A microfluidic droplet diameter intelligent prediction method and system based on machine learning

Abstract

The application discloses a kind of microfluidic droplet diameter intelligent prediction method and system based on machine learning, it relates to microfluidic droplet diameter intelligent prediction method and system.It is to solve the technical problems of the existing droplet measurement method, such as droplet diameter detection lag, low prediction accuracy, process parameter design depends on artificial experience, and model cannot evolve independently, the method: one, data acquisition;Two, coding and standardization;Three, model training;Four, model persistence;Five, two-way prediction;Six, data incremental update.The application can effectively predict after incremental sample, MAE≤0.02398mm, the accuracy is improved by 84% compared with empirical formula, and can be applied to the field of intelligent design and optimization of microfluidic droplet preparation process.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of microfluidic droplet preparation and artificial intelligence, specifically to a method and system for predicting the diameter of microfluidic droplets. Background Technology

[0002] Microfluidics, an advanced technology for manipulating microfluidics within micrometer-scale channels, has been widely applied in drug delivery, single-cell analysis, biosensoring, food science, and cosmetic preparation. Droplet diameter, as a core quality indicator in microfluidic droplet preparation, directly determines the encapsulation efficiency, release kinetics, and functional performance of the product, making it a critical parameter that requires strict control during the production process.

[0003] In the fabrication of microfluidic droplets, the droplet diameter is influenced by a combination of factors, including the properties of the dispersed phase (e.g., oil phase type), dispersed phase viscosity, continuous phase properties (e.g., aqueous phase and surfactant concentration), the flow rate ratio of the dispersed phase to the continuous phase, and chip channel size. Complex nonlinear relationships exist between these parameters. Traditional prediction methods based on empirical formulas or physical models (e.g., capillary number correlations) often suffer from prediction errors exceeding 15% because they neglect the interaction effects and dynamic changes between parameters. This makes it difficult to meet the precise design requirements for process parameters in high-precision droplet fabrication (e.g., monodisperse droplet CV value <3%).

[0004] Currently, the industry mainly uses offline microscopy to obtain droplet diameter data. This method has significant drawbacks: First, the detection cycle is long, requiring sampling, sample preparation, microscopic imaging, and image analysis, which cannot reflect the current preparation status in real time, leading to delays in process adjustment and a high risk of producing batches of unqualified products. Second, the detection process relies on manual operation, which introduces subjective measurement errors, and the accuracy is difficult to meet the production requirements of highly uniform droplets (target diameter deviation <5%). Third, process parameter debugging depends on the operator's experience. When raw material batches fluctuate or chip batches are changed, a large number of orthogonal experiments need to be repeated to optimize parameters, resulting in high trial-and-error costs and low efficiency, which seriously affects the new product development cycle.

[0005] Some existing technologies attempt to establish droplet diameter prediction models using linear regression or response surface methodology. However, due to the complex nonlinear behavior of two-phase fluid dynamics involved in the microfluidic droplet generation process, these methods are insufficient in characterizing the interactions between parameters, and the prediction accuracy is typically only ±15%~20%, failing to provide reliable guidance for precise design of process parameters. Furthermore, existing prediction models are all static models, unable to continuously optimize their performance using new experimental data, resulting in a decrease in model generalization ability with changes in raw material batches and equipment conditions. Summary of the Invention

[0006] This invention aims to address the technical problems of existing droplet measurement methods, such as lag in droplet diameter detection, low prediction accuracy, reliance on manual experience for process parameter design, and the inability of models to evolve autonomously. Instead, it provides a machine learning-based intelligent prediction method and system for microfluidic droplet diameter. The method is based on Stochastic Gradient Descent Regressor (SGDRegressor) for intelligent prediction of microfluidic droplet diameter. Through an incremental update mechanism, it achieves accurate prediction of droplet diameter and reverse optimization of process parameters, thereby enabling accurate prediction and intelligent control of the microfluidic droplet preparation process.

[0007] The intelligent prediction method for microfluidic droplet diameter based on machine learning of the present invention is carried out according to the following steps:

[0008] I. Data Acquisition: Standardized data was collected, forming six data fields. The data included dispersed phase properties, dispersed phase viscosity, continuous phase properties, dispersed phase to continuous phase flow rate ratio, chip channel size, and droplet diameter. The dispersed phase properties were specified as oil or water; the dispersed phase viscosity was in cP, ranging from 500 to 2000; the continuous phase properties were specified as water or oil; the dispersed phase to continuous phase flow rate ratio was formatted as "value:1", such as "0.8:1"; the chip channel size was in mm, ranging from 1 to 5; and the droplet diameter was in mm, representing the target variable for prediction. A cumulative data file was obtained.

[0009] II. Encoding and Standardization: The system automatically performs intelligent parsing of the velocity ratio field between the dispersed phase and the continuous phase in the data file library, and converts the "value:1" format or pure numerical format into floating-point features through a custom parsing function; discrete features are encoded using a one-hot encoder, and continuous features are incrementally adaptively standardized using RunningStandardScaler.

[0010] III. Model Training: The SGDRegressor prediction model from the scikit-learn library is used as the base model. The mean squared error loss function, L2 regularization, adaptive learning rate and maximum number of iterations are used for training to output the SGDRegressor prediction model; the residual standard deviation is calculated and saved for the quantification of prediction uncertainty.

[0011] IV. Model Persistence: The Joblib technology is used to save the encoded and standardized data, the trained SGDRegressor prediction model and the residual standard deviation as independent .pkl files; each component can be loaded and updated independently, which facilitates modular maintenance;

[0012] V. Two-way forecasting: Using forecasting models to perform forward and / or reverse forecasting;

[0013] The forward prediction method involves inputting five process parameters: dispersed phase properties, dispersed phase viscosity, continuous phase properties, flow rate ratio, and chip size. It then uses a prediction model to output a predicted droplet diameter and provides a 95% confidence interval (predicted value ± 1.96σ) based on the standard deviation of the training set residuals, quantifying the prediction uncertainty. The time taken to output the predicted droplet diameter using the prediction model is only 1 second.

[0014] Reverse prediction requires inputting the target droplet diameter, dispersed phase properties, and continuous phase properties, and providing 1-2 known values ​​from three numerical parameters: viscosity, flow rate ratio, and chip size. The prediction model enumerates all combinations of unknown parameters within a preset search space and calls the forward model to predict the droplet diameter one by one. It then filters out all combinations where the predicted droplet diameter deviates from the target value by less than the tolerance and outputs the corresponding viscosity, flow rate ratio, and chip size solutions in a concise list format, allowing users to select the optimal solution based on engineering constraints.

[0015] VI. Incremental Data Update: The data predicted in step five or from other sources is appended to the original data file library in step one. Then, the complete dataset is read and full training is performed. This more stable "full retraining" strategy ensures that the model is always based on all historical data and avoids parameter oscillation or forgetting problems that may occur in online learning.

[0016] Furthermore, in step two, the system uses pandas to achieve highly compatible reading of Excel files, supporting both .xlsx and .xls formats; RunningStandardScaler focuses on incremental standardization of continuous features.

[0017] Furthermore, in step three, the alpha value for L2 regularization is 0.0001;

[0018] Furthermore, in step three, the initial value of the adaptive learning rate is 0.01;

[0019] Furthermore, in step three, the maximum number of iterations is 900-1000.

[0020] Furthermore, in step three, the SGDRegressor prediction model supports the partial_fit method, which enables online learning for small batches or single samples.

[0021] Furthermore, the preset search space mentioned in step four has a viscosity range of 500~2000 with a step size of 50; a flow rate ratio of 0.5~2.6 with a step size of 0.1; and a chip size of 0.5~5 with a step size of 0.5.

[0022] Furthermore, the tolerance mentioned in step four is 0.005 mm;

[0023] Furthermore, the bidirectional prediction described in step four is based on Python Tkinter, which builds three independent interfaces: a forward prediction interface, a reverse prediction interface, and a model update interface. Each interface has a clearly designed input area, a result display area, and operation buttons, supporting parameter verification, one-click prediction, and data export from the forward prediction interface. This interface makes operation simple and suitable for laboratory technicians to use quickly. All system components are deployed locally, requiring no network connection, thus ensuring the security of experimental data.

[0024] Furthermore, the data update described in step six involves appending the newly added data Excel file to the data file library from step one, achieving permanent data accumulation; simultaneously, backing up the current SGDRegressor prediction model file to the "backup_models" directory; then retraining the new SGDRegressor prediction model based on the complete accumulated dataset; and recalculating and saving the model evaluation metrics, including mean absolute error (MAE) and coefficient of determination (R²). 2 The residual standard deviation σ is calculated; the model persistence file (.pkl) is updated. This mechanism ensures that the model continues to optimize as data accumulates, and version rollback can be achieved through backup files.

[0025] Compared with the prior art, the advantages of the present invention are as follows:

[0026] (1) Strong applicability to small samples: An effective model can be trained based on the initial 200 sets of experimental data (MAE is 0.03223, R... 2 The MAE was initially 0.6995, but with data accumulation (expanded to 11281 sets), the model's MAE can be reduced to 0.02398, and R... 2 The accuracy is 0.8120, and the precision continues to improve.

[0027] (2) High prediction accuracy: The average absolute error of the method of the present invention in predicting the droplet diameter is ≤0.04 mm, which is an improvement over the traditional empirical formula method (error is often >0.05 mm), and meets the engineering requirements for the preparation of monodisperse droplets (CV <3%).

[0028] (3) Two-way prediction improves process design efficiency: The reverse prediction mode can quickly provide multiple sets of feasible parameter combinations that meet the target diameter, replacing the traditional time-consuming large-scale orthogonal experiment, and significantly shortening the process development cycle of new products.

[0029] (4) Self-evolution capability: Through the update mechanism of "data appending-full retraining", the model can permanently remember all historical data and automatically optimize as the number of experiments increases without human intervention.

[0030] (5) Localized lightweight deployment: The system depends only on open source Python libraries, does not require commercial software licenses, can run on ordinary office computers, and is suitable for use in laboratories and small and medium-sized enterprises. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the forward prediction interface in Example 1;

[0032] Figure 2 This is a schematic diagram of the reverse prediction interface in Example 1;

[0033] Figure 3 This is a schematic diagram of the model update tool interface in Example 1;

[0034] Figure 4 This is a flowchart of the machine learning-based intelligent prediction method for microfluidic droplet diameter in Example 1. Detailed Implementation

[0035] The invention will be further described in detail below with reference to examples.

[0036] Example 1: The intelligent prediction method for microfluidic droplet diameter based on machine learning in this example is carried out according to the following steps:

[0037] I. Data Acquisition: Standardized data was collected, forming six data fields. The data included dispersed phase properties, dispersed phase viscosity, continuous phase properties, dispersed phase to continuous phase flow rate ratio, chip channel size, and droplet diameter. Among them, the dispersed phase property was oil phase; the dispersed phase viscosity was in cP, ranging from 500 to 2000; the continuous phase property was aqueous phase; the dispersed phase to continuous phase flow rate ratio was in the format "value:1", such as "0.8:1"; the chip channel size was in mm, ranging from 1 to 5; and the droplet diameter was in mm, with the droplet diameter being the target variable for prediction. 200 sets of droplet diameter data were collected, and the data format is shown in Table 1, resulting in a data file library.

[0038] Table 1. Data on the diameter of 200 droplets collected:

[0039] Dispersed phase properties Dispersed phase viscosity (cP) Continuous phase properties Flow rate ratio Chip size (mm) Droplet diameter (mm) oil phase 1000 Aqueous phase 0.8:1 2 0.043 oil phase 1000 Aqueous phase 1.0:1 2 0.056 …… …… …… …… …… ……

[0040] II. Encoding and Standardization: The system automatically performs intelligent parsing of the velocity ratio field of the dispersed phase and continuous phase in the data file library, and converts the "value:1" format into floating-point features through a custom parsing function; the oil / water phase of discrete features is encoded using a one-hot encoder, and the viscosity, velocity ratio value, and chip size of continuous features are incrementally adaptively standardized using RunningStandardScaler.

[0041] III. Model Training: The SGDRegressor prediction model from the scikit-learn library is used as the base model. Mean squared error loss function, L2 regularization, and adaptive learning rate are employed, with a maximum number of iterations set. Specifically, the L2 regularization alpha = 0.0001; the initial value of the adaptive learning rate is 0.01; and the maximum number of iterations is 1000. Training is then performed, outputting the SGDRegressor prediction model. After training, the following calculations are made: MAE = 0.03223 mm, R... 2 =0.6995, residual standard deviation σ=0.06mm; save the SGDRegressor prediction model file;

[0042] IV. Model Persistence: The Joblib technology is used to save the encoded and standardized data from step two, the trained SGDRegressor prediction model, and the residual standard deviation as separate .pkl files; each component can be loaded and updated independently, which facilitates modular maintenance.

[0043] V. Two-way forecasting: Using forecasting models to perform both forward and reverse forecasting;

[0044] The positive prediction method inputs five process parameters: the dispersed phase property is "oil phase", the dispersed phase viscosity is 1000, the continuous phase property is "water phase", the flow rate ratio is "1.2:1", and the chip size is 2 mm. The predicted droplet diameter output by the prediction model is 0.0972 ± 0.0098 mm (95% confidence interval). The predicted droplet diameter is obtained within 30 seconds, which is much faster than offline microscopic detection.

[0045] The reverse prediction method takes the target droplet diameter as input (0.07 mm), the dispersed phase as oil, the continuous phase as water, and the flow rate ratio as 1.2:1. The prediction model enumerates 88 combinations of viscosity (500-1550, step size 50) and chip size (1-4) within a preset search space. It then calls the forward model to predict the droplet diameter for each combination, selecting those where the predicted droplet diameter deviates from the target value by less than 0.005 mm. The corresponding viscosity and chip size schemes are output in a concise list format as follows:

[0046] Combination 1: Viscosity = 1000 cP; Chip size = 2 mm

[0047] Combination 2: Viscosity = 1050 cP; Chip size = 1 mm

[0048] Combination 3: Viscosity = 950 cP; Chip size = 3 mm

[0049] Technicians select experimental plans based on the actual situation;

[0050] VI. Incremental Data Update: After completing the new experiment, technicians will compile the new data into an Excel file and append it to the original data file library from step one. The old model will be backed up, and the SGDRegressor will be retrained based on all the data. The new model and the updated standard deviation will be saved. Upon completion, the interface will display: "Update successful! Current total data volume: 11281 records, MAE: 0.02398, R..." 2 The model standard deviation was 0.8120, and the model standard deviation was 0.04023 mm. All subsequent predictions were based on the updated model.

[0051] This invention collects standardized data including dispersed phase properties, viscosity, continuous phase properties, flow rate ratio, chip size, and droplet diameter; employs one-hot encoding and a custom incremental normalizer (RunningStandardScaler) for feature engineering; and calculates the residual standard deviation based on the SGDRegressor prediction model for uncertainty quantification. It provides two modes: forward prediction (input parameters, output diameter with a 95% confidence interval) and backward prediction (input target diameter and some known parameters, output combination of missing parameters); and supports continuous model evolution through data appending and full retraining. The system is built on Tkinter with three independent graphical interfaces and localized deployment. This invention can effectively predict with incremental samples, achieving a MAE ≤ 0.02398 mm, an 84% improvement in accuracy compared to empirical formulas, and can be widely applied to the intelligent design and optimization of microfluidic droplet fabrication processes.

Claims

1. A machine learning-based intelligent prediction method for microfluidic droplet diameter, characterized in that, This method is performed in the following steps: I. Data Acquisition: Collect standardized data to form six data fields, including dispersed phase properties, dispersed phase viscosity, continuous phase properties, dispersed phase to continuous phase flow rate ratio, chip channel size, and droplet diameter; among which, dispersed phase properties are such as oil phase or water phase; dispersed phase viscosity is in cP, ranging from 500 to 2000; continuous phase properties are water phase or oil phase; the dispersed phase to continuous phase flow rate ratio is in the format "value:1", such as "0.8:1"; Chip channel size, in mm, range 1~5; droplet diameter, in mm, droplet diameter is the target variable for prediction; cumulative data file is obtained; II. Encoding and Standardization: The system automatically performs intelligent parsing of the dispersed phase to continuous phase velocity ratio field in the data file library, and uses a custom parsing function to uniformly convert the "value:1" format or pure numerical format into floating-point features. Discrete features are encoded using a one-hot encoder, while continuous features are incrementally adaptively standardized using RunningStandardScaler. III. Model Training: The SGDRegressor prediction model in the scikit-learn library is used as the base model. The mean squared error loss function, L2 regularization, adaptive learning rate and maximum number of iterations are used for training to output the SGDRegressor prediction model. Calculate and save the residual standard deviation for use in predicting uncertainty quantification; IV. Model Persistence: The Joblib technology is used to save the encoded and standardized data, the trained SGDRegressor prediction model, and the residual standard deviation as independent .pkl files; each component can be loaded and updated independently, which facilitates modular maintenance. V. Two-way forecasting: Using forecasting models to perform forward and / or reverse forecasting; The forward prediction method involves inputting five process parameters: dispersed phase properties, dispersed phase viscosity, continuous phase properties, flow rate ratio, and chip size. It then uses a prediction model to output a predicted droplet diameter and provides a 95% confidence interval (predicted value ± 1.96σ) based on the standard deviation of the training set residuals, quantifying the prediction uncertainty. The time taken to output the predicted droplet diameter using the prediction model is only 1 second. Reverse prediction requires inputting the target droplet diameter, dispersed phase properties, and continuous phase properties, and providing 1-2 known values ​​from three numerical parameters: viscosity, flow rate ratio, and chip size. The prediction model enumerates all combinations of unknown parameters within a preset search space and calls the forward model to predict the droplet diameter one by one. It then filters out all combinations where the predicted droplet diameter deviates from the target value by less than the tolerance and outputs the corresponding viscosity, flow rate ratio, and chip size solutions in a concise list format, allowing users to select the optimal solution based on engineering constraints. VI. Incremental Data Update: The data predicted in step five or from other sources is appended to the original data file library in step one. Then, the complete dataset is read and full training is performed. This more stable "full retraining" strategy ensures that the model is always based on all historical data and avoids parameter oscillation or forgetting problems that may occur in online learning.

2. The intelligent prediction method for microfluidic droplet diameter based on machine learning according to claim 1, characterized in that, In step three, the alpha value for L2 regularization is 0.0001.

3. The intelligent prediction method for microfluidic droplet diameter based on machine learning according to claim 1 or 2, characterized in that, In step three, the initial value of the adaptive learning rate is 0.

01.

4. The intelligent prediction method for microfluidic droplet diameter based on machine learning according to claim 1 or 2, characterized in that, In step three, the maximum number of iterations is 900 to 1000.

5. A machine learning-based intelligent prediction method for microfluidic droplet diameter according to claim 1 or 2, characterized in that, The preset search space mentioned in step five has a viscosity range of 500~2000 with a step size of 50; a flow rate ratio of 0.5~2.6 with a step size of 0.1; and a chip size of 0.5~5 with a step size of 0.

5.

6. A machine learning-based intelligent prediction method for microfluidic droplet diameter according to claim 1 or 2, characterized in that, The tolerance mentioned in step five is 0.005 mm.

7. A machine learning-based intelligent prediction method for microfluidic droplet diameter according to claim 1 or 2, characterized in that, The data update described in step six involves appending the newly added data Excel file to the data file library from step one, achieving permanent data accumulation; simultaneously, backing up the current SGDRegressor prediction model file to the "backup_models" directory; then retraining the new SGDRegressor prediction model based on the complete accumulated dataset; and recalculating and saving the model evaluation metrics, including mean absolute error (MAE) and coefficient of determination (R²). 2 Set the residual standard deviation σ; update the model persistence file.

8. A microfluidic droplet diameter intelligent prediction system based on the method of any one of claims 1-7, characterized in that, The system comprises three graphical interface modules: forward prediction, backward prediction, and model update. Each module shares the same model file and preprocessor, enabling localized and lightweight deployment.

Images