Knowledge graph driven hydrogel formula prior and feasible region screening system

Abstract

The invention discloses a knowledge graph-driven hydrogel formula priori and feasible region screening system applied to the cross field of industrial intelligence and biological medicine, which comprises six modules, namely a structured analysis module, an ontology modeling module, a rule base construction module, a feasible region judgment module, a recommendation reasoning module and a man-machine closed loop module, and can be used for extracting formula triple knowledge from literatures and patents; constructing a directed weighted knowledge graph containing a collaboration / inhibition / toxicity / process relationship; dynamic mapping of forbidding rules is realized in connection with a multi-national supervision list; executing three-hop path reasoning on the user input component, judging whether the component exceeds a feasible region or not, and outputting a conflict basis; if so, recommending auxiliary components and an optimal matching process according to the collaborative strength; re-injecting experimental feedback data into the atlas to realize weight self-updating; through an executable prior rule network and a multi-hop reasoning mechanism, the invalid experiment rate and compliance risk are remarkably reduced, the research and development period is shortened, continuous evolution of the knowledge graph is supported, and the accuracy and safety of new material development are improved.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence and biomedicine, and in particular to a knowledge graph-driven hydrogel formulation prior and feasibility domain screening system. Background Technology

[0002] With the widespread application of hydrogel materials in high-end fields such as biomedicine, tissue engineering, and flexible electronics, the complexity of their formulation design and compliance requirements are increasing. Traditional hydrogel R&D relies on expert experience and literature searches. The prior knowledge of the formulation is scattered across a vast number of papers, patents, and process manuals, lacking structured expression and machine readability. This leads to a high dependence on manual trial and error in the R&D process, repeatedly exploring known ineffective combinations, resulting in low efficiency and high costs. More importantly, existing methods cannot systematically identify compliance risks such as prohibited monomers, reaction byproducts, or residual solvents, which can easily trigger safety red lines in later clinical or production stages, causing project rework or even legal liability.

[0003] One of the core challenges in hydrogel formulation design lies in constructing a multi-dimensional mapping relationship between monomers, reactions, network structures, performance, and compliance. An ideal solution must meet performance targets such as mechanical properties, swelling, and biocompatibility while avoiding toxic substances, ensuring sterilization stability, and complying with pharmacopoeia and FDA regulations. However, current technologies lack a unified modeling capability for these multi-dimensional constraints, and design space exploration often falls into local optima or compliance blind spots, making it difficult to maximize performance under controllable risks.

[0004] While existing materials informatics methods attempt to incorporate machine learning to predict performance, they generally neglect the hard boundaries of chemical rules and regulatory constraints. Model outputs often include thermodynamically infeasible or legally prohibited component combinations. Rule-based expert systems, on the other hand, suffer from rigid knowledge representation and outdated updates, making it difficult to cover emerging monomers and complex cross-linking mechanisms. This disconnect between the two leads to a breakdown in the "generation-verification" process, requiring candidate formulations to undergo multiple rounds of manual screening and experimental verification, a lengthy process prone to overlooking hidden risks. Especially in highly compliant scenarios such as implantable devices or drug delivery systems, the lack of capabilities in defining feasible domains and proactively intercepting risks has become a core bottleneck restricting the intelligent development of hydrogels. Summary of the Invention

[0005] The core of this invention lies in constructing a structured, machine-readable knowledge graph in the field of hydrogels, transforming the prior rules of formulations scattered in academic literature and patent texts into machine-executable logical constraints and material association networks, thereby achieving automated compliance screening and feasible domain boundary determination for hydrogel component combinations. This addresses the technical challenges of highly dispersed formulation knowledge, non-machine-readable rules, high costs of manual trial and error, and the risk of introducing prohibited substances in the current hydrogel material development process.

[0006] To solve the above problems, the present invention adopts the following technical solution.

[0007] The knowledge graph-driven hydrogel formulation prior and feasible domain screening system includes a hydrogel literature and patent structured analysis module, a knowledge graph ontology modeling and topology construction module, a prohibited substance dynamic rule base construction module, a formulation combination feasible domain boundary determination module, a formulation prior rule reasoning and recommendation module, and a human-machine collaborative verification and feedback closed loop module. The hydrogel literature and patent structured analysis module is used to extract hydrogel formulation components, preparation process parameters, performance indicators and material compatibility description information from publicly available academic papers and patent texts. Through named entity recognition and relation extraction algorithms, unstructured text is transformed into knowledge units in the form of triplets. Triplets include subject, predicate and object, where the subject is chemical substance or process step, the predicate is the material interaction relationship or process constraint, and the object is the performance threshold or compatibility conclusion. The knowledge graph ontology modeling and topology construction module is used to map triplet knowledge units to a unified semantic space based on an ontology framework defined by experts in the field of hydrogels. It establishes a directed weighted graph structure in which nodes are chemical substances, process parameters or performance indicators, and edges are material synergistic effects, inhibitory effects, toxicity taboos or process dependencies. The edge weights are jointly determined by literature support and experimental verification frequency. Relationship edges with support below a preset threshold are marked as unverified. The module for building a dynamic rule base for prohibited substances is used to connect with the latest lists of prohibited and restricted substances issued by the National Medical Products Administration, the European Chemicals Agency, and the U.S. Food and Drug Administration. It performs structured transformation of the list items according to the chemical structure code, molecular weight range, functional group characteristics, and application scenarios, and establishes a two-way mapping relationship with the substance nodes in the knowledge graph to realize the automatic triggering and real-time updating of prohibited rules. The formulation combination feasible domain boundary determination module is used to receive the target hydrogel component set input by the user, traverse all material nodes and relationship edges associated with each component in the knowledge graph, perform multi-hop reasoning path search, identify whether there is a path directly associated with prohibited substances, whether there are mutually inhibiting component pairs, and whether there are combinations that exceed the tolerance range of process parameters. If any illegal path or conflict relationship exists, it is determined that the component set is outside the feasible domain, and the conflicting material pairs and the basis for the violation are output. The formulation prior rule reasoning and recommendation module is used to generate a list of auxiliary components that can enhance the target performance based on the material nodes in the knowledge graph that have a positive synergistic effect with the target component when the component set is located within the feasible domain. The recommendation list is sorted by the strength of the synergistic effect and includes the optimal addition ratio range and process adaptation conditions of the recommended components. The human-machine collaborative verification and feedback closed-loop module is used to push the judgment results and recommendation list to the experimenter's terminal, receive feedback from the experimenter on the actual test results of the recommended components, including the performance improvement, compatibility performance and process stability evaluation, and inject the structured feedback data back into the knowledge graph to update the node attributes and edge weights, so as to realize the continuous evolution and accuracy improvement of the knowledge graph.

[0008] Furthermore, in the hydrogel literature and patent structured analysis module, named entity recognition adopts a joint model based on bidirectional long short-term memory network and conditional random field. The input is a text sequence that has been preprocessed with chemical terminology standardization, and the output is an entity sequence labeled with substance name, concentration unit, temperature parameter, time parameter and performance index category. Relation extraction adopts a graph neural network model based on attention mechanism. The input is the entity sequence and its context semantic vector, and the output is the semantic relationship type and confidence score between entity pairs.

[0009] Furthermore, in the knowledge graph ontology modeling and topology construction module, the ontology framework includes five types of core nodes: basic monomer nodes, crosslinking agent nodes, initiator nodes, solvent nodes, and performance index nodes, and four types of core relational edges: synergistic edges, chemical inhibition edges, toxicity prohibition edges, and process-dependent edges. Among them, synergistic edges connect substance pairs that can improve gel strength or biocompatibility, chemical inhibition edges connect substance pairs that can lead to polymerization failure or phase separation, toxicity prohibition edges connect substance pairs that are listed as prohibited by any regulatory agency, and process-dependent edges connect substance pairs that must be used synergistically within a specific temperature or time window.

[0010] Furthermore, in the module for constructing a dynamic rule base for prohibited substances, the chemical structure encoding of substances adopts simplified linear input standard strings, and the molecular weight range is divided into four levels: less than one hundred, one hundred to five hundred, five hundred to one thousand, and greater than one thousand. Functional group feature extraction adopts fingerprint vectors based on substructure matching. The application scenarios are divided into four categories: medical implantation, skin contact, oral sustained release, and environmental response. Each category corresponds to an independent subset of prohibited substances and a concentration upper limit threshold.

[0011] Furthermore, in the feasible domain boundary determination module of formulation combination, the multi-hop inference path search adopts the depth-first traversal algorithm, the maximum number of hops is set to three, and the path termination conditions include reaching the toxicity prohibition edge, reaching the chemical inhibition edge, the cumulative weight of the path being less than 0.5, or reaching the performance index node and the index value being less than the preset lower limit. The output format of the violation basis is the name of the conflicting substance pair, the conflict type, the literature source number and the regulatory agency clause number.

[0012] Furthermore, in the formulation prior rule reasoning and recommendation module, the synergistic effect strength is calculated by the weighted average of the paths between nodes. The weights are determined by the product of the edge type coefficient and the literature support. In the edge type coefficient, the synergistic effect edge is set to one, and the process dependence edge is set to 0.8. The optimal addition ratio range is obtained by fitting the concentration gradient experimental data of the component's contribution to the target performance in historical literature. The fitting method adopts a piecewise linear regression model. The process adaptation conditions include temperature range, stirring rate range, and light conditions.

[0013] Furthermore, in the human-machine collaborative verification and feedback closed-loop module, the performance improvement is quantified as a percentage. The compatibility performance is divided into four levels: complete compatibility, slight turbidity, local precipitation, and complete phase separation. The process stability evaluation is divided into four levels: no abnormalities, slight fluctuations, parameters needing adjustment, and inability to form. When feedback data is injected, the attribute update mechanism of the spectrum node is triggered. If the same substance pair shows synergistic effect in three independent experiments, its edge weight is increased by 0.2. If it shows chemical inhibition in all three experiments, a new chemical inhibition edge is added and assigned an initial weight of 0.8.

[0014] Furthermore, in the multi-hop inference path search of the feasible domain boundary determination module of the recipe combination, a message passing mechanism based on graph neural network is used to propagate constraint signals along the relation path, accumulate risk scores for recipe nodes, and accelerate high-frequency rule path matching based on the pre-built constraint rule index tree.

[0015] Furthermore, the formulation prior rule reasoning and recommendation module adopts a graph attention network architecture, which represents the formulation as a subgraph structure. Through a multi-layer attention mechanism, it dynamically calculates the contribution weight of each component and its interaction to the target performance index, and outputs the performance prediction range and the process parameter recommendation range.

[0016] Furthermore, the graph attention network architecture adopts a transfer learning strategy, first pre-training on a general polymer material knowledge graph, and then fine-tuning on a hydrogel-specific graph. The performance prediction range is optimized through a quantile regression loss function, and the process parameter range is searched for the Pareto optimal frontier in the historical case parameter space through a Bayesian optimization algorithm.

[0017] Compared with the prior art, the advantages of this invention are: (1) By constructing a structured, machine-readable knowledge graph in the field of hydrogels, this invention transforms the prior rules of formulations scattered in a large number of documents and patents into a logical constraint network that can be executed by machines, thereby realizing automated compliance screening and feasible domain boundary determination of hydrogel component combinations, and completely eliminating the inefficiency and human oversight problems caused by traditional manual literature review and comparison of prohibited lists.

[0018] (2) The system can accurately identify potential chemical inhibition relationships and toxicity incompatibilities between components through multi-hop reasoning path search, thus avoiding experimental failures and safety risks caused by incompatible substances or illegal additions from the source.

[0019] (3) At the same time, the system generates a list of recommended auxiliary components based on the ranking of synergistic effect strength, along with the optimal addition ratio and process adaptation conditions, which significantly shortens the formulation optimization cycle and improves R&D efficiency. The human-machine collaborative verification and feedback closed-loop mechanism ensures the continuous evolution and accuracy improvement of the knowledge graph, enabling the system to learn and adapt to new materials and processes.

[0020] (4) While ensuring compliance and safety, this invention significantly reduces the trial and error costs and time costs in the hydrogel R&D process, and provides system-level technical support for the efficient, accurate and compliant development of new materials. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical architecture of the knowledge graph-driven hydrogel formulation prior and feasible domain screening system proposed in this invention. Figure 2 This is a schematic diagram of the core principle framework of the knowledge graph ontology modeling and topology construction module in this invention; Figure 3 This is a logical flowchart of the hydrogel literature and patent structured analysis module in this invention; Figure 4 This is a logic framework diagram of multi-hop reasoning and conflict detection for the feasible domain boundary determination module of the recipe combination in this invention; Figure 5 This is a schematic diagram of the synergistic effect ranking and parameter adaptation framework of the recipe prior rule reasoning and recommendation module in this invention; Figure 6 This is a schematic diagram of the dynamic evolution mechanism of the knowledge graph in the human-machine collaborative verification and feedback closed-loop module of this invention. Detailed Implementation

[0022] The technical solutions will now be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention.

[0023] First implementation method: like Figure 1 The knowledge graph-driven hydrogel formulation prior and feasible domain screening system includes a knowledge graph ontology modeling and topology construction module, a hydrogel literature and patent structured analysis module, a prohibited substance dynamic rule base construction module, a formulation combination feasible domain boundary determination module, a formulation prior rule reasoning and recommendation module, and a human-machine collaborative verification and feedback closed loop module.

[0024] The hydrogel literature and patent structured analysis module is used to extract hydrogel formulation components, preparation process parameters, performance indicators, and material compatibility descriptions from publicly available academic papers and patent texts. Through named entity recognition and relation extraction algorithms, unstructured text is transformed into knowledge units in the form of triplets. Each triplet includes a subject, a predicate, and an object. The subject is a chemical substance or process step, the predicate is a material interaction relationship or process constraint, and the object is a performance threshold or compatibility conclusion. Named entity recognition employs a joint model based on a bidirectional long short-term memory network and a conditional random field. The input is a text sequence preprocessed with standardized chemical terminology, and the output is a sequence of entities labeled with substance names, concentration units, temperature parameters, time parameters, and performance indicator categories. Relation extraction uses a graph neural network model based on an attention mechanism. The input is an entity sequence and its contextual semantic vector, and the output is the semantic relationship type and confidence score between entity pairs.

[0025] The knowledge graph ontology modeling and topology construction module is used to map triplet knowledge units to a unified semantic space based on an ontology framework defined by experts in the hydrogel field. It establishes a directed weighted graph structure where nodes are chemical substances, process parameters or performance indicators, and edges are material synergistic effects, inhibitory effects, toxicity taboos or process dependencies. The edge weights are jointly determined by literature support and experimental verification frequency. Relationship edges with support below a preset threshold are marked as unverified. Named entity recognition adopts a joint model based on bidirectional long short-term memory network and conditional random field. The input is a text sequence that has been preprocessed with chemical terminology standardization, and the output is an entity sequence labeled with substance name, concentration unit, temperature parameter, time parameter and performance indicator category. Relation extraction adopts a graph neural network model based on attention mechanism. The input is an entity sequence and its context semantic vector, and the output is the semantic relationship type and confidence score between entity pairs.

[0026] The module for constructing a dynamic rule base for prohibited substances is used to connect with the latest lists of prohibited and restricted substances issued by the National Medical Products Administration, the European Chemicals Agency, and the U.S. Food and Drug Administration. It performs structured transformation of list items according to chemical structure codes, molecular weight ranges, functional group characteristics, and application scenarios, and establishes a bidirectional mapping relationship with substance nodes in the knowledge graph to achieve automatic triggering and real-time updates of prohibition rules. Among them, the chemical structure codes of substances adopt simplified molecular linear input standard strings, the molecular weight range is divided into four levels: less than 100, 100 to 500, 500 to 1000, and greater than 1000, the functional group feature extraction adopts fingerprint vectors based on substructure matching, and the application scenarios are divided into four categories: medical implantation, skin contact, oral sustained release, and environmental response. Each category corresponds to an independent subset of prohibited substances and upper limit threshold for concentration.

[0027] The formulation combination feasible domain boundary determination module receives the target hydrogel component set input by the user, traverses all material nodes and relation edges associated with each component in the knowledge graph, performs multi-hop reasoning path search, and identifies whether there are paths directly associated with prohibited substances, whether there are mutually inhibiting component pairs, and whether there are combinations that exceed the tolerance range of process parameters. If any illegal path or conflict relationship exists, the component set is determined to be outside the feasible domain, and the conflicting material pairs and violation evidence are output. The multi-hop reasoning path search adopts a depth-first traversal algorithm with a maximum number of hops of three. The path termination conditions include reaching a toxicity prohibition edge, reaching a chemical inhibition edge, the cumulative weight of the path being less than 0.5, or reaching a performance index node and the index value being less than a preset lower limit. The violation evidence output format is the name of the conflicting material pair, the conflict type, the literature source number, and the regulatory agency clause number.

[0028] The formulation prior rule reasoning and recommendation module is used to generate a list of auxiliary components that can enhance the target performance when the component set is located within the feasible region. This list is based on material nodes in the knowledge graph that have a positive synergistic effect with the target component. The recommendation list is sorted by the strength of the synergistic effect and includes the optimal addition ratio range and process adaptation conditions for the recommended components. The synergistic effect strength is calculated by the weighted average of the paths between nodes. The weights are determined by the product of the edge type coefficient and the literature support. In the edge type coefficient, the synergistic effect edge is set to 1, and the process dependence edge is set to 0.8. The optimal addition ratio range is obtained by fitting the concentration gradient experimental data of the component's contribution to the target performance in historical literature. The fitting method adopts a piecewise linear regression model. The process adaptation conditions include temperature range, stirring rate range, and light condition requirements.

[0029] The human-machine collaborative verification and feedback closed-loop module pushes the judgment results and recommendation list to the experimenter's terminal, receives feedback from the experimenter on the actual test results of the recommended components, including performance improvement, compatibility performance, and process stability evaluation. The feedback data is structured and injected back into the knowledge graph to update node attributes and edge weights, realizing the continuous evolution and accuracy improvement of the knowledge graph. Among them, the performance improvement is quantified in the form of percentage, compatibility performance is divided into four levels: completely compatible, slightly turbid, local precipitation, and completely phase separation, and process stability evaluation is divided into four levels: no abnormality, slight fluctuation, parameter adjustment required, and unable to form. When the feedback data is injected, the attribute update mechanism of the graph node is triggered. If the same substance pair shows synergistic effect in three independent experiments, its edge weight is increased by 0.2. If it shows chemical inhibition in all three experiments, a new chemical inhibition edge is added and assigned an initial weight of 0.8.

[0030] In practice, the modules interact with each other through standardized data interfaces and unified semantic protocols, forming a knowledge-driven R&D system that enables closed-loop iterative optimization.

[0031] like Figure 2 During system operation, the knowledge graph ontology modeling and topology construction module first defines the semantic framework of hydrogel formulation knowledge, establishing entity types, attribute relationships and constraint rules; then, the hydrogel literature and patent structured parsing module automatically extracts formulation components, preparation processes, performance indicators and interrelationships from massive amounts of scientific and technological literature and patent texts, and maps them to the preset ontology structure to complete the initial filling of the knowledge graph. The formulation combination feasible domain boundary determination module, based on the constructed knowledge graph, performs multi-hop semantic reasoning on the candidate formulations input by the user to identify whether they violate known physicochemical constraints, biocompatibility restrictions, or patent circumvention clauses. The formulation prior rule reasoning and recommendation module, based on historical successful cases and synergy rules, sorts and optimizes the formulations that pass the boundary screening and recommends suitable process parameter ranges. Finally, the human-machine collaborative verification and feedback closed-loop module receives experimental verification results and dynamically updates the confidence weights and rule validity in the knowledge graph to achieve continuous evolution of the system's knowledge base.

[0032] The knowledge graph ontology modeling and topology construction module serves to establish a unified semantic understanding foundation for the entire system. This module defines the core entity types in the hydrogel formulation knowledge system, including but not limited to monomers, crosslinking agents, initiators, solvents, additives, performance indicators, preparation conditions, biocompatibility levels, toxicity levels, patent status, and literature sources. Each entity type is equipped with standardized attribute fields. For example, monomer entities include attributes such as molecular weight, functional group type, polymerization activity, water solubility, and toxicity identifiers; crosslinking agent entities include attributes such as crosslinking density coefficient, reaction temperature range, and hydrolytic stability; and performance indicator entities include attributes such as tensile strength, elongation at break, swelling ratio, degradation period, and cytotoxicity level.

[0033] The types of relationships between entities are also clearly defined, including semantic relationships such as "participating in the reaction," "inhibiting polymerization," "enhancing mechanical properties," "reducing swelling rate," "existing patent conflict," and "document support level." This module uses a formal ontology language to logically model the aforementioned entities, attributes, and relationships, constructing a semantic network topology with reasoning capabilities. This topology supports attribute inheritance, relationship propagation, and constraint propagation. For example, if a monomer is labeled "highly cytotoxic," any formulation combination containing that monomer will automatically inherit the "biocompatibility risk" label; if a crosslinking agent and a specific solvent combination is documented as "causing phase separation," this combination relationship will be marked as "prohibited combination," triggering a conflict alert during subsequent screening. During ontology modeling, all entities and relationships are assigned unique Chinese identifiers, avoiding the use of any English abbreviations or numerical codes to ensure the purity and consistency of semantic expression.

[0034] like Figure 3 The hydrogel literature and patent structured analysis module automatically extracts formulation knowledge from unstructured scientific and technological texts and populates it into a knowledge graph. This module employs a hybrid analysis architecture combining multi-layer neural networks and a rule engine. Input data consists of pre-processed full-text hydrogel-related scientific and technological literature and patent specifications. The preprocessing stage includes text segmentation, paragraph semantic annotation, chemical nominated entity recognition, and standardization. Chemical nominated entity recognition uses a joint model based on a character-level bidirectional long short-term memory network and a conditional random field, accurately identifying key information such as chemical substance names, concentration values, temperature parameters, and time parameters appearing in the text.

[0035] The recognition results are then sent to the semantic role labeling module, which determines the functional role of each entity in the sentence based on a predefined semantic template, such as "used as a crosslinking agent", "dissolved in deionized water", "reacted at 60 degrees Celsius for two hours", etc.

[0036] After annotation, the system maps the extracted entities and their semantic roles to nodes and edges in the knowledge graph based on the entity and relation types in the ontology model. For example, when recognizing the sentence "acrylamide and N,N-methylenebisacrylamide polymerize in deionized water under the initiation of ammonium persulfate," the system will create nodes for "acrylamide," "N,N-methylenebisacrylamide," "ammonium persulfate," and "deionized water," and establish relation edges such as "acrylamide - participates in the reaction - N,N-methylenebisacrylamide," "ammonium persulfate - as an initiator - acrylamide," and "acrylamide - dissolved in - deionized water." For the extraction of performance indicators and experimental conditions, the system uses a numerical-unit joint recognition model to transform expressions such as "tensile strength is 0.5 MPa" and "swelling ratio reaches 300%" into structured numerical attributes and associate them with the corresponding formulation nodes.

[0037] All extracted results are accompanied by source document identifiers and confidence scores, which are determined by the output probability of the analytical model and the rule matching degree. This module supports incremental knowledge updates; new document analysis results will be automatically aligned with existing maps for entity and relationship fusion, avoiding knowledge redundancy and conflicts.

[0038] like Figure 4 The module for determining the feasible domain boundary of a formulation combination is designed to perform automated compliance screening and risk warning for candidate formulations proposed by the user. This module receives a list of formulation components and their proportions from the user. First, it locates the entity nodes corresponding to each component in the knowledge graph. Then, it initiates a multi-hop semantic reasoning engine to perform a multi-hop reasoning path search, conducting a depth-first traversal along relational edges to detect any paths that violate preset constraint rules.

[0039] Constraint rules are divided into two categories: hard boundary rules and soft risk rules.

[0040] Hard boundary rules include "toxic substance prohibition rules," "patent circumvention rules," and "chemical incompatibility rules," which directly determine that the formulation is infeasible once triggered. For example, if the formulation contains a component labeled as a "carcinogen" or contains a combination of core components explicitly protected in a patent claim, the system immediately returns the conclusion that "the formulation violates a hard boundary" and terminates subsequent analysis. Soft risk rules include "performance conflict rules," "process sensitivity rules," and "biological compatibility doubt rules," which do not directly reject the formulation when triggered, but generate a risk warning and lower its recommendation priority.

[0041] During multi-hop inference, the system employs a message-passing mechanism based on graph neural networks to propagate constraint signals along relational paths. For example, when the system detects that a formula contains both "polyethylene glycol diacrylate" and "chitosan," it propagates negative signals along the paths "polyethylene glycol diacrylate - may cause - phase separation" and "phase separation - affects - mechanical homogeneity," ultimately accumulating a risk score at the formula node. To improve inference efficiency, the system pre-builds a constraint rule index tree and caches frequently triggered rule paths to accelerate the process.

[0042] The boundary determination results are output in the form of a structured report, clearly listing each triggered rule, the combination of components involved, the source document or patent number of the rule, and avoidance suggestions. This module ensures that all output formulations are within the feasible domain defined by physicochemical laws, biosafety regulations, and intellectual property boundaries, eliminating invalid trial and error and legal risks from the source.

[0043] like Figure 5 The formulation prior rule reasoning and recommendation module is designed to predict the performance and optimize parameters of candidate formulations that have passed boundary screening. This module constructs a formulation performance prediction module based on historical success cases and synergistic effect rules stored in a knowledge graph. The prediction module employs a graph attention network architecture, representing the formulation as a subgraph structure, with nodes representing component entities and edges representing interactions between components. The model dynamically calculates the contribution weight of each component and its interactions to the target performance index through a multi-layer attention mechanism. For example, when predicting tensile strength, the model assigns higher attention weights to "high crosslinking density crosslinking agents" and "rigid main chain monomers"; when predicting swelling ratio, it assigns higher weights to the combination of "hydrophilic monomers" and "low crosslinking density". Prediction results are output in interval form, such as "tensile strength prediction interval is 0.3 to 0.8 MPa" and "swelling ratio prediction interval is 200 to 400%".

[0044] In addition to performance prediction, this module also recommends optimal process parameter combinations based on a synergistic effect rule base. The synergistic effect rule base contains empirically optimal ranges in a three-dimensional parameter space of "temperature-time-concentration," such as "when using ammonium persulfate initiator, the reaction temperature should be controlled between 60 and 70 degrees Celsius, the reaction time should be two to four hours, and the monomer concentration should be 10% to 20%." The recommendation process employs a case-based reasoning strategy. The system retrieves historical successful cases with similar formulation structures from the knowledge graph, extracts their process parameters as initial recommended values, and then fine-tunes them based on the component differences between the current formulation and the case formulations. Fine-tuning rules are jointly constructed by expert experience and machine learning, such as "if the monomer is replaced with a higher molecular weight version, the reaction time should be extended by 20%" and "if the crosslinking agent dosage is increased by 5%, the reaction temperature should be decreased by 5 degrees Celsius." The final output includes a complete formulation optimization report containing the performance prediction range, recommended process parameter range, references to similar historical cases, and explanations of the adjustment basis.

[0045] like Figure 6 The human-machine collaborative verification and feedback closed-loop module functions to feed experimental verification results back to the knowledge graph, driving the dynamic evolution and self-improvement of the system's knowledge base. This module provides a standardized experimental data input interface, allowing researchers to upload performance test results, process parameter records, and descriptions of abnormal phenomena from actual prepared hydrogel samples. The system automatically matches the uploaded data with the original predicted formulation, calculates the prediction error, and analyzes the source of the error. If there is a significant deviation between the measured performance and the predicted range, the system initiates an attribution analysis process to check for missing or incorrect component attributes, interaction relationships, or synergistic effect rules in the knowledge graph. For example, if the measured swelling ratio of a formulation is much higher than the predicted upper limit, the system will trace the "swelling contribution coefficient" attribute of all hydrophilic components in the formulation. If it finds that a component coefficient is not correctly defined, it will prompt the knowledge engineer to supplement the annotation. For newly discovered effective component combinations or process parameters, the system automatically generates new knowledge graph nodes and relationships and assigns an initial confidence score.

[0046] The confidence score increases with the number of repeated verifications. When the score exceeds a preset threshold, the knowledge item is included in the formal rule base for subsequent reasoning. For rules or attributes that are repeatedly falsified, the system automatically reduces their confidence weight until they are removed from the active reasoning path. This module also supports manual correction and knowledge annotation by experts. Researchers can directly modify entity attributes, add new relationships, or adjust rule weights in the system interface. All modifications are recorded in the version log to ensure the traceability of knowledge evolution. Through this closed-loop mechanism, the system can continuously absorb new knowledge and correct old errors, achieving a spiral increase in its recipe recommendation capabilities.

[0047] In this embodiment, the semantic network constructed by the knowledge graph ontology modeling and topology construction module supports dynamic expansion and version management in its topology structure. When introducing novel hydrogel systems such as self-healing hydrogels and conductive hydrogels, the system allows the addition of corresponding entity types and relationship types without disrupting the existing structure. For example, to support the development of self-healing hydrogels, a "dynamic bond type" entity can be added, including subtypes such as "hydrogen bond," "metal coordination bond," and "hydrophobic interaction," and a "component-formation dynamic bond-dynamic bond type" relationship can be established. The version management mechanism ensures that different R&D teams can construct domain-specific subgraphs based on the same basic ontology while maintaining the consistency of core semantics. The system provides an ontology difference comparison tool that can automatically identify structural changes and conflicts between different versions, assisting knowledge engineers in making merging decisions.

[0048] In this embodiment, the parsing accuracy of the hydrogel literature and patent structured analysis module is ensured through a dual verification mechanism. The first verification is post-processing by the rule engine. The system has hundreds of built-in chemical semantic verification rules, such as "the amount of initiator shall not exceed five percent of the total monomer amount" and "the reaction temperature shall not be lower than the solvent freezing point," to filter the neural network output results for logical rationality. The second verification is a manual sampling mechanism. The system randomly selects five percent of the parsing results daily for review by domain experts. The review results are used to retrain the neural network model, forming a positive feedback loop for improved accuracy. During the parsing process, for ambiguous chemical names such as "PEGDA," which may refer to polyethylene glycol diacrylate with different molecular weights, the system defaults to the most common molecular weight version and marks "There is molecular weight ambiguity; manual confirmation is recommended" in the output to ensure the rigor of knowledge representation.

[0049] In this embodiment, the multi-hop inference engine of the formulation combination feasible domain boundary determination module supports custom rule injection. Researchers can temporarily add project-specific constraint rules according to specific project needs. For example, when developing biomedical dressings, a rule such as "all components must pass ISO 10993 biocompatibility test" can be injected; when developing biodegradable hydrogels, a rule such as "degradation period must be less than 30 days" can be injected. Custom rules use the same semantic expression syntax as the system's preset rules, ensuring seamless compatibility of the inference engine. After rule injection, the system automatically recalculates the compliance status of all candidate formulations, realizing project-oriented dynamic boundary adjustment.

[0050] In this embodiment, the graph attention network architecture of the formulation prior rule reasoning and recommendation module is trained using historical formulation cases marked "experimentally verified successfully" from the knowledge graph. During training, the model learns the mapping function from the formulation subgraph structure to the performance index range. To address the data sparsity problem, the system employs a transfer learning strategy, first pre-training on a general polymer materials knowledge graph and then fine-tuning on a hydrogel-specific graph. The performance prediction range is optimized using a quantile regression loss function to ensure that the upper and lower prediction limits cover 95% and 5% of the measured data points, respectively. The recommended process parameter range is then searched for the Pareto optimal frontier in the historical case parameter space using a Bayesian optimization algorithm, balancing performance indicators and process feasibility.

[0051] In this embodiment, the confidence evolution mechanism of the human-machine collaborative verification and feedback closed-loop module adopts the Bayesian update formula. Let the initial confidence of a knowledge item be P0, and it be updated to P0 upon each successful verification. new =P old +(1-P old )*α, updated to P each time verification fails. new =P old *β, where α and β are preset learning rate parameters, typically α equal to 0.1 and β equal to 0.5. When the confidence level is below 0.3, the entry is marked as "to be verified"; when the confidence level is above 0.8, it is marked as "highly reliable". The system prioritizes using highly reliable entries for inference, and adds risk warnings to entries to be verified. This mechanism ensures that the knowledge base always reflects the most reliable scientific consensus.

[0052] This system, through the collaborative operation of the six modules described above, achieves systematic accumulation, automated reasoning, and intelligent recommendation of hydrogel formulation knowledge. It transforms traditional trial-and-error R&D, reliant on personal experience, into computationally driven R&D based on structured knowledge, significantly shortening the R&D cycle, reducing costs, and mitigating legal and safety risks. Each formulation report output by the system contains a complete knowledge traceability chain, from component selection criteria, performance prediction models, and recommended process parameters to risk avoidance suggestions, all traceable to original literature, patents, or experimental data, ensuring transparency and auditability in the R&D process. With the continuous expansion of the knowledge graph and the ongoing optimization of the model, the system will gradually cover a wider range of hydrogel application scenarios, becoming a core intelligent infrastructure in the field of materials research and development.

[0053] The above description is merely a preferred embodiment of the present invention; it encompasses all the protection scope of the present invention. Any equivalent substitutions or modifications made by those skilled in the art within the technical scope disclosed in the present invention, based on the technical solutions and improved concepts of the present invention, should be covered within the protection scope of the present invention.

Claims

1. A knowledge graph-driven prior and feasible domain screening system for hydrogel formulations, characterized in that: It includes modules for structured analysis of hydrogel literature and patents, knowledge graph ontology modeling and topology construction, construction of a dynamic rule base for prohibited substances, determination of the feasible domain boundary of formulation combination, reasoning and recommendation of formulation prior rules, and human-computer collaborative verification and feedback loop. The hydrogel literature and patent structured analysis module is used to extract hydrogel formulation components, preparation process parameters, performance indicators and material compatibility description information from publicly available academic papers and patent texts. Through named entity recognition and relation extraction algorithms, unstructured text is transformed into knowledge units in the form of triplets. Each triplet includes a subject, a predicate and an object, where the subject is a chemical substance or process step, the predicate is a material interaction relationship or process constraint, and the object is a performance threshold or compatibility conclusion. The knowledge graph ontology modeling and topology construction module is used to map the triplet knowledge units to a unified semantic space based on the ontology framework defined by experts in the field of hydrogels. It establishes a directed weighted graph structure in which nodes are chemical substances, process parameters or performance indicators, and edges are material synergistic effects, inhibitory effects, toxicity taboos or process dependencies. The edge weights are jointly determined by literature support and experimental verification frequency. Relationship edges with support below a preset threshold are marked as unverified. The prohibited substance dynamic rule base construction module is used to connect with the latest prohibited and restricted substance lists issued by the National Medical Products Administration, the European Chemicals Agency, and the U.S. Food and Drug Administration. It performs structured transformation of the list items according to the chemical structure code, molecular weight range, functional group characteristics, and application scenarios, and establishes a bidirectional mapping relationship with the substance nodes in the knowledge graph to realize the automatic triggering and real-time updating of the prohibited rules. The formulation combination feasible domain boundary determination module is used to receive the target hydrogel component set input by the user, traverse all material nodes and relationship edges associated with each component in the knowledge graph, perform multi-hop reasoning path search, identify whether there is a path directly associated with prohibited substances, whether there are mutually inhibiting component pairs, and whether there are combinations that exceed the tolerance range of process parameters. If any illegal path or conflict relationship exists, it is determined that the component set is outside the feasible domain, and the conflicting material pairs and the basis for the violation are output. The formulation prior rule reasoning and recommendation module is used to generate an auxiliary component recommendation list that can enhance the target performance based on the material nodes in the knowledge graph that have a positive synergistic effect with the target component when the component set is located within the feasible domain. The recommendation list is sorted according to the strength of the synergistic effect and includes the optimal addition ratio range and process adaptation conditions of the recommended components. The human-machine collaborative verification and feedback closed-loop module is used to push the judgment result and recommendation list to the experimenter's terminal, receive feedback from the experimenter on the actual test results of the recommended components, including performance improvement, compatibility performance and process stability evaluation, and inject the structured feedback data back into the knowledge graph to update node attributes and edge weights, so as to realize the continuous evolution and accuracy improvement of the knowledge graph.

2. The knowledge graph-driven hydrogel formulation prior and feasible domain screening system according to claim 1, characterized in that: In the hydrogel literature and patent structured analysis module, named entity recognition adopts a joint model based on bidirectional long short-term memory network and conditional random field. The input is a text sequence that has been preprocessed by chemical terminology standardization, and the output is an entity sequence labeled with substance name, concentration unit, temperature parameter, time parameter and performance index category. Relation extraction adopts a graph neural network model based on attention mechanism. The input is the entity sequence and its context semantic vector, and the output is the semantic relationship type and confidence score between entity pairs.

3. The knowledge graph-driven hydrogel formulation prior and feasible domain screening system according to claim 1, characterized in that: In the knowledge graph ontology modeling and topology construction module, the ontology framework includes five types of core nodes: basic monomer nodes, crosslinking agent nodes, initiator nodes, solvent nodes, and performance index nodes, and four types of core relational edges: synergistic edges, chemical inhibition edges, toxicity prohibition edges, and process-dependent edges. Synergistic edges connect substance pairs that can improve gel strength or biocompatibility, chemical inhibition edges connect substance pairs that can lead to polymerization failure or phase separation, toxicity prohibition edges connect substance pairs that are listed as prohibited by any regulatory agency, and process-dependent edges connect substance pairs that must be used synergistically within a specific temperature or time window. The edge weights are determined by a combination of literature support and experimental verification frequency.

4. The knowledge graph-driven hydrogel formulation prior and feasible domain screening system according to claim 3, characterized in that, In the prohibited substance dynamic rule base construction module, the chemical structure encoding of substances adopts simplified molecular linear input standard strings, the molecular weight range is divided into four levels: less than one hundred, one hundred to five hundred, five hundred to one thousand, and greater than one thousand. Functional group feature extraction adopts fingerprint vector based on substructure matching. The application scenarios are divided into four categories: medical implantation, skin contact, oral sustained release, and environmental response. Each category corresponds to an independent prohibited substance subset and concentration upper limit threshold.

5. The knowledge graph-driven hydrogel formulation prior and feasible domain screening system according to claim 4, characterized in that, In the feasible domain boundary determination module of the formulation combination, the multi-hop reasoning path search adopts the depth-first traversal algorithm, the maximum number of hops is set to three, and the path termination conditions include reaching the toxicity prohibition edge, reaching the chemical inhibition edge, the cumulative weight of the path being less than 0.5, or reaching the performance index node and the index value being less than the preset lower limit. The violation basis output format is the name of the conflicting substance pair, the conflict type, the literature source number and the regulatory agency clause number.

6. The knowledge graph-driven hydrogel formulation prior and feasible domain screening system according to claim 5, characterized in that, In the formulation prior rule reasoning and recommendation module, the synergistic effect strength is calculated by the weighted average of the paths between nodes. The weights are determined by the product of the edge type coefficient and the literature support. In the edge type coefficient, the synergistic effect edge is set to one, and the process dependence edge is set to 0.

8. The optimal addition ratio range is obtained by fitting the concentration gradient experimental data of the component's contribution to the target performance in historical literature. The fitting method adopts a piecewise linear regression model. The process adaptation conditions include temperature range, stirring rate range, and light conditions.

7. The knowledge graph-driven hydrogel formulation prior and feasible domain screening system according to claim 6, characterized in that, In the human-machine collaborative verification and feedback closed-loop module, the performance improvement is quantified as a percentage. The compatibility performance is divided into four levels: complete compatibility, slight turbidity, local precipitation, and complete phase separation. The process stability evaluation is divided into four levels: no abnormality, slight fluctuation, parameter adjustment required, and inability to form. When feedback data is injected, the attribute update mechanism of the spectrum node is triggered. If the same substance pair shows synergistic effect in three independent experiments, its edge weight is increased by 0.

2. If it shows chemical inhibition in all three experiments, a new chemical inhibition edge is added and assigned an initial weight of 0.

8.

8. The knowledge graph-driven hydrogel formulation prior and feasible domain screening system according to claim 1, characterized in that, In the multi-hop inference path search of the feasible domain boundary determination module of the recipe combination, a message passing mechanism based on graph neural network is used to propagate constraint signals along the relation path, accumulate risk scores for recipe nodes, and accelerate high-frequency rule path matching based on the pre-built constraint rule index tree.

9. The knowledge graph-driven hydrogel formulation prior and feasible domain screening system according to claim 1, characterized in that, The formulation prior rule reasoning and recommendation module adopts a graph attention network architecture, which represents the formulation as a subgraph structure. It dynamically calculates the contribution weight of each component and its interaction to the target performance index through a multi-layer attention mechanism, and outputs the performance prediction range and the process parameter recommendation range.

10. The knowledge graph-driven hydrogel formulation prior and feasible domain screening system according to claim 9, characterized in that, The graph attention network architecture adopts a transfer learning strategy. It is first pre-trained on a general polymer material knowledge graph and then fine-tuned on a hydrogel-specific graph. The performance prediction range is optimized by a quantile regression loss function, and the process parameter range is searched for the Pareto optimal frontier in the historical case parameter space by a Bayesian optimization algorithm.

Images