Intelligent and accurate design method for microfluidic liquid drops and diversified microspheres

Abstract

The invention discloses an intelligent accurate design method for microfluidic droplets and diversified microspheres, and relates to the field of microfluidic technology and artificial intelligence cross application, and the method comprises the following steps: S1, based on the material type and microsphere size input by a user, guiding reverse search and iterative correction by a microsphere size prediction model selected according to the material type, the system concentration and the droplet template size required for preparing the microspheres are obtained, and prediction from the microspheres to the droplets is completed; s2, based on the system concentration, obtaining corresponding physical property parameters from a physical property library; and S3, taking the size of the liquid drop template as a target value, taking the physical property parameter variable as a limit value, and adopting the liquid drop size prediction model to guide reverse search and iterative correction so as to obtain an experimental parameter combination for preparing the microspheres with the preset size. The prediction precision and generalization performance of the microfluidic liquid drops and the diversified microspheres are remarkably improved through a modular artificial intelligence framework based on coupling of the meta-model and continuous learning.

Description

Technical Field

[0001] The present invention relates to the cross-application field of microfluidics and artificial intelligence. More specifically, the present invention relates to an intelligent and precise design method for microfluidic droplets and diversified microspheres. Background Art

[0002] As a new type of functional material, microspheres have promoted innovation and development in many industries. In the pharmaceutical field, microspheres can be used as drug delivery carriers, achieving targeted delivery of specific drugs by controlling the release rate of drugs, while reducing drug degradation and loss, and reducing side effects. In the field of biological separation, microspheres can be used as chromatographic fillers. The uniform particle size and structure with functional groups on the surface can significantly improve the accuracy of chromatographic separation, thereby achieving the purpose of separation and purification. In the field of microchemicals, microspheres can be used as microreactors due to their miniaturization and high specific surface area, thereby significantly improving safety and reaction rates. Generally, the preparation of microspheres is mainly achieved through the solidification of droplet templates, and their quality characteristics are determined by key parameters such as the size and uniformity of the droplets.

[0003] Compared with traditional preparation methods, microfluidics provides an advanced technical platform for the precise and controllable preparation of droplets by precisely controlling the flow behavior of microscale multiphase fluids. The microfluidic preparation process of droplets involves a complex nonlinear multiphase flow process, which is affected by the coupling of multiple factors such as device geometric parameters, fluid physical properties, and operating conditions. Different application scenarios have different requirements for the diameter, generation rate, and flow type of droplets, and existing experimental design methods mainly rely on empirical optimization and resource-intensive iterative testing, resulting in a time-consuming preparation process and difficulty in accurately matching the target size required by users. Especially in the fields of life sciences and materials sciences, the controllable generation of droplets is often coupled with subsequent processes such as curing and functionalization, which further increases the complexity of experimental optimization.

[0004] Therefore, there is an urgent need to build a set of intelligent design methods that can reversely deduce the optimal experimental parameters based on user needs, reduce the trial and error process, improve experimental efficiency, and thus accelerate the precise manufacturing and industrial application of droplet micro-dispersion systems and diversified microspheres. Machine learning has significant advantages in dealing with complex, multivariable and nonlinear problems. By building data-driven models, machine learning can learn the key influencing factors in the droplet formation process from existing data sets, and then make predictions from target values ​​to experimental parameters. Compared with traditional experimental optimization methods, machine learning can not only improve prediction accuracy, but also actively recommend the optimal experimental parameter combination through optimization algorithms, thereby reducing experimental costs and time costs. Therefore, the introduction of machine learning to construct an intelligent design method can not only realize the intelligent prediction of droplet preparation parameters, but also expand its application range in different microfluidic devices and fluid systems.

[0005] At present, some studies have used machine learning to model the droplet generation process, such as using neural networks and regression models to predict droplet diameter and generation rate. However, the focus of most studies is still on the forward prediction from experimental conditions to droplet diameter, and there is a lack of reverse design from user needs to experimental conditions. In addition, most existing data come from a single microfluidic device or a specific experimental system, which limits the generalization ability of the model and makes it difficult to apply to different device structures and multiple fluid systems. At the same time, for microsphere curing and subsequent processing, existing research still lacks a systematic and accurate prediction model. Therefore, establishing a general and efficient machine learning model to achieve accurate prediction from droplet generation to microsphere curing, and then making a full-process intelligent design from user-required material type and microsphere size to reverse determination of experimental parameters, is still an important challenge for current research. Summary of the Invention

[0006] An object of the present invention is to solve at least the above problems and / or disadvantages and to provide at least the advantages which will be described hereinafter.

[0007] To achieve these and other advantages of the present invention, a method for intelligent and precise design of microfluidic droplets and diversified microspheres is provided, comprising:

[0008] S1. Based on the material type and microsphere size input by the user, the corresponding microsphere size prediction model is selected according to the material type to guide the reverse search and iterative correction to obtain the system concentration and droplet template size required for preparing microspheres using microfluidic equipment, and complete the prediction from microspheres to droplets;

[0009] S2. Based on the system concentration obtained in S1, obtain the corresponding physical property parameters from the physical property library;

[0010] S3, using the droplet template size obtained in S1 as the target value and the physical property parameter variables obtained in S2 as the limiting values, and using the droplet size prediction model to guide reverse search and iterative correction to obtain the experimental parameter combination for preparing microspheres of predetermined size using a microfluidic device;

[0011] Wherein, in S2, the physical property parameters refer to the viscosity and interfacial tension of the dispersed phase and the continuous phase fluid;

[0012] The experimental parameter combination includes geometric parameters and operating conditions.

[0013] Preferably, in S1, the droplet size prediction model is constructed by:

[0014] S10, constructing a numerical simulation model of a droplet microfluidic device, adjusting droplet construction parameters in the numerical simulation model to obtain corresponding droplet sizes through simulation calculations, and thereby obtaining a droplet simulation data set for droplet prediction;

[0015] S11, using the geometric parameters, physical properties, and operating conditions in the droplet simulation dataset as feature variables and the droplet size as the target variable to perform model training to obtain an XGBoost meta-model for droplet size prediction;

[0016] S12. Use experimental data to verify the XGBoost meta-model;

[0017] S13. Use the data generation algorithm to continuously learn the XGBoost meta-model to obtain a droplet size prediction model.

[0018] Preferably, in S10, the droplet construction parameters include: geometric parameters, physical parameters and operating conditions;

[0019] The droplet microfluidic device includes: a coordinated flow device, a T-type device and a flow focusing device.

[0020] Preferably, in S13, the data generation algorithm uses the WGAN-GP algorithm to generate a new data set with a distribution similar to that of the validation set data;

[0021] The continuous learning refers to adding new datasets to the training set of the droplet simulation dataset to fine-tune the XGBoost meta-model.

[0022] Preferably, in S3, the reverse search method is:

[0023] Based on the droplet template size obtained in S1, search for the five closest sets of data in the validation set data of the droplet prediction model;

[0024] Based on the physical property parameters obtained by S2, the operating conditions are optimized to obtain the corresponding experimental parameter combination in the droplet prediction model with the goal of minimizing the mean absolute percentage error of the droplet template size.

[0025] Preferably, in S3, the process of constructing the microsphere size prediction model includes:

[0026] S30, using a co-flow microfluidic device to prepare droplet templates of various biocompatible microspheres;

[0027] S31, solidifying the droplet templates obtained in S30 using corresponding solidification methods to obtain corresponding microsphere solidification data sets;

[0028] S32. Model training is performed based on the microsphere solidification dataset to obtain a variety of microsphere size prediction models that match a variety of biocompatible microspheres based on the XGBoost algorithm.

[0029] Preferably, the multiple biocompatible microspheres include: polyethylene glycol diacrylate microspheres, calcium alginate microspheres, chitosan microspheres, hyaluronic acid microspheres, gelatin microspheres, polycaprolactone microspheres, polylactic acid microspheres, polylactic acid-polyglycolic acid microspheres, poly-L-lactic acid microspheres, polylactic acid-polyethylene glycol microspheres, etc.

[0030] Preferably, the curing method includes: photocrosslinking method, ion crosslinking method, chemical crosslinking method, and solvent volatilization method.

[0031] The present invention has at least the following beneficial effects:

[0032] First, the present invention effectively separates the microfluidic control of diverse microspheres into two separate processes, droplet construction and microsphere solidification, through a modular artificial intelligence framework. This significantly reduces the complexity and dimensionality of the parameter space for obtaining microspheres through droplet solidification. This sparse parameter space enables the development of highly accurate prediction models (average error <10%) using only small experimental datasets.

[0033] Secondly, the present invention significantly improves the prediction accuracy and generalization ability of microfluidic droplets by training the model through adding a continuous learning strategy to the meta-model.

[0034] Third, the present invention innovatively proposes a microsphere reverse design method guided by microsphere material type and size targets, and through customized design and experimental verification of target-sized microspheres, the maximum error is 8.8%, realizing microfluidic precision intelligent manufacturing of multiphase single droplets and diversified biocompatible microspheres, providing a new way to guide the precise manufacturing of biomedical functional microsphere materials based on basic theoretical models.

[0035] Fourthly, the microsphere reverse design method proposed in the present invention can reversely deduce the optimal experimental parameters according to user needs, reduce the trial-and-error process, and improve experimental efficiency, thereby accelerating the precise manufacturing and industrial application of droplet micro-dispersion systems and diversified microspheres.

[0036] Other advantages, objectives and features of the present invention will be reflected in part from the following description and will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is the data distribution characteristics of multiple factors (dispersed phase flow Q1, continuous phase flow Q2) under the collaborative flow microfluidic structure of large-scale calculation;

[0038] Figure 2 It is the data distribution characteristics of multiple factors (dispersed phase viscosity μ1, continuous phase viscosity μ2) under the coordinated flow microfluidic structure of large-scale calculation;

[0039] Figure 3It is the data distribution characteristics of multiple factors (capillary diameter D, two-phase interfacial tension σ) under the coordinated flow microfluidic structure of large-scale calculation;

[0040] Figure 4 It is the data distribution characteristics of multiple factors (capillary diameters D1 and D3) under the coordinated flow microfluidic structure of large-scale calculation;

[0041] Figure 5 The experimental data under the cooperative flow microfluidic structure are used to verify the accuracy of the meta-model;

[0042] Figure 6 The experimental data under the flow-focusing microfluidic structure are used to verify the accuracy of the meta-model;

[0043] Figure 7 The experimental data under the T-type microfluidic structure are used to verify the accuracy of the meta-model;

[0044] Figure 8 is the validation error of the meta-model after continuous learning under the cooperative flow microfluidic structure;

[0045] Figure 9 is the validation error of the meta-model after continuous learning under the flow-focusing microfluidic structure;

[0046] Figure 10 is the verification error of the meta-model under the T-type microfluidic structure after continuous learning;

[0047] Figure 11 This is a schematic diagram of the prediction performance of the meta-model using multiple sets of data points from collaborative flow devices in the literature (Lan et al., Deng et al., and Gu et al.) as validation sets;

[0048] Figure 12 This is a graph of the prediction performance after continuous learning of multiple sets of data from collaborative flow devices in the literature (Lan et al., Deng et al., and Gu et al.);

[0049] Figure 13 Schematic diagram of the prediction performance of the meta-model using multiple data points from flow focusing devices in the literature (Mardani et al., Srikanth et al., and Sartipzadeh et al.) as validation sets;

[0050] Figure 14 The predicted performance graph is based on continuous learning of multiple sets of data from flow focusing devices in the literature (Mardani et al., Srikanth et al., and Sartipzadeh et al.);

[0051] Figure 15Schematic diagram of the prediction performance of the meta-model using multiple data points of T-type devices in the literature (Chen et al., Tarchichi et al., and Odera et al.) as the validation set;

[0052] Figure 16 This is a prediction performance graph after continuous learning of multiple sets of data from T-type devices in the literature (Chen et al., Tarchichi et al., and Odera et al.);

[0053] Figure 17 It is a schematic diagram of the process of intelligent design of user-defined target microspheres;

[0054] Figure 18 It is the result of intelligent design of microspheres with multiple target sizes in implementation cases 1 to 3;

[0055] Figure 19 The comparison between the experimental particle size of PLA microspheres intelligently designed and the particle size defined by the user;

[0056] Figure 20 The comparison between the experimental particle size of the intelligently designed PEGDA microspheres and the user-defined particle size;

[0057] Figure 21 Comparison of the experimental particle size of CaAlg microspheres designed intelligently and the user-defined particle size. DETAILED DESCRIPTION

[0058] The present invention will be described in further detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0059] The preparation of microspheres based on droplet microfluidics technology mainly includes the technical ideas of two key processes: droplet template construction and microsphere solidification. Since the size of the droplet template is mainly affected by factors such as the geometric size of the device, the physical properties of the fluid and the operating conditions, and the size of the microspheres after solidification is mainly determined by the droplet template size, the concentration of the dispersed phase and the concentration of the cross-linker, among which the physical properties of the fluid are significantly affected by the type and concentration of the microsphere material, the direct prediction process from experimental conditions to microsphere size is very complicated.

[0060] Therefore, the present invention establishes an intelligent design method for droplets and diversified microspheres based on a machine learning algorithm. It also establishes a predictive model for droplet template size and biocompatible microsphere size based on a machine learning algorithm, enabling reverse optimization design from user-required microsphere size and material system to experimental parameter determination. To improve prediction accuracy and unify the construction process for multiple microspheres, the present invention adopts a modular design approach. Specifically, a droplet size prediction model and a microsphere size prediction model are constructed within the prediction model based on key influencing factors. A physical property parameter library corresponding to different concentrations of the microsphere material system (referred to as the physical property library) is then established. The two are then coupled together to achieve accurate predictions from experimental conditions to diversified microspheres.

[0061] Specifically, the present invention first constructs a droplet simulation dataset and a microsphere solidification dataset, and based on the above datasets, combines nested cross-validation and Bayesian optimization algorithm to train machine learning models for droplet size and microsphere size prediction, uses experimental data and literature data to double-verify the performance of the model, and proposes a continuous learning mechanism to continuously optimize the generalization ability of the model and improve its prediction accuracy on different data distributions. Then, relying on the established high-precision prediction model, the reverse intelligent design method is further developed. According to the target microsphere size (D given by the user), the target microsphere size (D p-target ) and material type, and the required droplet template size (D d ) and curing conditions (C1, C2), the physical property parameter values ​​corresponding to the above concentrations (dispersed phase viscosity μ1, continuous phase viscosity μ2 and two-phase interfacial tension σ) were obtained from the experimental physical property library, and then compared with D d Together, they serve as input conditions for droplet construction, and the droplet generation model is used to reverse engineer the optimal geometric parameters and operating conditions. Finally, the optimized experimental conditions are used for experimental verification to achieve precise control of droplet preparation and microsphere solidification, providing an intelligent design solution for microsphere manufacturing. In actual operation, the specific steps are as follows:

[0062] Step 1: Construct a droplet simulation dataset containing droplet construction parameters and droplet size

[0063] Droplet construction parameters and droplet template size (D d ) are derived from large-scale numerical simulations. The model structures used for numerical simulations of microspheres include three types: a coordinated flow device, a T-type device, and a flow focusing device. The numerical model uses the volume of fluid method, which can obtain multivariate and wide-range data distributions. The dataset for the microdroplet solidification process was obtained through experimental methods using various biocompatible microspheres. This makes it easier to link experimental conditions with the final microsphere size.

[0064] Step 2: training a machine learning meta-model based on the data set to establish a meta-model for droplet size prediction;

[0065] Step 3: Validate the meta-model using experimental data from droplet microfluidics. This involves constructing monodisperse droplets under multiple conditions using a microfluidic device to validate the meta-model.

[0066] A new dataset with a distribution similar to the validation set data is generated using the WGAN-GP data generation algorithm. The meta-model is continuously learned using this new dataset, and the final droplet size prediction model is obtained by modifying the meta-model to improve its generalization ability. The modification objective is to modify the parameters while keeping the overall structure of the meta-model unchanged, thereby improving the verification accuracy.

[0067] This step constructs a data set between droplet size and characteristic variables through large-scale calculations in a wide range of feature space and distribution based on a high-confidence numerical model, establishes a machine learning model for quantitative prediction of microsphere size, and innovatively proposes a new model training strategy that combines a meta-model with a continuous learning mechanism, establishing a multiphase droplet size prediction model with high precision and strong generalization ability.

[0068] Step 4: Establish a microsphere solidification dataset of ten types of biocompatible microspheres, and train a microsphere size prediction model based on the microsphere solidification dataset (the microsphere size prediction model is trained using the XGBoost algorithm);

[0069] The microsphere solidification dataset is derived from the solidification experiments of ten microspheres, mainly including the droplet size (D d ) and curing conditions (solution concentration C1, crosslinking agent concentration C2, etc.) and the size of the microspheres after curing (D p ) target variable, the ten microspheres include: polyethylene glycol diacrylate (PEGDA) microspheres, calcium alginate (CaAlg) microspheres, chitosan microspheres, hyaluronic acid (HAMA) microspheres, gelatin (Gelatin) microspheres, polycaprolactone (PCL) microspheres, polylactic acid (PLA) microspheres, polylactic acid-polyglycolic acid (PLGA) microspheres, poly-L-lactic acid (PLLA) microspheres, polylactic acid-polyethylene glycol (PLA16k-b-PEG5k) microspheres, and the curing methods of the ten biocompatible microspheres are mainly divided into four categories, as follows:

[0070] (1) Photocrosslinking method

[0071] Polyethylene glycol diacrylate (PEGDA), hyaluronic acid (HAMA), and gelatin microspheres were prepared using the photocuring principle. A PEGDA aqueous solution containing HMPP (2-hydroxy-2-methyl-1-phenyl-1-propanone) photoinitiator served as the dispersed phase fluid, and soybean oil containing 4 wt% PGPR90 served as the continuous phase fluid. Monodisperse droplets were formed in a co-flow microfluidic device. UV light with a wavelength of 350 nm was applied downstream of the receiving tube to cure the PEGDA microspheres. Methacryl-coated hyaluronic acid and gelatin served as the dispersed phase fluid, and soybean oil containing 4 wt% PGPR90 served as the continuous phase fluid. Monodisperse droplets were formed in a co-flow microfluidic device using UV light with a wavelength of 350 nm downstream of the receiving tube to cure the hyaluronic acid (HAMA) and gelatin microspheres. The size of the droplet templates could be flexibly adjusted by adjusting the flow rates of the dispersed and continuous phases in the microfluidic device.

[0072] (2) Ionic crosslinking method

[0073] Calcium alginate microspheres were prepared using an ionic crosslinking method. A monodispersed droplet template was prepared in a co-flow microfluidic device using an aqueous solution of NaAlg and EDTA as the dispersed phase and soybean oil containing 4 wt% T154 as the continuous phase. Soybean oil containing glacial acetic acid was used as the receiving phase. When the droplet template entered the receiving phase, EDTA released Ca under the action of glacial acetic acid. 2+ , and then undergoes ionic cross-linking with NaAlg to form calcium alginate microspheres. The size of the droplet template can be flexibly adjusted by the flow rate of the dispersed phase and the continuous phase in the microfluidic device.

[0074] (3) Chemical cross-linking method

[0075] Monodisperse chitosan microspheres were prepared by chemical cross-linking. First, a chitosan droplet template was prepared in a co-flow microfluidic device. The dispersed phase was a chitosan solution, to which a certain amount of glacial acetic acid was added to promote the dissolution of chitosan. The continuous phase consisted of 1.5% terephthalaldehyde, 4wt% PGPR90 and soybean oil. Droplets of different sizes were flexibly constructed by adjusting the flow rates of the dispersed phase and the continuous phase. The droplet template was then collected in a culture dish containing the continuous phase, and the chitosan droplets were allowed to cross-link with the terephthalaldehyde in the continuous phase. During the cross-linking process, terephthalaldehyde reacted with the amino groups in the chitosan molecules to form a Schiff base to form a stable imine bond, thereby achieving the solidification of the droplets. The size of the droplet template can be flexibly adjusted by the flow rate of the dispersed phase and the continuous phase in the microfluidic device.

[0076] (4) Solvent evaporation method

[0077] PCL, PLA, PLGA, PLLA, and PLA16k-b-PEG5k microspheres were prepared by solvent evaporation. Taking the preparation of PCL microspheres as an example, first, PCL was dissolved in dichloromethane (DCM) as the dispersed phase fluid, and a 3wt% PVA solution was used as the continuous phase. Monodisperse droplet templates of different sizes were obtained by adjusting the flow rate in a microfluidic device. Since DCM is highly volatile, the concentration field around the droplets directly affects its evaporation rate. In order to ensure the stable solidification of the microspheres as much as possible, a rotating sample receiving method was adopted, that is, the receiving bottle was placed on a rotating table, and the receiving tube was inserted into the eccentric position of the rotating bottle. The received PCL droplet template was then left to stand and evaporated open for 48 hours to ensure sufficient solidification. Finally, PCL microspheres were obtained after washing with deionized water. PCL solutions of different concentrations were prepared according to Table 6-6 to obtain a data set on the effect of concentration on microsphere solidification. The preparation methods of PLA, PLGA, PLLA, and PLA16k-b-PEG5k microspheres are consistent with those of PCL, and the size of the droplet template can be flexibly adjusted by the flow rates of the dispersed phase and the continuous phase in the microfluidic device.

[0078] Step 5: Receive the target microsphere material and size input by the user. The target microsphere material and size input by the user are the initial conditions for intelligent design, and intelligent design is performed based on these conditions.

[0079] Step 6: Using the microsphere size prediction model as a guide, reverse search is performed to obtain the optimal droplet template size and system concentration required to prepare the microspheres;

[0080] Specifically, reverse searching for the optimal droplet template size and system concentration is the first step in intelligent design. That is, the corresponding droplet template size is first predicted through the microsphere model. Since different system concentrations correspond to different shrinkage rates, the target droplet size and system concentration need to be determined simultaneously based on the target microsphere size.

[0081] Step 7: Based on the system concentration determined in step 6, the physical property parameters such as the viscosity and interfacial tension of the dispersed phase and continuous phase fluids can be obtained by consulting the physical property library. These parameters are used as input conditions for the droplet prediction model to complete the physical property parameter determination (or physical property parameter update).

[0082] Step 8. Use the droplet template size obtained in step 6 as the target value, input the limit values ​​of the system's physical parameter variables such as viscosity and interfacial tension, and reversely search for experimental conditions such as optimal geometric parameters and operating conditions under the guidance of the droplet size prediction model. Optimize the operating conditions under the guidance of the droplet size prediction model and finally output the optimal droplet preparation experimental parameter combination (the experimental parameter combination is the experimental conditions for preparing microspheres of a specific size, including geometric dimensions, operating conditions, and system concentration).

[0083] Specifically, the reverse search for optimal geometric parameters and operating conditions is to search for five sets of data closest to the droplet template size in the validation set data of the droplet prediction model based on the required droplet template size, and then replace the physical properties with the physical properties corresponding to the system concentration predicted in the first step, and then input them into the droplet prediction model. The operating conditions are optimized with the goal of minimizing the average absolute percentage error of the droplet template size, and finally the optimal experimental parameter combination is output.

[0084] In the following examples, the droplet microfluidic devices used are T-type, flow focusing type and cooperative flow type. The corresponding droplet simulation data sets in the examples are mainly derived from large-scale numerical calculations. The fluid volume method is used to establish the numerical models of cooperative flow type, flow focusing type and T-type microfluidic devices, and the model reliability is verified. Figure 3 As shown, the experimental data are in good agreement with the simulation data, with an error range of less than 10%. Based on this, a wide range of large-scale numerical calculations were performed by adjusting geometric parameters, physical parameters, and operating conditions to obtain a relatively complete droplet construction dataset. 1008, 839, and 1123 data sets were calculated for the cooperative flow, flow focusing, and T-type flow regimes, respectively.

[0085] The various biocompatible microspheres prepared in the following examples mainly include but are not limited to: polyethylene glycol diacrylate (PEGDA) microspheres, calcium alginate (CaAlg) microspheres, chitosan microspheres, hyaluronic acid (HAMA) microspheres, gelatin (Gelatin) microspheres, polycaprolactone (PCL) microspheres, polylactic acid (PLA) microspheres, polylactic acid-polyglycolic acid (PLGA) microspheres, poly-L-lactic acid (PLLA) microspheres, polylactic acid-polyethylene glycol (PLA16k-b-PEG5k) microspheres, and their curing methods are mainly divided into four categories: photocuring, ion crosslinking, chemical reaction, and solvent evaporation.

[0086] Example 1

[0087] In this embodiment, a meta-model for droplet size prediction is trained based on ten machine learning algorithms, and the steps are as follows:

[0088] (1) Prepare the dataset

[0089] Prepare as Figure 1-Figure 4The droplet construction dataset for the co-flow microfluidic structure used for model training is shown. The scatter plot represents the correlation between features, and the histogram shows the distribution of individual features. It can be seen that the distribution of flow rate and viscosity is more dense in regions with smaller values, while data is less abundant in regions with larger values. This is because higher flow rate and viscosity are more likely to lead to jetting, resulting in poor monodispersity of the resulting droplets. Therefore, the research scope is mainly focused on the dripping range. The dataset construction methods for flow-focusing and T-type structures are consistent with those for the co-flow structure.

[0090] (2) Data preprocessing

[0091] From the original data set, geometric parameters (capillary diameter D1, D2, D3, channel width W1, W2, W3 and channel height H), physical parameters (dispersed phase viscosity μ1, continuous phase viscosity μ2 and two-phase interfacial tension σ), operating conditions (dispersed phase flow rate Q1, continuous phase flow rate Q2) are selected as characteristic variables, and droplet size D d As the target variable, 80% of the data is randomly selected as the training set, and the remaining 20% ​​of the data is used as the test set. To ensure the stability and comparability of the feature data, the feature variables are standardized so that the numerical ranges of different features remain consistent, thereby improving the convergence speed and generalization ability of the model.

[0092] (3) Model training

[0093] In order to avoid data leakage and provide more reliable model evaluation, the nested cross-validation method is used in each model training process, where the outer layer uses 10-fold cross-validation to evaluate the performance of the model, and the inner layer also uses 10-fold cross-validation for hyperparameter optimization during model training. Compared with general cross-validation, the introduction of the nested cross-validation method can reduce the risk of model overfitting and ensure that the performance of the final model on the independent test set is more representative. Usually, hyperparameters have a direct impact on the training and generalization ability of the model. In the inner model training, in order to find the best hyperparameter combination, the Bayesian optimization algorithm is used to optimize the learning rate, number of hidden layers, number of hidden nodes, tree depth, number of trees, subsampling rate, feature sampling rate and other hyperparameters in model training. Finally, the mean absolute percentage error (MAPE) is used to evaluate the predictive ability of the model, which is calculated as follows:

[0094]

[0095] In droplet size prediction, selecting the appropriate machine learning algorithm is crucial for model accuracy and generalization. To comprehensively compare the performance of different machine learning algorithms for droplet size prediction, ten machine learning algorithms, including multilayer perceptron, decision tree, ridge regression, stochastic gradient descent regression, support vector regression, lasso regression, K-nearest neighbor regression, LightGBM, random forest, and XGBoost, were used to train the inner prediction model. The XGBoost algorithm performed exceptionally well in both the training and test sets for the co-flow, flow focusing, and T-type devices. Specifically, XGBoost achieved MAPEs of 1.25%, 0.77%, and 1.96% on the training set, significantly lower than those of other algorithms, demonstrating its superiority in capturing data features. On the test set, XGBoost achieved MAPEs of 3.66%, 4.36%, and 4.7%, respectively, also significantly lower than those of other algorithms. Therefore, the prediction models for droplets and microspheres were developed based on the XGBoost algorithm.

[0096] In this embodiment, the droplet prediction meta-models under the collaborative flow type, T-type and flow focusing type structures trained by XGBoos have good training and testing accuracy.

[0097] Example 2

[0098] In this embodiment, a strategy for improving model generalization capability by combining a meta-model with a continuous learning mechanism is proposed. The steps are as follows:

[0099] (1) Metamodel Verification

[0100] Capillary assembly of the co-flow microfluidic device was used, and soft lithography was used to prepare the flow focusing and T-type microfluidic devices to construct a droplet validation set to verify the accuracy and generalization ability of the XGBoost meta-model trained in Example 1. The results are shown in Figure 2. Figures 5-7 As shown, the MAPE of the validation set is 15.04%, 71.6% and 17.9% in co-flow, flow focusing and T-type device, respectively.

[0101] (2) Data generation

[0102] To address these issues, we propose a continuous learning mechanism based on a data generation algorithm. This mechanism aims to fine-tune the meta-model by generating new datasets with a distribution similar to the validation set, thereby improving the model's adaptability. The specific steps are as follows: First, we use the WGAN-GP (Wasserstein GAN with Gradient Penalty) algorithm to generate a new dataset with a distribution similar to the validation set.

[0103] (3) Continuous learning

[0104] The generated new data is added to the original training set to fine-tune the XGBoost meta-model. During the fine-tuning process, the model gradually optimizes its internal parameters by learning the distribution characteristics of the new dataset. This method enables the meta-model to adapt to the new data distribution without leaking the validation set data, thereby improving its prediction accuracy. Finally, the validation set data is predicted on the fine-tuned model, and the results are as follows: Figure 8 As shown in the figure, the MAPE in the collaborative flow device dropped from the original 15.04% to 4.49%, and the predictive ability of the model was greatly improved. This result shows that by generating a new data set with a similar distribution to the validation set data, the model can better adapt to the new data distribution, thereby significantly improving its predictive ability. Similarly, the meta-model trained in the flow focusing and T-type microfluidic devices was retrained using the continuous learning method, and the predictive ability of the model was also greatly improved, as shown in the figure. Figure 9-10 As shown, MAPE dropped from 71.6% and 17.9% to 5.88% and 10.0% respectively.

[0105] In order to further evaluate the effect of continuous learning mechanism on improving the generalization ability of the model, multiple sets of data points from existing literature were collected as validation sets for model validation. Figure 11As shown in the figure, without introducing the continuous learning mechanism, the meta-model predicts the mean absolute percentage errors (MAPE) of Lan (Numerical and experimental investigation of dripping and jetting flow in a coaxial micro-channel [J]. Chemical Engineering Science, 2015, 134: 76–85.), Deng (Numerical and experimental study of oil-in-water (O / W) droplet formation in a co-flowing capillary device [J]. Colloids and Surfaces A: Physicochemical and Engineering Aspects, 2017, 533: 1–8.) and Gu (Theoretical analysis of 3D emulsion droplet generation by a device using coaxial glass tubes [J]. Sensors and Actuators A: Physical, 2011, 169 (2): 326–332.) data in the co-flow device with 47.4%, 38.6% and 31.1% respectively. These high error values ​​indicate that the meta-model has certain limitations in predicting across datasets, especially when the data distribution is quite different. However, by introducing a continuous learning mechanism, the prediction performance of the model is significantly improved, such as Figure 12 As shown, MAPE dropped to 15.2%, 10.9% and 13.2% respectively.

[0106] Similarly, if Figure 13-16As shown, the flow focusing device and the T-type device were validated against literature data. For the flow focusing device, the research data of Mardani (Mapping flow-focusing microfluidic droplet formation to determine high-throughput droplet generation configurations [J]. Results in Engineering, 2023, 18: 101125.), Srikanth (Experimental studies on droplet characteristics in a microfluidic flow focusing droplet generator: effect of continuous phase on droplet encapsulation [J]. The European Physical Journal E, 2021, 44 (8): 108.) and Sartipzadeh (Controllable size and form of droplets in microfluidic-assisted devices: Effects of channel geometry and fluid velocity on droplet size [J]. Materials Science and Engineering: C, 2020, 109: 110606.) were collected for model verification, and the MAPE decreased from the original 65.9%, 40.3% and 332.4% to 5.8%, 6.2% and 9.7%, respectively.For the T-type device, the research data of Chen et al. (CFD Simulation of Droplet Formation in a Wide-Type Microfluidic T-Junction [J]. Journal of Dispersion Science and Technology, 2012, 33(11): 1635–1641.), Tarchichi et al. (New regime of droplet generation in a T-shape microfluidic junction [J]. Microfluidics and Nanofluidics, 2013, 14(1–2): 45–51.), and Odera et al. (Droplet formation behavior in a microfluidic device fabricated by hydrogel molding [J]. Microfluidics and Nanofluidics, 2014, 17(3): 469–476.) were collected to verify the model. The MAPE decreased from the original 456.1%, 306.6% and 48.9% to 6.0%, 7.6% and 6.8%, respectively. This result fully proves that the continuous learning mechanism proposed in this embodiment significantly improves the generalization ability of the metamodel, and fully proves that the strategy of metamodel plus continuous learning mechanism has a strong ability in improving the generalization ability of the model.

[0107] Example 3

[0108] In this example, an experimental method was used to construct a data set of ten biocompatible microsphere solidifications and train a microsphere size prediction model. The steps are as follows:

[0109] (1) Droplet template preparation

[0110] Ten biocompatible microsphere droplet templates were prepared using a co-flow microfluidic device. The dispersed phase employed varying microsphere concentrations to investigate their effect on subsequent curing. By adjusting the flow rates of the dispersed and continuous phases, droplet templates of varying sizes could be flexibly constructed.

[0111] (2) Construction of microsphere solidification dataset

[0112] The droplet template prepared in step (1) is solidified using a corresponding solidification method to obtain microspheres.

[0113] (3) Microsphere size prediction model training

[0114] The final microsphere size is primarily determined by factors such as the initial droplet template size, droplet phase concentration, and crosslinker concentration. Therefore, using these influencing factors as feature variables and microsphere size as the target variable, the XGBoost algorithm described in Example 1 was used to train a prediction model for each microsphere. Similarly, nested cross-validation and Bayesian optimization were employed during model training, and model performance was ultimately evaluated using MAPE. The training results for ten biocompatible microsphere size prediction models are shown in Table 1.

[0115] Table 1

[0116]

[0117] The ten biocompatible microsphere size prediction models trained in this example all have good training and testing accuracy.

[0118] Example 4

[0119] In this embodiment, an intelligent design scheme with diverse microsphere sizes is proposed, and the operation flow chart is as follows: Figure 17 The specific steps are as follows:

[0120] (1) User input and model selection

[0121] The user enters the material type and target size (D p-target ), select the corresponding microsphere size prediction model according to the material type.

[0122] (2) Target size matching and data screening

[0123] Calculate the target value D p-target D p The MAPE of the five groups of samples with the smallest MAPE is selected. If the smallest MAPE among the five groups of samples is less than 1%, the optimal sample is directly output and enters the intelligent design link of the operating condition. Otherwise, the D of the five groups of samples is d Carry out optimized design.

[0124] (3) Droplet size optimization

[0125] The predicted value D p-predicted and target value D p-target The mean absolute percentage error (MAPE) is the objective function and iterative optimization is performed. First, in the D d Based on the above, the step size is increased or decreased, and then substituted into the microsphere prediction model to calculate D p-predicted With D p-target The process will have two directions, and then continue to change D in the direction that makes MAPE decrease. d, multiple iterations are performed until the MAPE is less than the set requirement (1%), then the optimization ends. If the condition is not met after the maximum number of iterations, the optimized sample with the smallest MAPE is output. It should be noted that the result of the first step screening has made the microsphere size (D p ) is close to the predicted value, so in order to simplify the process, only D is selected d The optimized design was performed while the droplet concentration and cross-linker concentration remained unchanged.

[0126] (4) Determination of physical property parameters

[0127] According to the dispersed phase and crosslinker concentrations of the optimal sample, the measured physical property parameter library is queried to determine the physical property parameters (viscosity and interfacial tension) of the dispersed phase and continuous phase, and the physical property parameters are substituted into the droplet size prediction model.

[0128] (5) Droplet simulation data set matching and optimization

[0129] The D of the optimal sample selected in step 3 d As the target value of the droplet prediction model, 10 groups of samples closest to the droplet size are selected from the dataset of the droplet prediction model training and verification, and the physical properties of the samples are replaced with the results obtained in step 4, and substituted into the model for prediction to calculate D d-predicted With D d If there is a MAPE less than 1%, the sample with the smallest MAPE (including geometric structure, operating conditions and physical parameters) is output, which is the D that meets the user's requirements. p-target If the optimal preparation conditions are met, conduct experimental design. Otherwise, further optimization of the operating conditions is required.

[0130] (6) Optimization design of operating conditions

[0131] The optimal design of dispersed phase flow rate Q1 and continuous phase flow rate Q2 is performed for the 10 best samples after replacing the physical property parameters in step 5. Similar to the optimization idea in step 3, the predicted value D d-predicted and target value D d The mean absolute percentage error (MAPE) is used as the objective function to iteratively optimize Q1 and Q2. Based on Q1 and Q2, the step size is increased or decreased, and then substituted into the droplet prediction model to calculate D d-predicted and target value D d MAPE, since two variables need to be optimized, and the step size of each variable is increased or decreased, there will be 4 results. Then continue to change Q1 or Q2 in the direction that makes the MAPE drop the most. The optimization ends after multiple iterations until the MAPE is less than the set requirement (1%). If the condition is still not met after the maximum number of iterations, the sample with the smallest MAPE after optimization is output.

[0132] (7) Experimental verification. According to the material type and target value D given by the user p-target The optimal sample given after the final intelligent design includes the geometric parameters of the device, solution concentration, and two-phase flow operating conditions. Experiments are carried out under these conditions, and the error between the final experimental microsphere size and the target microsphere size required by the user is measured. If it is within the allowable error range, the design is completed. Otherwise, return to step 3 and re-optimize the design.

[0133] The intelligent design scheme for diversified microsphere sizes proposed in this embodiment can meet the reverse intelligent design from user required sizes to experimental conditions.

[0134] Example 5

[0135] In this embodiment, PLA microspheres with a size of 70 μm are intelligently designed as follows:

[0136] (1) Target size matching and data screening

[0137] Calculate the target value of 70 μm and the D value of the PLA microsphere data set p The mean absolute percentage error (MAPE) is calculated and the five groups of samples with the smallest MAPE are selected. If the smallest MAPE among the five groups of samples is less than 1%, the optimal sample is directly output and enters the intelligent design link of the operating condition. Otherwise, the D of the five groups of samples is d Carry out optimized design.

[0138] (2) Droplet size optimization

[0139] The predicted value D p-predicted The MAPE of the target value of 70μm is used as the objective function and iterative optimization is performed. d Based on the above, the step size is increased or decreased, and then substituted into the microsphere prediction model to calculate D p-predicted Compared with the MAPE of 70μm, the process will have two directions, and then continue to change D in the direction that makes the MAPE decrease. d , iterate multiple times until the MAPE is less than the set requirement (1%) and the optimization ends. Since the result of the first step of screening has made the D of the optimal sample p Close to the predicted value, so in order to simplify the process, only D was selected d The optimized design was performed while the droplet concentration and cross-linker concentration remained unchanged. The optimized droplet template size was 238.5 μm and the PLA solution concentration was 2%.

[0140] (3) Determination of physical property parameters

[0141] According to the PLA solution concentration corresponding to the optimal sample, the experimentally measured physical property parameter library is queried to determine the physical property parameters (viscosity and interfacial tension) of the dispersed phase and the continuous phase, and the physical property parameters are substituted into the droplet size prediction model.

[0142] (4) Droplet simulation data set matching and optimization

[0143] The D of the optimal sample selected in step 2 d = 238.5 μm as the target value of the droplet prediction model. From the dataset of droplet prediction model training and validation, 10 groups of samples with the closest droplet size are selected. The physical properties of the samples are replaced with the results obtained in step 3 and substituted into the model for prediction. D is calculated. d-predicted With D d If there is a MAPE less than 1%, the sample with the smallest MAPE (including geometric structure, operating conditions and physical parameters) is output, which is the D that meets the user's requirements. p-target If the optimal preparation conditions are met, the experimental design is performed. Otherwise, the geometric parameters and physical properties of the optimal sample are fixed and the operating conditions are further optimized.

[0144] (5) Optimization design of operating conditions

[0145] The optimal 10 groups of samples with the physical property parameters replaced in step 4 are optimized for the dispersed phase flow rate Q1 and the continuous phase flow rate Q2. Similar to the optimization idea in step 2, the predicted value D d-predicted and target value D d The mean absolute percentage error (MAPE) is used as the objective function to iteratively optimize Q1 and Q2. Based on Q1 and Q2, the step size is increased or decreased, and then substituted into the droplet prediction model to calculate D d-predicted and target value D d MAPE, since two variables need to be optimized and the step size of each variable is increased or decreased, there will be 4 results. Then Q1 or Q2 is continued to be changed in the direction that maximizes the decrease in MAPE. The optimization ends after multiple iterations until MAPE is less than the set requirement (1%). The final Q1 and Q2 are 16.95μL / min and 22.64μL / min, respectively.

[0146] (6) Experimental verification

[0147] The geometric parameters determined according to the above steps were 30 μm, 70 μm, and 300 μm, respectively. The dispersed phase and continuous phase were 2% PLA solution and 3% PVA solution, respectively. The dispersed phase and continuous phase flow rates were 16.95 μL / min and 22.64 μL / min, respectively. PLA microspheres were prepared under these experimental conditions.

[0148] The experimental results of the intelligent design of 70 μm PLA microspheres in this embodiment are as follows Figure 18 As shown in the figure, the error between the final microsphere particle size of 73.2 μm and the user-required 70 μm is 4.57%, achieving the precise preparation of 70 μm PLA microspheres.

[0149] Example 6

[0150] In this embodiment, PLA microspheres with a size of 190 μm are intelligently designed as follows:

[0151] (1) Target size matching and data screening

[0152] The target value of 190 μm was calculated and the D value of the PLA microsphere data set was calculated. p The mean absolute percentage error (MAPE) is calculated and the five groups of samples with the smallest MAPE are selected. If the smallest MAPE among the five groups of samples is less than 1%, the optimal sample is directly output and enters the intelligent design link of the operating condition. Otherwise, the D of the five groups of samples is d Carry out optimized design.

[0153] (2) Droplet size optimization

[0154] The predicted value D p-predicted The MAPE of the target value of 190μm is used as the objective function and iterative optimization is performed. d Based on the above, the step size is increased or decreased, and then substituted into the microsphere prediction model to calculate D p-predicted With a MAPE of 190 μm, the same optimization strategy as in Example 5 was used for optimization. The optimized droplet template size was 432.9 μm and the concentration of the PLA solution was 8%.

[0155] (3) Determination of physical property parameters

[0156] According to the PLA solution concentration corresponding to the optimal sample, the experimentally measured physical property parameter library is queried to determine the physical property parameters (viscosity and interfacial tension) of the dispersed phase and the continuous phase, and the physical property parameters are substituted into the droplet size prediction model.

[0157] (4) Optimization design and experimental verification of operating conditions

[0158] The same strategy as in Example 5 was used to optimize the operating conditions. The final geometric parameters were 130 μm, 160 μm, and 700 μm, the dispersed phase and the continuous phase were 8% PLA solution and 3% PVA solution, respectively. The dispersed phase and the continuous phase flow rates were 10.27 μL / min and 653.62 μL / min, respectively. PLA microspheres were prepared under these experimental conditions.

[0159] The experimental results of the intelligent design of 190 μm PLA microspheres in this embodiment are as follows Figure 18As shown in the figure, the error between the final microsphere particle size of 177.1 μm and the user-required 190 μm is 6.77%, achieving the precise preparation of 190 μm PLA microspheres.

[0160] Example 7

[0161] In this embodiment, PLA microspheres with a size of 270 μm are intelligently designed as follows:

[0162] (1) Target size matching and data screening

[0163] The target value of 270 μm was calculated and the D value of the PLA microsphere data set was calculated. p The mean absolute percentage error (MAPE) is calculated and the five groups of samples with the smallest MAPE are selected. If the smallest MAPE among the five groups of samples is less than 1%, the optimal sample is directly output and enters the intelligent design link of the operating condition. Otherwise, the D of the five groups of samples is d Carry out optimized design.

[0164] (2) Droplet size optimization

[0165] The predicted value D p-predicted The MAPE of the target value of 270μm is used as the objective function and iterative optimization is performed. d Based on the above, the step size is increased or decreased, and then substituted into the microsphere prediction model to calculate D p-predicted With a MAPE of 270 μm, the same optimization strategy as in Example 5 was used for optimization. The optimized droplet template size was 561.6 μm and the concentration of the PLA solution was 10%.

[0166] (3) Determination of physical property parameters

[0167] According to the PLA solution concentration corresponding to the optimal sample, the experimentally measured physical property parameter library is queried to determine the physical property parameters (viscosity and interfacial tension) of the dispersed phase and the continuous phase, and the physical property parameters are substituted into the droplet size prediction model.

[0168] (4) Optimization design and experimental verification of operating conditions

[0169] The same strategy as in Example 5 was used to optimize the operating conditions. The final geometric parameters were 160 μm, 200 μm, and 750 μm, the dispersed phase and the continuous phase were 10% PLA solution and 3% PVA solution, respectively. The dispersed phase and the continuous phase flow rates were 11.09 μL / min and 144.04 μL / min, respectively. PLA microspheres were prepared under these experimental conditions.

[0170] The experimental results of the intelligent design of 270 μm PLA microspheres in this embodiment are as follows Figure 18As shown in the figure, the error between the final microsphere particle size of 258.6 μm and the user-required 270 μm is 4.21%, achieving the precise preparation of 270 μm PLA microspheres.

[0171] Example 8

[0172] In this embodiment, PEGDA microspheres with sizes of 300 μm, 350 μm, 400 μm, 450 μm, and 500 μm were intelligently designed as follows:

[0173] (1) Target size matching and data screening

[0174] Calculate the D of the PEGDA microsphere data set with target values ​​of 300μm, 350μm, 400μm, 450μm and 500μm respectively p The mean absolute percentage error (MAPE) of the five groups of samples is calculated, and the five groups of samples with the smallest MAPE are selected. If the smallest MAPE among the five groups of samples is less than 1%, the optimal sample is directly output and enters the intelligent design link of the operating condition. Otherwise, the D of the five groups of samples is d Carry out optimized design.

[0175] (2) Droplet size optimization

[0176] The predicted value D p-predicted The MAPE of target values ​​300μm, 350μm, 400μm, 450μm and 500μm is used as the objective function and iterative optimization is performed. d Based on the above, the step size is increased or decreased, and then substituted into the microsphere prediction model to calculate D p-predicted With MAPE of 300 μm, 350 μm, 400 μm, 450 μm and 500 μm, the optimization was performed according to the same optimization strategy in Example 5. The optimized droplet template sizes were 316.2 μm, 362.02 μm, 425.82 μm, 471.42 μm and 517.54 μm, the PEGDA solution concentrations were 70%, 80%, 70%, 90% and 80%, respectively, and the photoinitiator concentrations were 7%, 16%, 7%, 4.5% and 12%, respectively.

[0177] (3) Determination of physical property parameters

[0178] According to the PEGDA solution concentration corresponding to the optimal sample, the physical property parameter library measured experimentally was queried to determine the physical property parameters (viscosity and interfacial tension) of the dispersed phase and the continuous phase, and the physical property parameters were substituted into the droplet size prediction model.

[0179] (4) Optimization design and experimental verification of operating conditions

[0180] The same strategy as in Example 5 was used to optimize the operating conditions and intelligently design PEGDA microspheres of different sizes to obtain the final experimental conditions and experimental results shown in Table 2. PEGDA microspheres were prepared under these experimental conditions.

[0181] Table 2

[0182]

[0183] The experimental results of the intelligent design of PEGDA microspheres with sizes of 300 μm, 350 μm, 400 μm, 450 μm and 500 μm are as follows: Figure 20 As shown in the figure, the precise preparation of PEGDA microspheres of 300 μm, 350 μm, 400 μm, 450 μm and 500 μm was achieved.

[0184] Example 9

[0185] In this embodiment, calcium alginate microspheres with sizes of 50 μm, 60 μm, 70 μm, 80 μm and 90 μm are intelligently designed as follows:

[0186] (1) Target size matching and data screening

[0187] Calculate the D of the calcium alginate microsphere data set with target values ​​of 50μm, 60μm, 70μm, 80μm and 90μm respectively p The mean absolute percentage error (MAPE) of the five groups of samples is calculated, and the five groups of samples with the smallest MAPE are selected. If the smallest MAPE among the five groups of samples is less than 1%, the optimal sample is directly output and enters the intelligent design link of the operating condition. Otherwise, the D of the five groups of samples is d Carry out optimized design.

[0188] (2) Droplet size optimization

[0189] The predicted value D p-predicted The MAPE of target values ​​50μm, 60μm, 70μm, 80μm, 90μm and 100μm is used as the objective function and iterative optimization is performed. d Based on the above, the step size is increased or decreased, and then substituted into the microsphere prediction model to calculate D p-predictedThe MAPEs of 50 μm, 60 μm, 70 μm, 80 μm, 90 μm and 100 μm were optimized according to the same optimization strategy as in Example 5. The optimized droplet template sizes were 170.7 μm, 228.8 μm, 216.9 μm, 261.5 μm, 293.6 μm and 325.7 μm, and the concentrations of NaAlg solution were 1%, 1.5%, 2%, 2%, 2% and 2%, respectively, and the concentrations of EDTA were 2%, 1.5%, 2%, 3%, 2% and 3%, respectively.

[0190] (3) Determination of physical property parameters

[0191] According to the NaAlg solution concentration corresponding to the optimal sample, the experimentally measured physical property parameter library is queried to determine the physical property parameters (viscosity and interfacial tension) of the dispersed phase and the continuous phase, and the physical property parameters are substituted into the droplet size prediction model.

[0192] (4) Optimization design and experimental verification of operating conditions

[0193] The same strategy as in Example 5 was used to optimize the operating conditions and intelligently design CaAlg microspheres of different sizes to obtain the final experimental conditions and experimental results shown in Table 3. CaAlg microspheres were prepared under these experimental conditions.

[0194] Table 3

[0195]

[0196]

[0197] The experimental results of Example 9 are as follows: Figure 21 As shown, the precise preparation of calcium alginate microspheres was achieved.

[0198] The above solution is only an illustration of a preferred embodiment, but is not limited thereto. When implementing the present invention, appropriate replacements and / or modifications can be made according to user needs.

[0199] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and exemplary embodiments. They can be applied to a variety of fields suitable for the present invention. Further modifications will be readily apparent to those skilled in the art. Therefore, the present invention is not limited to the specific details and illustrations shown and described herein without departing from the general concept defined by the claims and their equivalents.

Claims

1. An intelligent and precise design method for microfluidic droplets and diversified microspheres, characterized in that: include: S1. Based on the material type and microsphere size input by the user, the corresponding microsphere size prediction model is selected according to the material type to guide the reverse search and iterative correction to obtain the system concentration and droplet template size required for preparing microspheres using microfluidic devices, and complete the prediction from microspheres to droplets; S2. Based on the system concentration obtained in S1, obtain the corresponding physical property parameters from the physical property library; S3, using the droplet template size obtained in S1 as the target value and the physical property parameter variables obtained in S2 as the limiting values, and using the droplet size prediction model to guide reverse search and iterative correction to obtain the experimental parameter combination for preparing microspheres of predetermined size using a microfluidic device; Wherein, in S2, the physical property parameters refer to the viscosity and interfacial tension of the dispersed phase and the continuous phase fluid; The experimental parameter combination includes: geometric parameters and operating conditions.

2. The intelligent and precise design method for microfluidic droplets and diversified microspheres according to claim 1, characterized in that: In S1, the droplet size prediction model is constructed as follows: S10, constructing a numerical simulation model of a droplet microfluidic device, adjusting droplet construction parameters in the numerical simulation model to obtain corresponding droplet sizes through simulation calculations, and thereby obtaining a droplet simulation data set for droplet prediction; S11, using the geometric parameters, physical properties, and operating conditions in the droplet simulation dataset as feature variables and the droplet size as the target variable to perform model training to obtain an XGBoost meta-model for droplet size prediction; S12. Use experimental data to verify the XGBoost meta-model; S13. Use the data generation algorithm to continuously learn the XGBoost meta-model to obtain a droplet size prediction model.

3. The intelligent and precise design method for microfluidic droplets and diversified microspheres according to claim 2, characterized in that: In S10, the droplet construction parameters include: geometric parameters, physical parameters and operating conditions; The droplet microfluidic device includes: a coordinated flow device, a T-type device and a flow focusing device.

4. The intelligent and precise design method for microfluidic droplets and diversified microspheres according to claim 2, characterized in that: In S13, the data generation algorithm uses the WGAN-GP algorithm to generate a new data set with a distribution similar to that of the validation set data; The continuous learning refers to adding new datasets to the training set of the droplet simulation dataset to fine-tune the XGBoost meta-model.

5. The intelligent and precise design method for microfluidic droplets and diversified microspheres according to claim 1, characterized in that: In S3, the reverse search method is: Based on the droplet template size obtained in S1, search for the five closest sets of data in the validation set data of the droplet prediction model; Based on the physical property parameters obtained by S2, the operating conditions are optimized to obtain the corresponding experimental parameter combination in the droplet prediction model with the goal of minimizing the mean absolute percentage error of the droplet template size.

6. The microsphere reverse design method for determining experimental parameters based on size and material system according to claim 1, characterized in that: In S3, the process of constructing the microsphere size prediction model includes: S30, using a co-flow microfluidic device to prepare droplet templates of various biocompatible microspheres; S31, solidifying the droplet templates obtained in S30 using corresponding solidification methods to obtain corresponding microsphere solidification data sets; S32. Model training is performed based on the microsphere solidification dataset to obtain a variety of microsphere size prediction models that match a variety of biocompatible microspheres based on the XGBoost algorithm.

7. The intelligent and precise design method for microfluidic droplets and diversified microspheres according to claim 6, characterized in that: The various biocompatible microspheres include: polyethylene glycol diacrylate microspheres, calcium alginate microspheres, chitosan microspheres, hyaluronic acid microspheres, gelatin microspheres, polycaprolactone microspheres, polylactic acid microspheres, polylactic acid-polyglycolic acid microspheres, poly-L-lactic acid microspheres, and polylactic acid-polyethylene glycol microspheres.

8. The intelligent and precise design method for microfluidic droplets and diversified microspheres according to claim 6, characterized in that: The curing methods include: photocrosslinking, ion crosslinking, chemical crosslinking, and solvent volatilization.

Images