A method and system for reverse design of hydrogel materials based on machine learning

Abstract

This invention discloses a machine learning-based reverse design method and system for hydrogel materials. The method includes: acquiring stress-strain curve data, material characteristics, and process parameters of hydrogel samples with different component ratios, and preprocessing them to construct a training dataset; establishing a performance prediction model based on the dataset regarding the relationship between hydrogel components and ductility, toughness, and maximum stress; introducing a target performance range to construct constraints, performing performance prediction and constraint screening on candidate component ratios; and comprehensively optimizing multiple performance indicators through joint error evaluation to obtain a hydrogel formulation scheme that meets the target performance range. This method combines an ensemble learning model with a range constraint optimization strategy to achieve reverse design with multi-performance synergistic optimization, improving material screening efficiency and design accuracy, reducing experimental trial-and-error costs, and is suitable for rapid formulation design and optimization of hydrogel materials and related biomedical materials.

Description

Technical Field

[0001] This invention belongs to the field of intelligent material design technology, specifically relating to a reverse design method and system for hydrogel materials based on machine learning. Background Technology

[0002] Hydrogels are widely used in regenerative medicine, cartilage repair, and 3D bioprinting due to their excellent biocompatibility and tunability. However, there is a complex nonlinear relationship between the properties of hydrogel materials and their component ratios, and it is often difficult to achieve the desired balance between different performance indicators (ductility, toughness, maximum stress).

[0003] Traditional research often relies on experimental trial and error to adjust the proportions. This method is not only inefficient and costly, but also makes it difficult to quickly obtain the optimal material system that meets multiple performance requirements in a high-dimensional composition space.

[0004] While some existing machine learning methods can be used to predict material properties, most are still limited to optimizing a single model or a single performance index, making it difficult to achieve synergistic optimization of multiple performance indices simultaneously. Consequently, they lack systematic reverse design capabilities, and it is particularly difficult to achieve stable and reliable automated material screening under multi-objective constraints.

[0005] For example, CN121862281A discloses a reverse design method for the microstructure of starch hydrogels. By collecting images of the micromorphology of starch hydrogels and their corresponding target performance index vectors, a standardized dataset is constructed. A nonlinear mapping relationship between macroscopic mechanical properties and microscopic topological structure is established by using a multi-scale self-attention mechanism and a two-stream physical consistency identification strategy. At the same time, by combining a differentiable physical performance prediction proxy model and a multi-dimensional physical consistency coupling loss function, structure generation and performance approximation based on adversarial iterative optimization are achieved.

[0006] However, this type of method is still mainly aimed at generative design tasks based on microstructure images. Its optimization process relies on adversarial generation and physical consistency constraints, which has certain limitations in practical material component ratio design problems. For example, it lacks the ability to directly map the experimental component parameter space, and in the hydrogel ratio optimization problem facing continuous component variables, it is difficult to directly support multi-performance collaborative screening and rapid engineering decision-making based on interval constraints.

[0007] Existing technologies still have the following shortcomings in the reverse design process for the component ratio of hydrogel materials: 1. Lack of direct reverse modeling capability for component ratio space; 2. Difficulty in achieving synergistic constraints and balance optimization among multiple performance indicators; 3. The efficiency of automated screening and optimization under multi-objective interval constraints still needs to be improved; 4. Existing methods have insufficient support for rapid decision-making on material ratios in engineering applications.

[0008] Therefore, there is an urgent need for a technical solution that can be used to achieve efficient reverse design and optimization screening under the synergistic constraints of multiple performance indicators in the hydrogel composition space. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a machine learning-based reverse design method and system for hydrogel materials.

[0010] The objective of this invention can be achieved through the following technical solutions: This invention provides a machine learning-based reverse design method for hydrogel materials, comprising the following steps: S1. Obtain hydrogel sample data with different group ratios, and preprocess the hydrogel sample data to obtain a training dataset; S2. Based on the training dataset, construct a hydrogel material performance prediction model and establish the mapping relationship between hydrogel component ratio and performance index. S3. Obtain the target performance range of the performance index, and construct constraints based on the target performance range; S4. Input the candidate group allocation ratio into the hydrogel material performance prediction model for performance prediction, and filter the prediction results based on the constraints to obtain candidate hydrogel formulation schemes that meet the target performance range. S5. Output the predicted performance results corresponding to the candidate hydrogel formulation scheme.

[0011] Furthermore, the hydrogel sample data includes stress-strain curve data, material characteristic parameters, and process characteristic parameters. The material characteristic parameters include molecular weight, degree of substitution, viscosity, swelling ratio, charge density, and glass transition temperature. The process characteristic parameters include printing temperature, ultraviolet light intensity, curing time, total solids content, and strain rate.

[0012] Furthermore, the component ratio includes the content ratio of GelMA, PEGDA and CMC.

[0013] Furthermore, the preprocessing of the hydrogel sample data to obtain the training dataset specifically includes: The hydrogel sample data were processed by removing missing values, filtering outliers, eliminating noise, and normalizing. Based on the preprocessed stress-strain curve data, the fracture strain, peak stress, and curve integral area are extracted, and the fracture strain, peak stress, and curve integral area are used as performance indicators, respectively. The integral area of ​​the curve is obtained by integrating the stress-strain curve, and the calculation formula is as follows: in, This indicates the toughness performance index. This represents the stress value corresponding to the strain. Indicates fracture strain; The performance indicators are associated with the corresponding component ratios, material characteristic parameters, and process characteristic parameters to generate a training dataset for model training.

[0014] Furthermore, the construction of the hydrogel material performance prediction model based on the training dataset specifically includes: Multiple machine learning models are constructed based on the training dataset, and the multiple machine learning models are used to train and predict each performance metric. The corresponding coefficient of determination is calculated based on the prediction results of each machine learning model. And according to the determination coefficient Target machine learning models; An ensemble learning prediction model is constructed based on the selected target machine learning model, and a mapping relationship between group allocation ratio and performance indicators is established.

[0015] Furthermore, the coefficient of determination The calculation formula is: in, Indicates the actual performance value. Indicates the predicted performance value. This represents the true average performance. Indicates the number of samples.

[0016] Furthermore, the ensemble learning prediction model is constructed by combining a base learner and a meta learner. The base learner includes at least one of the following: random forest model, extreme random tree model, gradient boosting tree model, XGBoost model, and LightGBM model. The meta learner adopts the Ridge regression model.

[0017] Furthermore, the construction of constraints based on the target performance range specifically includes: Obtain the target performance range corresponding to each performance index, and determine the upper limit and lower limit of the range for each performance index respectively; Performance constraints are constructed based on the upper and lower limits of each interval to limit the predicted performance value corresponding to the candidate group allocation ratio to be within the corresponding target performance interval. The performance constraints are expressed as follows: in, Indicates the first Predicted values ​​of each performance indicator Indicates the first Each performance metric corresponds to the lower limit of the target performance range. Indicates the first Each performance metric corresponds to the upper limit of the target performance range; The performance indicators include ductility, maximum stress, and toughness.

[0018] Furthermore, the step of inputting the candidate group allocation ratio into the hydrogel material performance prediction model for performance prediction, and filtering the prediction results based on the constraints to obtain candidate hydrogel formulation schemes that meet the target performance range, specifically includes: Multiple candidate group allocation ratios are generated, and each candidate group allocation ratio is input into the hydrogel material performance prediction model for performance prediction to obtain the predicted values ​​of ductility, maximum stress and toughness corresponding to each candidate group allocation ratio. Calculate the joint error value corresponding to the allocation ratio of each candidate group based on the constraints, and filter and sort the allocation ratio of each candidate group according to the joint error value; The formula for calculating the joint error value is as follows: in, Indicates the joint error value. Indicates the first The weighting coefficients corresponding to each performance indicator Indicates the first The interval deviation value corresponding to each performance indicator Indicates the number of performance indicators; The formula for calculating the interval deviation value is as follows: in, Indicates the first Predicted values ​​of each performance indicator Indicates the first The target performance range corresponding to each performance indicator; The candidate group allocation ratio with a joint error value lower than a preset threshold is determined as the candidate hydrogel formulation scheme that meets the target performance range.

[0019] Another aspect of the present invention provides a machine learning-based reverse design system for hydrogel materials, comprising: The data acquisition module is used to acquire hydrogel sample data with different component ratios, and to collect corresponding stress-strain curve data, material characteristic parameters and process characteristic parameters. The data processing module is used to preprocess the hydrogel sample data, extract performance indicators, and generate a training dataset. The model building module is used to build a hydrogel material performance prediction model based on the training dataset and establish the mapping relationship between component ratios and performance indicators. A constraint construction module is used to obtain the target performance range and construct constraint conditions based on the target performance range; The reverse optimization module is used to input the candidate group allocation ratio into the hydrogel material performance prediction model for performance prediction, and to filter the prediction results based on the constraints to obtain candidate hydrogel ratio schemes that meet the target performance range. The results output module is used to output the candidate hydrogel formulation scheme and its corresponding predicted performance results.

[0020] Compared with the prior art, the present invention has the following advantages: (1) In the prior art, there is a general lack of direct reverse modeling capability for the component ratio space in the reverse design process of hydrogel materials, which makes it difficult to deduce the corresponding material ratio from the target performance and still mainly relies on forward prediction or empirical trial and error for material design. This invention constructs a hydrogel material performance prediction model based on ensemble learning and establishes a performance mapping relationship with the component ratio as input variable. At the same time, combined with the automatic generation mechanism of candidate ratio space, it realizes the direct modeling and mapping relationship from the material component space to the performance space, thereby effectively improving the reverse modeling capability of the hydrogel component ratio space, enabling the automatic generation of material ratio driven by the target performance, and significantly reducing the reliance on traditional empirical trial and error.

[0021] (2) In the prior art, the problem of achieving coordinated constraints and balanced optimization among multiple performance indicators is quite prominent. When there are conflicts between different performance indicators, it is easy to have a single performance that is optimal but the overall performance does not meet the engineering requirements. This invention introduces interval constraints based on the target performance range on the basis of the performance prediction model, and further constructs a multi-performance joint error evaluation mechanism to uniformly constrain and comprehensively evaluate multiple performance indicators such as ductility, toughness and maximum stress. This enables multiple performance indicators to be coordinated and adjusted under the same optimization framework, thereby achieving balanced optimization among multiple performance indicators and improving the comprehensive adaptability and engineering usability of material design results in practical applications.

[0022] (3) In the prior art, the efficiency of automated screening and optimization under multi-objective interval constraints is low. It usually requires a large number of iterative calculations or relies on complex optimization strategies, resulting in high computational costs and insufficient convergence efficiency. This invention achieves efficient screening and optimization under a large-scale candidate space by inputting the candidate group allocation ratio into the ensemble learning prediction model for rapid batch prediction, and screening the prediction results in combination with interval constraints. At the same time, it uses the joint error function to sort and constrain the candidate schemes, thereby significantly improving the optimization efficiency under multi-objective constraints and reducing the computational cost of invalid candidate solutions.

[0023] (4) In the prior art, the ability to provide rapid decision support for the formulation of hydrogel materials in engineering applications is insufficient, making it difficult to output stable and reliable formulation schemes in a short period of time to meet the needs of practical applications. This invention constructs an ensemble learning prediction model that integrates multiple machine learning models, and combines a feature contribution analysis method based on SHAP values ​​to adaptively determine the weights of performance indicators. This enables the model to not only improve prediction accuracy but also enhance the interpretability and stability of the results. At the same time, with the inverse optimization screening mechanism, it realizes the ability to rapidly output decisions from target performance to material formulation, thereby significantly improving the system's response speed and decision reliability in engineering application scenarios. Attached Figure Description

[0024] Figure 1 This is a flowchart of the reverse design method for hydrogel materials according to an embodiment of the present invention; Figure 2 This is a performance comparison chart of the models in this embodiment of the invention; Figure 3 This is a distribution diagram of the SHAP Feature Impact of various performance aspects in embodiments of the present invention; Figure 4 This is a distribution chart showing the percentage distribution of feature importance in an embodiment of the present invention; Figure 5 This is a block diagram of the reverse design system for hydrogel materials according to an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0026] Example 1: This embodiment provides a machine learning-based reverse design method for hydrogel materials, such as... Figure 1 As shown, it includes the following steps: S1. Obtain hydrogel sample data with different group ratios and preprocess the hydrogel sample data to obtain the training dataset. In one specific implementation, experimental data of hydrogel samples with different component ratios are acquired, and a training dataset for model training is constructed based on the experimental data. Hydrogel samples are prepared using different proportions of GelMA, PEGDA, and CMC, wherein the ratios of GelMA, PEGDA, and CMC are adjusted to form hydrogel systems with different mechanical properties. Uniaxial compression experiments are performed on the prepared hydrogel samples, and the corresponding stress-strain curve data are collected and entered into a database.

[0027] In addition to stress-strain curve data, material characteristic parameters and process characteristic parameters related to material properties are also recorded. Material characteristic parameters include molecular weight, degree of substitution, viscosity, swelling ratio, charge density, and glass transition temperature, used to characterize the physicochemical differences between different raw materials. Process characteristic parameters include printing temperature, UV irradiation intensity, curing time, total solids content, and strain rate, used to reflect the impact of processing technology on material properties. Since hydrogel properties are affected not only by component ratios but also by processing conditions and the inherent properties of the material, simultaneously introducing material and process characteristic parameters helps improve the subsequent model's ability to learn complex nonlinear relationships, thereby enhancing the accuracy of performance prediction.

[0028] In the data preprocessing stage, the original experimental data were first subjected to missing value removal, outlier filtering, and noise reduction. Outlier filtering employed a statistical distribution-based screening method to remove abnormal sampling data caused by experimental errors. Noise reduction used a smoothing filter to reduce the impact of experimental sampling fluctuations on the stress-strain curve. Subsequently, data of different dimensions were normalized to avoid interference from differences in parameter scales during model training and to improve model convergence stability.

[0029] After pretreatment, fracture strain, peak stress, and area under the curve are extracted from the stress-strain curve and used as indicators of ductility, maximum stress, and toughness, respectively. Among them, fracture strain is used to characterize the maximum deformation capacity of the material before fracture, peak stress is used to reflect the maximum load that the material can withstand, and area under the curve is used to characterize the energy absorption capacity of the material during the stress process. Therefore, it can comprehensively reflect the overall mechanical properties of hydrogel materials.

[0030] The toughness performance index is obtained by integrating the stress-strain curve, and the calculation formula is as follows: in, This indicates the toughness performance index. This represents the stress value corresponding to the strain. Indicates fracture strain; After extracting the performance indicators, the component ratios, material characteristic parameters, process characteristic parameters, and performance indicators of each hydrogel sample are associated and stored to form a training dataset for model training. By establishing the correspondence between multi-dimensional input parameters and multiple performance indicators, basic data support can be provided for subsequent machine learning models to learn the nonlinear mapping relationship between hydrogel components and performance.

[0031] S2. Construct a hydrogel material performance prediction model based on the training dataset, and establish the mapping relationship between hydrogel component ratio and performance index; In one specific implementation, after constructing the training dataset, a hydrogel material performance prediction model is established based on the training dataset to construct a nonlinear mapping relationship between component ratios and performance indicators. Since the performance of hydrogel materials is simultaneously affected by component ratios, material characteristic parameters, and process characteristic parameters, and there are strong coupling relationships and nonlinear variation characteristics among these parameters, multiple machine learning models are used for training and prediction to improve the fitting ability and generalization ability of different performance indicators.

[0032] Specifically, we constructed Random Forest, ExtraTrees, Gradient Boosting, XGBoost, LightGBM, Ridge Regression, and ElasticNet models, respectively, and independently trained and validated them for three performance metrics: scalability, resilience, and maximum stress. Different models have different advantages in handling nonlinear relationships, feature coupling, and small sample data. By introducing a multi-model training mechanism, we can avoid the dependence of a single model on a specific data distribution, thereby improving the overall prediction stability.

[0033] During model training, the training dataset is divided into a training set and a test set, and the corresponding coefficient of determination R is calculated based on the prediction results of each model. 2 The coefficient of determination is used to evaluate the model's ability to fit performance metrics. The calculation formula is: in, Indicates the actual performance value. Indicates the predicted performance value. This represents the true average performance. Indicates the number of samples.

[0034] After selecting the target machine learning model, an ensemble learning prediction model is constructed based on the selection results. The ensemble learning prediction model is constructed by combining a base learner and a meta-learner. The base learner includes at least one of the following: random forest model, extreme random tree model, gradient boosting tree model, XGBoost model, and LightGBM model. The meta-learner is the Ridge regression model.

[0035] Among them, random forest and extreme random tree models can improve the model's ability to fit complex nonlinear relationships, gradient boosting tree, XGBoost and LightGBM models can enhance the model's ability to learn local feature changes, and Ridge regression model can reduce the collinearity effect between the output results of multiple base learners by introducing a regularization mechanism, thereby improving the stability of the model after fusion.

[0036] In the model fusion process, the Stacking ensemble learning framework is used to fuse the outputs of multiple base learners, and cross-validation is employed to generate meta-learner input data to reduce the risk of model overfitting. Compared to a single machine learning model, the ensemble learning prediction model can simultaneously integrate the predictive advantages of different models, improving the model's ability to express complex material performance relationships. This enhances the accuracy and generalization ability of hydrogel material performance prediction, providing more stable and reliable performance prediction results for subsequent inverse optimization.

[0037] S3. Obtain the target performance range of the performance index and construct constraints based on the target performance range; In one specific implementation, after constructing the hydrogel material performance prediction model, the target performance ranges corresponding to each performance index are further obtained, and constraints are constructed based on the target performance ranges to achieve targeted screening and reverse optimization of the candidate group allocation ratio. In practical applications, different application scenarios have different requirements for the performance of hydrogel materials. For example, cartilage repair materials usually need to simultaneously consider high ductility, appropriate toughness, and sufficient mechanical strength. Therefore, it is necessary to coordinate constraints on multiple performance indicators, rather than optimizing only a single performance.

[0038] In the specific implementation process, the target performance range corresponding to each performance index is input according to the actual application requirements. For example, the target range for ductility is set to 30% to 60%, and the target range for toughness is set to 2 to 6 MJ / m. 3 The maximum stress target range was set to 0.5–1.0 MPa. Subsequently, the upper and lower limits of each performance index were determined, and performance constraints were constructed based on these limits to ensure that the predicted performance values ​​corresponding to the candidate group allocation ratios were within the preset target range.

[0039] The performance constraints are expressed as follows: in, Indicates the first Predicted values ​​of each performance indicator Indicates the first Each performance metric corresponds to the lower limit of the target performance range. Indicates the first Each performance metric corresponds to the upper limit of the target performance range; The performance indicators include ductility, maximum stress, and toughness.

[0040] The aforementioned constraint method sets interval limits for each performance index, ensuring that the candidate material ratio not only meets individual performance requirements but also satisfies the synergistic needs between multiple performance indices. Compared to methods that only use single-objective extreme value optimization, the interval constraint method can avoid the problem of over-optimization of one performance index leading to a significant decline in other performances, thereby improving the balance of the material's overall performance and its engineering applicability.

[0041] Furthermore, during the candidate component ratio search process, a ternary component space composed of GelMA, PEGDA, and CMC is used as the search range, and an initial candidate ratio is generated by combining a grid search method. The grid search improves the completeness of the candidate space coverage by discretizing and traversing different component ratios, thereby avoiding the omission of potential preferred ratios in local areas.

[0042] Building upon this, Bayesian optimization is further employed to iteratively optimize the candidate group allocation ratio. Bayesian optimization dynamically adjusts the subsequent search direction using existing prediction results, gradually concentrating the search process towards regions where performance meets the target range. This reduces the number of invalid searches and improves search efficiency in complex component spaces. Compared to methods relying entirely on exhaustive search, this approach can quickly obtain the candidate group allocation ratio that meets the target performance range with fewer iterations, improving reverse engineering efficiency and reducing computational costs.

[0043] Building upon this foundation, Bayesian optimization is further employed to iteratively optimize the candidate group allocation ratio. Bayesian optimization uses a Gaussian process as a surrogate model, constructing a probability distribution model based on the relationship between the candidate group allocation ratio and the corresponding predicted performance. It then dynamically updates the subsequent search direction based on existing sampling results, gradually concentrating the search process towards regions that meet the target performance range. This reduces the number of invalid searches and improves search efficiency in complex component spaces.

[0044] The interval constraint error function is defined as the weighted sum of the deviations of each performance index from the target performance interval, and its calculation formula is as follows: in, Indicates the joint error value. Indicates the first The weighting coefficients corresponding to each performance indicator Indicates the first The interval deviation value corresponding to each performance indicator Indicates the number of performance indicators; The formula for calculating the interval deviation is as follows: in, Indicates the first Predicted values ​​of each performance indicator Indicates the first The target performance range corresponding to each performance indicator; The aforementioned joint error function calculates the degree to which each performance index deviates from the target range and assigns corresponding weights to different performance indices, thereby achieving a comprehensive evaluation of multiple target performances. When the predicted value is within the target performance range, the corresponding range deviation is 0; when the predicted value exceeds the target performance range, the greater the deviation, the higher the corresponding error value. This method can simultaneously consider the balance between multiple performance indices, avoiding over-optimization of a single performance that could lead to an imbalance in overall material performance.

[0045] Furthermore, the weighting coefficients corresponding to each performance index can be dynamically adjusted according to different application scenarios. For example, in cartilage repair applications, the weight corresponding to toughness can be appropriately increased to enhance the material's impact resistance; in 3D bioprinting applications, the weight corresponding to ductility can be appropriately increased to improve structural stability during the printing process. Through dynamic weight allocation, the adaptability of the reverse design process to different application requirements can be improved.

[0046] The Bayesian optimization process further incorporates the Expected Improvement (EI) criterion. The EI criterion selects the next sampling location by comprehensively considering the current optimal result and prediction uncertainty, thus maintaining a balance between "local performance improvement" and "global search capability." Compared to methods that only search based on the current optimal value, this approach effectively reduces the risk of getting trapped in local optima, thereby improving optimization efficiency and candidate ratio search accuracy in complex component spaces.

[0047] S4. Input the candidate group allocation ratio into the hydrogel material performance prediction model for performance prediction, and filter the prediction results based on the constraints to obtain the candidate hydrogel ratio scheme that meets the target performance range. In one specific implementation, based on the constructed ensemble learning prediction model, the candidate group allocation ratio is batch-based for performance prediction, and the prediction results are screened in combination with the constraints corresponding to the target performance range to obtain candidate hydrogel formulation schemes that meet the performance requirements.

[0048] In the specific implementation process, multiple candidate group allocation ratios are generated in the ternary component space composed of GelMA, PEGDA and CMC. The material characteristic parameters and process characteristic parameters corresponding to each candidate group allocation ratio are input into the integrated learning prediction model for performance prediction, thereby obtaining the predicted values ​​of ductility, maximum stress and toughness corresponding to each candidate group allocation ratio.

[0049] Since there are complex nonlinear relationships between different component proportions, ensemble learning prediction models can integrate the prediction results of different machine learning models, improve the performance prediction accuracy in complex component spaces, and reduce the impact of single model prediction fluctuations on the stability of results.

[0050] After obtaining the prediction results, the allocation ratios of each candidate group are further screened based on the pre-constructed performance constraints. For candidate group allocation ratios whose predicted performance values ​​are within the target performance range, they are retained as valid candidate solutions; for candidate group allocation ratios whose predicted performance values ​​exceed the target performance range, the joint error value is calculated based on the degree to which the corresponding performance index deviates from the target range, and the allocation ratios are sorted and filtered according to the joint error value.

[0051] Furthermore, the candidate group allocation ratios are sorted according to the joint error value, and the candidate group allocation ratios with joint error values ​​below a preset threshold are determined as candidate hydrogel formulation schemes that meet the target performance range. In some embodiments, the top K candidate group allocation ratios with the smallest joint error values ​​can be further selected as the preferred scheme output.

[0052] The above approach enables rapid screening of candidate material ratios that meet the synergistic requirements of multiple performance indicators within a large-scale component space, reducing the time and cost consumption caused by numerous invalid experiments in traditional experimental screening processes. Furthermore, the unified constraint of multiple performance indicators through a joint error evaluation mechanism helps improve the overall performance stability of candidate ratio schemes in practical engineering applications.

[0053] S5. Output the predicted performance results corresponding to the candidate hydrogel formulation schemes.

[0054] Example 2: This embodiment verifies the machine learning-based method for predicting and reverse designing the properties of hydrogel materials, and further analyzes the importance and interpretability of features in the model prediction process.

[0055] Hydrogel samples with different component ratios were prepared experimentally, including GelMA, PEGDA, and CMC. Uniaxial mechanical tests were performed on each sample to obtain stress-strain curve data. Simultaneously, the corresponding processing condition parameters and material physicochemical parameters were recorded. The processing condition parameters included ultraviolet light irradiation energy, strain rate, and total solid content, while the material physicochemical parameters included molecular weight, degree of substitution, glass transition temperature, storage modulus, swelling ratio, and charge density.

[0056] The collected multi-source data is input into the system for unified processing. During data processing, outliers and missing values ​​are first filtered out, and normalization is used to eliminate numerical differences between parameters of different dimensions, thereby improving the stability of subsequent model training. Based on this, feature engineering methods are further employed to generate multiple interactive feature terms, including combined features such as “UV×StrainRate”, “MolWt×StrainRate”, and “StrainRate×TotalSolid”.

[0057] The introduction of interactive features is mainly used to characterize the coupling relationship between material composition parameters and processing parameters. Since the mechanical properties of hydrogel materials are usually affected by multiple parameters, it is difficult to fully describe the complex nonlinear relationship using only a single feature. Therefore, constructing interactive terms can enhance the model's ability to express multivariate coupling effects and improve the model's prediction accuracy.

[0058] After feature construction, XGBoost, LightGBM, Random Forest, Extremely Random Tree, Gradient Boosting Tree, Ridge Regression and ElasticNet models were constructed and independently trained and optimized for three target performances: extensibility, toughness and maximum stress.

[0059] like Figure 2 As shown, during model training, the coefficient of determination R for each model is calculated on both the training and test sets. 2 To evaluate the predictive performance of different models.

[0060] Experimental results show that the XGBoost model, gradient boosting tree model, and Stacking ensemble learning model all have high coefficients of determination R in various performance predictions. 2 The test set determination coefficient R0 2 The values ​​are generally higher than 0.95, indicating that the model can effectively establish the complex nonlinear mapping relationship between the hydrogel composition ratio and performance indicators, and has good generalization ability and prediction stability.

[0061] To further analyze the internal decision-making mechanism of the model, this embodiment uses the SHAP (SHapley Additive ex Planations) algorithm to perform interpretive analysis on the model, in order to obtain the degree of influence and contribution relationship of each feature on different performance indicators.

[0062] like Figure 3 As shown, the SHAP analysis results indicate that the features with the greatest impact on toughness properties include “MolWt×StrainRate”, “StrainRate×TotalSolid”, and UV irradiation energy; the features with significant impact on ductility properties include the PEGDA / CMC ratio, “StrainRate×GelMA”, and “UV×GelMA”; and the features with significant impact on maximum stress properties mainly include “UV×StrainRate”, “MolWt×StrainRate”, and “log(StrainRate)”.

[0063] The above results indicate that the mechanical properties of hydrogel materials are influenced not only by the component ratio but also by the coupling effect between processing parameters and material parameters. Feature interpretation using the SHAP algorithm can identify key parameters that significantly impact performance prediction, thereby improving the interpretability of the model prediction process and its reliability in engineering applications.

[0064] Furthermore, such as Figure 4 As shown, the absolute values ​​of the SHAP values ​​corresponding to each feature were normalized to calculate the contribution ratio of different features to each performance index. The results show that for toughness performance, "MolWt×StrainRate" has the highest contribution rate, approximately 20%, the top three features contribute approximately 40% cumulatively, and the top ten features contribute approximately 68% cumulatively; for ductility performance, the PEGDA / CMC ratio has the highest contribution rate, approximately 14%, the top three features contribute approximately 28% cumulatively, and the top ten features contribute approximately 58% cumulatively; for maximum stress performance, "UV×StrainRate" has the highest contribution rate, approximately 10%, the top three features contribute approximately 23% cumulatively, and the top ten features contribute approximately 58% cumulatively.

[0065] The above results show that the model decisions for each performance index are mainly concentrated on a small number of key features. The top ten important features can explain about 60% to 70% of the model prediction results, indicating that the model has good stability and interpretability.

[0066] After completing model training and interpretive analysis, the system further generates comprehensive performance distribution results based on the output results of each performance predictor, and combines Bayesian optimization and interval constraints to screen and rank the candidate group allocation ratio.

[0067] During the candidate group allocation ratio search process, a Bayesian optimization algorithm is used to iteratively search the ternary component space composed of GelMA, PEGDA, and CMC, and a joint error function is used to coordinately constrain multiple performance indicators. This approach can complete the performance prediction and ranking of a large number of candidate combinations in a short time, thereby improving the efficiency of reverse engineering.

[0068] Experimental results show that the model predictions have a high degree of consistency with the actual test results, and the test set determination coefficient R0 is high. 2 The result is greater than 0.95, and the average prediction error is less than 5%. The results verify the accuracy, stability, and engineering application feasibility of the method of the present invention in the prediction and reverse design of hydrogel material properties.

[0069] Example 3: This embodiment provides a machine learning-based reverse design system for hydrogel materials, such as... Figure 5 As shown, it includes a data acquisition module, a data processing module, a model building module, a constraint building module, a reverse optimization module, and a result output module.

[0070] The data acquisition module is used to acquire hydrogel sample data with different component ratios and collect corresponding stress-strain curve data, material characteristic parameters, and process characteristic parameters. The component ratios include the content ratios of GelMA, PEGDA, and CMC. The material characteristic parameters include molecular weight, degree of substitution, viscosity, swelling ratio, charge density, glass transition temperature, and storage modulus. The process characteristic parameters include printing temperature, ultraviolet light intensity, curing time, total solids content, and strain rate.

[0071] The data acquisition module can communicate with uniaxial compression testing equipment, rheological testing equipment, and 3D bioprinting equipment to achieve automatic data collection and uploading. By simultaneously acquiring material composition information, processing technology information, and mechanical response information, the completeness and reliability of subsequent model training data can be improved.

[0072] The data processing module is used to preprocess the acquired hydrogel sample data, extract performance indicators, and generate a training dataset. Specifically, the data processing module performs missing value removal, outlier filtering, noise elimination, and normalization on the raw data, and extracts performance indicators such as fracture strain, maximum stress, and curve integral area based on the stress-strain curve.

[0073] Furthermore, the data processing module can also generate multiple interactive feature terms based on feature engineering methods, including combined features such as “UV×StrainRate”, “MolWt×StrainRate”, and “StrainRate×TotalSolid”, to enhance the model’s ability to express the coupling relationship between material parameters and process parameters.

[0074] The model building module is used to construct performance prediction models for hydrogel materials based on the training dataset, establishing the mapping relationship between component ratios and performance indicators. The module constructs random forest, extreme random tree, gradient boosting tree, XGBoost, LightGBM, Ridge regression, and ElasticNet models, respectively, and trains and predicts ductility, toughness, and maximum stress.

[0075] The model building module is based on the coefficient of determination R. 2 The target machine learning model is automatically selected, and an ensemble learning prediction model is further constructed using the Stacking ensemble learning approach to improve the performance prediction accuracy and generalization ability in complex component spaces.

[0076] The constraint construction module is used to obtain the target performance range and construct constraints based on the target performance range. Specifically, the constraint construction module determines the upper and lower limits of the range for each performance index and establishes the performance constraints.

[0077] By setting target performance ranges, multiple performance indicators such as ductility, toughness, and maximum stress can be constrained simultaneously, thereby improving the overall performance balance of materials.

[0078] The reverse optimization module is used to input the candidate group allocation ratio into the hydrogel material performance prediction model for performance prediction, and to filter the prediction results based on the constraints to obtain candidate hydrogel ratio schemes that meet the target performance range.

[0079] In the specific implementation process, the inverse optimization module first generates multiple candidate group allocation ratios in the ternary component space composed of GelMA, PEGDA, and CMC, and obtains the corresponding performance prediction results based on the ensemble learning prediction model. The prediction results are then comprehensively evaluated using a joint error function. The inverse optimization module further combines grid search and Bayesian optimization to search and update the candidate group allocation ratio. Bayesian optimization uses a Gaussian process as a surrogate model and combines iterative sampling of the candidate group allocation ratio with the expected improvement criterion, thereby improving the search efficiency in complex component spaces and reducing computational costs.

[0080] The results output module outputs candidate hydrogel formulations and their corresponding predicted performance results. This module can further generate performance radar charts, error distribution maps, and ternary phase diagrams, and supports data export in CSV, Excel, and Word formats for experimental verification and engineering application analysis.

[0081] The system provided in this embodiment can automate the entire process of hydrogel material processing, from data acquisition, model training, performance prediction to reverse optimization screening. It can significantly reduce the cost of traditional experimental trial and error, improve the efficiency of multi-performance synergistic optimization, and enhance the engineering design capabilities of complex hydrogel material systems.

[0082] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0083] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A reverse design method for hydrogel materials based on machine learning, characterized in that, Includes the following steps: S1. Obtain hydrogel sample data with different group ratios, and preprocess the hydrogel sample data to obtain a training dataset; S2. Based on the training dataset, construct a hydrogel material performance prediction model and establish the mapping relationship between hydrogel component ratio and performance index. S3. Obtain the target performance range of the performance index, and construct constraints based on the target performance range; S4. Input the candidate group allocation ratio into the hydrogel material performance prediction model for performance prediction, and filter the prediction results based on the constraints to obtain candidate hydrogel formulation schemes that meet the target performance range. S5. Output the predicted performance results corresponding to the candidate hydrogel formulation scheme.

2. The reverse design method for hydrogel materials based on machine learning according to claim 1, characterized in that, The hydrogel sample data includes stress-strain curve data, material characteristic parameters, and process characteristic parameters. The material characteristic parameters include molecular weight, degree of substitution, viscosity, swelling ratio, charge density, and glass transition temperature. The process characteristic parameters include printing temperature, ultraviolet light intensity, curing time, total solids content, and strain rate.

3. The reverse design method for hydrogel materials based on machine learning according to claim 1, characterized in that, The composition ratio includes the content ratio of GelMA, PEGDA and CMC.

4. The reverse design method for hydrogel materials based on machine learning according to claim 1, characterized in that, The preprocessing of the hydrogel sample data to obtain the training dataset specifically includes: The hydrogel sample data were processed by removing missing values, filtering outliers, eliminating noise, and normalizing. Based on the preprocessed stress-strain curve data, the fracture strain, peak stress, and curve integral area are extracted, and the fracture strain, peak stress, and curve integral area are used as performance indicators, respectively. The integral area of ​​the curve is obtained by integrating the stress-strain curve, and the calculation formula is as follows: in, This indicates the toughness performance index. This represents the stress value corresponding to the strain. Indicates fracture strain; The performance indicators are associated with the corresponding component ratios, material characteristic parameters, and process characteristic parameters to generate a training dataset for model training.

5. The reverse design method for hydrogel materials based on machine learning according to claim 1, characterized in that, The construction of the hydrogel material performance prediction model based on the training dataset specifically includes: Multiple machine learning models are constructed based on the training dataset, and the multiple machine learning models are used to train and predict each performance metric. The corresponding coefficient of determination is calculated based on the prediction results of each machine learning model. And according to the determination coefficient Target machine learning models; An ensemble learning prediction model is constructed based on the selected target machine learning model, and a mapping relationship between group allocation ratio and performance indicators is established.

6. The reverse design method for hydrogel materials based on machine learning according to claim 5, characterized in that, The coefficient of determination The calculation formula is: in, Indicates the actual performance value. Indicates the predicted performance value. This represents the true average performance. Indicates the number of samples.

7. The reverse design method for hydrogel materials based on machine learning according to claim 5, characterized in that, The ensemble learning prediction model is constructed by combining a base learner and a meta learner. The base learner includes at least one of the following: random forest model, extreme random tree model, gradient boosting tree model, XGBoost model, and LightGBM model. The meta learner adopts the Ridge regression model.

8. The reverse design method for hydrogel materials based on machine learning according to claim 1, characterized in that, The constraint conditions constructed based on the target performance range specifically include: Obtain the target performance range corresponding to each performance index, and determine the upper limit and lower limit of the range for each performance index respectively; Performance constraints are constructed based on the upper and lower limits of each interval to limit the predicted performance value corresponding to the candidate group allocation ratio to be within the corresponding target performance interval. The performance constraints are expressed as follows: in, Indicates the first Predicted values ​​of each performance indicator Indicates the first Each performance metric corresponds to the lower limit of the target performance range. Indicates the first Each performance metric corresponds to the upper limit of the target performance range; The performance indicators include ductility, maximum stress, and toughness.

9. The reverse design method for hydrogel materials based on machine learning according to claim 1, characterized in that, The step of inputting the candidate group allocation ratio into the hydrogel material performance prediction model for performance prediction, and filtering the prediction results based on the constraints to obtain candidate hydrogel formulation schemes that meet the target performance range, specifically includes: Multiple candidate group allocation ratios are generated, and each candidate group allocation ratio is input into the hydrogel material performance prediction model for performance prediction to obtain the predicted values ​​of ductility, maximum stress and toughness corresponding to each candidate group allocation ratio. Calculate the joint error value corresponding to the allocation ratio of each candidate group based on the constraints, and filter and sort the allocation ratio of each candidate group according to the joint error value; The formula for calculating the joint error value is as follows: in, Indicates the joint error value. Indicates the first The weighting coefficients corresponding to each performance indicator Indicates the first The interval deviation value corresponding to each performance indicator Indicates the number of performance indicators; The formula for calculating the interval deviation value is as follows: in, Indicates the first Predicted values ​​of each performance indicator Indicates the first The target performance range corresponding to each performance indicator; The candidate group allocation ratio with a joint error value lower than a preset threshold is determined as the candidate hydrogel formulation scheme that meets the target performance range.

10. A system for the reverse design method of hydrogel materials based on machine learning as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to acquire hydrogel sample data with different component ratios, and to collect corresponding stress-strain curve data, material characteristic parameters and process characteristic parameters. The data processing module is used to preprocess the hydrogel sample data, extract performance indicators, and generate a training dataset. The model building module is used to build a hydrogel material performance prediction model based on the training dataset and establish the mapping relationship between component ratios and performance indicators. A constraint construction module is used to obtain the target performance range and construct constraint conditions based on the target performance range; The reverse optimization module is used to input the candidate group allocation ratio into the hydrogel material performance prediction model for performance prediction, and to filter the prediction results based on the constraints to obtain candidate hydrogel ratio schemes that meet the target performance range. The results output module is used to output the candidate hydrogel formulation scheme and its corresponding predicted performance results.

Images