Thermosetting shape memory polymer formulation design method based on large language model
Patent Information
- Application Number
- CN202611062469.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-17
- Publication Date
- 2026-10-09
AI Technical Summary
[0006]本发明为解决现有TSMP配方设计技术存在设计过程缺乏目标性能直接驱动、候选空间搜索效率较低以及结果可控性不足的问题，进而提出一种基于大语言模型的热固性形状记忆聚合物配方设计方法
[0014]本发明的有益效果是：本发明能够充分利用来自文献的TSMP配方信息，并结合目标性能需求对配方进行定向优化的热固性形状记忆聚合物配方设计方法，以提高 TSMP 配方设计效率，降低盲目高通量筛选带来的计算和实验成本，并提升设计结果与目标性能需求之间的匹配程度。
Smart Images
Figure CN122889176A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a formulation design method for thermosetting shape memory polymers, belonging to the field of shape memory polymer inverse design technology. Background Technology
[0002] Thermoset shape memory polymers (TSMPs) are an important class of smart polymer materials capable of temporarily fixing their shape under external stimuli such as temperature and mechanical load, and recovering to their initial shape under specific triggering conditions. Due to their excellent shape-fixing performance, shape recovery performance, dimensional stability, and tunable mechanical and thermal properties, TSMPs have broad application prospects in aerospace, biomedical devices, flexible electronics, smart actuators, and deployable structures. The macroscopic properties of TSMPs are generally closely related to their monomer composition, molar ratio, functional group type, crosslinking density, and cured network structure. Therefore, how to reverse engineer TSMP formulations to meet specific performance requirements such as target glass transition temperature and rubber modulus has become an important research problem in the field of smart materials design.
[0003] Traditional TSMP (Structure-Driven Polymer Design) formulation design relies heavily on researchers' expertise, literature review, and repeated experimental verification. This approach typically starts with existing monomers or formulations, manually adjusting the types and proportions of monomers before experimental preparation and performance testing to determine if they meet target requirements. Due to the diverse component types, large proportioning space, and complex structure-property relationships in TSMP systems, traditional trial-and-error development methods often suffer from long development cycles, high costs, and low efficiency, making it difficult to meet the needs for rapid discovery and targeted design of high-performance, functionalized polymer materials. With the development of artificial intelligence, deep learning, and data-driven materials design technologies, establishing a mapping relationship between polymer formulations and properties using computational models and further conducting polymer inverse design has become an important direction for improving the efficiency of materials research and development.
[0004] Existing polymer inverse design methods are typically based on machine-readable polymer or molecular representations, such as PSMILES and BigSMILES, and are combined with machine learning models, deep generative models, or reinforcement learning methods to achieve property prediction and structure generation. For linear polymers with relatively regular structures and sufficient data, these methods have achieved some success. However, TSMP (Through-Straight-Mesh Polymer) systems are typical complex cross-linked polymer systems. Their final properties depend not only on the monomer molecular structure but also on factors such as the proportions of multiple components, functional group reaction relationships, and cross-linked network structure, making it difficult to fully characterize them using a single repeating unit or a simple linear representation. Therefore, inverse design methods for ordinary linear polymers are not directly applicable to TSMP formulation design tasks.
[0005] Currently, most formulation design studies for TSMP (Through-Self-Produced Polymer) employ a "forward prediction model combined with high-throughput screening" approach. This involves first compiling existing monomer and formulation data from relevant literature or databases, training a forward model to predict TSMP properties, then combining different monomers according to chemically feasible rules and enumerating different molar ratios, and finally predicting the performance of each candidate formulation, selecting the one closest to the target performance. While this method can reduce the experimental screening scope to some extent, it still relies on a large-scale candidate space traversal, consuming significant computational resources. Furthermore, since the candidate formulation generation process is not directly driven by the target performance requirements, its output often only approximates the target requirements, making it difficult to guarantee the controllability and search efficiency of the design process. Summary of the Invention
[0006] To address the problems of existing TSMP formulation design techniques, such as the lack of direct target performance driving the design process, low candidate space search efficiency, and insufficient controllability of results, this invention proposes a formulation design method for thermosetting shape memory polymers based on a large language model.
[0007] The technical solution adopted by the present invention to solve the above problems is as follows: The specific steps of the present invention include: Step 1: Search, download and organize literature containing TSMP formulations, and perform preprocessing; Step 2: Segment the preprocessed documents and extract recipes using a large language model; Step 3: Post-process the formulation extraction results and use the final processing results to construct the TSMP seed formulation library; Step 4: Based on the target nature and requirements of the TSMP formulation design task, retrieve the most suitable existing formulation from the seed formulation library as the object of formulation optimization and the starting point of the formulation design task. Step 5: Use a Monte Carlo tree search-based recipe optimization method to modify the recipe using three predefined actions. Use a forward model to evaluate the modified recipe until a TSMP recipe that meets the target property requirements within the error range is obtained.
[0008] Furthermore, the specific process of step 1 is as follows: Search for publicly available literature related to TSMP, filter the search results based on the abstract information, select and download PDF documents containing TSMP formulation data, convert them to Markdown format, extract the text and table content, and remove irrelevant content using regular expression matching.
[0009] Furthermore, the specific process of step 2 is as follows: Each preprocessed document is segmented into blocks according to natural paragraphs, and then recipe extraction is performed on each document; the steps for recipe extraction on each document include: Step 201: Call multiple large language models to perform relevance judgment on each segment, retain the segments that the large language models believe contain recipe information, filter out segments that are not of interest, and extract the TSMP names contained in the document from the retained segments. Step 202: Call the large language model to perform preliminary extraction of recipes with evidence-annotated references for each name. The preliminary extraction results should be output along with the references to each field of the extraction results. Step 203: Utilize the reflective capabilities of the large language model to perform two steps of "reflective filtering" and "omission supplementation" on the preliminary extraction results and corresponding evidence annotation information output in the previous step to obtain the final formula extraction results. The results include the name of each monomer and its corresponding molar ratio, as well as the values and units of the glass transition temperature and rubber modulus corresponding to the formula.
[0010] Furthermore, step 3 includes the following specific steps: Step 301: Based on the formulation monomer names in the extraction results, convert the monomer names into the corresponding SMILES strings using PubChem; Step 302: Normalize the molar ratio and standardize the values of glass transition temperature and rubber modulus. Step 303: Use RDKit to parse the monomer SMILES string in the formula and count functional group information at the formula level; Step 304 utilizes functional group information to construct a TSMP seed recipe library.
[0011] Furthermore, the specific process of step 4 is as follows: Based on the target property requirements and constraints of the TSMP formulation design task, the formulation that is closest to the target property requirements and meets the constraints is retrieved from the seed formulation library. Each target property requirement and constraint includes the target glass transition temperature and the target rubber modulus, and the included functional groups are optional constraints. This step and the next optimization process calculate the gap between the formulation and the target property requirements through the formulation scoring function.
[0012] Furthermore, the specific process of step 5 is as follows: Step 501: In each iteration, based on the existing statistical information in the current tree, recursively select child nodes starting from the root node until a node that has not yet been fully expanded is reached. Step 502: Expand at least one new child node at this node; Step 503: Starting from the newly expanded node, perform a random simulation according to the default strategy until the termination state is reached or the preset truncation condition is met. Step 504: Backpropagate the simulated rewards along the search path to update the visit count and value estimate of each node on the path; After numerous iterations, the action with the most visits at the root node is considered the current optimal decision; each node represents a TSMP recipe, and the optimal node is the optimal recipe obtained by the algorithm.
[0013] Furthermore, the three predefined actions in step 5 are ModifyRatio(i, r), Delete(i), and Replace(i, m), with the following meanings: ModifyRatio(i, r): Increases the proportion of monomer i in the current formulation by r, while maintaining the relative proportions of the remaining monomers unchanged; Delete(i): Deletes monomer i from the current formulation and normalizes the molar ratio of the remaining monomers; Replace(i, m): Replaces monomer i in the current recipe with monomer m. The monomer m used for replacement is generated by a large language model that has been fine-tuned by instructions.
[0014] The beneficial effects of this invention are: This invention can make full use of TSMP formulation information from the literature and combine it with the target performance requirements to optimize the formulation of thermosetting shape memory polymer formulations in a targeted manner, so as to improve the efficiency of TSMP formulation design, reduce the computational and experimental costs caused by blind high-throughput screening, and improve the matching degree between the design results and the target performance requirements. Attached Figure Description
[0015] Figure 1 This is a schematic diagram of the process of extracting thermosetting shape memory polymer formulations designed in this invention; Figure 2 This is a schematic diagram illustrating the process of using the reflective capabilities of a large language model to perform two stages of "reflective filtering" and "omission supplementation" on the initial formula extraction results. Figure 3 This is a schematic diagram illustrating the process of processing a recipe during the construction of a seed recipe library; Figure 4 This is a schematic diagram of the process of designing thermosetting shape memory polymer formulations and a schematic diagram of the Monte Carlo tree search process proposed in this invention. Detailed Implementation
[0016] Specific implementation method one: as follows Figures 1 to 4As shown, the steps of the thermosetting shape memory polymer formulation design method based on a large language model described in this embodiment include: Step 1: Search, download and organize literature containing TSMP formulations, and perform preprocessing; Search for publicly available literature related to TSMP, filter the search results based on the abstract information, select and download PDF documents containing TSMP formulation data, convert them to Markdown format, extract the text and table content, and remove irrelevant content using regular expression matching. Step 2: Segment the preprocessed documents and extract recipes using a large language model; Each preprocessed document was segmented into blocks according to natural paragraphs, and then recipe extraction was performed on each document. The steps for recipe extraction on each document included: Step 201: Call multiple large language models to perform relevance judgment on each segment, retain the segments that the large language models believe contain recipe information, filter out segments that are not of interest, and extract the TSMP names contained in the document from the retained segments. Step 202: Call the large language model to perform preliminary extraction of recipes with evidence-annotated references for each name. The preliminary extraction results should be output along with the references to each field of the extraction results. Step 203: Utilize the reflective capabilities of the large language model to perform two steps of "reflective filtering" and "omission supplementation" on the preliminary extraction results and corresponding evidence annotation information output in the previous step to obtain the final formula extraction results; the results include the name of each monomer and its corresponding molar ratio, as well as the values and units of the glass transition temperature and rubber modulus corresponding to the formula. Step 3: Post-process the formulation extraction results and use the final processed results to construct the TSMP seed formulation library; specific steps include: Step 301: Based on the formulation monomer names in the extraction results, convert the monomer names into the corresponding SMILES strings using PubChem; Step 302: Normalize the molar ratio and standardize the values of glass transition temperature and rubber modulus. Step 303: Use RDKit to parse the monomer SMILES string in the formula and count functional group information at the formula level; Step 304: Construct a TSMP seed recipe library using functional group information; Step 4: Based on the target nature and requirements of the TSMP formulation design task, retrieve the most suitable existing formulation from the seed formulation library as the object of formulation optimization and the starting point of the formulation design task. Based on the target property requirements and constraints of the TSMP formulation design task, the formulation that is closest to the target property requirements and meets the constraints is retrieved from the seed formulation library. Each target property requirement and constraint includes the target glass transition temperature and the target rubber modulus, and the included functional groups are optional constraints. This step and the next optimization process calculate the gap between the formulation and the target property requirements through the formulation scoring function. Formula scoring function The calculation formula is:
[0017]
[0018] in, x This is the current formula; T g * and Er * These are the glass transition temperature and rubber modulus, which are the target property requirements, respectively. and E It represents the user's preference weights for the glass transition temperature and rubber modulus, satisfying... + E =1 and all are non-negative; and These are the expected properties of the current formulation predicted by the forward model; Score(x) A higher value indicates a better formula. x The closer the expected nature is to the current target nature requirement; Step 5: Use the Monte Carlo tree search-based recipe optimization method to modify the recipe using three predefined actions, and use the forward model to evaluate the modified recipe until a TSMP recipe that meets the target property requirements within the error range is obtained. The specific process is as follows: Step 501: In each iteration, based on the existing statistical information in the current tree, recursively select child nodes starting from the root node until a node that has not yet been fully expanded is reached. Step 502: Expand at least one new child node at this node; Step 503: Starting from the newly expanded node, perform a random simulation according to the default strategy until the termination state is reached or the preset truncation condition is met. Step 504: Backpropagate the simulated rewards along the search path to update the visit count and value estimate of each node on the path; After numerous iterations, the action with the most visits at the root node is considered the current optimal decision; each node represents a TSMP recipe, and the optimal node is the optimal recipe obtained by the algorithm. The three predefined actions are ModifyRatio(i, r), Delete(i), and Replace(i, m), with the following meanings: ModifyRatio(i, r): Increases the proportion of monomer i in the current formulation by r, while maintaining the relative proportions of the remaining monomers unchanged; Delete(i): Deletes monomer i from the current formulation and normalizes the molar ratio of the remaining monomers; Replace(i, m): Replaces monomer i in the current recipe with monomer m. The monomer m used for replacement is generated by a large language model that has been fine-tuned by instructions.
[0019] The meanings of the four basic stages in the MCTS process are as follows: Selection: The selection phase weighs the prioritization of exploring action paths that are most likely to yield a formula that meets the target requirements against exploring unexplored action paths. A UCT-style selection formula is used for the current formula. x Every modification action tried a ∈ Calculate the selected value:
[0020] And select This is the next modification to be made. c >0 represents the exploration coefficient. N(x) Indicates the current formula x The number of times the formula was used for exploration during the process of gradually modifying it to meet the target properties. N(x, a) Indicating targeting x Modification actions a The number of times it was used for exploration. Q(x, a) This represents the average score of all recipes that can be obtained after the corresponding action.
[0021] Extension: During the extension phase, the highest priority action that has not yet been tried is selected from the list of available actions, and the selected recipe is modified once. This applies to the current recipe node. xIf there are still actions to be tried in the action list, the highest priority action is used to modify the recipe, and the modified recipe is used as the new recipe node for the next step. Then, according to the rules described in the previous section, an ordered action list is generated for the new recipe node for use when expanding that node in subsequent processes. To guide the algorithm to obtain the desired recipe through high-priority actions as much as possible, a gradual widening strategy is adopted to further control the expansion phase, prioritizing attempts to continue expanding the next recipe obtained through high-priority actions, rather than prematurely expanding low-priority actions that have not yet been tried for the current recipe.
[0022] Evaluation: This stage uses a forward model to predict and score the properties of the new formulation obtained in the expansion stage, and then... Score As its value V(x') ,Right now
[0023] Return: Obtain a new recipe x l rating V Then, the score needs to be fed back layer by layer, updating the access statistics and average score for each action edge along the recipe modification path. This applies to each action edge along the path. (x t , a t ) The following updates will be made:
[0024]
[0025]
[0026]
[0027] The algorithm first expands the root node corresponding to the seed recipe to obtain its list of possible actions, and then obtains the next recipe after applying the highest priority modification action in the list. Subsequently, the algorithm iterates continuously in the order of "selection-expansion-evaluation-return" until a recipe that meets the target requirements is found, or the number of iterations reaches the upper limit.
[0028] Example A literature search was conducted on the ChemRxiv platform using "thermoset shape memory polymer" as the keyword. A large language model was used to filter the literature based on abstract information, ultimately downloading 247 articles containing TSMP formulation records. These articles were processed using the formulation extraction method designed in this invention, resulting in 1,631 formulations. These were then used to construct a seed formulation library, serving as the basis for subsequent formulation design methods for thermoset shape memory polymers.
[0029] The target property requirement samples used in the experiment were derived from the property requirement samples used in the Prompt2Poly work and augmented using a large language model.
[0030] Instruction fine-tuning for generating large language models with replacement singletons By supervising and fine-tuning the SmileyLlama model proposed by Cavanagh et al., this method takes as input the SMILES of the monomer to be replaced and the required functional groups, and outputs SMILES that can be used for replacement. The SmileyLlama model is fine-tuned based on Llama-3.1-8B and supports outputting SMILES monomer strings based on the input prompt, which is basically consistent with the form of the current task.
[0031] The dataset used for fine-tuning synthesized monomers from a seed formulation library using publicly available molecular data from PubChem. In the synthesis dataset, monomers were substituted. m With the unit to be replaced i The following requirements must be met: 1. Monomer m With monomer i The main reactive functional groups remain the same, that is, the functional groups with the highest count are the same; 2. Monomer m Includes monomers i Unique functional groups; 3. Monomer m Satisfy the user-specified constraints on functional groups; 4. Monomer m With monomer i The similarity between Morgan's fingerprint and Tanimoto's fingerprint should not be lower than a certain threshold.
[0032] A total of 20,047 data points were synthesized for supervised fine-tuning of the SmileyLlama model.
[0033] Forward model technology selection Using the T proposed by Teimouri et al. for thermosetting shape memory polymers gThe prediction model for both Er and Er properties is used as the forward model. The model encodes the TSMP recipe into a latent vector representation using a variational autoencoder, and then inputs this latent vector into the property predictor for property prediction. The property predictor is a support vector regressor.
[0034] Evaluation indicators The thermosetting shape memory polymer formulation design method of this invention is a generative problem, using MAPE and R, which are oriented towards glass transition temperature and rubber modulus. 2 Five scores, including the average MAPE, were used as evaluation indicators. These indicators quantitatively measure the gap between the formulation design results and the target property requirements; the lower the MAPE, the lower the R... 2 The higher the value, the better the method is.
[0035] Evaluation of experimental results The experimental results in the table show that the formulation design method for thermosetting shape memory polymers based on a large language model proposed in this patent performs better than the high-throughput screening method and the method based on direct generation of large language models proposed by Das et al. Unlike the previous two methods, the method proposed in this chapter does not generate formulations from scratch. Instead, it first selects a suitable seed formulation as the starting point for the design task based on the target property requirements. The subsequent formulation optimization and improvement stages are driven entirely by the target property requirements, thereby obtaining better design results with less computational effort.
[0036] Compared to the method proposed in this paper, the method of Das et al. is not directly driven by the target property requirement. Instead, it finds the recipe closest to the target property requirement through permutation and combination followed by high-throughput screening. This is a commonly used paradigm for current TSMP recipe design tasks, but it suffers from low efficiency and insufficient controllability of results. While Prompt2Poly can directly output recipes based on the target property requirement by supervising and fine-tuning a large language model, the recipe design task itself is a one-to-many problem, making it difficult to complete with high quality through direct supervised fine-tuning of a large language model.
[0037] Table 1 Comparative Experimental Results
[0038] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.
Claims
1. A method for formulating thermosetting shape memory polymers based on a large language model, characterized in that, The specific steps include: Step 1: Search, download and organize literature containing TSMP formulations, and perform preprocessing; Step 2: Segment the preprocessed documents and extract recipes using a large language model; Step 3: Post-process the formulation extraction results and use the final processing results to construct the TSMP seed formulation library; Step 4: Based on the target nature and requirements of the TSMP formulation design task, retrieve the most suitable existing formulation from the seed formulation library as the object of formulation optimization and the starting point of the formulation design task. Step 5: Use a Monte Carlo tree search-based recipe optimization method to modify the recipe using three predefined actions. Use a forward model to evaluate the modified recipe until a TSMP recipe that meets the target property requirements within the error range is obtained.
2. The method for formulating thermosetting shape memory polymers based on a large language model according to claim 1, characterized in that, The specific process of step 1 is as follows: Search for publicly available literature related to TSMP, filter the search results based on the abstract information, select and download PDF documents containing TSMP formulation data, convert them to Markdown format, extract the text and table content, and remove irrelevant content using regular expression matching.
3. The method for formulating thermosetting shape memory polymers based on a large language model according to claim 1, characterized in that, The specific process of step 2 is as follows: each preprocessed document is cut into blocks according to natural paragraphs, and then formula extraction is performed on each document; The steps for formula extraction for each document include: Step 201: Call multiple large language models to perform relevance judgment on each segment, retain the segments that the large language models believe contain recipe information, filter out segments that are not of interest, and extract the TSMP names contained in the document from the retained segments. Step 202: Call the large language model to perform preliminary extraction of recipes with evidence-annotated references for each name. The preliminary extraction results should be output along with the references to each field of the extraction results. Step 203: Utilize the reflective capabilities of the large language model to perform two steps of "reflective filtering" and "omission supplementation" on the preliminary extraction results and corresponding evidence annotation information output in the previous step to obtain the final formulation extraction results. The results include the name of each monomer and its corresponding molar ratio, as well as the values and units of the glass transition temperature and rubber modulus corresponding to the formulation.
4. The method for formulating thermosetting shape memory polymers based on a large language model according to claim 1, characterized in that, Step 3 includes the following specific steps: Step 301: Based on the formulation monomer names in the extraction results, convert the monomer names into the corresponding SMILES strings using PubChem; Step 302: Normalize the molar ratio and standardize the values of glass transition temperature and rubber modulus. Step 303: Use RDKit to parse the monomer SMILES string in the formula and count functional group information at the formula level; Step 304 utilizes functional group information to construct a TSMP seed recipe library.
5. The method for formulating thermosetting shape memory polymers based on a large language model according to claim 1, characterized in that, The specific process of step 4 is as follows: Based on the target properties and constraints of the TSMP formulation design task, retrieve the formulation that is closest to the target properties and meets the constraints from the seed formulation library. One set of target property requirements and constraints includes the target glass transition temperature and the target rubber modulus, with the included functional groups as optional constraints. This step and the next optimization process use a formula scoring function to calculate the gap between the formula and the target property requirements.
6. The method for formulating thermosetting shape memory polymers based on a large language model according to claim 1, characterized in that, The specific process of step 5 is as follows: Step 501: In each iteration, based on the existing statistical information in the current tree, recursively select child nodes starting from the root node until a node that has not yet been fully expanded is reached. Step 502: Expand at least one new child node at this node; Step 503: Starting from the newly expanded node, perform a random simulation according to the default strategy until the termination state is reached or the preset truncation condition is met. Step 504: Backpropagate the simulated rewards along the search path to update the visit count and value estimate of each node on the path; After numerous iterations, the action with the most visits at the root node is considered the current optimal decision; each node represents a TSMP recipe, and the optimal node is the optimal recipe obtained by the algorithm.
7. The method for formulating thermosetting shape memory polymers based on a large language model according to claim 1, characterized in that, In step 5, the three predefined actions are ModifyRatio(i, r), Delete(i), and Replace(i, m), with the following meanings: ModifyRatio(i, r): Increases the proportion of monomer i in the current formulation by r, while maintaining the relative proportions of the remaining monomers unchanged; Delete(i): Deletes monomer i from the current formulation and normalizes the molar ratio of the remaining monomers; Replace(i, m): Replaces monomer i in the current recipe with monomer m. The monomer m used for replacement is generated by a large language model that has been fine-tuned by instructions.