A hybrid surfactant emulsion stability prediction method based on active learning

Abstract

This invention proposes a method for predicting the stability of mixed surfactant emulsions based on active learning, comprising the following steps: Step 1: Constructing a formulation space and preparing an initial dataset through stratified sampling; Step 2: Completing emulsion stability detection, data labeling, and feature dimensionality reduction; Step 3: Selecting a surrogate model and sampling strategy, and iteratively expanding the dataset through active learning; Step 4: Dividing the dataset and building multi-class machine learning models, and completing hyperparameter optimization training; Step 5: Evaluating model performance and analyzing the mechanism of action, and using the optimal model for prediction. This invention can accurately predict the stability level and emulsion layer height of emulsions under different formulations and temperature conditions while reducing experimental workload, and simultaneously analyzes the mechanism of action of various factors on stability, providing technical support for the rapid screening and optimization of industrial emulsion formulations.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of colloidal interface chemistry and machine learning, and in particular to a method for predicting the stability of mixed surfactant emulsions based on active learning. Background Technology

[0002] Emulsions are thermodynamically unstable dispersions formed by two immiscible liquids, and they have a wide range of applications in food processing, cosmetics, oil extraction, and pharmaceuticals. In actual production, the industry commonly uses mixed surfactant formulations to improve emulsion stability. This involves leveraging the synergistic effects of different surfactants to construct a dense interfacial film, thereby inhibiting droplet coalescence, stratification, and Ostwald ripening.

[0003] Currently, the formulation development of mixed surfactant emulsions largely relies on traditional experimental methods, commonly including single-factor experiments and orthogonal experiments. The emulsion system involves numerous influencing factors such as surfactant type, mixing ratio, concentration, oil phase type, oil-water ratio, and ambient temperature; the formulation space expands exponentially with the increase of variables. Furthermore, emulsion stability testing requires long-term static observation, resulting in a lengthy experimental cycle, relatively high trial-and-error costs, and high manpower consumption. Traditional experimental methods are difficult to use for rapid screening of large-scale formulations.

[0004] Machine learning technology is increasingly being applied to the field of emulsion performance prediction. This technology can effectively uncover the nonlinear correlation between formulation parameters and macroscopic performance; however, model effectiveness is highly dependent on large-scale, high-quality labeled data. Labeling mixed surfactant emulsion samples is challenging, and data acquisition efficiency is low. Conventional machine learning models often suffer from weak generalization ability and insufficient prediction accuracy. Active learning can autonomously select high-value unknown samples based on the model and supplement labeled data through iteration, rapidly improving the overall model performance under limited experimental resources. This technology has already been applied in drug development and materials screening, but a systematic prediction and optimization scheme has not yet been formed for mixed surfactant emulsion formulation systems.

[0005] To address the shortcomings of existing technologies, such as low experimental efficiency, high data acquisition costs, and limited model prediction accuracy, this invention proposes a method for predicting the stability of mixed surfactant emulsions based on active learning. By combining active learning data augmentation, multi-dimensional feature engineering, multi-model fusion, and model interpretation techniques, this method achieves high-precision prediction of emulsion stability and intelligent optimization of formulations. Summary of the Invention

[0006] This invention proposes a method for predicting the stability of mixed surfactant emulsions based on active learning. The aim is to overcome the shortcomings of traditional emulsion formulation development, such as long experimental cycles and high costs, as well as the high dependence of traditional machine learning on labeled data. The prediction method of this invention can accurately predict the stability level and emulsion layer height of emulsions under different formulations and temperature conditions while reducing the amount of experimental work. At the same time, it analyzes the mechanism of various factors on stability, providing technical support for the rapid screening and optimization of industrial emulsion formulations.

[0007] The present invention adopts the following technical solution.

[0008] A method for predicting the stability of mixed surfactant emulsions based on active learning is proposed. The overall process consists of five core steps: formulation space construction and initial dataset preparation, data testing and feature engineering, active learning iterative expansion of the dataset, machine learning modeling and hyperparameter optimization, model evaluation, and mechanism analysis. The prediction method includes the following steps;

[0009] Step 1: Construct the recipe space and prepare the initial dataset through stratified sampling;

[0010] Step 2: Complete emulsion stability testing, data labeling, and feature dimensionality reduction.

[0011] Step 3: Select the proxy model and sampling strategy, and expand the dataset iteratively through active learning;

[0012] Step 4: Divide the dataset and build multi-class machine learning models, and complete hyperparameter optimization training;

[0013] Step 5: Evaluate model performance and analyze its mechanism of action, and use the optimal model to make predictions.

[0014] Step 1 specifically includes:

[0015] Step 101: Select surfactant combination, concentration, compounding ratio, oil phase type, oil-water ratio, and temperature as variables, and set multiple gradient levels to construct a global unlabeled formulation pool;

[0016] Step 102: Stratify sampling according to component type and ratio to construct initial training set, validation set and test set respectively.

[0017] Step 101 completes the construction of the full-dimensional formulation space and the preparation of the initial dataset. Specifically, after determining the six core variables, namely surfactant combination, total surfactant concentration, surfactant mass ratio, oil phase type, oil-water volume ratio, and ambient temperature, multiple gradient variable levels are set so that the variables are combined to generate 1680 full formulations that constitute the unlabeled formulation pool.

[0018] Step 2 specifically includes:

[0019] Step 201: Prepare the emulsion under uniform process conditions and keep it at a constant temperature for static testing. Use the emulsion layer height CH as a quantitative evaluation index of stability.

[0020] Step 202: Divide the continuous emulsion layer height values ​​into multiple stability levels and complete the classification label construction.

[0021] Step 203: Extract molecular descriptors and process features, and perform feature dimensionality reduction through data cleaning, low variance filtering, and correlation analysis;

[0022] The CH value of the emulsion is calculated as follows:

[0023]

[0024] This refers to the height of the emulsion layer that separates after standing. This represents the total height of the emulsion system.

[0025] In step 201, the total volume of the emulsion is uniformly set to 5 mL. The mixed surfactant is completely dissolved in ultrapure water. After adding the corresponding oil phase, the emulsion is sheared at 10,000 rpm for 3 min using a high-speed homogenizer to complete the emulsion preparation.

[0026] In this step, Span 80, Tween 80, Triton X-100, dodecyl dimethyl betaine, hexadecyl trimethyl ammonium bromide, sodium dodecyl sulfate, dodecane, hexadecane, and ultrapure water can be used as experimental raw materials. All raw materials are analytical grade reagents.

[0027] As required, all variables and corresponding levels were set. Surfactant combinations could include NP10-BS12, NP10-CTAB, TX100-SDS, and TWEEN80-SPAN80. Total surfactant concentrations were set to 0.01%, 0.05%, 0.1%, 0.5%, and 1%. Surfactant mass ratios were set to 0:1, 1:9, 3:7, 5:5, 7:3, 9:1, and 1:0. Dodecane and hexadecane were used as the oil phase, with oil-water volume ratios set to 1:3, 1:1, and 3:1. Temperatures were set to 20 ℃ and 60 ℃. A total of 1680 formulations were obtained through all variable combinations, forming an unlabeled formulation pool.

[0028] Stratified sampling was performed according to surfactant type, formulation ratio, and oil phase. Within each stratum, 1-2 formulations were extracted to construct initial training sets, validation sets, and independent test sets, ensuring comprehensive sample coverage of the entire formulation space. With a fixed total emulsion volume of 5 mL, the surfactant was fully dissolved and added to the corresponding oil phase. The emulsion was prepared by shearing at 10000 rpm for 3 min. The emulsions were then incubated at 20 ℃ and 60 ℃ for 3 h, respectively, and the CH values ​​were calculated.

[0029] In step 202, the continuous CH values ​​are divided into 5 discrete stability levels to adapt to the classification task. CH > 0.8 corresponds to level 4 extremely stable, 0.6 < CH ≤ 0.8 corresponds to level 3 relatively stable, 0.4 < CH ≤ 0.6 corresponds to level 2 critically stable, 0.2 < CH ≤ 0.4 corresponds to level 1 relatively unstable, and CH ≤ 0.2 corresponds to level 0 extremely unstable.

[0030] In step 203, ChemDraw is used to draw various molecular structures and convert them into SMILES format. Then, RDKit is used to extract the geometric, electronic, and topological molecular descriptors of surfactants and oil phases, which are combined with the formulation process parameters to form an initial high-dimensional feature set.

[0031] Data cleaning and feature selection were completed sequentially. Features with a missing rate of over 50% were removed. The remaining features with missing values ​​were padded with zeros. Invalid features with minimal numerical fluctuations were removed by low variance filtering. Redundant features with correlation coefficients greater than or equal to 0.9 were removed by Spearman correlation analysis. Finally, the feature dimensions were reduced to 14, effectively eliminating the multicollinearity problem.

[0032] Step 3 specifically includes:

[0033] Step 301: Train multiple tree ensemble classification models and select a surrogate model by combining evaluation metrics; compare multiple sampling strategies to determine the optimal query method.

[0034] Step 302: Use the surrogate model to screen high-value unknown formulas and conduct experimental annotations, iteratively updating the model and training dataset.

[0035] In step 301, an active learning iterative framework is built to continuously expand the labeled dataset. Based on the initial training set, four tree ensemble classification models—RF, GBDT, XGBoost, and LightGBM—are trained respectively. The model performance is judged by combining the accuracy and F1 score on the validation and test sets, and the model with the best overall performance is selected as the active learning proxy model. The iterative effects of four sample query strategies—uncertainty sampling, entropy sampling, marginal sampling, and random sampling—are compared, and uncertainty sampling, which has the highest information gain and the fastest model convergence speed, is selected as the core query strategy.

[0036] In step 302, during each iteration, the surrogate model is used to predict all samples in the unlabeled formula pool. Combined with the selected sampling strategy, 10 groups of unknown formulas with the highest cognitive uncertainty and the greatest information gain are selected. Experiments are conducted on the selected samples to obtain the corresponding CH values ​​and stability level labels. The newly labeled samples are then incorporated into the original training set and the surrogate model is updated. The entire process is completed in 10 iterations, and the model accuracy, F1 score, AUC values ​​of each category, and overall prediction uncertainty of the unlabeled sample pool are monitored throughout the process.

[0037] Step 4 specifically includes:

[0038] Step 401: Merge all labeled samples to form the total training set, divide it into independent test sets, and use leave-one-out cross-validation to conduct internal verification;

[0039] Specifically, the process involves: dividing the dataset and constructing multiple types of machine learning models, and optimizing hyperparameters. The initial training set, newly added samples from active learning iterations, and the validation set are integrated into a total training set, while an independent test set is retained for the final unbiased evaluation of the model. Leave-one-out cross-validation is used to complete the internal validation of the model.

[0040] Step 402: Simultaneously build classification and regression models, covering linear regression and tree ensemble algorithms, and optimize model hyperparameters through grid search.

[0041] Specifically, two major systems were built: a classification model and a regression model. The classification model was used to determine the stability level of the emulsion, and the regression model was used to predict the height of the continuous emulsion layer (CH). The model system included GBDT, RF, XGBoost, and LightGBM.

[0042] Step 5 involves model performance evaluation and mechanism analysis, specifically including:

[0043] Step 501: Use accuracy, F1 score, and R... 2 MSE is used as the evaluation index to complete the model evaluation, and the optimal model is determined based on the evaluation results. For classification models, accuracy, F1 score, AUC value, and confusion matrix are used to evaluate the overall performance, while the coefficient of determination R is used for regression models. 2 Mean squared error (MSE) is used as an evaluation index;

[0044] Step 502: Analyze the influence of each feature on emulsion stability using feature importance and SHAP algorithm, and conduct experimental verification using unknown formulations;

[0045] Step 503: Based on the optimal model, complete the global formulation stability prediction, and correct the model bias by combining the experimental verification results;

[0046] Wherein, the determination coefficient R 2 The mean square error (MSE) is calculated as follows:

[0047]

[0048]

[0049] Indicates the actual value. This represents the model's predicted value. Indicates the number of samples.

[0050] The prediction method is used for oil-water two-phase mixed surfactant emulsion systems to predict the emulsion stability level and emulsion layer height under different temperature conditions, and is applied to emulsion formulation design in the food, cosmetics, and petrochemical industries.

[0051] In the prediction process, the optimal classification model and the optimal regression model are first determined by combining the evaluation results. Feature importance analysis combined with the SHAP interpretation algorithm is used to quantify the nonlinear influence of formulation parameters and molecular descriptors on emulsion stability, clarify the mechanism of action of water-oil ratio, surfactant concentration, compounding ratio, temperature and molecular microstructure. Then, representative candidate formulations are selected from the unlabeled formulation pool, and physical experiments are carried out under a high temperature of 60 ℃. The measured data are compared with the model prediction results to verify the model's generalization ability.

[0052] To address the blind spots in model cognition discovered during the experiment, corresponding samples were added to complete data iteration and correct model biases. Based on the optimal model, stability prediction was carried out on 1680 formulas across the entire domain, and high-stability formulas were screened in batches to achieve intelligent screening and optimization of emulsion formulas.

[0053] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0054] 1. This invention introduces an active learning strategy, which relies on the model to autonomously select high-value experimental samples. It can complete the efficient exploration of the entire formula space in just 10 iterations. Compared with the traditional full-scale experiment and random sampling modeling method, it significantly reduces the number of experimental samples and the static observation time, effectively reducing manpower and material costs and shortening the overall R&D cycle.

[0055] 2. This invention combines multi-dimensional feature engineering, multi-model fusion, and hyperparameter optimization techniques, resulting in excellent model prediction accuracy and generalization ability. The optimal XGBoost classification model achieves an accuracy of 0.7857 and an F1 score of 0.7711 on the test set, while the optimal RF regression model achieves a high accuracy on the test set. 2=0.8795, MSE=0.0103. In the high-temperature physical verification stage, the classification model's prediction accuracy can reach 81.25%, which can adapt to the stability prediction needs under different working conditions such as normal temperature and high temperature.

[0056] 3. This invention has good interpretability. Through feature importance and SHAP analysis, it can quantify the regulation of emulsion stability by macroscopic formulation parameters and microscopic molecular structure. While completing performance prediction, it can also reveal the physicochemical mechanism of the synergistic effect of mixed surfactants and the effect of temperature on emulsion stratification, providing data support for basic research in interface chemistry.

[0057] 4. This invention has a wide range of applications and can be adapted to various binary mixed surfactant systems, two types of oil phases, multiple oil-water ratios and temperature conditions. It can be directly applied to food, cosmetics, petroleum, pharmaceutical and other fields to complete high-throughput screening and iterative optimization of emulsion formulations. It has good engineering practicality and industrial application prospects. Attached Figure Description

[0058] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0059] Appendix Figure 1 This is a schematic diagram of the process for predicting the stability of mixed surfactant emulsions based on active learning in an embodiment of the present invention.

[0060] Appendix Figure 2 This is a schematic diagram illustrating the characteristic importance analysis for predicting the stability of mixed surfactant emulsions in an embodiment of the present invention.

[0061] Appendix Figure 3 This is a schematic diagram of the active learning iteration process in an embodiment of the present invention;

[0062] Appendix Figure 4 This is a schematic diagram of the model accuracy and F1 score changes during the iteration process in an embodiment of the present invention;

[0063] Appendix Figure 5 This is a bar chart illustrating the importance of the optimal model features in an embodiment of the present invention;

[0064] Appendix Figure 6 This is an embodiment of the present invention, showing the SHAP summary diagram and the SHAP dependency diagram of each factor;

[0065] Appendix Figure 7 This is a schematic diagram comparing the model prediction value and the experimental measured value in an embodiment of the present invention. Detailed Implementation

[0066] As shown in the figure, a method for predicting the stability of mixed surfactant emulsions based on active learning is presented. The overall process consists of five core steps: formulation space construction and initial dataset preparation, data testing and feature engineering, active learning iterative expansion of the dataset, machine learning modeling and hyperparameter optimization, model evaluation, and mechanism analysis. The prediction method includes the following steps;

[0067] Step 1: Construct the recipe space and prepare the initial dataset through stratified sampling;

[0068] Step 2: Complete emulsion stability testing, data labeling, and feature dimensionality reduction.

[0069] Step 3: Select the proxy model and sampling strategy, and expand the dataset iteratively through active learning;

[0070] Step 4: Divide the dataset and build multi-class machine learning models, and complete hyperparameter optimization training;

[0071] Step 5: Evaluate model performance and analyze its mechanism of action, and use the optimal model to make predictions.

[0072] Step 1 specifically includes:

[0073] Step 101: Select surfactant combination, concentration, compounding ratio, oil phase type, oil-water ratio, and temperature as variables, and set multiple gradient levels to construct a global unlabeled formulation pool;

[0074] Step 102: Stratify sampling according to component type and ratio to construct initial training set, validation set and test set respectively.

[0075] Step 101 completes the construction of the full-dimensional formulation space and the preparation of the initial dataset. Specifically, after determining the six core variables, namely surfactant combination, total surfactant concentration, surfactant mass ratio, oil phase type, oil-water volume ratio, and ambient temperature, multiple gradient variable levels are set so that the variables are combined to generate 1680 full formulations that constitute the unlabeled formulation pool.

[0076] Step 2 specifically includes:

[0077] Step 201: Prepare the emulsion under uniform process conditions and keep it at a constant temperature for static testing. Use the emulsion layer height CH as a quantitative evaluation index of stability.

[0078] Step 202: Divide the continuous emulsion layer height values ​​into multiple stability levels and complete the classification label construction.

[0079] Step 203: Extract molecular descriptors and process features, and perform feature dimensionality reduction through data cleaning, low variance filtering, and correlation analysis;

[0080] The CH value of the emulsion is calculated as follows:

[0081]

[0082] This refers to the height of the emulsion layer that separates after standing. This represents the total height of the emulsion system.

[0083] In step 201, the total volume of the emulsion is uniformly set to 5 mL. The mixed surfactant is completely dissolved in ultrapure water. After adding the corresponding oil phase, the emulsion is sheared at 10,000 rpm for 3 min using a high-speed homogenizer to complete the emulsion preparation.

[0084] In this step, Span 80, Tween 80, Triton X-100, dodecyl dimethyl betaine, hexadecyl trimethyl ammonium bromide, sodium dodecyl sulfate, dodecane, hexadecane, and ultrapure water can be used as experimental raw materials. All raw materials are analytical grade reagents.

[0085] As required, all variables and corresponding levels were set. Surfactant combinations could include NP10-BS12, NP10-CTAB, TX100-SDS, and TWEEN80-SPAN80. Total surfactant concentrations were set to 0.01%, 0.05%, 0.1%, 0.5%, and 1%. Surfactant mass ratios were set to 0:1, 1:9, 3:7, 5:5, 7:3, 9:1, and 1:0. Dodecane and hexadecane were used as the oil phase, with oil-water volume ratios set to 1:3, 1:1, and 3:1. Temperatures were set to 20 ℃ and 60 ℃. A total of 1680 formulations were obtained through all variable combinations, forming an unlabeled formulation pool.

[0086] Stratified sampling was performed according to surfactant type, formulation ratio, and oil phase. Within each stratum, 1-2 formulations were extracted to construct initial training sets, validation sets, and independent test sets, ensuring comprehensive sample coverage of the entire formulation space. With a fixed total emulsion volume of 5 mL, the surfactant was fully dissolved and added to the corresponding oil phase. The emulsion was prepared by shearing at 10000 rpm for 3 min. The emulsions were then incubated at 20 ℃ and 60 ℃ for 3 h, respectively, and the CH values ​​were calculated.

[0087] In step 202, the continuous CH values ​​are divided into 5 discrete stability levels to adapt to the classification task. CH > 0.8 corresponds to level 4 extremely stable, 0.6 < CH ≤ 0.8 corresponds to level 3 relatively stable, 0.4 < CH ≤ 0.6 corresponds to level 2 critically stable, 0.2 < CH ≤ 0.4 corresponds to level 1 relatively unstable, and CH ≤ 0.2 corresponds to level 0 extremely unstable.

[0088] In step 203, ChemDraw is used to draw various molecular structures and convert them into SMILES format. Then, RDKit is used to extract the geometric, electronic, and topological molecular descriptors of surfactants and oil phases, which are combined with the formulation process parameters to form an initial high-dimensional feature set.

[0089] Data cleaning and feature selection were completed sequentially. Features with a missing rate of over 50% were removed. The remaining features with missing values ​​were padded with zeros. Invalid features with minimal numerical fluctuations were removed by low variance filtering. Redundant features with correlation coefficients greater than or equal to 0.9 were removed by Spearman correlation analysis. Finally, the feature dimensions were reduced to 14, effectively eliminating the multicollinearity problem.

[0090] Step 3 specifically includes:

[0091] Step 301: Train multiple tree ensemble classification models and select a surrogate model by combining evaluation metrics; compare multiple sampling strategies to determine the optimal query method.

[0092] Step 302: Use the surrogate model to screen high-value unknown formulas and conduct experimental annotations, iteratively updating the model and training dataset.

[0093] In step 301, an active learning iterative framework is built to continuously expand the labeled dataset. Based on the initial training set, four tree ensemble classification models—RF, GBDT, XGBoost, and LightGBM—are trained respectively. The model performance is judged by combining the accuracy and F1 score on the validation and test sets, and the model with the best overall performance is selected as the active learning proxy model. The iterative effects of four sample query strategies—uncertainty sampling, entropy sampling, marginal sampling, and random sampling—are compared, and uncertainty sampling, which has the highest information gain and the fastest model convergence speed, is selected as the core query strategy.

[0094] In step 302, during each iteration, the surrogate model is used to predict all samples in the unlabeled formula pool. Combined with the selected sampling strategy, 10 groups of unknown formulas with the highest cognitive uncertainty and the greatest information gain are selected. Experiments are conducted on the selected samples to obtain the corresponding CH values ​​and stability level labels. The newly labeled samples are then incorporated into the original training set and the surrogate model is updated. The entire process is completed in 10 iterations, and the model accuracy, F1 score, AUC values ​​of each category, and overall prediction uncertainty of the unlabeled sample pool are monitored throughout the process.

[0095] Step 4 specifically includes:

[0096] Step 401: Merge all labeled samples to form the total training set, divide it into independent test sets, and use leave-one-out cross-validation to conduct internal verification;

[0097] Specifically, the process involves: dividing the dataset and constructing multiple types of machine learning models, and optimizing hyperparameters. The initial training set, newly added samples from active learning iterations, and the validation set are integrated into a total training set, while an independent test set is retained for the final unbiased evaluation of the model. Leave-one-out cross-validation is used to complete the internal validation of the model.

[0098] Step 402: Simultaneously build classification and regression models, covering linear regression and tree ensemble algorithms, and optimize model hyperparameters through grid search.

[0099] Specifically, two major systems were built: a classification model and a regression model. The classification model was used to determine the stability level of the emulsion, and the regression model was used to predict the height of the continuous emulsion layer (CH). The model system included GBDT, RF, XGBoost, and LightGBM.

[0100] Step 5 involves model performance evaluation and mechanism analysis, specifically including:

[0101] Step 501: Use accuracy, F1 score, and R... 2 MSE is used as the evaluation index to complete the model evaluation, and the optimal model is determined based on the evaluation results. For classification models, accuracy, F1 score, AUC value, and confusion matrix are used to evaluate the overall performance, while the coefficient of determination R is used for regression models. 2 Mean squared error (MSE) is used as an evaluation index;

[0102] Step 502: Analyze the influence of each feature on emulsion stability using feature importance and SHAP algorithm, and conduct experimental verification using unknown formulations;

[0103] Step 503: Based on the optimal model, complete the global formulation stability prediction, and correct the model bias by combining the experimental verification results;

[0104] Wherein, the determination coefficient R 2 The mean square error (MSE) is calculated as follows:

[0105]

[0106]

[0107] Indicates the actual value. This represents the model's predicted value. Indicates the number of samples.

[0108] The prediction method is used for oil-water two-phase mixed surfactant emulsion systems to predict the emulsion stability level and emulsion layer height under different temperature conditions, and is applied to emulsion formulation design in the food, cosmetics, and petrochemical industries.

[0109] In the prediction process, the optimal classification model and the optimal regression model are first determined by combining the evaluation results. Feature importance analysis combined with the SHAP interpretation algorithm is used to quantify the nonlinear influence of formulation parameters and molecular descriptors on emulsion stability, clarify the mechanism of action of water-oil ratio, surfactant concentration, compounding ratio, temperature and molecular microstructure. Then, representative candidate formulations are selected from the unlabeled formulation pool, and physical experiments are carried out under a high temperature of 60 ℃. The measured data are compared with the model prediction results to verify the model's generalization ability.

[0110] To address the blind spots in model cognition discovered during the experiment, corresponding samples were added to complete data iteration and correct model biases. Based on the optimal model, stability prediction was carried out on 1680 formulas across the entire domain, and high-stability formulas were screened in batches to achieve intelligent screening and optimization of emulsion formulas.

[0111] Example 1:

[0112] This paper proposes an active learning-based method for predicting the stability of mixed surfactant emulsions, belonging to the interdisciplinary field of colloidal interface chemistry and machine learning. Addressing the problems of numerous formulation variables in mixed surfactant emulsions, low efficiency of traditional experimental screening, and a lack of standardized machine learning labeled data, this invention selects six variables—surfactant combination, concentration, compounding ratio, oil phase type, oil-water ratio, and temperature—to construct a full-domain formulation space containing 1680 formulations. A stratified sampling method is used to divide the samples, simultaneously establishing an initial training set, an independent validation set, and an independent test set. Using emulsion layer height as a quantitative indicator of stability, continuous data is divided into five stability levels for label discretization. Feature cleaning and dimensionality reduction are performed in conjunction with molecular descriptors and process parameters. This invention uses a tree ensemble algorithm to screen active learning proxy models and an uncertainty sampling strategy. After 10 rounds of iteration to expand the labeled samples, the formulation space coverage increased from 94.1% to 99.2%. Two machine learning models, classification and regression, are constructed in parallel. Leave-one-out cross-validation and grid search are used to complete training and hyperparameter optimization. Accuracy, F1 score, and coefficient of determination R are used as the benchmarks. 2 The mean squared error (MSE) was used to evaluate the model, and the SHAP method and feature importance analysis were combined to reveal the regulatory mechanisms of various factors on emulsion stability. Test results show that the XGBoost classification model achieved an accuracy of 0.7857 and an F1 score of 0.7711 on the test set, while the RF regression model achieved a high accuracy of R0.05 on the test set. 2 The accuracy of the classification model prediction reached 81.25% in high-temperature physical verification, with a value of 0.8795 and an MSE of 0.0103. This invention effectively reduces the number of experimental trials and errors, accurately predicts the stability level and emulsion layer height under different working conditions, and enables intelligent screening and optimization of formulations across the entire range. It is applicable to food, cosmetics, petrochemical, pharmaceutical and other fields, and has high theoretical research value and engineering application prospects.

[0113] This example provides a method for predicting and optimizing the stability of mixed surfactant emulsions, which includes the following steps:

[0114] In this embodiment, Span 80, Tween 80, Triton X-100, dodecyl dimethyl betaine, hexadecyl trimethylammonium bromide, sodium dodecyl sulfate, dodecane, hexadecane, and ultrapure water were used as experimental materials. All materials were analytical grade reagents.

[0115] As required, all variables and corresponding levels were set. Surfactant combinations included NP10-BS12, NP10-CTAB, TX100-SDS, and TWEEN80-SPAN80. Total surfactant concentrations were set at 0.01%, 0.05%, 0.1%, 0.5%, and 1%, with surfactant mass ratios of 0:1, 1:9, 3:7, 5:5, 7:3, 9:1, and 1:0. Dodecane and hexadecane were used as the oil phase, with oil-water volume ratios of 1:3, 1:1, and 3:1. Temperatures were set at 20 ℃ and 60 ℃. A total of 1680 formulations were obtained through all variable combinations, forming an unlabeled formulation pool. Stratified sampling was performed according to surfactant type, mixing ratio, and oil phase. One to two formulations were extracted from each stratum to construct initial training sets, validation sets, and independent test sets, ensuring comprehensive coverage of the entire formulation space. With a fixed total emulsion volume of 5 mL, the surfactants were fully dissolved and added to the corresponding oil phase. The emulsion was prepared by shearing at 10000 rpm for 3 min. Samples were allowed to stand at 20 ℃ and 60 ℃ for 3 h respectively, and CH values ​​were calculated. Based on the CH values, five stability levels (0-4) were established, and the labels were discretized. Molecular descriptors for all components were extracted using RDKit, and an initial feature set was constructed using the formulation parameters. Missing data cleaning, low-variance filtering, and correlation redundancy removal were performed, ultimately retaining 14 effective features.

[0116] Initial model selection was completed based on the initial training set. Four classification models—RF, GBDT, XGBoost, and LGBM—were trained respectively. After comprehensively comparing various indicators, XGBoost was selected as the surrogate model. The iterative effects of the four query strategies were compared, and uncertainty sampling was determined to be the optimal sampling method. The entire process consisted of 10 rounds of active learning iterations. In each round, 10 sets of high-uncertainty recipes were selected for experimental labeling, and new samples were incorporated into the training set to update the model. After the iterations, the coverage of the recipe space by the training set increased from 94.1% to 99.2%, and the model's accuracy and F1 score on the validation set were significantly improved.

[0117] The initial training set, sampled data, and initial validation set were integrated into a training set, which was then tested on the initial test set. Leave-one-out cross-validation was used for internal verification. Four types of models—GBDT, RF, XGBoost, and LGBM—were built. Hyperparameters such as learning rate, number of trees, and maximum tree depth were optimized using grid search, completing the training of all models.

[0118] Model evaluation results show that XGBoost is the optimal classification model, with a test set accuracy of 0.7857 and an F1 score of 0.7711. RF is the optimal regression model, with R² on the test set. 2 =0.8795, MSE=0.0103. Combining SHAP and feature importance analysis, it can be seen that the proportion of aqueous phase and the total concentration of surfactant are the core factors affecting the stability of emulsion. A moderate compounding ratio can give full play to the synergistic effect of binary surfactants. Low temperature environment is more conducive to the stability of emulsion, while high temperature will accelerate droplet coalescence. Microscopic descriptors such as molecular charge and topological structure will directly affect the interfacial film strength.

[0119] Table 1 Evaluation results of classification and regression models

[0120]

[0121] Sixteen typical unknown formulations were selected and validation experiments were conducted at 60 ℃. The classification model achieved a prediction accuracy of 81.25%. Experimental samples were added to the feature intervals corresponding to the model's prediction bias to correct blind spots in the model's understanding. Based on the optimal model, stability predictions were performed on 1680 formulations across the entire domain, and multiple high-stability emulsion formulations were obtained through batch screening.

[0122] The experimental measurement results of the 16 selected sample sets are shown in Table 2.

[0123] Table 2 shows the experimental measurement results of the 16 selected sample sets.

[0124]

[0125] a This sequence number represents the sorting of samples within a 1680-formulation space;

[0126] b Type 1-4 represent NP10-BS12, NP10-CTAB, TX100-SDS, and TWEEN80-SPAN80, respectively.

[0127] c Oil 1-2 represent dodecane and hexadecane, respectively;

[0128] d The proportion of surfactants representing the former is calculated by mass ratio;

[0129] e Surfactant concentration is calculated as a percentage of water by mass (wt%).

[0130] In summary, this invention provides a method for predicting the stability of mixed surfactant emulsions based on active learning. It constructs a global formulation space by selecting multiple influencing factors, builds a dataset using hierarchical sampling and feature dimensionality reduction, it iteratively expands high-value experimental samples through active learning, and employs various machine learning algorithms to construct classification and regression prediction models. The model obtained by this invention exhibits excellent overall performance, with a classification model achieving an accuracy of 0.7857 and an F1 score of 0.7711 on the test set, and a regression model achieving a coefficient of determination R0 on the test set. 2 The accuracy rate reached 0.8795 with a mean square error of 0.0103, and the prediction accuracy under high-temperature conditions was 81.25% in practical verification. This invention combines leave-one-out cross-validation and grid search to complete model verification and hyperparameter optimization, and combines feature importance and SHAP algorithm to analyze the influence of each parameter on emulsion stability. This invention is applicable to oil-water two-phase mixed surfactant emulsion systems, and can accurately determine the stability level, calculate the emulsion layer height, and complete global formulation optimization. This method overcomes the limitations of traditional experimental screening, which is characterized by low efficiency and high trial-and-error costs, and provides a reliable technical solution for rapid research and development and iterative upgrading of emulsion formulations. It has significant application value and industrialization prospects in the fields of food, cosmetics, petrochemicals, and pharmaceuticals.

[0131] This invention is not limited to the above-described embodiments. Within the scope of the invention's technical concept, any conventional adjustments or equivalent substitutions made to formulation variables, model types, and active learning parameters, resulting in technical solutions, all fall within the protection scope of this invention. This method can be adapted to different emulsion formulation development needs across various industries by expanding the types of surfactants, oil phases, and temperature ranges according to actual application scenarios.

Claims

1. A method for predicting the stability of mixed surfactant emulsions based on active learning, characterized in that: Includes the following steps; Step 1: Construct the recipe space and prepare the initial dataset through stratified sampling; Step 2: Complete emulsion stability testing, data labeling, and feature dimensionality reduction. Step 3: Select the proxy model and sampling strategy, and expand the dataset iteratively through active learning; Step 4: Divide the dataset and build multi-class machine learning models, and complete hyperparameter optimization training; Step 5: Evaluate model performance and analyze its mechanism of action, and use the optimal model to make predictions.

2. The method for predicting the stability of mixed surfactant emulsions based on active learning according to claim 1, characterized in that: Step 1 specifically includes: Step 101: Select surfactant combination, concentration, compounding ratio, oil phase type, oil-water ratio, and temperature as variables, and set multiple gradient levels to construct a global unlabeled formulation pool; Step 102: Stratify sampling according to component type and ratio to construct initial training set, validation set and test set respectively.

3. The method for predicting the stability of mixed surfactant emulsions based on active learning according to claim 2, characterized in that: Step 101 completes the construction of the full-dimensional formulation space and the preparation of the initial dataset. Specifically, after determining the six core variables, namely surfactant combination, total surfactant concentration, surfactant mass ratio, oil phase type, oil-water volume ratio, and ambient temperature, multiple gradient variable levels are set so that the variables are combined to generate multiple full formulations that constitute the unlabeled formulation pool.

4. The method for predicting the stability of mixed surfactant emulsions based on active learning according to claim 1, characterized in that: Step 2 specifically includes: Step 201: Prepare the emulsion under uniform process conditions and keep it at a constant temperature for static testing. Use the emulsion layer height CH as a quantitative evaluation index of stability. Step 202: Divide the continuous emulsion layer height values ​​into multiple stability levels and complete the classification label construction; Step 203: Extract molecular descriptors and process features, and complete feature dimensionality reduction through data cleaning, low variance filtering and correlation analysis; The CH value of the emulsion is calculated as follows: This refers to the height of the emulsion layer that separates after standing. This represents the total height of the emulsion system.

5. The method for predicting the stability of mixed surfactant emulsions based on active learning according to claim 4, characterized in that: In step 201, the total volume of the emulsion is uniformly set, the mixed surfactant is completely dissolved in the ultrapure water, the corresponding oil phase is added, and the emulsion is prepared by using a high-speed homogenizer at a preset speed. In step 202, the continuous CH values ​​are divided into 5 discrete stability levels to adapt to the classification task. CH > 0.8 corresponds to level 4 extremely stable, 0.6 < CH ≤ 0.8 corresponds to level 3 relatively stable, 0.4 < CH ≤ 0.6 corresponds to level 2 critically stable, 0.2 < CH ≤ 0.4 corresponds to level 1 relatively unstable, and CH ≤ 0.2 corresponds to level 0 extremely unstable. In step 203, ChemDraw is used to draw various molecular structures and convert them into SMILES format. Then, RDKit is used to extract the geometric, electronic, and topological molecular descriptors of surfactants and oil phases, which are combined with the formulation process parameters to form an initial high-dimensional feature set.

6. The method for predicting the stability of mixed surfactant emulsions based on active learning according to claim 1, characterized in that: Step 3 specifically includes: Step 301: Train multiple tree ensemble classification models and select a surrogate model by combining evaluation metrics; compare multiple sampling strategies to determine the optimal query method. Step 302: Use the surrogate model to screen high-value unknown formulas and conduct experimental annotations, iteratively updating the model and training dataset.

7. The method for predicting the stability of mixed surfactant emulsions based on active learning according to claim 6, characterized in that: In step 301, an active learning iterative framework is built to continuously expand the labeled dataset. Based on the initial training set, four tree ensemble classification models—RF, GBDT, XGBoost, and LightGBM—are trained respectively. The model performance is judged by combining the accuracy and F1 score on the validation and test sets, and the model with the best overall performance is selected as the active learning proxy model. The iterative effects of four sample query strategies—uncertainty sampling, entropy sampling, marginal sampling, and random sampling—are compared, and uncertainty sampling, which has the highest information gain and the fastest model convergence speed, is selected as the core query strategy. In step 302, during each iteration, the surrogate model is used to predict all samples in the unlabeled formula pool. Combined with the selected sampling strategy, several groups of unknown formulas with the highest model cognitive uncertainty and the greatest information gain are selected. Experimental testing is carried out on the selected samples to obtain the corresponding CH value and stability level label. Then, the newly labeled samples are merged into the original training set and the surrogate model is updated. Several iterations are carried out, and the model accuracy, F1 score, AUC value of each category and the overall prediction uncertainty of the unlabeled sample pool are monitored throughout the process.

8. The method for predicting the stability of mixed surfactant emulsions based on active learning according to claim 1, characterized in that: Step 4 specifically includes: Step 401: Merge all labeled samples to form the total training set, divide it into independent test sets, and use leave-one-out cross-validation to conduct internal verification; Step 402: Simultaneously build classification and regression models, covering linear regression and tree ensemble algorithms, and optimize model hyperparameters through grid search.

9. The method for predicting the stability of mixed surfactant emulsions based on active learning according to claim 1, characterized in that: Step 5 specifically includes: Step 501: Use accuracy, F1 score, and R... 2 MSE is used as the evaluation index to complete the model evaluation, and the optimal model is determined based on the evaluation results. For classification models, accuracy, F1 score, AUC value, and confusion matrix are used to evaluate the overall performance, while the coefficient of determination R is used for regression models. 2 Mean squared error (MSE) is used as an evaluation index; Step 502: Analyze the influence of each feature on emulsion stability using feature importance and SHAP algorithm, and conduct experimental verification using unknown formulations; Step 503: Based on the optimal model, complete the global formulation stability prediction, and correct the model bias by combining the experimental verification results; Wherein, the determination coefficient R 2 The mean square error (MSE) is calculated as follows: Indicates the actual value. This represents the model's predicted value. Indicates the number of samples.

10. The method for predicting the stability of mixed surfactant emulsions based on active learning according to claim 9, characterized in that: The prediction method is based on an oil-water two-phase mixed surfactant emulsion system and is used to predict the emulsion stability level and emulsion layer height under different temperature conditions. In the prediction process, the optimal classification model and the optimal regression model are first determined by combining the evaluation results. Feature importance analysis combined with the SHAP interpretation algorithm is used to quantify the nonlinear influence of formulation parameters and molecular descriptors on emulsion stability, clarify the mechanism of action of water-oil ratio, surfactant concentration, compounding ratio, temperature and molecular microstructure. Then, representative candidate formulations are selected from the unlabeled formulation pool, and physical experiments are carried out under high temperature conditions. The measured data are compared with the model prediction results to verify the model's generalization ability.

Images