A method and system for predicting performance of confined catalytic water treatment based on machine learning

Abstract

The application discloses a kind of confined catalytic water treatment performance prediction method and system based on machine learning, the method includes the following steps: step S1, collection and construction include multiple data samples Electro-catalytic membrane performance dataset;Step S2, electro-catalytic membrane performance dataset is preprocessed, and standardization dataset is obtained;Step S3, standardization dataset is divided into training set and test set, machine learning model is trained using training set, and the optimized prediction model is obtained by hyperparameter optimization;Step S4, the performance of the optimized prediction model is evaluated using test set;Step S5, the input feature data to be predicted is input into the optimized and verified prediction model, and the predicted pollutant removal rate is obtained.The application realizes the fast, high-precision prediction of the performance of complex electro-catalytic membrane system, and provides an efficient and reliable tool for the design and process optimization of electro-catalytic membrane.

Description

Technical Field

[0001] This invention belongs to the field of water treatment technology and relates to a purely data-driven prediction method, specifically a confined catalytic water treatment performance prediction method and system based on machine learning. Background Technology

[0002] Electrocatalytic membrane technology couples advanced electrochemical oxidation with membrane filtration, demonstrating great potential in treating recalcitrant organic micropollutants due to its advantages such as enhanced mass transfer and rapid reaction kinetics. The core of this technology lies in generating strong oxidizing agents (such as hydroxyl radicals) through electrochemical reactions at the membrane electrode, thereby efficiently degrading pollutants. However, the performance of electrocatalytic membrane systems is influenced by a complex interplay of factors, including membrane material properties (such as conductivity, pore size, and thickness), operating conditions (such as current density, transmembrane pressure, and pH), and the physicochemical properties of the pollutants themselves (such as molecular weight and hydrophobicity). These factors collectively determine the electron transfer, mass transport, and interaction between pollutants and the membrane surface within the system, resulting in a complex internal mechanism and challenging performance control.

[0003] Currently, the research and development of electrocatalytic membranes mainly relies on experimental trial-and-error methods. Researchers design and prepare membrane materials with different structural parameters and conduct batch experiments under various operating conditions to screen for the best-performing combinations. This method has significant limitations: First, the experimental process is time-consuming and labor-intensive, often requiring several days or even weeks from material preparation to performance testing, and is costly, involving expensive chemical reagents, specialized equipment, and energy consumption. Second, due to the numerous and interdependent influencing factors, traditional experimental methods struggle to systematically and quantitatively clarify the contribution of each factor to the final removal performance, and find it difficult to quickly identify critical paths and potential bottlenecks for performance optimization. Faced with a large number of candidate membrane materials and complex operating conditions, the inefficiency of experimental trial-and-error methods has become a bottleneck restricting the rapid development and engineering application of electrocatalytic membrane technology.

[0004] Machine learning, as an important branch of artificial intelligence, possesses powerful data processing and high-dimensional nonlinear mapping capabilities, enabling it to uncover potential patterns and regularities from complex data. It has demonstrated outstanding predictive performance in multiple fields. In water treatment, machine learning has been successfully applied to predict the retention rates of specific pollutants by traditional organic membranes or to screen active sites in electrocatalytic materials. However, directly applying purely data-driven machine learning models to predict the performance of electrocatalytic membranes in pollutant removal remains challenging and unexplored. On the one hand, electrocatalytic membrane systems involve strong coupling of multiple phases and processes, with high feature dimensionality and complex nonlinear relationships in their datasets, placing higher demands on the generalization ability and robustness of the models. On the other hand, existing research often focuses on specific systems, lacking a universal, highly efficient, and data-driven model framework that can integrate multi-source features (membrane, operation, pollutants) for efficient prediction. Therefore, developing an efficient, accurate, and well-interpretable purely data-driven prediction method is of urgent technical necessity and significant practical value for accelerating the design, optimization, and practical application of electrocatalytic membranes. Summary of the Invention

[0005] To address the problems of high cost, long cycle time, and difficulty in systematically identifying key factors in existing technologies that rely on experimental trial-and-error methods to optimize electrocatalytic membrane performance, this invention provides a confined catalytic water treatment performance prediction method and system based on machine learning. This invention does not rely on prior knowledge modeling of specific physical mechanisms such as electron transfer, mass transfer, and adsorption. Instead, it directly learns the complex mapping relationship between input features (covering membrane, operation, and pollutants) and output performance (pollutant removal rate) from historical experimental data. This enables rapid and high-precision prediction of the performance of complex electrocatalytic membrane systems, providing an efficient and reliable tool for the design and process optimization of electrocatalytic membranes.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] A machine learning-based method for predicting the performance of confined catalytic water treatment includes the following steps:

[0008] Step S1: Collect and construct an electrocatalytic membrane performance dataset containing multiple data samples, wherein: each data sample includes input features related to electrocatalytic membrane performance and corresponding pollutant removal rates; the input features include at least membrane structure parameters, operating condition parameters, and pollutant physicochemical property parameters; the membrane structure parameters include anode and cathode material type, membrane thickness, anode membrane pore size, porosity, contact angle, and conductivity; the operating condition parameters include electrolyte concentration, pH value, temperature, transmembrane pressure, current density or voltage, membrane flux, and electrode arrangement order; the pollutant physicochemical property parameters include pollutant concentration, molecular weight, octanol-water partition coefficient, number of rotatable bonds, and number of molecular rings.

[0009] Step S2: Preprocess the electrocatalytic membrane performance dataset to obtain a standardized dataset. The specific steps are as follows:

[0010] Step S21: Impute missing values ​​in the data sample;

[0011] Step S22: Encode the classification features in the data sample. The classification features include electrode material type, electrolyte type and electrode arrangement order. Encode them using a unique thermal coding method.

[0012] Step S23: Normalize or standardize all numerical features;

[0013] Step S24: Eliminate highly correlated features by calculating the Pearson correlation coefficient matrix;

[0014] Step S3: Divide the standardized dataset into a training set and a test set. Use the training set to train a machine learning model and obtain an optimized prediction model through hyperparameter optimization. The machine learning model is selected from at least one of extreme gradient boosting tree model, random forest model, and classification gradient boosting tree model. The hyperparameter optimization method adopts Bayesian optimization combined with cross-validation strategy.

[0015] Step S4: Evaluate the performance of the optimized prediction model using the test set;

[0016] Step S5: Input the input feature data to be predicted into the optimized and validated prediction model to obtain the predicted pollutant removal rate.

[0017] A machine learning-based confined catalytic water treatment performance prediction system includes a data acquisition and storage module, a data preprocessing module, a model training and optimization module, and a performance prediction module, wherein:

[0018] The data acquisition and storage module is used to store electrocatalytic membrane performance datasets;

[0019] The data preprocessing module is used to preprocess the electrocatalytic membrane performance dataset;

[0020] The model training and optimization module is used to train and optimize the machine learning prediction model based on the output of the preprocessing module;

[0021] The performance prediction module is used to load the optimized prediction model generated by the model training and optimization module, and output the predicted pollutant removal rate based on the user-input features to be predicted.

[0022] The prediction system also includes a model interpretation module, which performs feature importance analysis and visualization interpretation of prediction results based on the SHAP value method.

[0023] The model training and optimization module has a variety of optional machine learning algorithms built in, allowing users to select and configure models according to their needs;

[0024] The performance prediction module provides an interactive interface through which users can input or select membrane structure parameters, operating condition parameters, and pollutant physicochemical property parameters.

[0025] The data acquisition and storage module supports importing and updating electrocatalytic membrane performance data from external databases or literature data.

[0026] Compared with the prior art, the present invention has the following advantages:

[0027] 1. For the first time, a pure data-driven prediction model for pollutant removal by electrocatalytic membranes was constructed, avoiding the high cost and long cycle of traditional experimental trial and error methods. It can achieve performance prediction at the second or minute level, greatly improving R&D efficiency.

[0028] 2. By integrating multi-source features and advanced machine learning algorithms, the model can effectively capture nonlinear relationships in complex systems, with high prediction accuracy and strong generalization ability, providing a general framework for performance prediction of different systems.

[0029] 3. Even as a pure data model, this invention can be combined with interpretable artificial intelligence technologies such as SHAP to quantify the contribution of each input feature to the prediction results, thereby revealing the key factors affecting performance, providing data insights for mechanism research and experimental optimization, and enhancing the transparency and practicality of the model. Attached Figure Description

[0030] Figure 1 This is a flowchart of the overall process for predicting the performance of confined catalytic water treatment based on machine learning.

[0031] Figure 2 Scatter plot comparing the performance of different pure data-driven machine learning models (such as multiple linear regression (MLR), random forest (RF), and extreme gradient boosting tree (XGBoost)) in predicting pollutant removal rates on the test set. Detailed Implementation

[0032] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.

[0033] This invention provides a machine learning-based method for predicting the performance of confined catalytic water treatment. Its core lies in constructing a comprehensive, high-quality dataset and using it to train a high-performance, purely data-driven prediction model. The method mainly includes dataset construction, data preprocessing, model training and optimization, performance evaluation, and application prediction. The specific steps are as follows:

[0034] First, experimental data on electrocatalytic membrane treatment of different pollutants were collected extensively from publicly available academic literature, experimental reports, or databases to construct a multi-source feature dataset. Each data sample is treated as an instance, and its input features should reflect the experimental conditions as comprehensively as possible, including at least three categories: (1) membrane structure parameters, such as anode and cathode material types (after encoding), membrane thickness, anode membrane pore size, porosity, contact angle, conductivity, etc.; (2) operating condition parameters, such as electrolyte concentration, pH value, temperature, transmembrane pressure, current density or voltage, membrane flux, electrode arrangement order (after encoding), etc.; (3) physicochemical property parameters of pollutants, such as pollutant concentration, molecular weight, octanol-water partition coefficient (log K). ow The output variables include the number of rotatable bonds, the number of molecular rings, etc. The output variable is the corresponding pollutant removal rate.

[0035] Secondly, rigorous preprocessing of the constructed original dataset is crucial for ensuring model performance. Preprocessing steps include: handling missing values ​​in features using methods such as linear interpolation and mean imputation; converting categorical features such as "cathode and cathode materials" and "electrolyte type" into numerical forms using one-heat encoding for model processing; normalizing or standardizing all numerical features to eliminate the influence of differences in feature dimensions and numerical ranges; and feature filtering, such as calculating the Pearson correlation coefficient matrix to remove highly correlated features (e.g., correlation coefficients greater than 0.7) to reduce feature redundancy and multicollinearity, thereby improving model efficiency and generalization ability. The preprocessed dataset is then standardized.

[0036] Then, the standardized dataset is randomly divided into training and test sets according to a preset ratio (e.g., 8:2). Using the training set data, one or more machine learning models are selected and trained. Preferred models include, but are not limited to, ensemble learning models such as extreme gradient boosting trees, random forests, and classification gradient boosting trees, which perform well in handling high-dimensional, non-linear data. During training, Bayesian optimization and grid search strategies are used to optimize the model's hyperparameters, and K-fold cross-validation is used to evaluate the model performance under different parameter combinations to prevent overfitting, ultimately obtaining the prediction model with optimal generalization ability. Using the reserved test set data, the prediction accuracy and robustness of the optimized model are evaluated using metrics such as coefficient of determination, root mean square error, and mean absolute error.

[0037] Finally, the optimized and validated prediction model is deployed in the application phase. When it is necessary to predict the performance of a new set of membrane structures, operating conditions, and pollutant combinations, the corresponding feature data only needs to be preprocessed according to the same process as in the training phase, and then input into the trained model to quickly obtain the predicted pollutant removal rate, thereby providing forward-looking guidance for experimental design or process parameter adjustment.

[0038] Example 1

[0039] This embodiment provides a method for predicting the pollutant removal performance of electrocatalytic membranes based on machine learning. The overall process is as follows: Figure 1 As shown, starting from the collection of raw data, through data preprocessing, dataset partitioning, model training and optimization, performance evaluation, and finally to application prediction, a complete and closed-loop machine learning modeling and application process is formed, which specifically includes the following steps:

[0040] Step S101: Data Collection and Construction. Search academic databases such as Web of Science and Google Scholar using keywords such as "electrocatalytic membrane," "anodic oxidation," and "micropollutantremoval" to select relevant research papers published between 2010 and 2025. Extract experimental data manually or using digital tools (such as WebPlotDigitizer) from the eligible literature. The constructed dataset should include detailed conditions for each experiment. Specifically, the input features can be designed with 48 dimensions, covering three main categories: 1. Membrane structure parameters: anode material (e.g., Ti / RuO2-IrO2, Ti / SnO2-Sb, etc., which need to be coded), cathode material, anode membrane thickness, anode membrane conductivity, anode membrane average pore size, anode membrane porosity, anode membrane contact angle, anode-cathode distance, etc. 2. Operating condition parameters: Electrolyte type (e.g., Na2SO4, NaCl, etc., which need to be coded), electrolyte concentration, solution pH, temperature, transmembrane pressure, applied current density or voltage, anolyte flux, electrode arrangement (e.g., cathode-anode CA, or anode-cathode AC, which need to be coded), etc. 3. Physicochemical properties of pollutants: Pollutant name, molecular weight, initial concentration (usually log10 transformed), octanol-water partition coefficient log Kow, number of rotatable bonds, number of aromatic rings, etc. The output variable is the pollutant removal rate (percentage) under the corresponding conditions. Finally, this embodiment constructed a dataset containing 1036 valid data samples.

[0041] Step S102: Data preprocessing. The dataset obtained in step S101 is processed as follows:

[0042] 1. Missing value handling: For a few missing numerical features (such as membrane thickness not reported in some literature), the median of the feature in all samples is used to fill in the missing values.

[0043] 2. Classification Feature Encoding: For the four classification features "anodine material", "cathode material", "electrolyte type", and "electrode arrangement order", unique thermal encoding is used to convert them into binary vectors. For example, if there are 3 types of anode materials, they are encoded as a 3-dimensional vector.

[0044] 3. Feature Normalization: For all numerical input features (including the encoded portion) and the output variable "removal rate", a minimum-maximum normalization method is used to map the values ​​to the [0, 1] interval. The formula is: ,in It is the normalized value. These are the original data values. It is the minimum value of that feature (or output variable) across the entire dataset. It is the maximum value of that feature (or output variable) across the entire dataset.

[0045] 4. Feature Selection: Calculate the Pearson correlation coefficient matrix among all numerical features. For any pair of features, if their absolute correlation coefficient is greater than 0.7, one of them is removed based on domain knowledge or the initial correlation between the feature and the output variable to reduce redundancy.

[0046] Step S103: Dataset Splitting. The preprocessed, standardized dataset is randomly shuffled and then split into two parts: a training set (80%) and a test set (20%). That is, the training set contains approximately 828 samples, and the test set contains approximately 208 samples. This split ensures that the model is evaluated on unseen data.

[0047] Step S104: Model Training and Hyperparameter Optimization. The XGBoost model is selected as the base predictor. On the training set, Bayesian optimization is used to tune the key hyperparameters of XGBoost, with the optimization objective being the average R-value under five-fold cross-validation. 2 The optimized hyperparameters include: learning rate ([0.01, 0.3]); maximum tree depth ([3, 10]); subsample ratio ([0.5, 1.0]); feature sampling ratio (colsample_bytree) ([0.5, 1.0]); and regularization parameter lambda ([0, 5]). Bayesian optimization is performed for 30-100 iterations to obtain an optimal combination of hyperparameters. Using this optimal combination, the final XGBoost prediction model is retrained on all training data.

[0048] Step S105: Model Performance Evaluation. The final model obtained in Step S104 is evaluated using the reserved test set data. The coefficient of determination R of the model on the test set is calculated. 2 The root mean square error (RMSE) and mean absolute error (MAE) were also measured. For comparison, a purely data-driven multiple linear regression model and a random forest model (using default parameters or a simple grid search) were trained simultaneously and evaluated on the same test set. Performance comparison results are shown below. Figure 2 As shown. Figure 2 The predictive performance of different pure data-driven machine learning models (MLR, RF, CatBoost, XGBoost) on the test set was compared using scatter plots. The horizontal axis represents the experimentally measured pollutant removal rate, and the vertical axis represents the model-predicted removal rate; ideally, the prediction should fall on the diagonal. By observing the closeness of the scatter plots to the diagonal, the prediction accuracy of each model can be visually compared. Figure 2 As shown, the XGBoost model's predicted points cluster most closely around the diagonal, exhibiting the least dispersion. Its test set R... 2 The RMSE and MAE of the algorithm are above 0.85, while the prediction points of the MLR model are more dispersed. The RMSE and MAE of the algorithm are significantly lower than those of the MLR model, and it also has certain advantages compared with the random forest model. This figure not only strongly demonstrates the superiority of the machine learning algorithm preferred by this invention in solving this prediction problem, but also proves the effectiveness of the model and process used in this invention.

[0049] Step S106: Apply prediction and model interpretation. Save the trained optimal XGBoost model. When it is necessary to predict a novel titanium-based coated anode film (film thickness 200 μm, pore size 0.5 μm, porosity 40%), under the following conditions: electrolyte 50 mM Na2SO4, pH=7, transmembrane pressure 1 bar, current density 10 mA / cm². 2 Under the given conditions, when treating pollutants with an initial concentration of 10 μM (i.e., log10(concentration) = 5.0), a molecular weight of 200 g / mol, and log Kow = 3.0, the aforementioned feature values ​​are first normalized according to the same preprocessing procedure in step S102 (calling the saved normalization parameters). Then, the processed feature vector is input into the saved XGBoost model, which outputs a normalized removal rate prediction value. Reverse normalization yields the predicted removal rate. Furthermore, the SHAP tool can be used to interpret the model. By calculating and ranking the SHAP average absolute values ​​of each feature, it is clear that features such as current density and pollutant concentration have the highest global importance for the model's prediction. This provides direct data support for researchers to optimize process parameters (such as prioritizing current density adjustment).

[0050] Example 2

[0051] This embodiment provides a system for implementing the method described in Embodiment 1 above. This system is a web-based application and mainly includes the following modules:

[0052] 1. Data Management Backend: As a data acquisition and storage module, it provides a data upload interface, supporting users to upload their collected experimental data in CSV or Excel format. Data fields must match preset feature templates. The system includes the benchmark dataset described in Example 1.

[0053] 2. Automated Preprocessing Pipeline: As a data preprocessing module, after users upload data or select a built-in dataset, the system automatically performs operations such as missing value imputation (median), one-hot encoding of categorical features, and feature normalization. Users can view data statistics before and after preprocessing on the interface.

[0054] 3. Model Factory: As a model training and optimization module, it provides a variety of machine learning algorithms (XGBoost, Random Forest, CatBoost, SVM, etc.) for users to choose from. Users can configure the training / test set split ratio and choose whether to enable automatic hyperparameter optimization (Bayesian optimization). After the training task is submitted, it is executed asynchronously in the server background, and a model performance report is returned upon completion.

[0055] 4. Online Prediction and Interpretation Platform: This platform serves as the core of the performance prediction and model interpretation modules. Users input or select membrane parameters, operating conditions, and pollutant properties through a form-based interface. After submission, the system automatically invokes the preprocessing pipeline and the trained optimal model to perform predictions, returning the predicted removal rate in real time. Simultaneously, the system generates a SHAP force interpretation map for this prediction, visually demonstrating how each feature pushes the predicted value from the baseline to the final result, as well as the specific contribution value of each feature.

[0056] This system encapsulates professional data analysis and modeling processes into easy-to-use tools, greatly lowering the barrier for electrocatalytic membrane researchers to apply machine learning technology.

Claims

1. A machine learning-based method for predicting the performance of confined catalytic water treatment, characterized in that... The method includes the following steps: Step S1: Collect and construct an electrocatalytic membrane performance dataset containing multiple data samples; Step S2: Preprocess the electrocatalytic membrane performance dataset to obtain a standardized dataset; Step S3: Divide the standardized dataset into a training set and a test set, use the training set to train the machine learning model, and obtain the optimized prediction model through hyperparameter optimization. Step S4: Evaluate the performance of the optimized prediction model using the test set; Step S5: Input the input feature data to be predicted into the optimized and validated prediction model to obtain the predicted pollutant removal rate.

2. The confined catalytic water treatment performance prediction method based on machine learning according to claim 1, characterized in that... In step S1, each data sample includes input features related to the performance of the electrocatalytic membrane and the corresponding pollutant removal rate. The input features include at least membrane structure parameters, operating condition parameters, and pollutant physicochemical property parameters. The membrane structure parameters include anode and cathode material type, membrane thickness, anode membrane pore size, porosity, contact angle, and conductivity. The operating condition parameters include electrolyte concentration, pH value, temperature, transmembrane pressure, current density or voltage, membrane flux, and electrode arrangement order. The pollutant physicochemical property parameters include pollutant concentration, molecular weight, octanol-water partition coefficient, number of rotatable bonds, and number of molecular rings.

3. The method for predicting the performance of confined catalytic water treatment based on machine learning according to claim 1, characterized in that... The specific steps of step S2 are as follows: Step S21: Impute missing values ​​in the data sample; Step S22: Encode the classification features in the data sample. The classification features include electrode material type, electrolyte type and electrode arrangement order. Encode them using a unique thermal coding method. Step S23: Normalize or standardize all numerical features.

4. The machine learning-based method for predicting the performance of confined catalytic water treatment according to claim 3, characterized in that... Step S2 further includes step S24, which involves calculating the Pearson correlation coefficient matrix to remove highly correlated features.

5. The method for predicting the performance of confined catalytic water treatment based on machine learning according to claim 1, characterized in that... In step S3, the machine learning model is selected from at least one of the extreme gradient boosting tree model, random forest model, and classification gradient boosting tree model; the hyperparameter optimization method adopts Bayesian optimization combined with cross-validation strategy.

6. A machine learning-based confined catalytic water treatment performance prediction system for implementing the method of any one of claims 1-5, characterized in that... The prediction system includes a data acquisition and storage module, a data preprocessing module, a model training and optimization module, and a performance prediction module, wherein: The data acquisition and storage module is used to store electrocatalytic membrane performance datasets; The data preprocessing module is used to preprocess the electrocatalytic membrane performance dataset; The model training and optimization module is used to train and optimize the machine learning prediction model based on the output of the preprocessing module; The performance prediction module is used to load the optimized prediction model generated by the model training and optimization module, and output the predicted pollutant removal rate based on the user-input features to be predicted.

7. The confined catalytic water treatment performance prediction system based on machine learning according to claim 6, characterized in that... The prediction system also includes a model interpretation module, which performs feature importance analysis and visualization interpretation of prediction results based on the SHAP value method.

8. The confined catalytic water treatment performance prediction system based on machine learning according to claim 6, characterized in that... The model training and optimization module incorporates a variety of machine learning algorithms, allowing users to select and configure models according to their needs.

9. The confined catalytic water treatment performance prediction system based on machine learning according to claim 6, characterized in that... The performance prediction module provides an interactive interface through which users can input or select membrane structure parameters, operating condition parameters, and pollutant physicochemical property parameters.

10. The confined catalytic water treatment performance prediction system based on machine learning according to claim 6, characterized in that... The data acquisition and storage module supports importing and updating electrocatalytic membrane performance data from external databases or literature data.

Images