Membrane material screening and process parameter optimization method and system based on machine learning

Abstract

The invention belongs to the field of guiding air-gap membrane distillation to recycle organic solvents, and discloses a method and a system for membrane material screening and process parameter optimization based on machine learning, and the method comprises the following steps: collecting pervaporation document data and air-gap membrane distillation experimental data, cleaning and aligning, and constructing a heterogeneous data set; generating an SMILES sequence code and a molecular fingerprint of the solvent, and carrying out standardization processing on process parameters to generate multi-modal input data; based on a deep learning regression model of a Transform architecture, executing a pre-training task based on pervaporation data and a transfer learning fine tuning task based on air gap type membrane distillation data; and performing high-throughput prediction on the chemical molecule library, and outputting an optimal solvent list and recommended process parameters. According to the method, historical data prediction and reverse optimization are utilized through machine learning, so that the experiment cost and trial and error times of membrane distillation organic solvent separation process optimization are remarkably reduced.

Description

Technical Field

[0001] This invention belongs to the field of guided air-gap membrane distillation for the recovery of organic solvents, specifically involving a method and system for membrane material screening and process parameter optimization based on machine learning. Background Technology

[0002] Polytetrafluoroethylene (PTFE) is a polymer material polymerized from tetrafluoroethylene monomers. Due to the extremely high bond energy of the carbon-fluorine bonds in its molecular structure, PTFE possesses excellent chemical stability, superior resistance to high and low temperatures, and significant hydrophobic properties, enabling it to resist the erosion of strong acids, strong alkalis, oils, and various organic solvents. This makes PTFE widely used in membrane distillation, and it is one of the most ideal membrane materials in organic solvent recovery and dehydration processes. Utilizing PTFE membrane modules for the separation, enrichment, and reuse of waste solvents has become a key technological pathway for achieving green manufacturing.

[0003] Despite the excellent properties of PTFE membrane materials, their processing and molding are difficult, and the multistage air-gap membrane distillation (MAMD) process involves complex heat and mass transfer processes, with the selection of process parameters having a significant impact on separation performance. For a long time, researchers have mainly relied on traditional methods such as orthogonal experimental design or single-factor analysis for parameter optimization, but these methods have significant drawbacks in practical applications. (1) High cost and long cycle: It requires a large number of physical experiments, which consumes huge human, material and time costs; (2) Insufficient exploration of interaction: It is difficult to effectively reveal the complex nonlinear interaction between multiple process parameters; (3) Membrane distillation processes usually face a trade-off between transmembrane flux and rejection rate, that is, an increase in flux is often accompanied by a decrease in rejection rate. Traditional methods can usually only optimize for a single objective and cannot provide a series of optimal solutions covering different preferences, making it difficult to meet the different emphasis requirements of industrial sites for yield or purity.

[0004] In recent years, although some researchers have attempted to use machine learning models to replace physical experiments and construct surrogate models between process parameters and performance, existing technologies have the following shortcomings for multi-stage air-gap membrane distillation systems for recovering organic solvents: (1) Large deviation in physical mechanism model: The MAMD process involves a multi-level series structure. Each air gap layer not only has complex heat conduction, but also multi-component molecular diffusion phase transition under non-isothermal conditions. Traditional Knudsen diffusion or molecular diffusion models are difficult to accurately describe the complex heat-mass coupling and transfer behavior in the air gap layer. Moreover, the physical properties of the organic solvent system change drastically with concentration and temperature, resulting in a large deviation in the estimation of the mass transfer coefficient. The model prediction accuracy is difficult to meet engineering requirements.

[0005] (2) Poor robustness of single machine learning models: Existing data-driven research often uses a single, shallow machine learning model. Due to the difficulty in obtaining experimental data for MAMD, the sample size is usually small, and a single model is extremely sensitive to data noise, easily falling into overfitting or local optima. In addition, existing models mostly only use process parameters as numerical inputs, ignoring the deep structure-property relationship between solvent molecular structure and membrane material microstructure, resulting in extremely weak generalization ability of the model when facing new solvent systems.

[0006] (3) Disconnect between optimization efficiency and model integration: The optimization algorithm and the prediction model are disconnected: Existing technologies usually only build performance prediction models in a single direction, lacking deep integration of high-precision prediction models with multi-objective evolutionary algorithms. This results in researchers being able to predict the results, but unable to quickly and automatically reverse the optimization, making it difficult to locate the optimal combination of process parameters that simultaneously satisfies high throughput and high enrichment factor in a huge parameter space at low cost. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a method and system for membrane material screening and process parameter optimization based on machine learning. Specifically, by introducing a transfer learning strategy, using mature pervaporation data to assist training, and combining the Transformer deep learning architecture to extract the deep chemical structure features of solvent molecules and separated membrane monomers, high-precision prediction of air-gap membrane distillation performance and reverse screening of process parameters are achieved under small sample conditions.

[0008] To achieve the above objectives, this invention provides a method for membrane material screening and process parameter optimization based on machine learning, the method comprising: S1: Collect literature data on pervaporation and experimental data on air-gap membrane distillation, and clean and align them to construct a heterogeneous dataset; S2: Generate the SMILES sequence encoding and molecular fingerprint of the solvent, extract statistical features using the SMILES sequence, extract the deep structural features of the separation membrane and solvent using the molecular fingerprint, and standardize the process parameters to generate multimodal input data. S3: A deep learning regression model based on the Transformer architecture, which performs pre-training tasks based on pervaporation data and transfer learning fine-tuning tasks based on air-gap membrane distillation data. S4: Perform high-throughput prediction on chemical molecular libraries and output a list of optimal solvents and recommended process parameters.

[0009] Preferably, the heterogeneous dataset in step S1 includes: a source domain dataset and a target domain dataset; The source domain dataset contains operating parameters for the pervaporation process, including feed temperature, downstream membrane pressure, feed concentration, and membrane material type. The target domain dataset contains source domain process parameter feature variables including: top surface temperature, bottom surface temperature, feed temperature, air gap width, number of stages, feed water flow rate, temperature difference, and power. The target domain process parameters include: hot side temperature, cold side temperature, air gap width, membrane pore size, and membrane thickness. To address the inconsistency in the dimensions of operating parameters between the source and target domains, alignment is performed using parameter mapping or zero-value filling methods.

[0010] Preferably, step S2, which uses SMILES sequences to extract statistical features, specifically includes: using the cheminformatics tool RDkit to normalize the SMILES strings of organic solvents and membrane materials; merging the SMILES vectors output by the Transformer encoder, membrane material parameters, and process parameter numerical vectors to form a joint feature vector that is input to the fully connected layer; including the operating parameters of the air-gap membrane distillation process; performing SMILES serialization encoding and molecular fingerprint calculation on organic solvent molecules as sequence input to the Transformer model.

[0011] Preferably, step S2, which utilizes molecular fingerprinting to extract deep structural features of the separation membrane and solvent, specifically includes: setting the iteration radius and bit vector length, and using the Morgan fingerprint algorithm to generate a binary fingerprint vector of the organic solvent; the binary fingerprint vector serves as a global structural feature to characterize the local chemical environment and topological structure of the molecule.

[0012] Preferably, the construction of multimodal input in step S2 specifically involves: constructing a feature fusion layer, which concatenates the deep semantic vector of the SMILES sequence extracted by the deep learning model, the molecular fingerprint vector, and the normalized process parameter numerical vector to form a joint feature vector.

[0013] Preferably, the deep learning regression model based on the Transformer architecture in step S3 has the following architecture: Embedding layer: used to convert the SMILES numerical index sequence into a high-dimensional vector representation; Encoder: contains multiple stacked coding blocks, each composed of a multi-head self-attention mechanism and a feedforward neural network, used to capture long-range dependencies between solvent molecule atoms; Regression prediction head: composed of a multi-layer sensing mechanism, receives the joint feature vector, and outputs predicted values ​​of transmembrane flux and separation factor.

[0014] Preferably, the pre-training and fine-tuning process in step S3 specifically includes: Pre-training stage: using pervaporation data from the source domain to perform full-parameter supervised training on the model, using mean squared error as the loss function, so that the model learns the general physicochemical relationship between solvent molecular structure and membrane separation performance; Fine-tuning stage: freezing the weight parameters of the model embedding layer and the bottom layer of the Transformer encoder, and only updating the parameters of the top layer encoder and regression prediction head; inputting air-gap membrane distillation data from the target domain into the model for training to adapt to the characteristics of air-gap mass transfer resistance and thermal efficiency.

[0015] Preferably, the high-throughput screening in step S4 specifically includes: constructing an optimized variable space using a mixed integer encoding strategy, wherein the membrane material type and solvent type are mapped as discrete integer variables, and the membrane synthesis conditions and process operation parameters are defined as continuous real variables; the deep learning regression model based on the Transformer architecture is encapsulated as a fitness evaluation function of the non-dominated sorting genetic algorithm NSGA-II; performing evolutionary search, in each generation of population evolution, the decoded discrete features, including SMILES sequences and molecular fingerprints and continuous process parameters, are input into the model to predict transmembrane flux and separation factor as fitness values, and the population is iteratively updated through non-dominated sorting, crowding distance calculation, and genetic operations; finally, a Pareto optimal front solution set is output, which contains multiple optimal membrane material selection and process parameter coupling schemes.

[0016] The present invention also provides a system for screening membrane materials and optimizing process parameters based on machine learning. The system is used to implement the aforementioned method and includes: a data construction module, a preprocessing module, a model training module, and a screening and optimization module. The data construction module is used to collect pervaporation literature data and air-gap membrane distillation experimental data, and to clean and align them to construct a heterogeneous dataset. The preprocessing module is used to generate the SMILES sequence code and molecular fingerprint of the solvent, and to standardize the process parameters to generate multimodal input data. The model training module is used to perform pre-training tasks based on pervaporation data and transfer learning fine-tuning tasks based on air-gap membrane distillation data for deep learning regression models based on the Transformer architecture. The screening and optimization module is used to perform high-throughput prediction of chemical molecular libraries and output a list of optimal solvents and recommended process parameters.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The method and framework provided by this invention employ an integrated reverse design framework that utilizes transfer learning for evolutionary optimization, significantly reducing R&D costs. Addressing the pain point of scarce experimental data in air-gap membrane distillation, this invention innovatively introduces a transfer learning strategy, utilizing mature data from the pervaporation field to assist training, effectively solving the problems of model cold start and overfitting. Furthermore, this invention encapsulates a high-precision Transformer prediction model as a black-box fitness function, deeply coupling it with a non-dominated sorting genetic algorithm to construct an automated closed-loop optimization system. Compared to traditional single-factor experiments or orthogonal designs, this framework can significantly shorten the process optimization cycle and save on expensive membrane materials and reagent costs.

[0018] 2. This invention constructs a Transformer deep learning model based on a self-attention mechanism and introduces SMILES sequences and molecular fingerprints as multimodal inputs. Unlike existing shallow models (such as random forests and stacked ensembles) that only utilize numerical process parameters for simple regression, this model can capture the long-range dependence and nonlinear mapping between the deep chemical structure of solvent molecules (such as functional groups and topological structures) and membrane separation performance. It still exhibits strong generalization ability and predictive stability when facing novel solvent systems or membrane materials. It overcomes the limitations of traditional numerical models by deeply analyzing structure-property relationships based on multimodal Transformer.

[0019] 3. This invention employs a mixed-integer encoding strategy, analyzing and predicting the separation mechanism of target organic solvents by membrane distillation based on data analysis from different experimental systems. Data transfer provides insights into new membrane selection and experimental conditions for optimizing membrane distillation, reducing the number of experiments. A Whitney test (P-value less than 0.001) confirms that the process combination derived from the reverse design of this invention significantly outperforms the randomized experimental group, and this is not a random, accidental result. The Kolmogorov-Smirnov test further quantifies the distributional differences, indicating that the probability of this optimal combination falling at the tail end of a random distribution is extremely low. This fully demonstrates that this invention can identify superior membrane separation schemes with extremely high confidence and efficiency. Attached Figure Description

[0020] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are 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.

[0021] Figure 1 This is a schematic diagram of the machine learning-based method flow according to an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] Example 1 This invention aims to save experimental costs and overcome the limitations of existing experience by providing a machine learning-based method for membrane material screening and process parameter optimization. It utilizes machine learning-assisted reverse guidance based on historical experimental data, employs a transfer learning strategy, uses permeate-evaporation data as the source domain to assist in training the MGMD target domain model, and combines it with the NSGA-II algorithm to achieve simultaneous reverse design of membrane materials and process parameters. This provides insights into the interaction effects between multiple factors affecting membrane performance, significantly reducing the cost of trial-and-error experiments, decreasing the number of experimental verifications, and improving the efficiency of process optimization.

[0025] like Figure 1 As shown, the technical method adopted to achieve the above objectives is: S1: Collect literature data on pervaporation and experimental data on air-gap membrane distillation, and clean and align them to construct a heterogeneous dataset; S2: Generate the SMILES sequence encoding and molecular fingerprint of the solvent, extract statistical features using the SMILES sequence, extract the deep structural features of the separation membrane and solvent using the molecular fingerprint, and standardize the process parameters to generate multimodal input data. S3: A deep learning regression model based on the Transformer architecture, which performs pre-training tasks based on pervaporation data and transfer learning fine-tuning tasks based on air-gap membrane distillation data. S4: Perform high-throughput prediction on chemical molecular libraries and output a list of optimal solvents and recommended process parameters.

[0026] In this embodiment, the heterogeneous dataset in step S1 includes: a source domain dataset and a target domain dataset; The source domain dataset contains operating parameters for the pervaporation process, including feed temperature, downstream membrane pressure, feed concentration, and membrane material type. The target domain dataset contains source domain process parameter feature variables including: top surface temperature, bottom surface temperature, feed temperature, air gap width, number of stages, feed water flow rate, temperature difference, and power. The target domain process parameters include: hot side temperature, cold side temperature, air gap width, membrane pore size, and membrane thickness. To address the inconsistency in the dimensions of operating parameters between the source and target domains, alignment is performed using parameter mapping or zero-value filling methods.

[0027] In this embodiment, step S2, which uses SMILES sequences to extract statistical features, specifically includes: using the cheminformatics tool RDkit to normalize the SMILES strings of organic solvents and membrane materials; merging the SMILES vectors output by the Transformer encoder, membrane material parameters, and process parameter numerical vectors to form a joint feature vector that is input to the fully connected layer; including the operating parameters of the air-gap membrane distillation process; performing SMILES serialization encoding and molecular fingerprint calculation on organic solvent molecules as sequence input to the Transformer model.

[0028] In this embodiment, step S2, which uses molecular fingerprinting to extract deep structural features of the separation membrane and solvent, specifically includes: setting the iteration radius and bit vector length, and using the Morgan fingerprint algorithm to generate a binary fingerprint vector of the organic solvent; the binary fingerprint vector serves as a global structural feature to characterize the local chemical environment and topological structure of the molecule.

[0029] In this embodiment, the construction of multimodal input in step S2 specifically involves: constructing a feature fusion layer, which concatenates the deep semantic vector of the SMILES sequence extracted by the deep learning model, the molecular fingerprint vector, and the normalized process parameter numerical vector to form a joint feature vector.

[0030] In this embodiment, the deep learning regression model based on the Transformer architecture in step S3 has the following architecture: Embedding layer: used to convert the SMILES numerical index sequence into a high-dimensional vector representation; Encoder: contains multiple stacked coding blocks, each composed of a multi-head self-attention mechanism and a feedforward neural network, used to capture long-range dependencies between solvent molecule atoms; Regression prediction head: composed of a multi-layer sensing mechanism, receives the joint feature vector, and outputs predicted values ​​of transmembrane flux and separation factor.

[0031] In this embodiment, the pre-training and fine-tuning process in step S3 specifically includes: Pre-training stage: using pervaporation data from the source domain to perform full-parameter supervised training on the model, using mean squared error as the loss function, so that the model learns the general physicochemical relationship between solvent molecular structure and membrane separation performance; Fine-tuning stage: freezing the weight parameters of the model embedding layer and the bottom layer of the Transformer encoder, and only updating the parameters of the top layer encoder and regression prediction head; inputting air-gap membrane distillation data from the target domain into the model for training to adapt to the characteristics of air-gap mass transfer resistance and thermal efficiency.

[0032] In this embodiment, the high-throughput screening in step S4 specifically includes: constructing an optimized variable space using a mixed integer encoding strategy, where membrane material type and solvent type are mapped as discrete integer variables, and membrane synthesis conditions and process operation parameters are defined as continuous real variables; the deep learning regression model based on the Transformer architecture is encapsulated as a fitness evaluation function of the non-dominated sorting genetic algorithm NSGA-II; performing evolutionary search, in each generation of population evolution, the decoded discrete features, including SMILES sequences and molecular fingerprints and continuous process parameters, are input into the model to predict transmembrane flux and separation factor as fitness values, and the population is iteratively updated through non-dominated sorting, crowding distance calculation, and genetic operations; finally, a Pareto optimal front solution set is output, which contains multiple optimal membrane material selection and process parameter coupling schemes.

[0033] Example 2 Regarding the same inventive concept as Embodiment 1, this embodiment also discloses a method for membrane material screening and process parameter optimization based on machine learning. The technical method employed is as follows: S1: Obtain data from literature related to pervaporation separation of organic solvents to construct a source domain dataset.

[0034] S2: Preprocess the source domain dataset.

[0035] S3: Construct a deep learning model, validate its predictive performance on the source domain dataset, and tune its hyperparameters. Build a deep learning model comprising an embedding layer, a multi-layer self-attention encoder, and a multi-layer perceptron regression head. Perform fully parameter-supervised training on the model using a large source domain dataset to learn the general physicochemical laws governing the relationship between molecular structure and separation performance.

[0036] S4: Using the target organic solvent dataset as the fitness function, coupled with NSGA-II, to solve the multi-objective Pareto front.

[0037] S5: Perform statistical verification and experimental confirmation of the solution results of the model.

[0038] Specifically, it includes the following steps: S1. Based on the membrane separation process of pervaporation, data from relevant literature on the separation of organic solvents by pervaporation were collected to obtain the membrane monomers and manufacturing conditions, as well as the membrane characteristic parameters, including membrane monomers, water contact angle, pore size, membrane thickness, zeta potential, membrane material type (e.g., PDMS, PVA, PTFE), organic solvent type, operating temperature, downstream pressure, and corresponding flux and separation factor. Using publicly available chemical databases, the chemical and physical properties of different organic solvents were obtained, including molecular descriptors, dipole moments, solubility coefficients, molecular weights, dielectric constants, solubility parameters, and feed concentrations, to establish a dataset. Based on a five-stage air-gap membrane distillation device, experimental data on the removal or enrichment of organic solvents (e.g., DMF, ethanol) were collected, recording process parameters such as hot-side temperature, cold-side temperature, feed flow rate, air-gap width, and heating power. To address the inconsistency between the source and target domain feature spaces, an asymmetric feature filling strategy was adopted to fill the source domain features with zero or mean values, constructing a heterogeneous dataset of the same dimension.

[0039] S2. The SMILES strings of solvent and membrane material monomers are processed using the RDkit toolkit. The Morgan fingerprint algorithm is employed to generate binary vectors representing the molecular topology. The SMILES strings are then serialized and encoded to generate a numerical index sequence, which serves as the input to the embedding layer of the Transformer model. Continuous variables such as temperature, flow rate, and power are Z-score standardized to eliminate dimensional differences.

[0040] S3. Develop a deep learning regression model based on the Transformer architecture, dividing the dataset into training, validation, and test sets in a 6:2:2 ratio. The training set serves as the foundation for model training, and the model's R-value on the validation set is... 2 MAE and MSE are used as evaluation metrics for model parameter tuning. The test set is used to verify the generalization prediction performance of the trained final model on external data. The underlying parameters of the model are frozen, and the top-level regression head is fine-tuned using small sample data from the target domain. Mean squared error (MSE) and L2 regularization are used as the loss function, and the AdamW optimizer is used for iterative updates until the coefficient of determination (R²) on the validation set is achieved. 2 The coefficient of performance should not be less than 0.85. Five-fold cross-validation was used to evaluate model performance, with root mean square error (RMSE) and mean absolute error (MAE) as evaluation metrics.

[0041] S4. Construct a population containing discrete and continuous variables. Discrete individuals correspond to indices of membrane materials and solvents, while continuous individuals correspond to synthesis conditions and process parameters. Use the trained Transformer model as the fitness function for NSGA-II. During evolution, decode the discrete indices of individuals, call SMILES / fingerprint features, and combine them with continuous parameters input into the model to predict transmembrane flux and separation factor. Perform non-dominated sorting, crowding distance calculation, and genetic operations, simulating binary crossover (SBX) and polynomial mutation. After the algorithm converges, output the Pareto front solution set.

[0042] S5. The performance distributions of the optimized group and the randomized experimental group were compared using the Mann-Whitney U test. The p-value was <0.001, demonstrating that the optimization results were statistically significant. The KS test further confirmed that the probability of the optimized solution falling at the tail end of the random distribution was extremely low, verifying the effectiveness of the reverse design.

[0043] The membrane separation process is a five-stage air-gap membrane distillation organic solvent enrichment system.

[0044] The feature variables of the machine learning model adopt multimodal input, specifically including: SMILES sequences and molecular fingerprints characterizing the microstructure of membrane materials and solvents, continuous variables characterizing the membrane synthesis process including crosslinking agent concentration and temperature, and macroscopic process conditions of air-gap membrane distillation including hot-side temperature, flow rate, and air-gap width.

[0045] The model described is a multivariate, multi-objective regression prediction model based on the Transformer architecture, which uses mean squared error (MSE) combined with an L2 regularization term as the loss function.

[0046] The hyperparameters in step S3 are automatically tuned using Bayesian optimization. The hyperparameters during training include the number of layers in the Transformer encoder, the number of heads in the multi-head attention mechanism, the embedding layer dimension, the positional encoding parameters, the learning rate, the batch size, and the sequence length. The optimizer used is AdamW, and iterative training is performed until the coefficient of determination (R²) on the validation set is not less than 0.85.

[0047] The cross-validation in step S3 is a K-fold strategy, where K=5.

[0048] A mixed-variable co-optimization method based on NSGA-II was adopted to simultaneously reverse-engineer the polymer membrane material selection, synthesis conditions, and membrane distillation process parameters using a trained model, thereby obtaining the optimal coupling scheme of materials and process operations, and applying it to the separation and enrichment system of organic solvents.

[0049] Compared with the prior art, the present invention has the following obvious and prominent substantive features and significant advantages: 1. The method and framework provided by this invention construct an integrated reverse design framework for prediction and optimization, significantly reducing R&D costs. Addressing the pain point of scarce experimental data in air-gap membrane distillation, this invention innovatively introduces a transfer learning strategy, utilizing mature data from the pervaporation field to assist training, effectively solving the problems of model cold start and overfitting. Furthermore, this invention encapsulates a high-precision Transformer prediction model as a fitness function, deeply coupling it with the NSGA-II algorithm. Compared to traditional single-factor experiments or simple model prediction, this framework achieves global collaborative optimization from discrete membrane material selection to continuous process parameter control, significantly shortening the R&D cycle and saving experimental costs.

[0050] 2. Breaking through the limitations of traditional numerical models, this invention utilizes a multimodal Transformer to deeply analyze structure-property relationships. Unlike existing shallow models that only employ numerical parameters, this invention is based on the Transformer architecture and incorporates SMILES sequence and molecular fingerprinting techniques. It employs theory to guide the screening and synthesis of membrane monomers, enabling rapid and accurate identification of key functional groups and uncovering complex, multi-dimensional relationships between materials and membrane properties from actual experimental data. The model leverages a self-attention mechanism to capture deep chemical structural features between solvent molecules and membrane materials at the atomic level, including functional group interactions and steric hindrance, and integrates these features with macroscopic process parameters across modes. This design allows the model to more accurately uncover complex nonlinear interactions between multidimensional variables and exhibits stronger generalization ability when facing new solvent systems.

[0051] 3. This invention utilizes the Mann-Whitney U test. Based on different experimental systems, this invention analyzes and predicts the separation mechanism of target organic solvents by membrane distillation through data analysis. It leverages data transfer to provide new insights into membrane selection and experimental conditions for optimizing membrane distillation, and introduces rigorous statistical tests to evaluate the optimization results. The Whitney test shows a p-value less than 0.001, statistically rejecting the null hypothesis. This indicates that the process combination selected through machine learning reverse design in this invention significantly outperforms the randomized experimental group, and is not a random, accidental result. Furthermore, the KS test further confirms a significant difference between the distribution of the optimized solution set and the random distribution, demonstrating that this method can identify superior membrane materials and process schemes with extremely high confidence.

[0052] Example 3 This embodiment details the specific implementation methods for synergistic optimization of a specific system (such as DMF / water mixture separation) using the above method, including the following steps: S1. Based on literature on the separation of organic solvents using pervaporation polymer membranes, monomer, polymer concentration, crosslinking agent type, crosslinking agent concentration, and experimental temperature were collected as input variables, while membrane flux and separation factor were used as evaluation indicators. A virtual chemical molecular library was constructed containing 50 commonly used hydrophobic polymer membrane materials, including PTFE, PVDF, PP and their modified derivatives, and 20 common organic solvents. Process parameter boundaries were set as follows: hot-side temperature [40, 90]℃, cold-side temperature [10, 30]℃, air gap width [1, 10]mm, and feed flow rate [50, 500]L / h. The optimization objective was to balance the two indicators and maximize them within the experimental range. For categorical features, unique thermal coding was used to create feature variables.

[0053] S2. Randomly generate the initial population (size N=100). Each individual contains: [membrane material ID, solvent ID, hot-side temperature, cold-side temperature, air gap width, feed flow rate]. For example, an individual might be: [12 (PTFE), 5 (DMF), 75.5℃, 20.0℃, 3.5mm, 200L / h].

[0054] S3. The system converts ID=12 and ID=5 into corresponding SMILES sequences (e.g., FC(F)(F)C(F)(F)) and Morgan fingerprints. These structural features are then concatenated with the standardized process parameters and input into the Transformer model trained in step S3 to obtain the model's predicted output: flux J = 15.2 kg / m³. 2 h, segregation factor α = 98.5. These two values ​​are directly used as the fitness of the individual.

[0055] S4. Based on non-dominated sorting and crowding distance, individuals in Rank 1 and sparse regions are preferentially selected. For discrete variables such as membrane material ID, integer crossover is used; for continuous variables such as temperature / flow rate, SBX crossover is used to generate new variation values. A constraint is added: if the boiling point of the generated solvent is lower than the hot-side temperature, a poor fitness score is imposed as a penalty.

[0056] S5. After 200 iterations, the system outputs the Pareto front. Different schemes are selected based on requirements: High-flux scheme: Modified PTFE membrane is selected, process conditions are 85℃, 5mm air gap, predicted flux 22.0, separation factor 80. High-purity scheme: PVDF-HFP membrane is selected, process conditions are 60℃, 2mm air gap, predicted flux 10.5, separation factor 150. Experimental verification shows that the membrane module prepared using the high-purity scheme exhibits an actual separation performance error of less than 5% compared to the predicted value in five-stage air-gap membrane distillation, fully demonstrating the effectiveness of the process improvement.

[0057] Example 4 The present invention also provides a system for screening membrane materials and optimizing process parameters based on machine learning. The system is used to implement the method described in Embodiment 1. The system includes: a data construction module, a preprocessing module, a model training module, and a screening and optimization module. The data construction module is used to collect literature data on pervaporation and experimental data on air-gap membrane distillation, and to clean and align them to build a heterogeneous dataset. The preprocessing module is used to generate the SMILES sequence code and molecular fingerprint of the solvent, and to standardize the process parameters to generate multimodal input data. The model training module is used for deep learning regression models based on the Transformer architecture to perform pre-training tasks based on pervaporation data and transfer learning fine-tuning tasks based on air-gap membrane distillation data. The screening and optimization module is used for high-throughput prediction of chemical molecular libraries, and outputs a list of optimal solvents and recommended process parameters.

[0058] The specific implementation process includes: The data construction module includes methods for obtaining penetration testing datasets, including: Using literature data in the field of pervaporation as the source domain dataset, and collecting experimental data of air gap membrane distillation as the target domain dataset, a heterogeneous dataset is constructed.

[0059] Preferably, the data points need to include a clear polymer membrane type, solvent type, operating temperature, and corresponding flux value; the target domain dataset is obtained based on actual measurements of a five-stage air-gap membrane distillation device. Preferably, to address the issue of inconsistency between the feature spaces of the source and target domains, an asymmetric feature filling strategy is adopted: for the air gap-related features missing in the source domain, they are set as fixed constants or zero vectors; for the percolation-evaporation-specific pressure features missing in the target domain, mean-based filling based on neighboring samples is used to ensure the uniformity of the input dimensions.

[0060] Preferably, the membrane separation process is a five-stage air-gap membrane distillation system.

[0061] The preprocessing module is used to preprocess the experimental dataset. It uses SMILES sequences and molecular fingerprints to extract deep structural features of the separation membrane and solvent, and constructs multimodal inputs in combination with process parameters to preprocess the dataset.

[0062] Preferably, during the molecular fingerprint extraction process, the Morgan fingerprint algorithm is used, with an iteration radius of 2 and a bit vector length of 1024 or 2048 bits. This parameter setting generates a fingerprint equivalent to ECFP4, effectively balancing feature sparsity with the expressive power of chemical substructures, and accurately capturing the local chemical environment and topological structure information of solvent molecules.

[0063] Preferably, in the preprocessing of process parameters, continuous variables such as temperature, flow rate, and power are standardized using Z-Score to convert the data into a distribution with a mean of 0 and a standard deviation of 1, so as to eliminate the influence of different physical dimensions on the convergence speed of gradient descent.

[0064] The model training module builds a deep learning neural network regression model based on the self-attention mechanism, using source domain data for pre-training and inputting target domain data to fine-tune the model.

[0065] Preferably, the encoder of the Transformer regression model comprises 2 to 6 stacked self-attention modules, and each module's multi-head attention mechanism includes 4 to 8 parallel attention heads. This deep multi-head structure can simultaneously capture local chemical bond information and global molecular polarity features in the molecular sequence.

[0066] Preferably, a differentiated learning rate strategy is adopted in the fine-tuning stage: for the bottom embedding layer and the first few encoder layers, a lower learning rate (1e-5) is set to retain the general chemical knowledge learned during pre-training; for the top fully connected regression head, a higher learning rate (1e-3) is set to quickly adapt to the unique mass transfer characteristics of air-gap membrane distillation.

[0067] Preferably, the loss function incorporates an L2 regularization term, which uses the weighted sum of the mean squared error (MSE) and the model weight norm as the total loss to prevent overfitting during fine-tuning with small samples.

[0068] The screening and optimization module is used to perform the inverse collaborative optimization task based on the non-dominated sorting genetic algorithm (NSGA-II). This module uses a hybrid encoding strategy to construct the search space, encapsulates the trained Transformer regression model into a fitness evaluation function, drives the population to evolve towards the Pareto optimal frontier, and outputs the optimal membrane material selection and process parameter coupling scheme.

[0069] Preferably, the discrete variables of the search space are mapped to a virtual chemical molecular library constructed through the PubChem database, which contains no less than 1,000 commonly used organic solvents and a variety of candidate membrane materials; the continuous variables correspond to preset membrane synthesis conditions and air-gap membrane distillation process parameter ranges.

[0070] Preferably, a physical constraint filter is integrated into the optimization process to automatically detect and eliminate infeasible solutions that violate thermodynamic constraints (boiling point below feed temperature) or material compatibility constraints (solvent corrosion of polymer film based on Hansen solubility parameters) during the individual evaluation stage, ensuring the engineering validity of the evolution results.

[0071] Preferably, the module outputs a Pareto optimal frontier solution set and supports a weighted scoring mechanism based on decision preferences, that is, it automatically recommends a unique best compromise solution from the frontier solution set according to the weight coefficients of the transmembrane flux and the logarithmic separation factor set by the user.

[0072] A hybrid variable co-optimization method based on NSGA-II was adopted to achieve simultaneous reverse design of polymer membrane material selection and air gap membrane distillation process parameters, thereby obtaining the optimal multi-scale coupling scheme, which was then directly applied to the membrane distillation separation and enrichment system of organic solvents.

[0073] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for machine learning based membrane material screening and process parameter optimization, characterized in that, The method includes: S1: Collect literature data on pervaporation and experimental data on air-gap membrane distillation, and clean and align them to construct a heterogeneous dataset; S2: Generate the SMILES sequence encoding and molecular fingerprint of the solvent, extract statistical features using the SMILES sequence, extract the deep structural features of the separation membrane and solvent using the molecular fingerprint, and standardize the process parameters to generate multimodal input data. S3: A deep learning regression model based on the Transformer architecture, which performs pre-training tasks based on pervaporation data and transfer learning fine-tuning tasks based on air-gap membrane distillation data. S4: Perform high-throughput prediction on chemical molecular libraries and output a list of optimal solvents and recommended process parameters.

2. The method of claim 1, wherein, The heterogeneous dataset in step S1 includes: the source domain dataset and the target domain dataset; The source domain dataset contains operating parameters for the pervaporation process, including feed temperature, downstream membrane pressure, feed concentration, and membrane material type. The target domain dataset contains source domain process parameter feature variables including: top surface temperature, bottom surface temperature, feed temperature, air gap width, number of stages, feed water flow rate, temperature difference, and power. The target domain process parameters include: hot side temperature, cold side temperature, air gap width, membrane pore size, and membrane thickness. To address the inconsistency in the dimensions of operating parameters between the source and target domains, alignment is performed using parameter mapping or zero-value filling methods.

3. The method of claim 1, wherein, Step S2, which utilizes SMILES sequences to extract statistical features, specifically includes: using the cheminformatics tool RDkit to normalize the SMILES strings of organic solvents and membrane materials; merging the SMILES vectors output by the Transformer encoder, membrane material parameters, and process parameter numerical vectors to form a joint feature vector that is input to the fully connected layer; including the operating parameters of the air-gap membrane distillation process; performing SMILES serialization encoding and molecular fingerprint calculation on organic solvent molecules; and using these as the sequence input to the Transformer model.

4. The method of claim 1, wherein, Step S2, which utilizes molecular fingerprinting to extract deep structural features of the separation membrane and solvent, specifically includes: setting the iteration radius and bit vector length, and using the Morgan fingerprinting algorithm to generate a binary fingerprint vector of the organic solvent; the binary fingerprint vector serves as a global structural feature to characterize the local chemical environment and topological structure of the molecule.

5. The method of claim 1, wherein, In step S2, the construction of multimodal input specifically involves: constructing a feature fusion layer, which concatenates the deep semantic vector of the SMILES sequence extracted by the deep learning model, the molecular fingerprint vector, and the normalized process parameter numerical vector to form a joint feature vector.

6. The method of claim 5, wherein, The deep learning regression model based on the Transformer architecture in step S3 has the following architecture: Embedding layer: used to convert the SMILES numerical index sequence into a high-dimensional vector representation; Encoder: contains multiple stacked coding blocks, each composed of a multi-head self-attention mechanism and a feedforward neural network, used to capture long-range dependencies between solvent molecule atoms; Regression prediction head: composed of multiple sensing mechanisms, receives the joint feature vector, and outputs predicted values ​​of transmembrane flux and separation factor.

7. The method of claim 1, wherein, The pre-training and fine-tuning process in step S3 specifically includes: Pre-training stage: using pervaporation data from the source domain to perform full-parameter supervised training on the model, using mean squared error as the loss function, so that the model learns the general physicochemical relationship between solvent molecular structure and membrane separation performance; Fine-tuning stage: freezing the weight parameters of the model embedding layer and the bottom layer of the Transformer encoder, and only updating the parameters of the top layer encoder and regression prediction head; inputting air-gap membrane distillation data from the target domain into the model for training to adapt to the characteristics of air-gap mass transfer resistance and thermal efficiency.

8. The method of claim 1, wherein, Step S4, the high-throughput screening, specifically includes: constructing an optimized variable space using a mixed integer encoding strategy, where membrane material type and solvent type are mapped as discrete integer variables, and membrane synthesis conditions and process operation parameters are defined as continuous real variables; the deep learning regression model based on the Transformer architecture is encapsulated as a fitness evaluation function of the non-dominated sorting genetic algorithm NSGA-II; performing evolutionary search, in each generation of population evolution, the decoded discrete features, including SMILES sequences and molecular fingerprints, and continuous process parameters are input into the model to predict transmembrane flux and separation factor as fitness values, and iteratively updating the population through non-dominated sorting, crowding distance calculation, and genetic operations; finally, outputting a Pareto optimal front solution set, which contains multiple optimal membrane material selection and process parameter coupling schemes.

9. A system for membrane material screening and process parameter optimization based on machine learning, said system being used to implement the method described in any one of claims 1-8, characterized in that, The system includes: a data construction module, a preprocessing module, a model training module, and a screening and optimization module; The data construction module is used to collect pervaporation literature data and air-gap membrane distillation experimental data, and to clean and align them to construct a heterogeneous dataset. The preprocessing module is used to generate the SMILES sequence code and molecular fingerprint of the solvent, and to standardize the process parameters to generate multimodal input data. The model training module is used to perform pre-training tasks based on pervaporation data and transfer learning fine-tuning tasks based on air-gap membrane distillation data for deep learning regression models based on the Transformer architecture. The screening and optimization module is used to perform high-throughput prediction of chemical molecular libraries and output a list of optimal solvents and recommended process parameters.

Images