Performance prediction method based on membrane distillation membrane material standardization database
By constructing a standardized database and performance prediction model, the problem of integrating membrane distillation data was solved, enabling efficient screening of membrane materials and optimization of operating conditions, thereby improving the comparability and efficiency of the membrane distillation process.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DALIAN UNIV OF TECH
- Filing Date
- 2026-04-13
- Publication Date
- 2026-07-14
AI Technical Summary
Existing technologies struggle to effectively integrate and compare membrane distillation data from different literature sources, leading to difficulties in membrane material selection and operating condition optimization, and a lack of interpretability and operability.
A standardized database is constructed to unify the conversion of performance indicators, perform feature classification and sample processing, establish a performance prediction model, and realize the screening of membrane materials and optimization of operating conditions.
It improves the comparability and utilization efficiency of membrane distillation data, reduces reliance on human experience, and increases the efficiency of membrane material selection and operating condition optimization.
Smart Images
Figure CN122392735A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of membrane separation process modeling, literature data mining and intelligent optimization technology, and specifically relates to a performance prediction method based on a standardized database of membrane distillation membrane materials. Background Technology
[0002] Crystallization engineering serves a wide range of fields, including electronic chemicals, fine chemicals, energy chemicals, pharmaceuticals, and the preparation of high-purity materials. As products upgrade from industrial-grade to electronic-grade, ultrapure, and functionalized grades, the requirements for crystal purity, particle size distribution, and crystal stability significantly increase. Membrane distillation and its coupled membrane crystallization processes, due to their ability to achieve selective transfer of volatile components at lower temperatures and to regulate mass transfer and crystallization behavior through hydrophobic membrane interfaces, hold significant promise for applications in high-salinity wastewater treatment, salt resource recovery, thermosensitive solution concentration, and fine crystal preparation.
[0003] However, the performance of membrane distillation processes is influenced by a multitude of factors, including membrane material properties, membrane structural parameters, surface wetting characteristics, modification methods, feed properties, and operating conditions. Taking membrane materials as an example, the polymer matrix, pore size, porosity, membrane thickness, contact angle, liquid inlet pressure (LEP), and surface roughness all affect vapor transfer resistance, interfacial wetting risk, and fouling behavior. Regarding operating conditions, feed temperature, cooling temperature, flow rate, vacuum level, temperature difference, and solution concentration collectively determine the transmembrane driving force, concentration polarization, and long-term operational stability. For membrane crystallization equipment, it is also necessary to consider seed nucleation, continuous growth, and the membrane surface's resistance to salt fouling. Therefore, relying solely on experience to select membrane materials often fails to balance flux, anti-wetting ability, and crystallization quality.
[0004] Existing research has reported numerous experimental results for commercial and modified membranes in different membrane distillation configurations. Materials such as PTFE, PVDF, PP, PES, and ceramic membranes have been applied in direct contact membrane distillation, air-gap membrane distillation, vacuum membrane distillation, and scavenging membrane distillation. Some studies further indicate that the hydrophilicity / hydrophobicity control of the membrane surface and the pore structure design not only affect permeation performance but may also influence seed nucleation and crystal sieving behavior during membrane crystallization. Nevertheless, due to significant differences in the apparatus structures, feed systems, operating conditions, and performance characterization methods used in various studies, data from different publications often exhibit contradictory meanings or are not directly comparable.
[0005] First, the calculation methods for mass transfer driving forces differ across membrane distillation configurations. If raw flux is directly used as the evaluation metric, results from different configurations such as DCMD, AGMD, PGMD, VMD, and SGMD will be difficult to compare. Even with flux data, variations in feed concentration, temperature, and vacuum conditions across different studies render simple numerical comparisons meaningless. Second, the expression of membrane parameters in the literature is highly heterogeneous. The same parameter may have different nomenclature, units, or descriptive granularities; modification methods and modifier categories are often described in text, making them unsuitable for direct use in subsequent statistical learning. Third, some studies only report flux without providing the driving pressure differential or directly usable permeability, while others lack crucial temperature, vacuum, or feed information, making it impossible to directly input raw data into the model.
[0006] Currently, some data-driven methods attempt to use machine learning to predict membrane separation performance. However, most methods directly fit mixed data, lacking configuration identification, unified conversion of driving pressure difference, and scenario comparability screening for membrane distillation scenarios. They also lack interpretable analysis of low-frequency parameters, interaction relationships, and local sample contributions, making it difficult to translate model conclusions into actionable material screening rules. Furthermore, existing methods often remain at the "prediction" level and have not yet formed a complete closed loop from automatic literature expansion, performance standardization, model training, interpretation and analysis to candidate membrane scheme screening, operational window optimization, and updates based on new literature / experimental feedback.
[0007] Therefore, there is an urgent need for a dedicated method for membrane distillation and membrane crystallization scenarios that can uniformly extract, standardize the expression, and perform physical comparability conversion of scattered literature data. Based on this, an interpretable performance prediction model should be established, and the prediction results should be truly used for membrane material screening and operating condition optimization. Summary of the Invention
[0008] To address the aforementioned problems, this invention provides a performance prediction method based on a standardized database of membrane distillation materials. This method does not simply stack existing literature data to train a model; instead, it focuses on membrane distillation configurations and mass transfer mechanisms, designing modular approaches for data acquisition, performance standardization, feature classification, and model training. This ultimately forms an executable technical route for membrane material selection and operational optimization. This method can be used for membrane material evaluation in membrane distillation processes, as well as for preliminary screening of candidate membrane materials, optimization of operating parameters, and determination of modification directions in membrane crystallization equipment.
[0009] To achieve the above objectives, the present invention adopts the following technical solution: A performance prediction method based on a standardized database of membrane distillation materials is proposed. First, a standardized database is constructed. Second, the original performance indicators are uniformly converted according to different membrane distillation configurations to obtain standardized performance indicators that can be compared horizontally. Then, membrane-related data are structured and the samples are screened to form a standardized sample set. Finally, a performance prediction model is established to predict the performance of membrane materials under given conditions, thereby enabling membrane material selection and optimization of operating conditions. The specific steps are as follows: Step 1: Standardized Database Construction Experimental data extracted from the literature were standardized in naming, unitization, and field mapping, integrating information scattered across tables, text, and figure captions into a structured database. The database includes: pore size, porosity, contact angle, LEP, solute concentration, modifier quantity, feed temperature, cooling temperature, feed flow rate, coolant flow rate, vacuum pressure, temperature difference, membrane distillation type, membrane geometry, polymer matrix, commercial membrane name, solute type, modification method, and modifier category.
[0010] Step 2: Configuration Identification and Standardized Performance Conversion First, determine the membrane distillation configuration corresponding to each sample, and then determine the transmembrane vapor pressure difference based on the configuration. The calculation method. The standardized performance indicator uses penetration rate. ,in, To standardize penetration rate, This represents the original flux.
[0011] For direct contact membrane distillation (DCMD), air gap membrane distillation (AGMD), and permeate gap membrane distillation (PGMD) configurations, the preferred method is... For vacuum membrane distillation (VMD) configuration, the preferred method is... For the scavenging membrane distillation (SGMD) configuration, the preferred method is... .in, For the feed side Saturated vapor pressure at temperature Cooling side temperature The saturated vapor pressure below; The water activity on the feed side; This refers to the vacuum pressure on the permeate side. The saturated vapor pressure is... The water activity of a saline system can be calculated using the Antoine equation. It can be estimated using empirical formulas for solute mole fraction, activity coefficient models, or corresponding formulas from literature. Through configuration identification and unified conversion of driving pressure difference, the raw fluxes from different literature can be transformed into standard outputs with cross-comparability.
[0012] Step 3: Feature Classification and Sample Processing For the samples obtained in step 1, pore size, porosity, contact angle, LEP, solute concentration, modifier quantity, feed temperature, cooling temperature, feed flow rate, coolant flow rate, vacuum pressure, and temperature difference are defined as numerical features, and normalization or standardization is used to eliminate the influence of dimensions. Whether it is a commercial membrane and whether it is modified are defined as Boolean features, and logical mapping is used to convert them into numerical representations of 0 or 1. Membrane distillation type, membrane geometry, polymer matrix, commercial membrane name, solute type, modification method, and modifier category are defined as categorical features, and One-Hot encoding or Multi-Hot encoding is used for processing. Samples with missing key physical quantities that cannot be reliably inferred are removed. For samples with missing non-key auxiliary features, group median, mode, or independent missing value indicators can be used for processing. Outliers are preferably identified and removed using IQR statistics, Z-score method, or a combination of both; finally, a standardized sample set is obtained.
[0013] Step 4: Training the performance prediction model Based on the standardized sample set obtained in step 3, a machine learning model for predicting membrane material performance is established. Specifically, the numerical, Boolean, and encoded categorical features processed in step 3 are used as the input feature matrix, and the standardized permeability is used as the input feature matrix. To output the target value, model weights are fitted using the training set data. Specifically, Random Forest, Gradient Boosting, XGBoost, and LightGBM models are constructed and compared through training, validation, and test set splitting, K-fold cross-validation, and hyperparameter optimization. Model selection criteria include the coefficient of determination R0. 2 The model with the smallest root mean square error (RMSE) and mean absolute error (MAE) is used as the final prediction model; the performance of the membrane distillation membrane material is judged based on the predicted permeability.
[0014] Compared with the prior art, the present invention has the following beneficial effects: First, by constructing a standardized database and uniformly converting performance indicators under different membrane distillation configurations, the comparability of data from different sources is improved. Second, by structurally expressing and screening sample parameters, the utilization efficiency of membrane distillation data is improved. Third, it reduces the reliance on manual experience in membrane material screening and operating condition optimization, thereby improving evaluation and screening efficiency. This method can significantly improve the comparability and utilization efficiency of multi-source membrane distillation data, and is particularly suitable for evaluating membrane materials, screening membrane schemes, and optimizing operating conditions for membrane crystallization equipment. Attached Figure Description
[0015] Figure 1 This is a flowchart of the overall process of the method of the present invention. Detailed Implementation
[0016] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.
[0017] The basic process of the performance prediction method based on a standardized database of membrane distillation membrane materials in this embodiment is as follows: Figure 1 As shown, the specific steps are as follows: Step 1: Standardized performance conversion and feature classification: This embodiment first identifies the membrane distillation configuration of the samples based on the device description and experimental procedure in the literature. For DCMD, AGMD, and PGMD samples, the driving force is calculated using the difference between the saturated vapor pressure on the feed side and the saturated vapor pressure on the cooling side; for VMD samples, the driving force is calculated using the difference between the saturated vapor pressure on the feed side and the vacuum pressure on the permeate side; for SGMD samples, the effective driving force is calculated using the product of the saturated vapor pressure on the feed side and the water activity of the feed. After conversion, the driving force is... As a preferred standardized output, the original flux of different documents is uniformly mapped to the same physical quantity.
[0018] In terms of feature engineering, this embodiment organizes input variables according to a structure of "numerical + Boolean + categorical". During actual processing, 21 original features were extracted and expanded into 68 coded features usable for training, including 12 numerical features, 2 Boolean features, and 7 categorical features. The numerical features include pore size, porosity, contact angle, LEP, solute concentration, modifier quantity, feed temperature, cooling temperature, feed flow rate, coolant flow rate, vacuum pressure, and temperature difference. The Boolean features include whether the membrane is commercial and whether it is modified. The categorical features include membrane distillation type, membrane geometry, polymer matrix, commercial membrane name, solute type, modification method, and modifier category.
[0019] For single-value variables in categorical features, such as membrane distillation configuration, membrane geometry, and polymer matrix, One-Hot coding is preferred. For variables that may have multiple values, such as modifier type and composite modification pathway, Multi-Hot coding is preferred. For missing values, if they belong to fields related to flux, configuration, critical temperature, or driving force, they are judged as non-comparable samples and removed; if they belong to commercial membrane names or auxiliary text fields, independent missing value indicators can be used. For extreme outliers, it is preferred to use IQR statistics combined with Z-score method for dual identification to reduce the misleading effect of non-physical outlier records on the model.
[0020] Step Two: Performance Prediction Model Establishment and Material Screening The dataset, consisting of 7671 samples after standardization and feature engineering in step one, is received. It is then divided into training and testing sets in an 8:2 ratio. This embodiment uses the XGBoost algorithm to construct a membrane material performance prediction model, taking 68 features (including membrane material, modifier type, normalized pore size, porosity, feed temperature, and temperature difference) after One-Hot and Multi-Hot encoding as model inputs. Standardized permeability is also included. As the target for prediction output, to ensure the model conforms to the physical laws of membrane distillation, positive monotonic constraints were applied to features such as "feed temperature," "porosity," and "temperature difference," while a negative monotonic constraint was applied to "membrane thickness" in the model parameter settings. After optimizing the tree depth and learning rate through K-fold cross-validation, excellent prediction performance was achieved on the test set, with a coefficient of determination R0. 2 It reaches 0.85 or higher.
[0021] In the membrane material screening and optimization application stage: a given operating condition for high-salt wastewater treatment is set (e.g., feed temperature 70℃, cooling temperature 20℃, 5 wt% NaCl solution). The parameters of 10 pre-selected candidate membrane materials with different pore sizes and surface modification schemes from the literature library are input into a trained XGBoost prediction model. The model quickly outputs the predicted normalized permeability of each candidate membrane under this specific operating condition. The "coated PTFE membrane" with the highest predicted permeability and hydrophobic modification characteristics is selected as the optimal screening result, thus replacing the traditional high-cost trial-and-error experiment and achieving efficient and intelligent screening of membrane materials.
Claims
1. A performance prediction method based on a standardized database of membrane distillation membrane materials, characterized in that, First, a standardized database is constructed. Second, the original performance indicators are uniformly converted according to different membrane distillation configurations to obtain standardized performance indicators that can be compared horizontally. Subsequently, the membrane-related data are expressed in a structured manner, and the samples are screened to form a standardized sample set. Establish a performance prediction model to predict the performance of membrane materials under given conditions, and use this model to screen membrane materials and optimize operating conditions.
2. The performance prediction method based on a standardized database of membrane distillation membrane materials according to claim 1, characterized in that, The specific steps are as follows: Step 1: Standardized Database Construction The experimental data extracted from the literature were standardized in terms of naming, units, and fields, and the information scattered in tables, text, and figure captions was integrated into a structured database. The database contains: pore size, porosity, contact angle, LEP, solute concentration, modifier quantity, feed temperature, cooling temperature, feed flow rate, coolant flow rate, vacuum pressure, temperature difference, membrane distillation type, membrane geometry, polymer matrix, commercial membrane name, solute type, modification method, and modifier category. Step 2: Configuration Identification and Standardized Performance Conversion First, determine the membrane distillation configuration corresponding to each sample, and then determine the transmembrane vapor pressure difference based on the configuration. The calculation method; Standardized performance indicators use penetration rate ,in, To standardize penetration rate, This is the original flux; For direct contact membrane distillation (DCMD), air gap membrane distillation (AGMD), and permeate gap membrane distillation (PGM) configurations, the following are adopted: For the vacuum membrane distillation (VMD) configuration, the following is adopted: For the scavenging membrane distillation SGMD configuration, the following is adopted: ;in, For the feed side Saturated vapor pressure at temperature Cooling side temperature The saturated vapor pressure below; The water activity on the feed side; This refers to the vacuum pressure on the permeation side. Step 3: Feature Classification and Sample Processing For the samples obtained in step 1, pore size, porosity, contact angle, LEP, solute concentration, modifier quantity, feed temperature, cooling temperature, feed flow rate, coolant flow rate, vacuum pressure, and temperature difference are defined as numerical features, and normalization or standardization is used to eliminate the influence of dimensions. Whether it is a commercial membrane and whether it is modified are defined as Boolean features, and logical mapping is used to convert them into numerical representations of 0 or 1. Membrane distillation type, membrane geometry, polymer matrix, commercial membrane name, solute type, modification method, and modifier category are defined as categorical features, and One-Hot encoding or Multi-Hot encoding is used for processing. Samples with missing key physical quantities that cannot be reliably inferred are removed. For samples with missing non-key auxiliary features, group median, mode, or independent missing value indicators are used for processing. Outliers are preferably identified and removed using IQR statistics, Z-score method, or a combination of both. Finally, a standardized sample set is obtained. Step 4: Training the performance prediction model Based on the standardized sample set obtained in step 3, a machine learning model for predicting membrane material performance is established; the numerical, Boolean, and encoded categorical features processed in step 3 are used as the input feature matrix, with standardized permeability as the input feature matrix. To output the target value, the model weights are fitted using the training set data to construct a Random Forest, Gradient Boosting, XGBoost, or LightGBM model, and the performance of the membrane distillation membrane material is judged based on the predicted permeability.
3. The performance prediction method based on a standardized database of membrane distillation membrane materials according to claim 2, characterized in that, In step 4, the models are compared through training set, validation set, and test set partitioning, K-fold cross-validation, and hyperparameter optimization; the model selection criterion is the coefficient of determination R. 2 The model with the smallest root mean square error (RMSE) and mean absolute error (MAE) is used as the final prediction model.