A method for predicting performance of a thin film nanocomposite membrane based on optimized machine learning

Abstract

The application relates to the technical field of film material performance prediction, and discloses a thin film nanocomposite membrane performance prediction method based on optimized machine learning, which comprises the following steps: thin film nanocomposite membrane sample data is divided into discrete attribute features and continuous action features, and an input feature matrix is constructed; the contribution information of each input feature to water flux and salt rejection rate is determined respectively, the input features are double-target cooperatively screened, and a coupled feature subset is obtained; a performance prediction model is constructed and trained, the optimized performance prediction model is determined according to the comprehensive loss of water flux and salt rejection rate; the discrete attribute features, the continuous action features and the coupled features between the two in the coupled feature subset are verified, the coupled feature subset is iteratively updated, the performance prediction model is retrained, and the target performance prediction model is determined until the verification result meets the rationality requirement. The application establishes the coupling mapping relationship between the material features and the membrane performance, and improves the accuracy of the membrane performance prediction.

Description

Technical Field

[0001] This invention relates to the field of membrane material performance prediction technology, and more specifically, to a method for predicting the performance of thin film nanocomposite membranes based on optimized machine learning. Background Technology

[0002] Thin film nanocomposite (TFN) membranes have become an important development direction in membrane separation materials due to their excellent separation performance in water treatment, resource recovery, and high-salinity wastewater treatment. By introducing inorganic nanomaterials into the polyamide separation layer, the microstructure, interfacial properties, and mass transfer channels of the membrane can be adjusted, thereby affecting the membrane's water flux and salt rejection rate.

[0003] However, the performance of thin-film nanocomposite membranes is not determined by a single factor, but is jointly related to factors such as the chemical composition, functional group type, material morphology, dosage, pore size, membrane contact angle, surface roughness, polyamide layer thickness, feed concentration, and operating pressure of the inorganic nanomaterials. These factors include discrete material properties such as chemical elements, functional groups, and material morphology, as well as continuous structural and operational parameters such as dosage, pore size, pressure, and concentration. Furthermore, there are complex coupling relationships between different types of parameters, making the performance prediction and optimization design of thin-film nanocomposite membranes highly challenging.

[0004] Current research on the performance of thin-film nanocomposite membranes mainly relies on experimental testing and empirical analysis. This typically involves observing the impact of single parameters such as the amount of material added, pore size, surface properties, or operating pressure on water flux or salt rejection rate. While these methods can reveal local patterns, they struggle to reflect the nonlinear coupling relationships between material composition, membrane structure, and operating conditions. Furthermore, they are costly, time-consuming, and difficult to rapidly screen a large number of candidate inorganic nanomaterials. In recent years, machine learning methods have been used to predict membrane material performance. By collecting existing literature or experimental data, they establish mapping relationships between material parameters and membrane performance, thereby assisting in the design of thin-film nanocomposite membranes. These methods have improved prediction efficiency to some extent.

[0005] However, from the perspective of the directional design requirements of thin-film nanocomposite membranes, existing technologies still have significant shortcomings. On the one hand, the chemical elements, functional group types, and material morphology in thin-film nanocomposite membranes are not ordinary numerical parameters; they reflect the compositional origin of inorganic nanomaterials and the basis of interfacial interactions. On the other hand, continuous parameters such as the amount added, pore size, membrane contact angle, surface roughness, polyamide layer thickness, feed concentration, and operating pressure directly affect the membrane structure and operating state. Existing methods often simply incorporate these two types of characteristics into the same input space, failing to further analyze how discrete material properties affect membrane performance through continuous structural and operating parameters. As a result, although the model can provide predictive results, it is difficult to provide a clear basis for material composition selection and process parameter adjustment. On the other hand, water flux and salt rejection rate are usually not unidirectional optimization objectives in the design of thin film nanocomposite membranes. Some parameters may increase water flux but have an adverse effect on salt rejection rate, while some parameters may not contribute significantly to a single objective but may play an important role in the balance between the two. Existing models mostly predict water flux and salt rejection rate separately or screen based solely on the importance of a single feature, which can easily lead to the accidental deletion of features that are valuable for the synergistic optimization of the two objectives, and also to the retention of features that are only beneficial to a single objective but destroy the overall separation performance.

[0006] There are currently no effective solutions to the problems in the relevant technologies. Summary of the Invention

[0007] To address the problems in related technologies, this invention proposes a method for predicting the performance of thin-film nanocomposite membranes based on optimized machine learning, in order to overcome the aforementioned technical problems existing in the current related technologies.

[0008] Therefore, the specific technical solution adopted by the present invention is as follows:

[0009] A method for predicting the performance of thin-film nanocomposite membranes based on optimized machine learning is disclosed. This method includes: dividing thin-film nanocomposite membrane sample data into discrete attribute features and continuous interaction features, and constructing an input feature matrix; determining the contribution information of each input feature to water flux and salt rejection rate based on the input feature matrix, and using the contribution information to perform dual-objective collaborative screening of the input features to obtain a coupled feature subset; constructing and training a performance prediction model based on the coupled feature subset, and determining the optimized performance prediction model based on the combined loss of water flux and salt rejection rate; verifying the discrete attribute features, continuous interaction features, and the coupling features between them in the coupled feature subset based on the optimized performance prediction model; if the verification results do not meet the reasonableness requirements, iteratively updating the coupled feature subset and retraining the performance prediction model until the verification results meet the reasonableness requirements; determining the performance prediction model when the verification results meet the reasonableness requirements as the target performance prediction model; the target performance prediction model is used to predict the water flux and salt rejection rate of the thin-film nanocomposite membrane based on the inorganic nanomaterial composition, structural parameters, membrane properties, and operating conditions.

[0010] The beneficial effects of this invention are as follows:

[0011] (1) The present invention divides the sample data of thin film nanocomposite membrane into discrete attribute features and continuous action features and processes them separately to generate coupling features between the two. This allows the input features to reflect not only the compositional attributes of inorganic nanomaterials such as chemical elements, functional group types and material morphology, but also the influence of continuous factors such as the amount of inorganic nanomaterials added, membrane structure parameters and operating conditions on membrane performance. Compared with the method of simply merging various parameters as model input, the present invention can better express the correlation between material composition, membrane structure and operating conditions, providing a more complete feature basis for predicting water flux and salt rejection rate, thereby improving the pertinence of the performance prediction results of thin film nanocomposite membranes.

[0012] (2) Based on the contribution information of each input feature to water flux and salt rejection rate, the present invention performs dual-objective collaborative screening, and determines the coupled feature subset by combining the interaction contribution degree of coupled features, the dual-objective contribution direction and feature redundancy. This can avoid the accidental deletion of features that are valuable for the collaborative optimization of water flux and salt rejection rate when screening only according to the importance of a single objective feature. The present invention can retain key structure-property information while taking into account the two performance objectives of water flux and salt rejection rate, making the selected feature subset more suitable for the comprehensive performance prediction and optimization design of thin film nanocomposite membranes.

[0013] (3) Based on the coupling feature subset, the present invention establishes water flux prediction branch and salt cutoff rate prediction branch respectively, and determines the optimized performance prediction model according to the comprehensive loss of the two objectives. This makes the model training process no longer rely on a single performance index, but considers the prediction errors of water flux and salt cutoff rate at the same time. By normalizing and weighting the errors of different dimensions, the present invention can more objectively evaluate the overall prediction ability of the model for the two types of performance indicators, thereby reducing the situation where the large range of a certain objective value dominates the direction of model optimization.

[0014] (4) The present invention also verifies the coupling features and their corresponding discrete attribute features and continuous action features in the coupling feature subset, and iteratively updates the coupling feature subset and performance prediction model when the verification results do not meet the preset rationality requirements; by comparing the performance response curves corresponding to the continuous action features under different discrete attribute values, the present invention can determine whether the retained coupling features form an effective coupling effect, avoid the model only retaining statistically relevant features but lacking stable response significance, not only can it output the prediction results of water flux and salt interception rate, but it also helps to shorten the experimental screening cycle and reduce the research and development cost. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating an embodiment of the present invention;

[0017] Figure 2 This is a comparison chart of the predicted and actual values ​​of water flux by the model according to an embodiment of the present invention;

[0018] Figure 3 This is a comparison chart of the predicted and actual values ​​of the salt interception rate by the model according to an embodiment of the present invention;

[0019] Figure 4 This is a feature importance contribution map of the output target being water flux, according to an embodiment of the present invention;

[0020] Figure 5 This is a feature importance contribution map with the salt rejection rate as the output target according to an embodiment of the present invention;

[0021] Figure 6 These are specific implementation diagrams according to embodiments of the present invention;

[0022] Figure 7This is a diagram illustrating the nonlinear influence of key continuous parameters on water flux according to an embodiment of the present invention;

[0023] Figure 8 This is a diagram illustrating the nonlinear influence of key continuous parameters on the salt rejection rate according to an embodiment of the present invention. Detailed Implementation

[0024] To further illustrate the various embodiments, the present invention provides accompanying drawings. These drawings are part of the disclosure of the present invention and are mainly used to illustrate the embodiments, and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. Those skilled in the art, in conjunction with the above content, can understand other possible implementations and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols generally represent similar components.

[0025] According to an embodiment of the present invention, a method for predicting the performance of thin film nanocomposite films based on optimized machine learning is provided.

[0026] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a method for predicting the performance of thin-film nanocomposite membranes based on optimized machine learning is provided. This method includes: S1, dividing the sample data of the thin-film nanocomposite membrane into discrete attribute features and continuous action features, and constructing an input feature matrix; S2, determining the contribution information of each input feature to water flux and salt rejection rate based on the input feature matrix, and using the contribution information to perform dual-objective collaborative screening of the input features to obtain a coupled feature subset; S3, constructing and training a performance prediction model based on the coupled feature subset, and then predicting the performance based on water flux and salt rejection rate. S4. Based on the optimized performance prediction model, the discrete attribute features, continuous action features, and coupling features between the two in the coupled feature subset are verified. If the verification results do not meet the reasonableness requirements, the coupled feature subset is iteratively updated and the performance prediction model is retrained until the verification results meet the reasonableness requirements. S5. The performance prediction model when the verification results meet the reasonableness requirements is determined as the target performance prediction model. The target performance prediction model is used to predict the water flux and salt rejection rate of the thin film nanocomposite membrane based on the composition, structural parameters, membrane characteristics, and operating conditions of the inorganic nanomaterial.

[0027] It should be noted that, as Figure 6As shown, this invention discloses a method for predicting the performance of thin-film nanocomposite membranes based on optimized machine learning. This method addresses the problems of long experimental screening cycles, difficulty in expressing feature coupling relationships, and insufficient basis for dual-objective prediction and optimization in traditional TFN membrane performance research. It constructs a multi-dimensional feature system that integrates the composition, structure, and operating conditions of inorganic nanomaterials. Based on the composition, structural characteristics, and operating conditions of inorganic nanomaterials, it can predict the water flux and salt rejection rate of TFN membranes, thereby guiding the optimized design and performance improvement of TFN membrane materials. This method includes: acquiring sample data of thin-film nanocomposite membranes; performing feature analysis on the sample data based on membrane material composition, nanofiller structural characteristics, and operating conditions; dividing the input variables into discrete attribute features characterizing material category differences and continuous interaction features characterizing continuous change patterns; constructing a multi-dimensional feature matrix; calculating the contribution of each feature to different performance indicators based on dual performance targets of water flux and salt rejection rate; and conducting synergistic screening based on the interaction relationships between features to obtain a coupled feature subset that can characterize the relationship between membrane material composition, structure, and performance; constructing a machine learning prediction model; training the model using the coupled feature subset; and iteratively optimizing the model based on the deviation between the prediction results and experimental results, as well as the changes in feature interaction contributions; further verifying the stability of the contributions of discrete attribute features, continuous interaction features, and their coupled features to the prediction results; updating the feature subset and model parameters until a target performance prediction model that meets the requirements of prediction accuracy and physical rationality is obtained.

[0028] Specifically, this invention distinguishes between discrete attribute features such as chemical elements, functional group types, and material morphology, and continuous interaction features such as the amount of inorganic nanomaterials added, pore size, membrane contact angle, surface roughness, polyamide layer thickness, feed concentration, and operating pressure. Coupled features are generated by combining discrete attribute features and continuous interaction features, enabling the model input to simultaneously characterize the correlation between material composition, membrane structure, and operating conditions. Based on the input feature matrix, this invention determines the contribution of each input feature to water flux and salt rejection rate. It then combines dual-objective collaborative screening, interactive contribution judgment, and redundant feature removal to obtain a subset of coupled features. This reduces the interference of low-impact and repetitive features on model training while retaining key features that contribute to the comprehensive prediction of water flux and salt rejection rate.

[0029] Specifically, this invention employs a CatBoost gradient boosting decision tree model to construct water flux prediction branches and salt rejection rate prediction branches, respectively. A comprehensive loss is formed by normalizing the water flux error and salt rejection rate error to determine the optimized performance prediction model. Based on this, the invention verifies the responses of coupling features and their corresponding discrete attribute features and continuous action features within the coupling feature subset. The differences in performance response curves corresponding to continuous action features under different discrete attribute values ​​are used to determine whether the coupling features meet the preset rationality requirements. When the verification results do not meet the preset rationality requirements, the coupling feature subset is updated and the performance prediction model is retrained until the retained coupling features have stable response significance and the comprehensive loss meets the requirements. Therefore, this invention can establish a target performance prediction model for predicting the water flux and salt rejection rate of thin-film nanocomposite membranes, and provides an actionable analytical basis for the selection of inorganic nanomaterial composition, membrane structure regulation, and optimization of operating conditions.

[0030] It should be noted that the method includes the following steps:

[0031] Step 1: Construction of Input Feature (Parameter) System: Collect data on the composition, structure, membrane properties and operating conditions of inorganic nanomaterials for thin film nanocomposite membranes from the literature, and perform one-hot encoding on discrete features and standardization on continuous features to form an input feature matrix of uniform scale, and construct an input feature system that includes discrete and continuous features.

[0032] Step 2, Input Feature Screening: The Shapley value-based model interpretability analysis (SHAP) method based on game theory is used to evaluate the global contribution of the input features, screen out the key features that affect water flux and salt rejection rate, and remove low contribution and redundant input features.

[0033] Step 3: Preliminary construction of performance prediction model: Based on the selected key features, a CatBoost model based on gradient boosting decision tree is constructed to predict the water flux and salt rejection rate of thin film nanocomposite membrane.

[0034] Step 4: Performance Prediction Model Training and Optimization: Using the prediction model, the membrane performance under different combinations of inorganic nanomaterial composition, structural parameters and operating conditions is predicted, and the model performance is optimized through iterative training and early stopping mechanism.

[0035] Step 5: Verification of the rationality of the structure-activity relationship of the performance prediction model: Based on the completed prediction model, interpretability analysis method is introduced to verify the rationality of the structure-activity relationship of the performance prediction model from the perspectives of global feature importance, local feature effect, univariate nonlinear relationship and bivariate interaction.

[0036] Step Six: Performance Prediction Model Establishment and Application: Based on the above training, optimization, and rationality verification steps, a further improved performance prediction model is formed. By inputting the characteristics of multivariate inorganic nanomaterials, the performance prediction of TFN membrane water flux and salt rejection rate is realized.

[0037] In one embodiment, the thin-film nanocomposite membrane sample data includes inorganic nanomaterial composition data, inorganic nanomaterial structure data, membrane property data, operating condition data, and corresponding water flux and salt rejection rate; the input feature matrix includes a first feature matrix and a second feature matrix; constructing the input feature matrix includes: performing unique thermal encoding on the chemical elements, functional group types, and material morphology in the inorganic nanomaterial composition data to obtain discrete attribute features in binary vector form; standardizing the inorganic nanomaterial addition amount, feed concentration, and operating pressure in the inorganic nanomaterial structure data, membrane property data, and operating condition data to obtain continuous action features with uniform dimensions; horizontally concatenating the discrete attribute features and continuous action features in the same sample column-wise to obtain a first feature matrix with a preset feature dimension; generating coupling features based on the combination relationship between discrete attribute features and continuous action features, and concatenating the coupling features with the first feature matrix column-wise to obtain a second feature matrix; wherein, the coupling features are obtained by multiplying the binary value of each discrete attribute feature with the standardized value of each continuous action feature.

[0038] It should also be noted that in this invention, the presence or absence of a material property and the magnitude of a continuous effect factor are combined into the same feature through coupling features. For example, when a sample contains Ti, the binary value corresponding to Ti is 1. This binary value is multiplied by the standardized value of the amount of inorganic nanomaterial added to obtain a coupling feature representing the effect of the amount added when Ti is present. When the sample does not contain Ti, the binary value corresponding to Ti is 0. The resulting coupling feature is 0, indicating that the continuous effect factor does not form a corresponding combined effect when the material property is absent. Thus, coupling features can express the conditional correlation between discrete property features and continuous effect features, enabling the model to not only identify the influence of a single element, functional group, or material morphology itself, but also its influence on water flux and salt rejection rate when combined with continuous factors such as addition amount, pore size, membrane contact angle, and operating pressure. This provides a feature basis for subsequent dual-objective collaborative screening and coupling feature verification.

[0039] Specifically, in step one, discrete features include chemical elements, functional group types, and material morphology. Chemical elements include C, H, O, N, S, Ti, F, Si, Na, Al, and Ca. Functional group types include -NH, -OH, -COOH, -SO3H, COC, C=O, -F, Si-OH, Si-O-Si, CF, Ph, and Al-Si-O. Material morphology includes nanoparticles, nanosheets, layered structures, and nanotubes. Continuous features include the amount of inorganic nanomaterials added, the pore size of inorganic nanomaterials, the membrane contact angle, surface roughness, polyamide layer thickness, feed concentration, and operating pressure. For chemical elements and functional group types, if a sample contains multiple chemical elements or multiple functional groups, the corresponding category is recorded as 1, and categories not included are recorded as 0. For material morphology, the category corresponding to the main material morphology of the sample is recorded as 1, and other material morphology categories are recorded as 0.

[0040] Specifically, in step one, the feature system is processed using one-hot encoding to form a multi-dimensional feature space. The dimension of the feature space is determined by the encoding dimension of discrete attribute features and the number of continuous interaction features. With the inclusion of chemical elements, functional group types, material morphology, and continuous interaction features, the first feature matrix comprises 27 dimensions of discrete attribute features and 7 dimensions of continuous interaction features, totaling 34 dimensions. Continuous interaction features are standardized. During standardization, the mean and standard deviation of each continuous interaction feature are first calculated based on the training samples. Then, the original value of the continuous interaction feature in each sample is subtracted from the corresponding mean and divided by the corresponding standard deviation, transforming continuous interaction features with different dimensions and numerical ranges to a comparable scale. For subsequent test samples or samples to be predicted, the mean and standard deviation already determined in the training samples are used for the same processing to avoid inconsistencies in standards between different datasets.

[0041] In one embodiment, based on the input feature matrix, the contribution information of each input feature to water flux and salinity cutoff is determined, including: constructing a first contribution evaluation model with the second feature matrix as input and water flux as the output target; constructing a second contribution evaluation model with the second feature matrix as input and salinity cutoff as the output target; both the first and second contribution evaluation models adopt the CatBoost gradient boosting decision tree architecture and are trained until convergence using the root mean square error as the loss function; for the first contribution evaluation model, each sample in the test set is traversed, and the Shapley value of each input feature for each sample is calculated; the Shapley value is obtained by enumerating the feature. The marginal contribution of the model output before and after adding different feature subsets is obtained by weighted averaging of all possible subsets; the absolute values ​​of the Shapley values ​​of the same input feature in all samples are averaged to obtain the global contribution of the input feature to water flux; for the second contribution evaluation model, each sample in the test set is traversed, and the Shapley value of the corresponding salinity cutoff output target is calculated for each input feature of each sample, and the absolute values ​​of the absolute values ​​of the same input feature in all samples are averaged to obtain the global contribution of the input feature to the salinity cutoff rate; the global contribution of each input feature to water flux and the global contribution to salinity cutoff rate are respectively used as the contribution information of the input feature.

[0042] In one embodiment, contribution information is used to perform dual-objective collaborative screening of input features to obtain a subset of coupled features, including: weighting and summing the global contribution of each input feature to water flux and its global contribution to salinity interception rate according to preset weights to obtain the comprehensive feature contribution of each input feature; the sum of the preset weights is one; calculating the Shapley value of each coupled feature in the first contribution evaluation model and the second contribution evaluation model respectively, and taking the average absolute value of the Shapley value under the water flux target as the first interaction contribution, and taking the average absolute value of the Shapley value under the salinity interception rate target as the second interaction contribution; based on the comprehensive feature contribution, The first and second interaction contribution values ​​are used to determine the comprehensive collaborative contribution of each input feature. All input features are sorted from high to low according to their comprehensive collaborative contribution. Input features with a comprehensive collaborative contribution value lower than a preset contribution threshold are marked as candidate features. A dual-objective collaborative retention judgment is performed on the candidate features, and candidate features that do not meet the dual-objective collaborative retention conditions are removed. The correlation coefficient between features is calculated pairwise. When the absolute value of the correlation coefficient between two features is greater than a preset redundancy threshold, the feature with the lower comprehensive collaborative contribution value is marked as a redundant feature and removed. The discrete attribute features, continuous action features, and coupled features retained after removal are combined into a coupled feature subset.

[0043] In one embodiment, performing a dual-objective collaborative retention judgment on candidate features includes: extracting the Shapley sign distribution of each candidate feature for water flux and the Shapley sign distribution for salinity cutoff; when the contribution directions of the same candidate feature to the Shapley values ​​of water flux and salinity cutoff have opposite trends, and the proportion of samples with opposite trends is greater than a preset conflict proportion threshold, marking the candidate feature as a dual-objective conflict feature; calculating the contribution ranking of each dual-objective conflict feature under the water flux objective and the contribution ranking under the salinity cutoff objective, and determining the dual-objective collaborative ranking based on the two contribution rankings; when the dual-objective collaborative ranking is better than a preset retention ranking threshold, retaining the dual-objective conflict feature to the coupled feature subset; if the candidate feature is a coupled feature, and at least one of its first interaction contribution and second interaction contribution is greater than a preset interaction threshold, then retaining the coupled feature to the coupled feature subset.

[0044] It should be noted that in this invention, the first contribution evaluation model is a CatBoost regression model with each input feature in the second feature matrix as independent variables and water flux as the single output target; the second contribution evaluation model is a CatBoost regression model with each input feature in the second feature matrix as independent variables and salt rejection rate as the single output target. The input feature columns of the first and second contribution evaluation models are the same, both including discrete attribute features, continuous action features, and coupling features obtained by multiplying the two. However, their output targets are different, thus enabling them to learn the influence relationship between each input feature and water flux and salt rejection rate respectively. Specifically, in the construction, each row in the second feature matrix is ​​treated as a thin-film nanocomposite membrane sample, and the corresponding feature columns are used as model inputs. The label of the first contribution evaluation model is the water flux corresponding to the sample, and the label of the second contribution evaluation model is the salt rejection rate corresponding to the sample. Each CatBoost regression model consists of multiple sequentially generated decision trees. Each subsequent decision tree is used to correct the error between the combined prediction results of the preceding decision trees and the true label. The final model output is the sum of the initial predicted value and the output values ​​of each decision tree. By setting the number of decision tree iterations, learning rate, tree depth, regularization coefficient, random seed, and number of early stopping rounds, the model learns the nonlinear mapping relationship between features and target performance on the training set, and monitors the generalization error on the validation set.

[0045] Specifically, in this invention, when training to convergence using root mean square error (RMSE) as the loss function, the initial predicted value of the model is first determined based on the overall level of the target variable in the training set. Then, in each iteration, the error between the current predicted value and the true value is calculated, and the newly generated decision tree fits the descent direction corresponding to this error. After each iteration, the RMSE is calculated on the validation set to determine whether the newly added decision tree continues to improve the model's prediction performance. When the RMSE on the validation set does not continue to decrease within a preset number of early stopping rounds, or reaches a preset maximum number of iterations, training stops, and the model parameters corresponding to the lowest RMSE on the validation set are selected as the contribution evaluation model parameters after convergence.

[0046] Specifically, in calculating the dual-objective collaborative ranking, this invention first sorts the input features from high to low based on their global contribution to water flux, obtaining a contribution ranking for the water flux objective; then, it sorts the input features from high to low based on their global contribution to salinity interception, obtaining a contribution ranking for the salinity interception objective, where a higher contribution corresponds to a smaller ranking value. For candidate features marked as dual-objective conflict features, their contribution rankings under the water flux objective and salinity interception objective are averaged or weighted according to preset weights to obtain the dual-objective collaborative ranking. When water flux and salinity interception are equally important, the two rankings use the same weight; when a certain objective has a higher priority in practical applications, the weight of that objective's ranking is increased. The smaller the value of the dual-objective collaborative ranking, the higher the comprehensive importance of the candidate feature in the two objectives; when the value of the dual-objective collaborative ranking is less than or equal to a preset retention ranking threshold, it is considered that although the feature has differences in the direction of contribution to the two objectives, it still has retention value for the collaborative optimization of water flux and salinity interception.

[0047] Specifically, step two includes: 1) Based on the input feature matrix constructed in step one, establish a first contribution evaluation model with water flux as the output target and a second contribution evaluation model with salinity interception rate as the output target; 2) Calculate the Shapley value of each input feature to the model output result using the SHAP method to obtain the contribution of each feature to the prediction results of water flux and salinity interception rate; 3) Perform statistical analysis on the absolute values ​​of the Shapley values ​​of all input features to obtain the global contribution of each input feature under the targets of water flux and salinity interception rate; 4) Determine the comprehensive synergistic contribution based on the global contribution, the first interaction contribution and the second interaction contribution corresponding to the coupled features, and combine the preset contribution threshold, the dual-target synergistic retention conditions and the correlation analysis results between features to remove low contribution features and redundant features, and retain discrete attribute features, continuous action features and coupled features that have an effect on the comprehensive prediction of water flux and salinity interception rate; 5) Reconstruct the coupled feature subset based on the retained features after screening for subsequent performance prediction model training and optimization.

[0048] Specifically, the interaction consistency coefficient of each input feature is determined based on the first interaction contribution and the second interaction contribution; the interaction consistency coefficient increases with the combined increase of the first interaction contribution and the second interaction contribution, and decreases with the increase of the difference between the two; the interaction consistency coefficient is used to enhance the contribution of the comprehensive feature, thereby obtaining the comprehensive collaborative contribution of each input feature. In this embodiment, the calculation expression for the comprehensive collaborative contribution is:

[0049] ;

[0050] In the formula, , These represent the global contributions of the input features to water flux and salt rejection rate, respectively. , These represent the contribution of the first interaction and the contribution of the second interaction, respectively. and These are preset weight parameters.

[0051] In one embodiment, the performance prediction model includes a coupled feature input layer, a water flux prediction branch, and a salinity cutoff prediction branch. The coupled feature input layer is a feature column mapping layer, used to extract corresponding feature columns based on the column positions of each feature in the second feature matrix within the coupled feature subset, and form feature vectors input to the water flux prediction branch and the salinity cutoff prediction branch. The water flux prediction branch is used to predict water flux based on the features output by the coupled feature input layer. The salinity cutoff prediction branch is used to predict salinity cutoff based on the features output by the shared coupled feature input layer. Both the water flux prediction branch and the salinity cutoff prediction branch adopt a CatBoost gradient boosting decision tree architecture.

[0052] It should be added that the coupled feature input layer in this invention is an input processing module used to perform feature column filtering and feature column mapping. Specifically, during construction, the column name or column number of each retained feature in the coupled feature subset is first recorded in the second feature matrix. Then, the corresponding feature column is extracted from the second feature matrix according to the column name or column number, and concatenated into a unified input feature vector in a preset order. Both the water flux prediction branch and the salinity cutoff prediction branch use this input feature vector as input, but they correspond to different output labels. The water flux prediction branch uses the measured water flux of the sample as the training label, and the salinity cutoff prediction branch uses the measured salinity cutoff of the sample as the training label. Thus, the two prediction branches share the same coupled feature input, ensuring that both objectives are predicted based on the same features, while simultaneously learning the nonlinear mapping relationship between water flux and salinity cutoff through their respective independent CatBoost regression models.

[0053] Specifically, in step three, for the chemical elements, functional group types, or material morphology fields recorded as string lists in the original data, they are first parsed into multiple independent category identifiers, and then one-hot encoded according to a preset category set. If a sample lacks a corresponding category or the corresponding attribute is not detected, the corresponding encoding position is filled with 0. The discrete attribute features obtained through one-hot encoding, the standardized continuous interaction features, and the coupled features obtained by multiplying the two together form a complete feature set. In this way, the variable-length composition information in the original sample can be converted into a fixed-length feature vector, which is convenient for the CatBoost model to directly read and train.

[0054] Specifically, in step three, an independent random seed is assigned to each target variable. For example, a first random seed is assigned to the water flux prediction branch, and a second random seed is assigned to the salinity cutoff prediction branch. The random seed controls the random processes during sample partitioning, model initialization, and decision tree training, ensuring consistent training results can be obtained repeatedly under the same data and parameter conditions. During sample partitioning, a test set can be first partitioned from all samples according to a preset ratio, and then a validation set and training set can be partitioned from the remaining samples. In this embodiment, 20% of the samples are first partitioned as the test set, and then 15% of the remaining training samples are partitioned as the validation set. In the water flux prediction branch and the salinity cutoff prediction branch, their respective random seeds are used to complete the partitioning and model training, thereby reducing the randomness caused by identical random processes in the two target models while maintaining reproducibility of the results.

[0055] Specifically, in step three, the hyperparameters of the CatBoost model include the number of decision tree iterations, learning rate, tree depth, leaf node L2 regularization coefficient, random perturbation strength, Bayesian bootstrap temperature, early stopping method, number of early stopping rounds, and loss function type. The number of decision tree iterations limits the maximum number of decision trees the model can generate; the learning rate controls the contribution of each decision tree to the final prediction result; the tree depth controls the nonlinear complexity that a single decision tree can express; the regularization coefficient suppresses overfitting of the model to the training samples; the number of early stopping rounds terminates training when the validation set loss no longer improves; and the loss function type limits the direction of model optimization. In this embodiment, root mean square error can be used as the loss function.

[0056] Specifically, in step four, iterative training involves sequentially adding decision trees and continuously refining the model based on the prediction residuals from the previous round. For both the water flux prediction branch and the salinity cutoff branch, the error between the current predicted value and the true value is calculated in each iteration, and the newly added decision tree is fitted with the correction amount of this error in the negative gradient direction of the loss function. The outputs of multiple decision trees are gradually accumulated according to the learning rate, allowing the model to gradually approximate the complex nonlinear mapping relationship between the input features and water flux and salinity cutoff, thereby forming the corresponding CatBoost ensemble prediction model.

[0057] Specifically, in step four, the early stopping mechanism is used to prevent the model from continuing to reduce errors on the training set but experiencing a decline in generalization ability on the validation set. During training, after each iteration, the losses of the water flux prediction branch and the salinity cutoff prediction branch on the validation set are calculated separately, and the normalized comprehensive loss is further calculated. If the comprehensive loss does not decrease within a preset number of early stopping rounds, or if it rebounds compared to the historical best comprehensive loss, it is determined that continuing to add decision trees cannot effectively improve the model's generalization performance, and training is terminated. At the same time, the model parameters of the water flux prediction branch and the salinity cutoff prediction branch are rolled back to the historical iteration with the lowest comprehensive loss.

[0058] Specifically, in step four, the coefficient of determination and the normalized root mean square error (RMSE) are used as the core evaluation indicators for the model's prediction accuracy. The coefficient of determination is used to evaluate the degree of fit between the predicted values ​​and the trends of the actual values, while the RMSE is used to evaluate the relative prediction error under different dimensions of the target variable. The mean square error, root mean square error, and mean absolute error can be used as auxiliary evaluation indicators to record the model error level, but they are not necessary constraints for determining the optimized performance prediction model.

[0059] Specifically, a performance loss normalization mechanism is adopted to unify the scaling of prediction errors in the water flux prediction branch and the salinity cutoff prediction branch, thereby improving the comparability between the two output target loss values. The normalized root mean square error (NRMSE) is used as the model performance evaluation index. By normalizing the model prediction error relative to the range of target variable values, the impact of differences in the dimensions of different target variables on the comprehensive loss calculation results is reduced. The calculation method is as follows: ;in, For the true value, For predicted values, For the sample size, and These are the maximum and minimum values ​​of the corresponding target variable used for normalization. The maximum and minimum values ​​can be determined based on the range of values ​​of the target variable in the validation or test set, but should remain consistent in the same model evaluation.

[0060] It should be noted that the multi-objective joint importance analysis process in this invention includes: ① First, the feature importance of the water flux prediction target and the salt interception rate prediction target are calculated separately.

[0061] Specifically, for the first Each input feature is used to obtain its global contribution to the water flux prediction result using the feature interpretation algorithm in the prediction model. and the global contribution to the salt cutoff rate prediction results The target contribution of all input features is normalized, and the overall contribution level of water flux target is calculated separately. and the overall contribution level of the salt interception rate target The calculation method is as follows:

[0062] ;

[0063] ;

[0064] in, The number of input features; the target contribution level is used to represent the sensitivity of different prediction targets in the current dataset to changes in input features.

[0065] ② Calculate the prediction confidence correction factor based on the prediction errors of the water flux prediction model and the salinity interception rate prediction model on the validation set. Calculate the error response values ​​of the two prediction branches based on the root mean square error (RMSE) of the validation set: ; ;in, To prevent correction constants where the denominator is zero.

[0066] Specifically, the target contribution level and the predicted confidence response value are jointly mapped to obtain the joint weight of the two objectives: ; ;

[0067] ③ Normalize the joint weights to obtain the target weight coefficient α for water flux and the target weight coefficient β for salinity interception: ; Where α+β=1.

[0068] Specifically, to avoid bias in the overall evaluation due to excessively high local contributions from a single prediction target, this invention sets constraint adjustment rules for α and β. When any weight coefficient exceeds a preset threshold range, a secondary correction is performed based on the contribution level of another target, so that: ; This ensures a balance of feature contributions in the dual-objective prediction task.

[0069] ④ Calculate the overall contribution of the input features based on the corrected target weights: ;in, Indicates the first The joint contribution of each input feature to the task of predicting the overall performance of thin film nanocomposite films.

[0070] In one embodiment, the training process of the performance prediction model includes: assigning independent random seeds to the water flux prediction branch and the salinity cutoff prediction branch respectively; dividing the data corresponding to the coupled feature subset into a training set and a test set according to a preset ratio, and then dividing the training set into a validation set according to a preset ratio; setting the number of decision tree iterations, learning rate, tree depth, regularization coefficient, and number of early stopping waiting rounds for the water flux prediction branch and the salinity cutoff prediction branch; training the two prediction branches based on the training set, and using the validation set to monitor the loss changes of the two prediction branches to obtain the model parameters of the water flux prediction branch and the salinity cutoff prediction branch in each iteration round.

[0071] It should also be noted that, in this invention, the independent random seed refers to the initial random number value set for the water flux prediction branch and the salinity cutoff prediction branch, respectively, used to control random processes such as random sample partitioning, model parameter initialization, feature sampling, or sample sampling. In this embodiment, a first random seed is set for the water flux prediction branch, and a second random seed is set for the salinity cutoff prediction branch; during the training of each prediction branch, the corresponding random seed is written into the data partitioning program and the CatBoost model training parameters. The purpose of this is twofold: firstly, it ensures that the same branch yields reproducible training results during repeated experiments; secondly, it allows the two output targets to be trained independently under their respective random conditions, avoiding unnecessary coupling between the two branches due to identical random processes.

[0072] Specifically, the Sobolev distance is used for the coupling response distance verification process. When calculating the coupling response distance, not only are the numerical differences in the predicted values ​​of water flux and salinity interception at each interpolation point under different discrete attribute values ​​compared, but also the differences in the increments of predicted values ​​between adjacent interpolation points. The former reflects the difference in the numerical level of the performance response curves, while the latter reflects the difference in the trend of change of the performance response curves. Therefore, even if two response curves have similar values ​​at some interpolation points, as long as their trends of change with the continuous action feature are different, they can still be identified through the coupling response distance, thus more accurately determining whether there is an effective coupling effect between discrete attribute features and continuous action features.

[0073] In one embodiment, determining the optimized performance prediction model based on the combined loss of water flux and salinity cutoff rate includes: iteratively training the water flux prediction branch and the salinity cutoff rate prediction branch respectively; in each iteration, sequentially adding a decision tree, the fitting target of which is the value of the prediction residual of the previous round along the negative gradient direction of the loss function; after each iteration, calculating the root mean square error of the water flux prediction branch on the validation set and the root mean square error of the salinity cutoff rate prediction branch on the validation set respectively, and then weighted summing them after normalization to eliminate dimensional differences to obtain the combined loss; the normalization process involves dividing the prediction branches... The root mean square error of the branch is divided by the range of the corresponding target variable in the test set. If the overall loss does not decrease or increases within a preset number of early stop waiting rounds, training is terminated, and the model parameters of the water flux prediction branch and the salinity cutoff prediction branch are rolled back to the historical iteration round with the lowest overall loss. The prediction accuracy of the water flux prediction branch and the salinity cutoff prediction branch on the test set is evaluated using the coefficient of determination and the normalized root mean square error, respectively. When the prediction accuracy of both the water flux prediction branch and the salinity cutoff prediction branch meets the preset accuracy conditions, the current performance prediction model is used as the optimized performance prediction model.

[0074] Specifically, the comprehensive loss is used to simultaneously measure the prediction performance of the water flux prediction branch and the salinity cutoff prediction branch on the validation set. During calculation, the predicted values ​​of the two branches under the same validation set partitioning condition are first obtained and compared with the corresponding true values ​​to obtain the root mean square error (RMS) of water flux and the RMS of salinity cutoff. Then, the two RMS errors are divided by the range of the corresponding target variable to convert the water flux error and salinity cutoff error into dimensionless errors. Finally, the two dimensionless errors are weighted and summed according to preset weights to obtain the comprehensive loss for the current iteration. If water flux and salinity cutoff are equally important, their weights can be the same; if the actual application scenario focuses more on one target, the weight of the loss corresponding to that target can be increased. During training, the comprehensive loss and the corresponding model parameters of the two prediction branches are recorded for each iteration. When the comprehensive loss reaches its historical minimum value, the current parameters are saved; when the early stopping condition is triggered, the parameters corresponding to the historical minimum comprehensive loss are used as the optimized performance prediction model parameters. Subsequently, the determination coefficients and normalized root mean square errors of the two prediction branches are calculated on the test set. If both meet the preset accuracy conditions, the current model is determined as the optimized performance prediction model; if the preset accuracy conditions are not met, the model can be retrained after adjusting the coupled feature subset or model hyperparameters.

[0075] In one embodiment, the verification of discrete attribute features, continuous action features, and the coupling features between them in the coupling feature subset includes: using the coupling feature as the verification unit, determining the discrete attribute features and continuous action features corresponding to the coupling feature; setting several interpolation points within the range of inorganic nanomaterial addition amount, feed liquid concentration, and operating pressure according to the continuous action features; calculating the water flux prediction value and salt rejection rate prediction value output by the optimized performance prediction model at each interpolation point under different values ​​of the discrete attribute features, and obtaining the performance response curves corresponding to different discrete attribute values; obtaining the coupling response difference value based on the morphological differences of the performance response curves under different discrete attribute values; if the coupling response difference value is greater than or equal to the significance threshold, the coupling feature is determined to have passed the verification; otherwise, the coupling feature is determined to have failed the verification.

[0076] Specifically, coupling feature verification is used to determine whether the selected coupling features truly reflect the effective combined influence between discrete attribute features and continuous action features. For any coupling feature, the corresponding discrete attribute feature and continuous action feature are first determined according to its generation method. In the above embodiment, a certain coupling feature is obtained by multiplying the binary value of Ti element by the standardized value of inorganic nanomaterial addition amount. Therefore, the discrete attribute feature corresponding to this coupling feature is Ti element, and the continuous action feature is the amount of inorganic nanomaterial addition amount. Multiple interpolation points are set within the sample value range of this continuous action feature, while other features can be kept as sample mean, median, or preset representative values. Then, discrete attribute features are set to existent and non-existent states respectively, and the corresponding coupling feature value is regenerated at each interpolation point. This value is input into the optimized performance prediction model to obtain the predicted values ​​of water flux and salt rejection rate. The predicted values ​​obtained under different discrete attribute value states with changes in interpolation points are connected to form a performance response curve, and the numerical differences and trend differences of each response curve are compared to determine whether the discrete attribute feature changes the way the continuous action feature affects membrane performance.

[0077] In one embodiment, the coupling response difference value is obtained based on the morphological differences of the performance response curves under different discrete attribute values. This includes: forming a two-dimensional performance response vector sequence by arranging the predicted water flux and salt rejection rate corresponding to the same coupling feature in the order of interpolation points; performing endpoint normalization on the two-dimensional performance response vector sequences under different discrete attribute values; calculating the coupling response distance between the two-dimensional performance response vector sequences corresponding to any two discrete attribute values, wherein the coupling response distance is jointly determined by the square integral of the difference between the response vectors at the same interpolation point and the square integral of the difference between the response increments between adjacent interpolation points; and taking the maximum value among the coupling response distances as the coupling response difference value.

[0078] Specifically, in this invention, the two-dimensional performance response vector sequence refers to a sequence of vectors formed by combining the predicted values ​​of water flux and salinity cutoff at each interpolation point into a two-dimensional vector, and arranging them in ascending order of the interpolation points. Endpoint normalization can be understood as using the start and end points of the response curve as normalization references to transform the response curves under different discrete attribute values ​​to a comparable scale, thereby reducing the impact of differences in the absolute numerical ranges of water flux and salinity cutoff on the curve shape comparison. When calculating the coupled response distance, on the one hand, the difference between the two-dimensional vectors corresponding to the two discrete attribute values ​​at the same interpolation point is calculated, and the squares of the differences at all interpolation points are accumulated or integrated; on the other hand, the difference in the increments of the two-dimensional vectors between adjacent interpolation points is calculated, and the squares of the increment differences between all adjacent intervals are accumulated or integrated. The results of the above two parts can be added according to preset weights and the square root is taken to obtain the coupled response distance. In this embodiment, the coupled response distance adopts the Sobolev distance, which considers both the difference in the predicted values ​​of the response curves and the difference in the first-order trend of the response curves. If the same coupling feature has multiple discrete attribute values, then calculate the coupling response distance between any two states and take the maximum value as the coupling response difference value of the coupling feature.

[0079] In one embodiment, iteratively updating the coupling feature subset and retraining the performance prediction model includes: when the difference value of the coupling response of a coupling feature is less than a significance threshold, determining that the coupling feature has not formed an effective coupling effect, and removing the coupling feature from the coupling feature subset; when the discrete attribute feature or continuous action feature corresponding to the removed coupling feature no longer corresponds to any retained coupling feature, and the comprehensive synergistic contribution of the discrete attribute feature or continuous action feature is lower than a preset contribution threshold, removing the corresponding discrete attribute feature or continuous action feature from the coupling feature subset; reconstructing and training the performance prediction model based on the updated coupling feature subset, and recalculating the comprehensive loss of water flux and salt interception rate; repeatedly performing coupling feature verification, coupling feature subset update, and performance prediction model training until the verification result meets the preset rationality requirements; wherein, the preset rationality requirements include: the proportion of the number of verified coupling features to the current total number of coupling features reaches a preset pass rate, and the current comprehensive loss is not higher than the previous round of comprehensive loss.

[0080] Further, in step five, SHAP analysis and performance response curve analysis are used to verify the rationality of the relationship between changes in input features and the model output response. By changing the continuous action feature corresponding to the coupling feature to be verified while keeping other variables constant, the response curve of the model output as a function of the input variables is obtained. The response curve is then compared with known physicochemical laws in the field of membrane separation to verify the rationality of the model's prediction results. Specifically, if the model predicts an overall increasing trend in water flux when the operating pressure increases, the model is determined to conform to the law of pressure-driven water transport. If the model predicts a nonlinear change or a threshold change trend in salt rejection rate when the feed concentration increases, the model is determined to conform to the laws of concentration polarization and interfacial mass transfer. Furthermore, by combining the contribution direction and magnitude of SHAP features, it is verified whether the direction of the influence of key input features on membrane performance is consistent with the membrane separation mechanism.

[0081] It should be noted that in step six, the core parameters of the target performance prediction model can be determined based on the sample size and validation set loss. In this embodiment, the decision tree iteration count is set to 1000, the learning rate to 0.05, the tree depth to 6, the leaf node L2 regularization coefficient to 3, the random perturbation intensity to 1, the Bayesian bootstrap temperature to 1, the early stopping method to stop early according to the iteration round, the early stopping waiting rounds to 50, and the loss function to be root mean square error. Different random seeds are set for the water flux prediction branch and the salinity cutoff prediction branch. The above parameters can be adjusted according to the sample size, feature dimension, and validation set error. During the construction of the input feature system, a data processing program with matrix operation capabilities can be used to achieve one-hot encoding, standardization, feature concatenation, and coupled feature generation.

[0082] Specifically, after model training, linear regression is performed on the predicted and actual values ​​on the test set to obtain the slope and intercept of the regression line. Furthermore, the confidence interval and prediction interval of the predicted values ​​are calculated based on the residual standard error to quantitatively assess the consistency and reliability of the model's predictions. This process can serve as an auxiliary evaluation step for the target performance prediction model to demonstrate the consistency between the predicted and actual results. However, the optimized performance prediction model is still determined based on whether the coefficients of determination and normalized root mean square error of the water flux prediction branch and the salinity cutoff branch on the test set meet the preset accuracy conditions.

[0083] Specifically, the target performance prediction model encapsulates the performance prediction model after training optimization and rationality verification into a callable module. By inputting a multivariate combination of the chemical composition, structural parameters, membrane properties and operating conditions of inorganic nanomaterials, it first constructs the input feature vector of the sample to be predicted according to the unique thermal coding category, standardized parameters and coupling feature generation rules determined in the training stage. Then, it inputs the water flux prediction branch and the salt rejection rate prediction branch respectively, and simultaneously outputs the predicted values ​​of water flux and salt rejection rate of the thin film nanocomposite membrane.

[0084] To facilitate understanding of the above-mentioned technical solution of the present invention, the following is a detailed explanation using the thin film nanocomposite membrane performance prediction process based on publicly available experimental data as an example:

[0085] Example 1

[0086] In this embodiment, experimental data of thin film nanocomposite films reported in published literature were used as the sample source, and SiO2-TFN, TiO2-TFN, GO-TFN, N-GOQDs-TFN and Ti3C2T were selected. x - TFN and other typical membrane systems were analyzed, resulting in 194 experimental samples with complete performance records. Sample information included parameters such as inorganic nanomaterial composition, material morphology, functional group type, membrane contact angle, pore size, surface roughness, polyamide layer thickness, addition amount, and operating pressure, with water flux and salt rejection rate as output targets.

[0087] During data processing, discrete information such as chemical elements, functional groups, and material morphology is converted into coded features. Continuous parameters such as dosage, membrane contact angle, pore size, surface roughness, polyamide layer thickness, and operating conditions are standardized to form the input data for training the CatBoost model. After training, the model is used to predict and validate water flux and salt rejection rates. Figure 2 As shown, the predicted water flux values ​​are mostly distributed near the actual values, and the fitted line of the test set is in good agreement with the ideal fitted line, indicating that the model can stably reflect the variation law of water flux of TFN membrane.

[0088] like Figure 3 As shown, there is a high correlation between the predicted and actual salt cutoff rates. The vast majority of sample points are concentrated in the high salt cutoff rate range and distributed near the ideal fitting line, indicating that the model's prediction error for the salt cutoff rate is small. Combined with... Figure 2 and Figure 3 It can be seen that the established CatBoost model can not only predict performance indicators with a wide range of variation, such as water flux, but also maintain a good recognition ability for performance indicators with a relatively concentrated numerical distribution, such as salt rejection rate, thus verifying the applicability of the model for TFN membrane performance prediction.

[0089] Example 2

[0090] In this embodiment, feature contribution analysis is further performed on the model obtained in Example 1 to determine the degree of influence of each input feature on the prediction results of water flux and salt rejection rate. Specifically, after the model training is completed, the SHAP contribution value of each feature is calculated for the water flux prediction target and the salt rejection rate prediction target, respectively, and the influence direction of different material parameters on membrane performance is determined based on the magnitude and positive / negative distribution of the contribution values. This process does not change the original experimental data and is only used to identify the response pattern of the model to key factors.

[0091] like Figure 4 As shown, under the target of water flux prediction, the amount of additive, membrane contact angle, operating pressure, surface roughness, and characteristics of some chemical elements have a significant impact on the prediction results. Among them, the contributions of the amount of additive and the membrane contact angle are more concentrated, indicating that they have a strong regulatory effect on flux changes. Figure 4 The different colored sample points reflect the direction of the influence of feature values ​​on the model output. For example, some high-value samples will increase the predicted value of water flux, while some parameters will decrease the prediction result when the value is high. This can provide a basis for subsequent screening of key features.

[0092] like Figure 5 As shown, under the salt rejection rate prediction target, operating pressure, membrane contact angle, surface roughness, additive dosage, and pore size all exhibit significant contributions. However, the influence direction of some continuous parameters on the salt rejection rate is not entirely consistent with their influence direction on water flux. This result indicates that in TFN membrane performance prediction, water flux and salt rejection rate are not simultaneously determined by a single factor, but rather by the combined effects of material composition, membrane structure, and operating conditions. Figure 4 and Figure 5 By comparing the features with those of low contribution and poor stability, we can filter out features that contribute less and are less stable, while retaining key features that are valuable for explaining both objectives.

[0093] Example 3

[0094] In this embodiment, after completing the SHAP contribution analysis, the input features are filtered and the model is retrained to verify the improvement effect of feature filtering on prediction performance. During filtering, features that contribute significantly to water flux or salt rejection are prioritized, while considering the coordination between the two output objectives to avoid over-retaining features that are effective for only one objective but significantly interfere with the other. After filtering, the model input is adjusted from the initial feature set to an optimized feature set including key material parameters, membrane structure parameters, and operational parameters.

[0095] After retraining, the model's prediction accuracy improved. In the initial CatBoost model, the coefficient of determination (COP) for water flux prediction was 0.793, and for salinity cutoff prediction it was 0.858. After SHAP filtering, the COP for water flux prediction increased to 0.813, and for salinity cutoff prediction it increased to 0.911. This result indicates that removing low-contribution or redundant features allows the model to more effectively learn the changes directly related to membrane performance, reducing the interference of invalid features on the training process and improving the prediction stability for unknown samples.

[0096] Building upon this foundation, differentiated weights were introduced based on the varying importance of key features, allowing the model to pay closer attention to parameters significantly impacting membrane performance during training. After weight optimization, the coefficient of determination for water flux prediction increased to 0.841, and the coefficient of determination for salt rejection rate prediction increased to 0.951. This demonstrates that SHAP contribution analysis can not only interpret model prediction results but also serve as a basis for feature selection and weight adjustment, transforming the model from a simple data fitter into a performance prediction tool that incorporates material properties.

[0097] Example 4

[0098] In this embodiment, to further illustrate the nonlinear relationship between key continuous parameters and membrane performance, the Partial Dependence Plot (PDP) analysis method was used to verify the response behavior of the screened model. This analysis focuses on examining the effects of continuous parameters such as additive amount, pore size, surface roughness, membrane contact angle, and polyamide layer thickness on water flux and salt rejection. Unlike ranking based on single feature contributions, PDP curves can show the changing trends of parameters across different value ranges in the predicted results, making them more suitable for observing the response behavior of membrane performance to changes in process parameters.

[0099] like Figure 7 As shown, key continuous parameters exhibit different forms of nonlinear influence on water flux. Specifically, as the addition amount increases, the water flux generally shows an upward trend, but the increase slows down in the higher addition range; the effect of pore size on water flux exhibits a plateau range; and as surface roughness, membrane contact angle, and polyamide layer thickness increase, the water flux generally decreases or decreases in a stepwise manner. These results correspond to the changes in pore size, surface wettability, and polyamide layer resistance during the TFN membrane mass transfer process, indicating that the laws learned by the model have a certain degree of physical rationality.

[0100] like Figure 8As shown, the effects of key continuous parameters on the salt rejection rate are relatively gradual, but significant range-specific differences can still be observed. After the addition amount increases in some ranges, the salt rejection rate decreases slightly; after the pore size increases to a certain range, the salt rejection rate shows a decreasing trend; the effects of surface roughness, membrane contact angle, and polyamide layer thickness on the salt rejection rate exhibit local fluctuations and slow changes. Combined with... Figure 7 and Figure 8 It is known that the same continuous parameter may have different effects on water flux and salt rejection rate. Therefore, when optimizing the TFN membrane structure, it is necessary to take into account both flux improvement and salt rejection rate.

[0101] according to Figure 7 and Figure 8 The revealed nonlinear response relationship was used to adjust the weights of parameters such as membrane contact angle, polyamide layer thickness, and pore size, and the model was retrained. The final model achieved a coefficient of determination of 0.854 for water flux prediction and 0.963 for salt rejection. These results demonstrate that PDP analysis can provide a basis for the weight allocation of key parameters, enabling the model to better reflect the coupling effects between TFN membrane material structure, interfacial properties, and operating conditions while maintaining prediction accuracy.

[0102] Table 1 Comparison of TFN membrane performance prediction results at different optimization stages

[0103]

[0104] As shown in Table 1, with the gradual introduction of SHAP feature selection, differential weight optimization, and PDP-guided secondary optimization, the model's prediction accuracy for water flux and salt rejection rate has been continuously improved. Figure 2 and Figure 3 The consistency between the model predictions and experimental results was verified. Figure 4 and Figure 5 This illustrates the differences in the contribution of key features to different output targets. Figure 7 and Figure 8 This further reveals the nonlinear control law of continuous parameters on membrane performance. Therefore, the method of this invention can realistically reflect the influence of material composition, membrane structure, and operating conditions on TFN membrane performance, and can be used to guide the performance prediction and parameter optimization of thin-film nanocomposite membranes.

[0105] According to another embodiment of the present invention, a thin-film nanocomposite membrane performance prediction system based on optimized machine learning is also provided. This system includes: a feature selection module for dividing thin-film nanocomposite membrane sample data into discrete attribute features and continuous action features, and constructing an input feature matrix; a collaborative selection module for determining the contribution information of each input feature to water flux and salt rejection rate based on the input feature matrix, and using the contribution information to perform dual-objective collaborative selection of the input features to obtain a coupled feature subset; and a model optimization module for constructing and training a performance prediction model based on the coupled feature subset, and optimizing the model based on water flux and salt rejection rate. The comprehensive loss of salt rejection rate determines the optimized performance prediction model; the iterative verification module is used to verify the discrete attribute features, continuous action features, and coupling features between them in the coupled feature subset based on the optimized performance prediction model. If the verification results do not meet the reasonableness requirements, the coupled feature subset is iteratively updated and the performance prediction model is retrained until the verification results meet the reasonableness requirements; the target determination module is used to determine the performance prediction model when the verification results meet the reasonableness requirements as the target performance prediction model; the target performance prediction model is used to predict the water flux and salt rejection rate of the thin film nanocomposite membrane based on the inorganic nanomaterial composition, structural parameters, membrane properties, and operating conditions.

[0106] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the performance of thin-film nanocomposite films based on optimized machine learning, characterized in that, include: The sample data of thin film nanocomposite membranes are divided into discrete property features and continuous interaction features, and an input feature matrix is ​​constructed. Based on the input feature matrix, the contribution information of each input feature to water flux and salt rejection rate is determined, and the contribution information is used to perform dual-objective collaborative screening of the input features to obtain a coupled feature subset; A performance prediction model is constructed and trained based on a coupled feature subset, and the optimized performance prediction model is determined based on the combined loss of water flux and salt rejection rate. Based on the optimized performance prediction model, the discrete attribute features, continuous action features, and coupling features between the two in the coupled feature subset are verified. If the verification results do not meet the reasonableness requirements, the coupled feature subset is iteratively updated and the performance prediction model is retrained until the verification results meet the reasonableness requirements. The performance prediction model that meets the rationality requirements of the verification results is determined as the target performance prediction model; the target performance prediction model is used to predict the water flux and salt rejection rate of the thin film nanocomposite membrane based on the inorganic nanomaterial composition, structural parameters, membrane characteristics and operating conditions.

2. The method for predicting the performance of thin-film nanocomposite films based on optimized machine learning according to claim 1, characterized in that, The sample data of the thin film nanocomposite membrane includes inorganic nanomaterial composition data, inorganic nanomaterial structure data, membrane characteristic data, operating condition data, and corresponding water flux and salt rejection rate; The input feature matrix includes a first feature matrix and a second feature matrix; The construction of the input feature matrix includes: The chemical elements, functional group types, and material morphology in the composition data of inorganic nanomaterials are subjected to unique thermal encoding to obtain discrete attribute features in binary vector form; The inorganic nanomaterial addition amount, feed concentration and operating pressure in the inorganic nanomaterial structure data, membrane characteristic data and operating condition data are standardized to obtain continuous action characteristics with uniform dimensions. The discrete attribute features and continuous action features in the same sample are horizontally concatenated column by column to obtain a first feature matrix with a preset feature dimension. Based on the combination relationship between discrete attribute features and continuous action features, coupled features are generated, and the coupled features are concatenated with the first feature matrix column by column to obtain the second feature matrix. The coupling feature is obtained by multiplying the binary value of each discrete attribute feature with the standardized value of each continuous action feature.

3. The method for predicting the performance of thin-film nanocomposite films based on optimized machine learning according to claim 2, characterized in that, The step of determining the contribution information of each input feature to water flux and salt rejection rate based on the input feature matrix includes: A first contribution evaluation model is constructed using the second feature matrix as input and water flux as output target; a second contribution evaluation model is constructed using the second feature matrix as input and salt cutoff rate as output target; both the first and second contribution evaluation models are trained until convergence using root mean square error as the loss function. For the first contribution evaluation model, each sample in the test set is traversed, and the Shapley value of each input feature of each sample is calculated. The Shapley value is obtained by enumerating the marginal contribution of the model output before and after adding the feature to different feature subsets, and averaging it according to the weighted average of all possible subsets. The global contribution of the input feature to water flux is obtained by averaging the absolute values ​​of the Shapley values ​​of the same input feature across all samples. For the second contribution evaluation model, each sample in the test set is traversed, and the Shapley value of the corresponding salt cutoff rate output target is calculated for each input feature of each sample. The absolute values ​​of the Shapley values ​​of the same input feature in all samples are averaged to obtain the global contribution of the input feature to the salt cutoff rate. The global contribution of each input feature to water flux and the global contribution to salt interception rate are respectively used as the contribution information of that input feature.

4. The method for predicting the performance of thin-film nanocomposite films based on optimized machine learning according to claim 1, characterized in that, The method of using contribution information to perform dual-objective collaborative screening of input features to obtain a coupled feature subset includes: The global contribution of each input feature to water flux and its global contribution to salt interception rate are weighted and summed according to preset weights to obtain the comprehensive feature contribution of each input feature; the sum of the preset weights is one. The Shapley value of each coupling feature is calculated in the first contribution evaluation model and the second contribution evaluation model respectively. The average absolute value of the Shapley value under the water flux target is taken as the first interaction contribution, and the average absolute value of the Shapley value under the salt interception rate target is taken as the second interaction contribution. Based on the comprehensive feature contribution, the first interaction contribution, and the second interaction contribution, the comprehensive collaborative contribution of each input feature is determined. All input features are sorted from high to low according to their overall collaborative contribution; input features with an overall collaborative contribution below a preset contribution threshold are marked as candidate features. Perform a dual-objective collaborative retention judgment on candidate features and remove candidate features that do not meet the dual-objective collaborative retention conditions; The correlation coefficient between features is calculated pairwise. When the absolute value of the correlation coefficient between two features is greater than the preset redundancy threshold, the feature with the lower overall synergistic contribution is marked as a redundant feature and removed. The discrete attribute features, continuous action features, and coupling features retained after removal are combined to form the coupling feature subset.

5. The method for predicting the performance of thin-film nanocomposite films based on optimized machine learning according to claim 4, characterized in that, The process of performing a dual-objective collaborative preservation judgment on candidate features includes: Based on each candidate feature, extract its Shapley sign distribution with respect to water flux and its Shapley sign distribution with respect to salt cutoff rate; When the same candidate feature contributes to the Shapley value of water flux and salt interception in opposite directions, and the proportion of samples with opposite trends is greater than the preset conflict proportion threshold, the candidate feature is marked as a dual-objective conflict feature. Based on the conflict characteristics of each dual objective, its contribution ranking under the water flux objective and its contribution ranking under the salt interception objective are calculated respectively, and the dual objective synergy ranking is determined based on the two contribution rankings. When the dual-objective collaborative ranking is better than the preset ranking retention threshold, the dual-objective conflict feature is retained in the coupling feature subset. If a candidate feature is a coupled feature, and at least one of its first interaction contribution and second interaction contribution is greater than a preset interaction threshold, then the coupled feature is retained in the coupled feature subset.

6. The method for predicting the performance of thin-film nanocomposite films based on optimized machine learning according to claim 1, characterized in that, The performance prediction model includes a coupled feature input layer, a water flux prediction branch, and a salt rejection rate prediction branch. The coupled feature input layer is a feature column mapping layer, which is used to extract the corresponding feature columns according to the column positions of each feature in the second feature matrix and form feature vectors that are input to the water flux prediction branch and the salt rejection rate prediction branch. The water flux prediction branch is used to predict water flux based on the features output by the coupled feature input layer. The salt cutoff prediction branch is used to predict the salt cutoff rate based on the features output by the shared coupling feature input layer. The water flux prediction branch and the salt rejection rate prediction branch both adopt the CatBoost gradient boosting decision tree architecture. The training process of the performance prediction model includes: Independent random seeds were assigned to the water flux prediction branch and the salinity cutoff prediction branch, respectively. The data corresponding to the coupled feature subset is first divided into a training set and a test set according to a preset ratio, and then a validation set is divided from the training set according to a preset ratio. Set the number of decision tree iterations, learning rate, tree depth, regularization coefficient, and number of early stopping wait rounds for the water flux prediction branch and the salinity interception rate prediction branch; The two prediction branches are trained based on the training set, and the loss changes of the two prediction branches are monitored using the validation set to obtain the model parameters of the water flux prediction branch and the salt interception rate prediction branch in each iteration round.

7. The method for predicting the performance of thin-film nanocomposite films based on optimized machine learning according to claim 1, characterized in that, The optimized performance prediction model, determined based on the combined losses of water flux and salt rejection rate, includes: The water flux prediction branch and the salt rejection rate prediction branch are trained iteratively. In each iteration, a decision tree is added sequentially. The fitting target of the decision tree is the value of the prediction residual of the previous round along the negative gradient direction of the loss function. After each iteration, the root mean square error of the water flux prediction branch and the root mean square error of the salt cutoff rate prediction branch on the validation set are calculated respectively. After eliminating the dimensional differences through normalization, the weighted sum is obtained to obtain the comprehensive loss. The normalization process is to divide the root mean square error of each prediction branch by the range of values ​​of the corresponding target variable in the test set. If the overall loss does not decrease or increases within the preset number of early stop waiting rounds, training will be terminated, and the model parameters of the water flux prediction branch and the salt cutoff prediction branch will be rolled back to the historical iteration round with the lowest overall loss. The prediction accuracy of the water flux prediction branch and the salt cutoff rate prediction branch on the test set is evaluated by the coefficient of determination and the normalized root mean square error, respectively. When the prediction accuracy of both the water flux prediction branch and the salt cutoff rate prediction branch meets the preset accuracy conditions, the current performance prediction model is used as the optimized performance prediction model.

8. The method for predicting the performance of thin-film nanocomposite films based on optimized machine learning according to claim 1, characterized in that, The verification of discrete attribute features, continuous action features, and coupling features between the two in the coupled feature subset includes: Using the coupling feature as the verification unit, the discrete attribute features and continuous action features corresponding to the coupling feature are determined; Based on the continuous action characteristics, several interpolation points were set for the amount of inorganic nanomaterials added, the concentration of feed liquid, and the operating pressure within their respective ranges. Under different values ​​of discrete attribute features, the water flux prediction value and salt interception rate prediction value of the optimized performance prediction model are calculated at each interpolation point to obtain the performance response curves corresponding to different discrete attribute values. The coupling response difference value is obtained based on the shape difference of the performance response curve under different discrete attribute values. If the difference in the coupling response is greater than or equal to the significance threshold, the coupling feature is deemed to have passed the verification; otherwise, the coupling feature is deemed to have failed the verification.

9. The method for predicting the performance of thin-film nanocomposite films based on optimized machine learning according to claim 8, characterized in that, The step of obtaining the coupling response difference value based on the morphological differences of the performance response curves under different discrete attribute values ​​includes: The predicted values ​​of water flux and salt rejection rate corresponding to the same coupling feature are arranged into a two-dimensional performance response vector sequence according to the interpolation point order; Endpoint normalization is performed on the two-dimensional performance response vector sequence under different discrete attribute values; Calculate the coupling response distance between two-dimensional performance response vector sequences corresponding to any two discrete attribute values. The coupling response distance is determined by the square integral of the difference in response vectors at the same interpolation point and the square integral of the difference in response increments between adjacent interpolation points. The maximum value among the various coupling response distances is taken as the coupling response difference value.

10. The method for predicting the performance of thin-film nanocomposite films based on optimized machine learning according to claim 1, characterized in that, Iteratively update the coupled feature subset and retrain the performance prediction model, including: When the difference in the coupling response of a coupling feature is less than the significance threshold, it is determined that the coupling feature has not formed an effective coupling effect, and the coupling feature is removed from the subset of coupling features. When the discrete attribute feature or continuous action feature corresponding to the removed coupling feature no longer corresponds to any retained coupling feature, and the comprehensive synergistic contribution of the discrete attribute feature or continuous action feature is lower than a preset contribution threshold, the corresponding discrete attribute feature or continuous action feature is removed from the coupling feature subset. The performance prediction model was rebuilt and trained based on the updated coupled feature subset, and the combined loss of water flux and salt rejection rate was recalculated. Repeat the coupling feature verification, coupling feature subset update, and performance prediction model training until the verification results meet the preset reasonableness requirements; The preset rationality requirements include: the proportion of the number of verified coupled features to the total number of current coupled features reaches a preset pass rate, and the current comprehensive loss is not higher than the comprehensive loss of the previous round.

Images