting the cleaned repeating unit structure from SMILES representation to SELFIES representation, and augmenting the structure samples in the training set using random equivalent enumeration. Based on the augmented training set, SELFIES tokens are statistically analyzed to construct a structure containing... <pad>(Aligning sequences of different lengths within a batch during training) <bos>(The starting point of rational time) <eos>(The generated termination signal) and <unk>A structured vocabulary of rare tokens not covered by the training vocabulary during inference; Step S3: Calculate the mean, standard deviation, minimum, maximum and a preset number of temperature bucket boundaries based on the glass transition temperature values ​​in the training set. In subsequent training and inference processes, the target Tg is simultaneously converted into a standardized continuous value and a discrete bucket number. Step S4: Construct a conditional autoregressive generative model, including a continuous scalar coding path, a discrete bucket coding path, a conditional memory coding unit, and a sequence decoding unit; Among them, the continuous scalar encoding path and the discrete bucket encoding path perform vector mapping on the normalized Tg and Tg bucket number respectively. After the two vectors are added together and normalized, a fusion condition vector is obtained. The fusion condition vector further forms a condition memory representation, which is called by the sequence decoding generation unit through cross attention during the generation process. Step S5: Using the SELFIES token sequence as the generation target, the conditional autoregressive generation model is trained. During training, a teacher-forced method is used, with the current token prefix sequence and the target Tg condition as inputs, and the next position token as the supervision signal. The model parameters are optimized through cross-entropy loss so that the model learns the probability distribution of the polyimide repeating unit structure under the target Tg condition. Step S6: In the inference phase, input the target glass transition temperature, first standardize it using the statistical parameters saved in the training phase and map it to bucket numbers, then... <bos>Using this as a starting marker, SELFIES tokens are generated bit-by-bit through an autoregressive sampling strategy until the output is reached. <eos>Or reach the maximum length to obtain a candidate polyimide repeating unit structure; Step S7: Decode the generated SELFIES sequence into SMILES structure and perform multi-level structure screening in sequence, including: molecular resolvability verification based on chemical toolkit, structural rule verification based on imide skeleton and five-membered imide ring pattern, and heuristic screening based on molecular weight range, heavy atom number range and predefined unstable group pattern. Step S8: Output the candidate polyimide repeating units that have passed the multi-level structure screening, and record the corresponding structure identifier, SMILES representation, SELFIES representation, and structure screening index. In the above technical solution, the continuous scalar encoding path preferably uses a multilayer perceptron to map the standardized Tg, and the discrete bucket encoding path preferably uses an embedding layer to map the interval to which Tg belongs, so as to simultaneously retain the fine numerical information of the target performance and the coarse-grained interval semantics. In the above technical solution, the sequence decoding generation unit preferably adopts a Transformer decoder structure, which includes a mask self-attention module and a cross-attention module. The former is used to model the internal dependency relationship of the generated structure prefix, and the latter is used to introduce the target Tg condition memory to realize the performance condition to guide the structure generation process bit by bit. In the above technical solution, the autoregressive sampling strategy in the inference stage is preferably a random sampling strategy, which retains a subset of tokens with a cumulative probability quality not lower than a preset threshold and performs random sampling within the subset, so as to take into account both the rationality and diversity of the candidate structure. In the above technical solution, the multi-level structure screening preferably adopts a three-layer verification mechanism. The first layer is molecular resolvability verification based on RDKit, the second layer is structural rule verification based on imide skeleton and five-membered imide ring, and the third layer is heuristic filtering based on molecular weight, number of heavy atoms and list of prohibited unstable groups. To make the technical means, innovative features and objectives of this invention clearer, the invention will be further explained below with reference to the figures and specific implementation. like Figure 1 As shown, this invention proposes a deep learning-based method for generating polyimide repeating units at a specific glass transition temperature. This framework aims to achieve reverse generation of polyimide SMILES formulas at a given glass transition temperature, and the process includes three key stages. Phase 1: Data Preprocessing and Vocabulary Construction. The historical polyimide repeating unit structure and its glass transition temperature data are normalized and cleaned. SMILES representations are converted to SELFIES representations. Random equivalent enumeration augmentation is performed on the training set. A structural vocabulary is statistically constructed, and statistical parameters of the glass transition temperature and bucket boundaries are calculated. The second stage: target performance conditional encoding and autoregressive generation. The target glass transition temperature is simultaneously mapped to continuous numerical embedding and discrete interval embedding, which are fused to form a conditional memory vector, driving a Transformer-based decoder to generate candidate repeating unit SELFIES sequences bit by bit through autoregression; The third stage: multi-level structure screening and candidate output. The generated SELFIES sequences are decoded into SMILES, and then subjected to chemical resolvability verification, imide backbone rule verification, and molecular property and stability screening in sequence, outputting candidate polyimide repeating units that meet the rule constraints.

[0019] The data preprocessing and vocabulary construction process of the present invention includes the following steps: Step S1: Collect sample data containing repeating unit structures and glass transition temperatures from polyimide literature databases and related public datasets. Use cheminformatics tools (RDKit) to normalize the representation of repeating units SMILES, remove invalid structures and duplicate samples, and obtain the cleaned basic dataset. Divide it into training set, validation set and test set in a ratio of 8:1:1. Step S2, SELFIES Transformation and Training Set Augmentation: The cleaned repeating unit structure is transformed from SMILES representation to SELFIES representation. Random equivalent enumeration augmentation is performed only on repeating units in the training set to extend the robustness of the model to different writing forms of the same structure. Step S3, Vocabulary Construction: Based on the enhanced training set, count all occurrences of the SELFIES tokens to construct a structured vocabulary. The vocabulary must contain at least the padding tokens. <pad>, start mark <bos>Termination mark <eos>and unknown markers <unk>And convert each training sample into <bos>beginning, <eos>The final token sequence; Step S4: Calculation of temperature statistical parameters: Calculate the mean value based on the glass transition temperature values ​​in the training set. and standard deviation : Where N is the number of samples in the training set. Let be the glass transition temperature of the i-th sample, and simultaneously record the minimum temperature of the training set. and ; The number of temperature chambers is K. The equally spaced chamber boundaries are constructed using the following formula: For any target temperature Tg, the number of discrete barrels C is determined according to the following formula: Where, digitize means to locate the temperature range according to the boundary array, and clip means to truncate the number to the legal range [0, K-1]. In this invention, K=20. like Figure 2 The diagram shows the structure of the target Tg continuous scalar encoding path and discrete bucket encoding path of the present invention, including a target performance conditional encoding unit, a conditional memory encoding unit, and a sequence decoding generation unit. The specific implementation includes the following steps: Step S1: Continuous scalar encoding. The target glass transition temperature Tg is standardized using the statistical parameters obtained in step S4 to obtain a continuous scalar input. Where Tg represents the target glass transition temperature. This represents the average glass transition temperature of the training set. The z-axis represents the standard deviation of the glass transition temperature of the training set. Then, z is input into a two-layer perceptron to obtain a continuous conditional vector. : Here, GELU represents the activation function.

[0020] Step S2: Discrete bucket encoding. Input the discrete bucket number c obtained in step S5 into the embedding layer to obtain the discrete condition vector. : Where Emb() is the learnable embedding matrix, and the embedding dimension is the same as the output dimension of the continuous path, which is 1. ; Step S3, Dual-path fusion and conditional memory encoding, involves summing and fusing continuous and discrete conditional vectors and then normalizing them: Then Expanded into a sequence of length 1, and refined by the Transformer Encoder, the conditional memory M is obtained: in, Conditional memory for decoder cross-attention invocation, in this invention =256, encoder layer count num_enc_layers=2; Step S4, Sequence Decoding Generation Unit, assuming the target SELFIES token sequence is... Each token is first mapped to the sum of its token embedding and its learnable position embedding: Where E() is the token embedding matrix, For positional embedding, the sequence prefix is ​​input into a Transformer decoder with a causal mask, and combined with the conditional memory M, to obtain the hidden state for each position: The decoder performs mask self-attention and cross-attention computation simultaneously at each generation step to predict the conditional probability distribution of the token at the next position: in, This represents the hidden state of the decoder at position t. To output the projection matrix, and = In this invention, the number of decoder layers num_dec_layers=4, the number of multi-head attention heads nhead=8, and the feedforward layer dimension... =1024, dropout is 0.1, maximum sequence length is 300; Step S5: Model training. The input and supervision sequences are constructed using a teacher-forced method, where the last token is removed from the input sequence and the first token is removed from the supervision sequence. For the set of unfilled positions Ω, the loss function is: Where |V| is the vocabulary size, Let i be the probability that sample i is the v-th token in the vocabulary at position t. To supervise the distribution, this invention uses a label smoothing coefficient of label_smoothing=0.1; Step S6: Use the AdamW optimizer with an initial learning rate of 1×10⁻⁶. -4 The weight decay coefficient is 1×10 -2 The gradient clipping threshold is 1.0. The learning rate scheduling employs a strategy combining warm-up and cosine annealing. The total number of training steps is set to... The number of preheating steps is Then the learning rate scaling factor λ(s) is: like Figure 3 The diagram shown is a flowchart of the multi-level structure screening process generated by this invention. To ensure the chemical validity and material system consistency of the generated candidate structures, the multi-level structure screening process includes the following steps: Step S1, SELFIES Decoding: Decode the SELFIES sequence generated during the inference stage into SMILES representation using the SELFIES library, and use it as input for subsequent filtering; Step S2, First-level screening: Molecular resolvability verification. Use the RDKit chemical toolkit to determine whether the decoded SMILES can be successfully used to construct a molecular object. Structures that cannot be resolved are discarded directly. Step S3, Second-level screening: Imidine skeleton rule verification. For candidate structures that pass the first-level screening, use SMARTS substructure search to determine whether they meet the following two rules: An imide skeleton mode exists [#6](=O)-[#7]-[#6](=O); The structure containing a five-membered imide ring is discarded if any rule is not satisfied. Step S4, Third-level screening: Molecular property and stability screening. For candidate structures that have passed the second-level screening, the following constraint verifications are performed sequentially: Molecular weight range screening: Where MW is the molecular weight (unit: g / mol). The lower limit of 150 g / mol ensures the minimum chemical requirement that the repeating unit contains a complete imide ring structure, while the upper limit of 2000 g / mol covers complex repeating units containing polyaromatic rings and bridging groups, while ensuring the applicability of the Tg prediction model and the feasibility of synthesis; Screening for heavy atom number range: ,in The number of heavy atoms; Unstable group blacklist filtering: SMARTS search is performed on the candidate structure to check whether it contains predefined unstable groups. The unstable groups include, but are not limited to, nitro, azide, peroxide, nitrite, heavy metal ligand, epoxy peroxide, continuous triazine and thiourea, etc. Structures containing any unstable group are discarded, and structures that pass all three constraints enter the output stage. Candidate structure output: Retain candidate polyimide repeating units that have passed all three layers of screening, and record their structural information such as structural identifier, SMILES and SELFIES representation.

[0021] like Figure 4 As shown, to verify the effectiveness of the present invention, 1000 candidate structures were generated at each of the three temperature points with target Tg of 200 ℃, 300 ℃ and 400 ℃.

[0022] Experimental results show that in the experiment with a target Tg of 200 ℃, all 1000 candidates yielded resolvable structures, with a resolvability rate of 100%. Among them, 737 candidates were further screened through a three-layer rule, with a pass rate of 73.7%. In the experiments with target Tg of 300 ℃ and 400 ℃, the resolvability rates were 100% and 99.6%, respectively, and the rule pass rates were 76.7% and 71.8%, respectively.

[0023] like Figure 5 The diagram shown is a schematic of the polyimide structure generated by the design model of this invention.

[0024] The above results show that the present invention can stably generate candidate repeating units with polyimide structural characteristics under different target performance guidance, and the generated structure exhibits excellent performance in terms of chemical efficiency and material system consistency.

[0025] The features of the present invention have been described in detail above. Although the present invention has been described with reference to the above embodiments, the present invention is not limited thereto. Various changes and variations can be made without departing from the spirit of the present invention and the scope of the appended claims. Those skilled in the art will understand that, without departing from the core idea of ​​the present invention, adaptive adjustments can be made to the conditional input variables, the number of temperature bins, the dimension of hidden layers, the number of attention heads, the number of encoder layers, the number of decoder layers, the dimension of feedforward layers, the upper limit of sequence length, the filtering rule set, the candidate output format, and the sampling strategy. These equivalent substitutions or variations should all fall within the protection scope of the present invention.< / eos> < / bos> < / unk> < / eos> < / bos> < / pad> < / eos> < / bos> < / unk> < / eos> < / bos> < / pad> < / eos> < / bos>

Claims

1. A method for generating repeating polyimide units at specific glass transition temperatures based on deep learning, characterized in that, The raw data of high-performance polymer repeating units are preprocessed, and a conditional autoregressive generation model is constructed with SELFIES sequences as the generation object. The target glass transition temperature is mapped into a continuous conditional representation and a discrete interval conditional representation. Based on the conditional memory vector formed by the fusion of the continuous conditional representation and the discrete interval conditional representation, candidate polymer repeating unit structures are generated bit by bit. The generated candidate polymer repeating unit structure results are subjected to chemical resolvability verification, target skeleton structure rule verification, and molecular property stability screening. Finally, candidate polymer repeating units that meet the rule constraints are output.

2. The method for generating repeating polyimide units with a specific glass transition temperature based on deep learning according to claim 1, characterized in that, Raw data of high-performance polymer repeating units containing repeating unit structures and glass transition temperatures were collected, and missing values ​​were removed, structures were normalized, invalid structures were removed, and duplicate samples were removed. The normalized SMILES representation was converted into SELFIES representation, and the training set and validation set were divided according to a preset ratio. Random equivalent enumeration enhancement was performed only on the training set. A SELFIES vocabulary was constructed and the mean, standard deviation, and binning interval boundary values ​​of the glass transition temperatures were statistically analyzed.

3. The method for generating repeating polyimide units with a specific glass transition temperature based on deep learning according to claim 2, characterized in that, The mapping constructs a dual-path conditional representation model of the target glass transition temperature, simultaneously mapping the target glass transition temperature into a standardized continuous scalar and a discrete bucket number. The standardized continuous scalar is input into a multilayer perceptron through a continuous scalar encoding path to obtain a continuous conditional vector, and the discrete bucket number is input into an embedding layer through a discrete bucket encoding path to obtain a discrete conditional vector. The two conditional vectors are then fused to form a conditional memory representation for subsequent generation model calls. Standardized continuous scalar Calculate using the following formula: ； T g Indicates the target glass transition temperature. This represents the average glass transition temperature of the training set. This represents the standard deviation of the glass transition temperature of the training set.

4. The method for generating repeating polyimide units with a specific glass transition temperature based on deep learning according to claim 3, characterized in that, The conditional autoregressive generative model employs a conditional encoding method that combines continuous and discrete paths. Standardized continuous scalars are input into a multilayer perceptron, and discrete bucket numbers are input into the embedding layer. The two conditional vectors are fused to form a conditional memory representation. Then, a Transformer-based sequence decoding unit, guided by this conditional memory representation, generates a high-performance polymeric repeating unit SELFIES sequence. The fusion of the conditional vectors from the continuous and discrete paths is achieved through the following formula: Where z represents a normalized continuous scalar, c represents the discrete barrel number corresponding to the target glass transition temperature, MLP() represents a continuous conditional coding network, and Emb() represents a discrete interval embedding layer. The fused condition vector is further extended into a condition memory sequence of length 1 and input into the condition memory encoding unit.

5. The method for generating repeating polyimide units with a specific glass transition temperature based on deep learning according to claim 4, characterized in that, The sequence decoding unit of the conditional autoregressive generative model adopts a Transformer-based decoding architecture. It uses a representation that adds token embeddings and learnable position embeddings to the input SELFIES token sequence. In each generation step, it simultaneously performs mask self-attention calculation on the generated SELFIES prefix sequence and cross-attention calculation on the conditional memory sequence to predict the probability distribution of the next position token.

6. The method for generating repeating polyimide units with a specific glass transition temperature based on deep learning according to claim 5, characterized in that, The output of the conditional autoregressive generative model includes a layer normalization layer and an output projection layer. The weights of the output projection layer are shared with the weights of the token embedding layer. The model training adopts a teacher-forced approach, using the real structure prefix sequence and the target glass transition temperature condition as inputs, and the next position token as the supervision signal. The model parameters are optimized by ignoring the cross-entropy loss of the padding bit.

7. The method for generating repeating polyimide units with a specific glass transition temperature based on deep learning according to claim 6, characterized in that, From the initial label during the reasoning phase <bos>The process begins by generating SELFIES tokens one by one. First, the vocabulary probabilities output from the current generation step are temperature-scaled. Then, the tokens are sorted from highest to lowest probability. The set of tokens with the smallest cumulative probability reaching a preset threshold is retained. Within this retained set, tokens are re-normalized and then randomly sampled until a termination token is generated. <eos> Or it may reach a preset maximum length; the temperature scaling is calculated according to the following formula:< / eos> < / bos> ； in, This represents the output logit of the j-th token in the current generation step. Indicates the current generation step word in the word list. The output logit for each token This is the index of each candidate token in the vocabulary; Indicates the temperature coefficient. This represents the probability of the corresponding token after temperature scaling; subsequently, random sampling is performed only within a subset of tokens that meet the cumulative probability threshold condition.

8. The method for generating repeating polyimide units with a specific glass transition temperature based on deep learning according to claim 7, characterized in that, After decoding the generated SELFIES sequences into SMILES, a multi-level structure screening mechanism is constructed, including: using cheminformatics tools to determine whether the SMILES can be successfully resolved into molecular objects; judging whether the candidate structures meet the structural requirements of the set high-performance polymer repeating units based on the target skeleton pattern; and performing heuristic filtering of candidate structures based on molecular weight range, heavy atom number range, and predefined unstable group patterns.

9. The method for generating repeating polyimide units with a specific glass transition temperature based on deep learning according to claim 8, characterized in that, In the multi-level structure screening mechanism, the structure rule verification can be configured with one or a combination of the following skeleton modes for different polymer systems: (1) Polyimide: imide skeleton -C(=O)-NC(=O)- and five-membered imide ring; (2) Polybenzoxazole: benzoxazole ring and corresponding precursor hydroxyamide structure; (3) Polybenzimidazole: benzimidazole ring skeleton; (4) Polyamide imide: amide bond and imide ring coexisting structure; The molecular property stability screening includes at least the following constraints: Where MW represents molecular weight, with units of g / mol. The range indicates the number of heavy atoms; the above range is determined based on the typical structural characteristics of polyimide repeating units, with the lower limit of molecular weight corresponding to the smallest effective imide unit and the upper limit covering complex repeating units containing multiple aromatic rings and bridging groups.

10. The method for generating repeating polyimide units at specific glass transition temperatures based on deep learning according to claim 9, characterized in that, The trained conditional autoregressive generative model generates multiple candidate polymer structure repeating units for the input target glass transition temperature, and outputs the selected candidate structures and their corresponding structural identifiers, SMILES and SELFIES information.

Images