Reverse design and precise preparation method of micro-spiral driven by visual identification and machine learning

Abstract

The invention belongs to the field of micro-spiral design and preparation, and provides a visual identification and machine learning driven micro-spiral reverse design and precise preparation method, which comprises the following steps: collecting micro-fluid pictures under a series of operation conditions, putting a micro-fluid picture set marked with a flow state into a YOLO model to obtain a flow state identification model, inputting the micro-spiral picture set with the marked size into a YOLO model to obtain a size detection model, inputting a database of flow state data and operation working conditions into an SVM algorithm to obtain a flow state prediction model, and inputting a database of size data and working condition parameters into an XGBoost algorithm to obtain a size prediction model; reverse design of the micro-spiral is realized by combining the size prediction model and the Pareto algorithm, and the optimal operation condition is inversely calculated according to the target size of the micro-spiral, so that accurate preparation of the micro-spiral is realized. According to the method, a traditional micro-spiral forward circulation design method is completely innovated, and high-efficiency and high-precision reverse design and precise preparation of the micro-spiral are achieved.

Description

Technical Field

[0001] This invention belongs to the field of microhelical design and fabrication, and relates to a method for reverse design and precise fabrication of microhelices driven by visual recognition and machine learning. Background Technology

[0002] Microhelices, exhibiting periodically twisted three-dimensional microstructures at the microscale, demonstrate significant value in numerous fields due to their unique geometry (Nature, 2023, 613: 7945). In biomedicine, the helical morphology of microhelices, similar to the natural structure of microvessels, reduces their resistance to movement, enabling targeted delivery of drugs or cells within complex biological networks and providing an ideal carrier for precision medicine (Engineering, 2024, 41: 172-178). In microcomputer systems, microhelices, with their excellent elastic deformation capabilities and programmable mechanical response characteristics, are ideal building blocks for constructing miniature sensors and actuators, enabling intelligent sensing and actuation of external environments such as magnetic fields and temperature (Small, 2025, 18: 21). In the development of novel functional materials, microhelices with specific structural parameters can serve as basic units for photonic crystals or supermaterials. By precisely controlling these structural parameters, the propagation characteristics of light and sound waves can be manipulated, laying the foundation for next-generation optical devices (Adv. Funct. Mater., 2019, 29, 1807934). Custom-made and precisely fabricated microhelices with specific structural parameters is key to achieving these applications. The structural parameters of microhelices, such as pitch, diameter, and amplitude, are controlled by factors such as the channel size of the microfluidic device, the viscosity of the fluid used in fabrication, the flow rate ratio, and interfacial tension. These factors exhibit strong nonlinear and multi-physics coupling relationships. The structural parameters of microhelices significantly impact their mechanical strength, deformation characteristics, and motion performance. For example, in drug delivery, an excessively small pitch leads to low delivery efficiency, while an excessively small diameter results in insufficient drug loading. In microcomputer systems, the structural parameters of microhelices directly affect their propulsion efficiency and control precision in specific fluid environments, while the amplitude determines their adaptability and stability within the manipulation space. Therefore, the precise control of microhelical structure parameters is not merely a matter of manufacturing precision, but a crucial issue connecting the fabrication process, end-use performance, and application success. Thus, developing methods for precisely fabricating microhelices is of great significance for improving their performance and effectiveness.

[0003] Currently, researchers have developed various methods for preparing microhelices based on microfluidic technology. Tottori et al. (RSC Adv., 2015, 5: 33691) precisely prepared calcium alginate (CaAlg) microhelices by adding sodium citrate (TSC) to the internal phase solution to regulate the reaction rate of sodium alginate (NaAlg) and calcium chloride (CaCl2). However, because the formation process of the microhelices involves transient processes with strong coupling of multiple physical fields, such as multiphase flow, interfacial instability, and rapid solidification, the microhelices prepared by this method have poor performance in terms of helical morphology and uniformity. Xu et al. (Adv. Mater., 2017, 29:1701664) constructed CaAlg microhelices using the ectopic coiling effect and established an empirical formula for flow rate and size. This method can provide rapid result prediction within a certain operating window and avoid complex theoretical calculations. However, due to its strong nonlinearity, it is difficult to apply to complex operating conditions and cannot be used to guide the design of microhelices with multi-parameter coordinated changes. Ma et al. (ACS Appl. Mater. Interfaces, 2021, 13: 59392-59399) enhanced the focused flow of the internal NaAlg microfluidic phase by designing a sleeve in the injection tube, thus improving the stability of microspiral preparation. While the analysis of the flow mechanism provides guidance for microspiral size control, it cannot be applied to precise size design. Furthermore, conducting full-parameter combined scanning experiments faces the problem of high cost. Liu et al. (Engineering, 2024, 41: 172-178) constructed CaAlg microspirals based on a two-step crosslinking method. They precisely controlled the microspiral size using the traditional "preparation-characterization-re-preparation" model. However, this method relies on empirical "trial and error" experiments, which has significant drawbacks such as long cycle time and high cost, making it difficult to achieve active design. In particular, existing methods focus on "forward prediction" from operating conditions to microhelical structure parameters, while the ability to "reverse design"—solving for optimal operating conditions based on target dimensions—is generally lacking. This leads to inefficiency in practical applications, requiring repeated iterative trial and error to approximate the design target, failing to balance prediction accuracy and design efficiency. This severely restricts the efficient development and application of microhelical devices for specific high-performance requirements. Therefore, reverse-engineering the target structure of microhelices to design operating conditions for precise fabrication of microhelices presents a significant challenge. Summary of the Invention

[0004] Traditional microhelix fabrication methods cannot reverse engineer optimal operating conditions based on target size to achieve reverse design of the target microhelix. In practical applications, repeated iterative trial and error are required to approximate the design target, which cannot balance prediction accuracy and design efficiency. This invention provides a visual recognition and machine learning-driven method for reverse design and precise fabrication of microhelices. The method automatically detects the flow state and size of microhelices using a visual recognition (YOLO) model, establishes a microhelix flow state and size prediction model using a support vector machine classification (SVM) algorithm and an extreme gradient boosting (XGBoost) algorithm, and then reverse engineers the operating conditions for the target size based on a multi-objective optimization algorithm to achieve high-efficiency and high-precision reverse design of microhelices.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by the present invention is as follows:

[0006] A reverse design method for microhelices driven by visual recognition and machine learning includes the following steps:

[0007] S1, Acquisition of microfluidic images

[0008] The microfluidic device is operated under different operating conditions to construct microfluidics with different flow regimes and acquire optical images of the microfluidics; the operating conditions include the viscosity of the inner phase fluid, the viscosity of the outer phase fluid, the inner diameter of the injection tube cone, the inner diameter of the receiving tube, the flow velocity of the inner phase fluid, and the flow velocity of the outer phase fluid; the flow regimes include blockage, spiral, and jet.

[0009] S2, Construction of the Flow Recognition Model

[0010] Images are collected according to the method in step S1 and then fed into CVAT software for flow regime annotation. The annotated images are then fed into the YOLO object detection model for training to obtain a flow regime recognition model. The flow regime recognition model is then used to perform flow regime recognition on the images collected in step S1, and images with a spiral flow regime are selected.

[0011] S3, Construction of the Dimension Inspection Model

[0012] The images of the spiral microfluidic selected in step S2 are input into CVAT software for size annotation. The annotated images are then input into the YOLO object detection model for training to obtain the size detection model.

[0013] S4, Construction of the Flow Prediction Model

[0014] The flow pattern recognition model is used to identify the flow pattern of the images collected according to the method in step S1, and different flow pattern values ​​are assigned to the blockage, spiral and jet flow patterns. The operating conditions and the corresponding flow pattern values ​​are constructed into flow pattern data bars. All flow pattern data bars are fed into the SVM algorithm for training to obtain the flow pattern prediction model.

[0015] S5, Construction of Size Prediction Model

[0016] The images of the helical microfluidic flow selected in step S2 are input into the size detection model for size detection to obtain the pitch, diameter and amplitude data of the micro-helical flow. The operating conditions are respectively constructed into pitch, diameter and amplitude data bars with the corresponding pitch, diameter and amplitude size data. All pitch data bars, diameter data bars and amplitude data bars are respectively input into the XGBoost algorithm for training to obtain pitch prediction model, diameter prediction model and amplitude prediction model.

[0017] S6, reverse design of micro-helices

[0018] S61, set the target size and target size range of the micro-spiral, set the boundary conditions of the operating conditions, and standardize the operating conditions; based on the pitch prediction model, diameter prediction model and amplitude prediction model, combined with the multi-objective optimization algorithm, back-calculate the operating conditions that can achieve the target size, and put the back-calculated operating conditions into the flow prediction model for flow verification.

[0019] S62, if an operating condition passes the flow regime verification, then the operating condition that passes the flow regime verification is used as a candidate operating condition; if no operating condition passes the flow regime verification, then the boundary conditions of the operating condition are modified, the operating condition is standardized again, and the operation of step S61 is repeated until a candidate operating condition that can pass the flow regime verification is obtained.

[0020] S63, input the candidate operating conditions into the size prediction model to predict the size, calculate the total average absolute percentage error between the predicted size and the target size, and take the candidate operating conditions with the total average absolute percentage error less than the set threshold as the feasible operating conditions.

[0021] In the above-mentioned reverse design method, during flow regime verification, the back-calculated operating conditions are input into the flow regime prediction model for fluid prediction. If the flow regime prediction result is a spiral, then the operating condition passes the flow regime verification.

[0022] In the above reverse engineering method, step S61 includes the following steps:

[0023] S611 sets the target size and target size range of the micro-spiral according to user needs, sets the boundary conditions of the operating conditions, and standardizes the operating conditions.

[0024] S612 uses pitch prediction model, diameter prediction model and amplitude prediction model as surrogate models. It uses Pareto algorithm to search in the parameter space of operating conditions and back-calculates the optimal solution set of operating conditions that can achieve the target size. The operating conditions in the optimal solution set are then put into the flow prediction model for flow verification.

[0025] In the above reverse design method, the target dimensions of the micro-spiral include pitch, diameter, and amplitude, and the target size range includes pitch range, diameter range, and amplitude range.

[0026] In the above reverse design method, when standardizing the operating conditions in step S61, the mean and standard deviation of the operating conditions are set.

[0027] In the above reverse engineering method, the operation of the total average absolute percentage error between the predicted size and the target size in step S63 is as follows: The candidate operating conditions are input into the pitch prediction model, diameter prediction model, and amplitude prediction model to predict the pitch, diameter, and amplitude of the micro-spiral flow; the average absolute percentage error between the predicted pitch and the target pitch of the micro-spiral flow is calculated; the average absolute percentage error between the predicted diameter and the target diameter of the micro-spiral flow is calculated; and the average absolute percentage error between the predicted amplitude and the target amplitude of the micro-spiral flow is calculated. These three average absolute percentage errors are summed to obtain the total average absolute percentage error between the predicted size and the target size.

[0028] In step S63 of the above reverse design method, the feasible operating conditions are sorted in ascending order according to the total average absolute percentage error. The smaller the total average absolute percentage error, the better the corresponding feasible operating condition.

[0029] In step S63 of the above reverse engineering method, the set threshold is determined according to actual application requirements. Typically, the set threshold refers to the total average absolute percentage error being 10%. Further, in step S63, the feasible working condition where the total average absolute percentage error is less than the set threshold and the total average absolute percentage error is minimized is taken as the optimal working condition.

[0030] In the above reverse engineering method, when training the flow recognition model and size detection model in steps S2 to S3, the number of images required for training is at least 1000; when training the flow prediction model, pitch prediction model, diameter prediction model and amplitude prediction model in steps S4 to S5, the number of flow data bars, pitch data bars, diameter data bars and amplitude data bars required for training is at least 1000.

[0031] In the above reverse design method, step S1 involves setting up orthogonal experiments under different operating conditions when constructing microfluidics with different flow states using a microfluidic device.

[0032] In the above reverse design method, when constructing a microfluidic with a spiral flow pattern in step S1, the construction is carried out in the microfluidic device based on the principle of the rope-winding effect caused by the viscosity difference and velocity difference between the inner and outer phase fluids.

[0033] In the aforementioned reverse engineering approach, after training the flow regime recognition model, size detection model, flow regime prediction model, and size prediction model, the trained models are validated. A model is considered to have satisfactory performance when its performance in real-world application scenarios meets expectations. Typically, when training the flow regime recognition model using the YOLO object detection model, the model with the largest average mean accuracy (mAP) and the largest diagonal value of the chaos matrix is ​​used as the flow regime recognition model; when training the size detection model using the YOLO object detection model, the model with the largest average mean accuracy (mAP) and the largest diagonal value of the chaos matrix is ​​used as the size detection recognition model; when training the flow regime prediction model using the SVM algorithm, the model with the highest average test accuracy is used as the flow regime prediction model; when training the pitch prediction model, diameter prediction model, and amplitude prediction model using the XGBoost algorithm, the model with the smallest mean absolute percentage error (MAPE) and the largest coefficient of determination (R²) is used. 2 Larger models are size prediction models.

[0034] In the above reverse design method, in step S5, before training begins, the pitch, diameter, and amplitude data bars in the pitch, diameter, and amplitude database are preprocessed by normalization; in step S63, after obtaining the feasible operating conditions, the feasible operating conditions are inversely normalized to obtain the final feasible operating conditions.

[0035] Based on the above-mentioned reverse design method, this invention also provides a precise preparation method for microhelices driven by visual recognition and machine learning, comprising the following steps:

[0036] The feasible operating conditions are obtained by using the above reverse design method. The microfluidic device is operated under the feasible operating conditions to construct a microspiral flow. Ultraviolet light is applied to the position where the microspiral flow in the microfluidic device has reached a stable state to induce the photopolymerization reaction of the photopolymerizable polymer monomers in the microspiral flow, thus obtaining a microspiral.

[0037] The operating conditions include the viscosity of the inner phase fluid, the viscosity of the outer phase fluid, the inner diameter of the injection tube cone, the inner diameter of the receiving tube, the flow rate of the inner phase fluid, and the flow rate of the outer phase fluid; the inner phase fluid is an aqueous solution containing photopolymers, photoinitiators, and polymers, and the outer phase fluid is an aqueous solution of water or small molecule substances.

[0038] In the above-mentioned precise preparation method, the viscosity of the inner phase fluid is greater than that of the outer phase fluid. The viscosity of the inner phase fluid is adjusted by regulating the concentration of polymers and photopolymers in the inner phase fluid, while the viscosity of the outer phase fluid is adjusted by regulating the concentration of small molecules in the outer phase fluid.

[0039] In the above-mentioned precise preparation method, the polymer includes sodium alginate, sodium carboxymethyl cellulose, or polyvinyl alcohol; the small molecule includes ethylene glycol or glycerol.

[0040] In the above-mentioned precise preparation method, when constructing the micro-spiral flow, the principle of the rope-winding effect caused by the viscosity difference and velocity difference between the inner and outer phase fluids is used to construct the micro-spiral flow in a free-flow state in the microfluidic device.

[0041] To verify the accuracy of the reverse design model, 10 sets of target structural dimensions were randomly selected. Microhelices of various structural dimensions were prepared in the PEGDA / NaAlg-H2O system based on the optimal operating conditions obtained from the solution. The prepared microhelices exhibited good helical morphology and uniform structure with small errors, demonstrating the excellent design accuracy of the reverse design method described in this invention. The deep integration of machine learning and microfluidics technologies not only achieved real-time identification of microhelical flow patterns and accurate detection of structural dimensions, but also enabled the prediction of microfluidic flow patterns and the structural dimensions of microhelical flows. Furthermore, it enabled the reverse design of target microhelices, effectively solving the prominent problems of low design efficiency, long design cycles, and high costs associated with current microhelice designs. This invention is expected to provide a technical path for the efficient development of microhelical devices, promote the deep application of microhelices in biomedicine and other fields, and provide a valuable reference for the widespread application of machine learning in the field of microfluidics.

[0042] Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0043] 1. This invention provides a reverse design method for micro-spirals driven by visual recognition and machine learning. This method automatically detects the flow regime and size of micro-spiral flows using the YOLO model, and simultaneously establishes predictive models of the flow regime and size of micro-spiral flows using SVM and XGBoost algorithms. Finally, it performs reverse design based on the Pareto algorithm for feasible operating conditions of the target size. This method overcomes the problems of high time consumption, low efficiency, and low accuracy associated with traditional manual recognition and detection. Furthermore, by using a coupled model, it comprehensively reflects the influence and extent of various operating conditions on the structural size of the micro-spiral flow, thus significantly improving research efficiency and experimental progress. The use of an intelligent design model solves the trial-and-error problem in traditional design and has important guiding significance for the identification and modeling of multi-target features. At the same time, this method completely revolutionizes the traditional forward loop design method, greatly improving the design accuracy and research efficiency of micro-spirals.

[0044] 2. The flow pattern recognition model constructed in this invention demonstrates excellent precision and confidence in recognizing blockage, spiral, and jet flow patterns. Coupled with high-speed camera software, this model can monitor the online flow pattern of microfluidics in real time, potentially enabling real-time early warning and automatic error correction to reduce the risk of blockage. Simultaneously, the size detection model constructed in this invention also demonstrates excellent precision and confidence in detecting pitch, diameter, and amplitude in micro-spiral flow images. In particular, the size detection model can directly extract size parameters from large-scale images; this automation strategy significantly improves detection efficiency and reduces the cost of manual inspection.

[0045] 3. The flow prediction model constructed in this invention can be used for preliminary verification of design results, especially for high-risk clogging flow patterns. Preliminary verification can avoid operating conditions that easily lead to blockage in microchannels, thus preventing the occurrence of microchannel blockage. This can significantly improve design efficiency and the normal utilization rate of the reverse design method device for microfluidic vision recognition and machine learning driven micro-spirals. In addition, the boundary equation of the flow spectrum can also be established based on the fluid prediction model to study the law of flow pattern distribution.

[0046] 4. This invention experimentally verifies that the constructed pitch, diameter, and amplitude prediction model has good prediction accuracy. It can achieve real-time and rapid size prediction for helical flow within the boundary conditions of the operating conditions. The pitch, diameter, and amplitude prediction model greatly improves the design efficiency of micro-helices. By making predictions in advance, it can solve the problems of low efficiency and high cost in traditional trial-and-error research, effectively save experimental costs and human resources, and also play an important role in improving design accuracy.

[0047] 5. Based on the reverse design method of this invention, this invention also provides a precise preparation method for microhelices driven by visual recognition and machine learning. Following the reverse design method described in this invention, feasible operating conditions are obtained. A microfluidic device is operated under these feasible operating conditions to construct a microhelical flow. Ultraviolet light is applied to a position within the microfluidic device where the microhelical flow has reached a stable morphology to induce photopolymerization of the photopolymerizable polymer monomers in the microhelical flow, thereby obtaining microhelices. This method can achieve precise preparation of microhelices with specific structural dimensions, providing an important approach for customized microhelices, and especially offering a crucial solution for the application of microhelices in the biomedical field. Attached Figure Description

[0048] Figure 1This is a schematic diagram of the reverse design method of micro-spirals driven by visual recognition and machine learning as described in this invention, wherein (a) Model_SD in the figure represents the flow recognition model, (b) Model_DD in the figure represents the size detection model, (c) Model_SP in the figure represents the flow prediction model, and (d) Model_DP in the figure represents the size prediction model.

[0049] Figure 2 The rheological characteristics are shown in sodium alginate (NaAlg, Figure a) and glycerol (Glycerol, Figure b) solutions.

[0050] Figure 3 This document describes the construction of a fluid dynamics recognition model and the analysis of its recognition performance. Figure a shows microfluidic images under normal conditions and under conditions of exposure, staining, impurities, low light, scaling, and deflection. Figure b shows the results of fluid dynamics prediction using the fluid dynamics recognition model on images in the fluid dynamics validation image set. Figure c shows the distribution of the precision detection rate for fluid dynamics recognition on images in the fluid dynamics validation image set. Figure d shows the confidence statistics of images where the fluid dynamics recognition model can correctly identify the fluid dynamics type.

[0051] Figure 4 This diagram illustrates the impact of model parameters used during the training of the fluid recognition model on the training effect. Figure a shows the influence of training batch and training epoch on the training effect (mAP value), with Lr0 = 0.001. Figures b to d show the influence of different learning rates (Lr0) on the training effect (mAP value) of the fluid recognition model. The Lr0 values ​​corresponding to figures b to d are 0.005, 0.010, and 0.025, respectively. In the figures, the size of the bubbles represents the high and low levels of mAP values ​​between 0.80 and 0.93.

[0052] Figure 5 This relates to the impact of the training sample size on the training performance of the fluid recognition model.

[0053] Figure 6 The results show the detection performance of the size detection model. Figure a shows the size detection performance of micro-spiral images at different scales. The blue line in the figure represents the pitch, the green line represents the diameter, and the red line represents the amplitude. Figures b to d show the errors between the actual size and the predicted size of the micro-spiral flow in terms of pitch, diameter, and amplitude.

[0054] Figure 7The figures show the prediction performance of the flow pattern prediction model and the size prediction model trained using the XGBoost algorithm. Figure a shows the flow pattern distribution and precision rate of the flow pattern prediction model on the validation dataset. Figure b shows the confidence statistics of the images in which the flow pattern prediction model can correctly predict the flow pattern type. Figure c shows the MAPE values ​​of the pitch, diameter, and amplitude prediction model trained using the XGBoost algorithm for micro-helical flow, comparing the predicted values ​​with the experimental values. Figures d to f show the deviation analysis and error statistics of the predicted values ​​of pitch, diameter, and amplitude prediction model trained using the XGBoost algorithm for micro-helical flow, comparing the predicted values ​​with the experimental values.

[0055] Figure 8 The R-squared values ​​of pitch, diameter, and amplitude of micro-helical flow are compared with experimental values ​​by a pitch, diameter, and amplitude prediction model trained based on the XGBoost algorithm. 2 value.

[0056] Figure 9 The figures show the data distribution of the operating conditions in the experimental design. Figure a shows the viscosity distribution of the internal and external phase fluids, Figure b shows the distribution of the inner diameter of the injection tube cone and the inner diameter of the receiving tube, and Figure c shows the flow velocity distribution of the internal and external phase fluids.

[0057] Figure 10 This is the size distribution of the micro-spiral flow constructed in Example 5, where figures a to c represent pitch, diameter, and amplitude, respectively.

[0058] Figure 11 The comparison shows the size prediction performance of pitch, diameter, and amplitude prediction models obtained from 10 different machine learning algorithms on micro-spiral flow. Figures a to c represent pitch, diameter, and amplitude, respectively.

[0059] Figure 12 These are SHAP analysis diagrams showing the influence of operating conditions on the structural parameters of the micro-helical flow. Diagrams a, c, and e are SHAP analysis diagrams showing the influence of operating conditions on pitch, diameter, and amplitude, while diagrams b, d, and f are weighted ranking diagrams showing the influence of operating conditions on pitch, diameter, and amplitude.

[0060] Figure 13 This is a comprehensive analysis of the impact of operating conditions on the structural parameters of the micro-spiral flow. Figure a shows the effect of the internal and external fluid velocities on the pitch; Figure b shows the effect of the internal and external fluid viscosities on the pitch; Figure c shows the effect of the inner diameter of the injection tube cone and the viscosity of the internal fluid on the diameter; and Figure d shows the effect of the inner diameter of the injection tube cone and the inner diameter of the receiving tube on the amplitude.

[0061] Figure 14Figure a shows the distribution of predicted pitch and diameter values ​​for each candidate operating condition, as well as the total average absolute percentage error (Sum of MAPE(%)) for each candidate operating condition. Figure 14 Figure b shows the distribution of predicted pitch and amplitude values ​​for each candidate operating condition, as well as the total mean absolute percentage error (Sum of MAPE (%)) for each candidate operating condition. The data points in the figure represent different candidate operating conditions. The darker the color of the data point, the smaller the total mean absolute percentage error, and the better the effect of the corresponding operating condition.

[0062] Figure 15 The screenshots show the interfaces of the four models constructed from graph a, where graphs a1 to a4 correspond to the flow regime recognition model, size detection model, flow regime prediction model, and size prediction model, respectively. Figure 15 Image b is an optical image of the microspiral fabricated under a series of operating conditions. Figure 15 The c-to-e diagram is an error analysis of the experimental values ​​and target values ​​of pitch, diameter, and amplitude under a series of operating conditions.

[0063] Figure 16 This is a uniformity analysis of a group of microspirals prepared in Example 7, where figures a to c show the size distribution of pitch, diameter, and amplitude, respectively. Detailed Implementation

[0064] The following examples further illustrate the reverse design and precise fabrication method of microhelices driven by visual recognition and machine learning provided by this invention. It should be noted that the following examples are only for further illustration of this invention and should not be construed as limiting the scope of protection of this invention. Any non-essential improvements and adjustments made by those skilled in the art based on the above description of the invention to implement it are still within the scope of protection of this invention.

[0065] In the following embodiments, the primary microfluidic device used is a capillary coaxial device, including an injection tube, a receiving tube, and a connecting tube. The outlet of the injection tube is tapered, and the outlet of the injection tube is inserted into the inlet of the receiving tube. The injection tube and the receiving tube are connected by the connecting tube. The injection tube, connecting tube, and receiving tube are arranged coaxially and used in conjunction with an injection pump and an ultraviolet point light source. The injection tube is a cylindrical glass capillary tube. The tail of the cylindrical glass capillary tube is drawn into a conical shape using a needle puller, and then sanded on sandpaper until the inner diameter of the conical opening is approximately 60~130 μm (i.e., the inner diameter at the outlet of the injection tube or the inner diameter of the conical opening of the injection tube is 60~130 μm). The outer diameter of its cylindrical section is 960 μm, and the inner diameter is 700 μm. The connecting tube is a square glass tube. After the square glass tube is cut to the required length, both ends are sanded smooth. A square through hole with a cross-sectional dimension of 1×1 mm is provided in the center of the tube. The receiving tube is a cylindrical glass capillary. After cutting the cylindrical glass capillary to the required length, both ends are sanded smooth. After preparing the injection tube, connecting tube, and receiving tube, they need to be ultrasonically vibrated in anhydrous ethanol and deionized water for 60 seconds to clean impurities, and then dried with nitrogen. The tail end of the injection tube is inserted into the head of the receiving tube and connected by the connecting tube. The injection tube, receiving tube, and connecting tube are arranged coaxially. The non-inlet end of the connecting tube is sealed with AB glue. A flat-tipped needle is fixed at the inlet end of the connecting tube with AB glue, and a flat-tipped needle is fixed at the inlet end of the injection tube with AB glue. The flat-tipped needle is connected to the injection pump through a polyethylene tube. An ultraviolet point light source is placed in the middle of the receiving tube to apply ultraviolet light to irradiate the receiving tube, initiating a photopolymerization reaction of the polymer monomers in the micro-spiral flow inside the receiving tube. The micro-spirals formed after photopolymerization are naturally discharged from the outlet end of the receiving tube, and then cleaned with deionized water.

[0066] Example 1

[0067] In this embodiment, a method for constructing microfluidics is provided, comprising the following steps:

[0068] (1) Prepare internal phase fluid and external phase fluid

[0069] Preparation of internal phase fluid: Sodium alginate (NaAlg) and Rhodamine B are dissolved in deionized water to obtain a high-viscosity internal phase fluid; in the internal phase fluid, the mass fraction of NaAlg is 1.0 ~ 2.2% and the mass fraction of Rhodamine B is 0.1%.

[0070] Preparation of external phase fluid: Deionized water is used as the external phase fluid, or glycerol is dissolved in deionized water to form the external phase fluid; the mass fraction of glycerol in the external phase fluid is 0~50%.

[0071] Figure 2Figure a shows the effect of NaAlg concentration on the viscosity of the internal phase fluid. Figure 2 Figure b shows the effect of Glycerol concentration in the external phase fluid on the viscosity of the external phase fluid, reflecting the rheological characteristics of the internal and external phase fluids.

[0072] (2) Constructing micro-spiral flow

[0073] A microfluidic system is constructed using a single-stage microfluidic device as described above. The device is vertically positioned with the injection tube above the receiving tube, and a receiving container filled with the receiving liquid placed below the receiving tube, ensuring the outlet of the receiving tube is below the liquid surface in the receiving container. The internal and external phase fluids are continuously injected into the injection and receiving tubes respectively using a constant-flow syringe pump to form the microfluidic system. Depending on the operating conditions, the resulting microfluidic system exhibits flow patterns including blocked, helical, and straight.

[0074] In this step, orthogonal experiments were designed to simulate different operating conditions, including internal phase fluid viscosity, external phase fluid viscosity, internal phase fluid flow rate, external phase fluid flow rate, and the inner diameter of the receiving tube and the inner diameter of the injection tube conical orifice. The internal phase fluid flow rate was controlled at 10–200 μL / min, and the external phase fluid flow rate at 100–1000 μL / min. When gradient adjustment of the internal or external phase fluid flow rates was required, the external phase fluid flow rate was typically adjusted by gradient adjustment while keeping the internal phase fluid flow rate constant, or vice versa. When gradient adjustment of the internal and external phase fluid flow rates, the internal phase fluid flow rate gradient was 10 μL / min, and the external phase fluid flow rate gradient was 50 μL / min. The inner diameters of the injection tube conical orifice were 60, 70, 80, 90, 100, 110, 120, and 130 μm, and the inner diameters of the receiving tube were 700, 1000, and 1200 μm, respectively.

[0075] In subsequent embodiments, when the prepared microfluidic does not require solidification, the microfluidic is constructed according to the method of this embodiment.

[0076] Example 2

[0077] In this embodiment, a method for constructing a flow regime recognition model is provided, and the steps are as follows:

[0078] (1) Acquisition of microfluidic images

[0079] Microfluids with different flow regimes were constructed under different operating conditions according to the method of Example 1, and optical images of the microfluids formed in the collection tube under each operating condition were captured. The flow regimes included blockage, spiral, and jet.

[0080] (2) Annotation and database construction of flow diagrams

[0081] The images acquired in step (1) were used in CVAT (Computer Vision Annotation Tool) software for flow regime annotation, establishing flow regime labels, specifically including clogging, spiral, and jet labels. The flow regimes of the microfluidics were rigorously defined. The critical state of spiral and clogging was defined as when two microfluidic coils were in complete contact, with a corresponding size relationship of P = D. The critical state of spiral and jet was defined as when the microfluidic coil was stretched to a two-dimensional fold, with a corresponding size relationship of P = 4D. These relationships show slight differences in micro-spiral flows at different scales, but do not affect the research conclusions.

[0082] Specifically, the visible flow region inside the receiving tube in the image is taken as the effective region. The Rectangle tool in CVAT software is used to annotate the flow state of each image. The annotated images are exported as txt files in Ultralytics YOLODetection format. All the exported txt files are combined into a txt file set. All the annotated images are combined into a png file set. The png file set and the corresponding txt file set are used to create a data database file (data.yaml format configuration file).

[0083] In step (2), during the construction of the data database file, in addition to the three conventional flow regimes of clogging, spiral, and jet, experiments were specially designed to accommodate different backgrounds, angles, light fields, and other external influencing factors. Images of the microfluidic stream were collected under exposure, dyeing, impurity, dark, scaling, and deflection conditions, such as... Figure 3 As shown in Figure a, this expands the training scope of step (3) and extends the generalization of the training results.

[0084] In step (2), during the process of constructing the data database files, data database files containing different numbers of images were constructed. The number of images in each data database file were 300, 450, 600, 750, 900, 1050, 1200, 1350, and 1500, respectively.

[0085] (3) Training of the flow regime recognition model

[0086] Place the data database file on a computer with the YOLO environment configured, and start the training program using Python code.

[0087] Download the YOLOv10m model from the YOLO official website, configure the running environment on the computer, input the data database file constructed in step (2) for training, the training sample size is 800, orthogonally design the learning rate (Lr0), training batch (Batch), and training epochs (Epochs) and train the model, check the confusion matrix and mAP value in the training results, take the YOLOv10m model with the largest mAP value obtained from training as the flow recognition model, and record its model parameters.

[0088] During training, the effects of learning rate, training batches, and training epochs on the model training performance were examined by adjusting the internal parameters of the YOLOv10m model, such as... Figure 4 As shown. Figure 4 This shows the impact of YOLOv10m model parameters on training performance. The size of the bubbles in the figure represents the mAP value between 0.80 and 0.93. The larger the bubble, the higher the mAP value. Figure 4 Figure a shows the effect of batch size and epochs on the mAP value under the condition that Lr0 = 0.001. Figure 4 The b~d graphs show the impact of batch size and epochs on the mAP value under the conditions of Lr0 being 0.005, 0.010, and 0.025. The optimal configuration of the YOLOv10m model's internal parameters was ultimately determined to be: Lr0 = 0.005, Batch = 16, Epochs = 502, corresponding to an mAP value of 0.92. The YOLOv10m model was trained using this optimal configuration of internal parameters, and the trained YOLOv10m model (the best.pt file in the calculation results) was used as the flow regime recognition model.

[0089] (4) Validation of the flow regime recognition model

[0090] One hundred randomly selected, untrained images collected in step (1) were used to create a flow regime verification image set. The flow regime recognition model trained in step (3) was then used to perform flow regime recognition on each image in the flow regime verification image set. The results showed that the flow regime recognition model exhibited excellent recognition performance for all three flow regimes, especially accurately identifying the boundaries of the flow regime regions, such as... Figure 3 As shown in Figure b, statistical analysis of the flow regime identification results revealed that the precision of the flow regime identification model for the three flow regimes—blockage, spiral, and jet—was α, α, α, and α, respectively. B = 98%, α H = 96%, α S = 91%, such as Figure 3As shown in Figure c. Furthermore, confidence scores were calculated for images where the flow regime recognition model correctly identified the flow regime types. The results showed that the average confidence scores of the flow regime recognition model for the three flow regimes—blockage, spiral, and jet—were ε0 and ε1, respectively. B = 0.83, ε H = 0.94, ε S = 0.91, as Figure 3 As shown in Figure d, the overall confidence level is relatively high, indicating that the flow state recognition results are reliable.

[0091] (5) Examine the impact of training sample size on the recognition performance of the fluid state recognition model.

[0092] Replace the training sample size in step (3) with 300, 450, 600, 750, 900, 1050, 1200, 1350, and 1500 respectively. That is, a series of flow recognition models are trained based on the data database file containing different numbers of images constructed in step (2). According to the operation in step (4), this series of flow recognition models are used to perform flow recognition on each image in the verification image set, and the recognition results are statistically analyzed. The results are as follows. Figure 5 As shown, when the number of images in the data database file reaches 750, especially 1050 or more, the mAP value of the fluid recognition model and its precision rate for fluid recognition can be stabilized at a level above 0.90. In subsequent embodiments, the fluid recognition model trained with 1200 training samples in this embodiment is used for fluid recognition.

[0093] Example 3

[0094] In this embodiment, a method for constructing a size detection model is provided, and the steps are as follows:

[0095] (1) Acquisition of microfluidic images

[0096] Microfluids with different flow regimes were constructed under different operating conditions according to the method of Example 1, and optical images of the microfluids formed in the collection tube under each operating condition were captured. The flow regimes included blockage, spiral, and jet.

[0097] (2) Flow pattern identification of micro-spiral flow

[0098] The flow recognition model trained in Example 1 is used to perform flow recognition on the images collected in step (1) and images with spiral flow are selected.

[0099] (3) Labeling and database creation of sized images

[0100] The images of spiral flow patterns selected in step (2) are entered into CVAT software for dimension annotation, and dimension labels are created, including pitch, diameter and amplitude.

[0101] Specifically, the visible flow region inside the receiving tube in the image is considered the effective region. The Rectangle tool in CVAT software is used to annotate each image. More specifically: two pitch annotation boxes are set, located at the first pitch on both sides of the micro-helical flow, with a width of 70 px; the diameter annotation box is located at the point where the micro-helical flow begins to bend significantly from the conical opening, i.e., the point of minimum curvature, with a width of 20 px; the upper edge of the amplitude annotation box is located at the starting position of the first pitch, and the lower edge is located at the bottom of the effective region of the micro-helical flow. The two sides of the amplitude annotation box overlap with the left and right extreme positions of the two pitch annotation boxes. The annotated images are exported as txt files in Ultralytics YOLODetection format. All exported txt files are combined into a txt file set, and all annotated images are combined into a png file set. The png file set and the corresponding txt file set are then used to create a data database file (data.yaml format configuration file).

[0102] (4) Training of the size detection model

[0103] Place the data database file on a computer with the YOLO environment configured, and start the training program using Python code.

[0104] Download the YOLOv10m model from the YOLO official website, configure the running environment on the computer, input the data database file constructed in step (3) for training, the training sample size is 1000, orthogonally design the learning rate (Lr0), training batch (Batch), and training epochs (Epochs) and train the model, check the confusion matrix and mAP value, and use the YOLOv10m model with the largest mAP value obtained from training as the size detection model.

[0105] (5) Validation of the size inspection model

[0106] A size verification image set is constructed from 100 randomly selected untrained images with a spiral flow pattern. The size detection model trained in step (4) is used to perform size detection on each image in the size verification image set.

[0107] Figure 6Figure a shows the size detection results for images of micro-spiral flows at different scales. The pitch is represented by the blue line, the diameter by the green line, and the amplitude by the red line. The boundaries of the detection boxes in the figure show a high degree of overlap with the features, and the number of labels meets expectations. Statistical analysis of the size detection results shows that the dimensions detected by the size detection model are converted into actual physical dimensions according to a scale and compared with the corresponding dimensions measured manually in Photoshop. The results show that the dimensions detected by the size detection model (detected values) and the actual measured dimensions (measured values) exhibit a high degree of agreement over a large range. The Mean Absolute Percentage Error (MAPE) is used to evaluate the detection capability of the size detection model. The formula for calculating MAPE is as follows:

[0108] MAPE = (1 / n) × ∑|(X expt -X pred ) / X expt |×100%

[0109] In the above formula, X expt This is the measured value of the micro-spiral flow size, X. pred These are the measured values ​​of the micro-spiral flow size obtained through a size detection model.

[0110] The MAPE values ​​for pitch, diameter, and amplitude of the micro-helical flow were 3.3%, 4.4%, and 1.1%, respectively. Figure 6 As shown in Figure b, all values ​​are less than 5%, indicating that the size reference of the micro-spiral flow automatically extracted from the image using the size detection model has high accuracy.

[0111] Example 4

[0112] In this embodiment, a method for constructing a flow regime prediction model is provided, and the steps are as follows:

[0113] (1) Acquisition of microfluidic images

[0114] Microfluids with different flow regimes were constructed under different operating conditions according to the method of Example 1, and optical images of the microfluids formed in the collection tube under each operating condition were captured. The flow regimes included blockage, spiral, and jet.

[0115] (2) Identification and assignment of microfluidic flow states

[0116] The flow recognition model trained in Example 1 is used to perform flow recognition on the images collected in step (1) and assign different flow values ​​to the blockage, spiral, and jet flow states. The blockage, spiral, and jet flow states are assigned the values ​​0, 1, and 2, respectively.

[0117] (3) Training of reservoir construction and flow pattern prediction models

[0118] In Excel, the operating conditions and corresponding flow regime values ​​are constructed into flow regime data bars, specifically in the format of "internal phase fluid viscosity - external phase fluid viscosity - injection tube conical inlet diameter - receiving tube inlet diameter - internal phase fluid velocity - external phase fluid velocity - flow regime value". All 5000 flow regime data bars are then compiled into a flow regime database. This database is then used to train a Support Vector Machine (SVM) classification algorithm to obtain a flow regime prediction model (the SVM.pkl model file in the calculation results). Before training, each flow regime data bar in the database needs to be preprocessed using normalization.

[0119] (4) Validation of the flow pattern prediction model

[0120] Operating condition data was constructed according to the format "inner phase fluid viscosity - outer phase fluid viscosity - inner diameter of injection tube cone - inner diameter of receiving tube - inner phase fluid velocity - outer phase fluid velocity". A validation dataset was then created by randomly selecting 100 untrained operating condition data points. This validation dataset was then fed into the flow prediction model obtained in step (3) for flow prediction, and the flow patterns were compared with those of the microfluidics prepared from the operating conditions in the validation dataset. The results showed that the precision detection rates of the flow prediction model for spirals, blockages, and jets were respectively: α H = 87%, α B = 86% and α S = 93%, such as Figure 7 As shown in Figure a. Furthermore, confidence level statistics were performed on images where the flow pattern prediction model correctly predicted the flow pattern type. The average confidence levels of the flow pattern prediction model for spirals, blockages, and jets were found to be: ε H = 0.87, ε B = 0.90 and ε S = 0.93, indicating a relatively high overall distribution level. This suggests that the flow prediction model trained in step (3) has high accuracy in predicting the flow of microfluidics.

[0121] Example 5

[0122] In this embodiment, a method for constructing a size prediction model is provided, and the steps are as follows:

[0123] (1) Acquisition of microfluidic images

[0124] Microfluids were constructed under different operating conditions according to the method of Example 1, and optical images of the microfluids formed in the collection tube under each operating condition were taken. Three images of the microfluids were taken for each operating condition.

[0125] This step involved controlling the operating conditions over a wide range during the construction of the microfluidic system; specifically, the viscosity of the internal phase fluid (μ) was adjusted. iThe control range is 100 ~ 800 mPa·s, and the viscosity of the external phase fluid (μ) is... o The control range is 1 ~ 6 mPa·s, and the inner diameter of the injection tube conical orifice (D) i The diameter of the receiving tube is 60 ~ 130 μm, and the inner diameter of the receiving tube (D) is 60 ~ 130 μm. o The internal phase fluid velocity (u) is 700 ~ 1200 μm. i The velocity of the external phase fluid is 0.1 ~ 0.5 m / s, and the velocity of the external phase fluid (u) is... o The value is 0.002 ~ 0.025 m / s. Figure 9 This shows the data distribution of the operating conditions in this step. Figure a shows the distribution of the viscosity of the internal phase fluid and the viscosity of the external phase fluid. Figure b shows the distribution of the inner diameter of the injection tube cone and the inner diameter of the receiving tube. Figure c shows the distribution of the flow velocity of the internal phase fluid and the flow velocity of the external phase fluid.

[0126] (2) Flow pattern identification of microfluidics

[0127] The flow recognition model trained in Example 1 is used to perform flow recognition on the images collected in step (1) and images with spiral flow are selected.

[0128] exist Figure 9 Under the operating conditions shown, micro-spiral flows with diverse structures were constructed. The pitch of the constructed micro-spiral flows ranged from approximately 250 to 1150 μm, the diameter ranged from approximately 150 to 300 μm, and the amplitude ranged from approximately 500 to 1000 μm. Figure 10 The diagram shows the size distribution of the constructed micro-spiral flow, with figures a to c showing the distribution of pitch, diameter, and amplitude, respectively.

[0129] (3) Size detection and library construction of micro-spiral flow

[0130] Images of helical flow patterns selected in step (2) are input into the size detection model trained in Example 3 for size detection, obtaining the pitch, diameter, and amplitude data of the micro-helical flow prepared under each operating condition. The pitch, diameter, and amplitude data of three images of micro-helical flow taken under the same operating condition are averaged to obtain the average pitch, average diameter, and average amplitude. The average pitch, average diameter, and average amplitude are used for subsequent processing to more accurately reflect the influence of the operating condition on the size of the micro-helical flow.

[0131] In Excel, combine the operating conditions and corresponding average pitch into data bars. Specifically, construct pitch data bars in the format "Internal phase fluid viscosity - External phase fluid viscosity - Inner diameter of injection tube cone - Inner diameter of receiving tube - Internal phase fluid velocity - External phase fluid velocity - Average pitch," and create a pitch database from all 1200 pitch data bars. Similarly, in Excel, combine the operating conditions and corresponding average diameter into data bars. Specifically, construct diameter data bars in the format "Internal phase fluid viscosity - External phase fluid viscosity - Inner diameter of injection tube cone - Inner diameter of receiving tube - Internal phase fluid velocity - External phase fluid velocity - Average diameter," and create a diameter database from all 1200 diameter data bars. In Excel, combine the operating conditions and the corresponding average amplitude into data bars. Specifically, construct amplitude data bars in the format of "inner phase fluid viscosity - outer phase fluid viscosity - inner diameter of injection tube cone - inner diameter of receiving tube - inner phase fluid velocity - outer phase fluid velocity - average amplitude", and build an amplitude database with all amplitude data bars (a total of 1200).

[0132] (4) Training of the size prediction model

[0133] The pitch database constructed in step (3) is trained using the Extreme Gradient Boosting (XGBoost) algorithm to obtain the pitch prediction model (xgbmodel_P.json model file in the calculation results). The diameter database constructed in step (3) is trained using the XGBoost algorithm to obtain the diameter prediction model (xgbmodel_D.json model file in the calculation results). The amplitude database constructed in step (4) is trained using the XGBoost algorithm to obtain the amplitude prediction model (xgbmodel_A.json model file in the calculation results).

[0134] To examine the training effects of other machine learning algorithms on the size prediction model, the pitch database was trained using Decision Tree (DT), Lasso (Minimum Absolute Value Shrinkage and Selection), Ridge Regression (RR), Stochastic Gradient Descent with Restart (SGDR), k-Nearest Neighbor (kNN), Support Vector Regression (SVR), Random Forest (RF), Lightweight Gradient Boosting Machine (LGBM), and Multilayer Perceptron (MLP), resulting in a series of pitch prediction models based on different machine learning algorithms. Similarly, the diameter database was trained using DT, Lasso, RR, SGDR, kNN, SVR, RF, LGBM, and MLP, yielding a series of diameter prediction models based on different machine learning algorithms. Finally, the amplitude database was trained using DT, Lasso, RR, SGDR, kNN, SVR, RF, LGBM, and MLP, resulting in a series of amplitude prediction models based on different machine learning algorithms.

[0135] In this step, before training begins, the pitch, diameter, and amplitude data bars in the pitch, diameter, and amplitude database need to be normalized and preprocessed.

[0136] (5) Validation of the size prediction model

[0137] The operating condition data is constructed according to the format of "inner phase fluid viscosity - outer phase fluid viscosity - inner diameter of injection tube cone - inner diameter of receiving tube - inner phase fluid velocity - outer phase fluid velocity". 50 random untrained operating condition data are used to form a verification operating condition dataset. The verification operating condition dataset is then fed into the pitch prediction model, diameter prediction model and amplitude prediction model obtained in step (4) to predict the pitch, diameter and amplitude.

[0138] The actual pitch, diameter, and amplitude data of the micro-spiral flow prepared under the operating conditions of the validation dataset were measured and recorded as experimental values. Pitch, diameter, and amplitude prediction models based on the above 10 machine learning algorithms were used to predict these values, which were recorded as predicted values. The predicted values ​​were compared and analyzed with the experimental values, and the MAPE value and R0 of the predicted and experimental values ​​were calculated. 2 Value and 1-R 2 value.

[0139] Comparing the differences in the prediction performance of pitch, diameter, and amplitude prediction models obtained from the above 10 machine learning algorithms on the size of micro-helical flows, the results are as follows: Figure 11 As shown in the figure, figures a to c represent the prediction performance of prediction models based on different machine learning algorithms for pitch, diameter, and amplitude. Compared to the other nine machine learning algorithms, the pitch, diameter, and amplitude prediction model trained based on the XGBoost algorithm shows better prediction performance in MAPE and 1-R. 2 They all showed superior performance. Therefore, the pitch, diameter, and amplitude prediction models obtained based on the XGBoost algorithm were subsequently used for size prediction.

[0140] Figure 7 Figure c shows the MAPE values ​​for the pitch, diameter, and amplitude of micro-helical flow, compared with experimental values, obtained by the pitch, diameter, and amplitude prediction model trained based on the XGBoost algorithm. Figure 8 The R-squared values ​​of pitch, diameter, and amplitude of micro-helical flow are compared with experimental values ​​by a pitch, diameter, and amplitude prediction model trained based on the XGBoost algorithm. 2 value. Figure 7 The d-to-f graph is an analysis of the deviation between the predicted and experimental values ​​of pitch, diameter, and amplitude of micro-helical flow based on the pitch, diameter, and amplitude prediction model trained using the XGBoost algorithm, and includes error statistics. Figure 7From graph d to f, it can be seen that: the pitch prediction model trained based on the XGBoost algorithm has a small deviation from the experimental values ​​for the pitch of the micro-helical flow, with a deviation range of +10% to -8% and a MAPE of 3.4%; the diameter prediction model trained based on the XGBoost algorithm has a relatively smaller deviation from the experimental values ​​for the diameter of the micro-helical flow, with a deviation range of +6% to -5% and a MAPE of 1.3%; the amplitude prediction model trained based on the XGBoost algorithm has a smaller deviation from the experimental values ​​for the amplitude of the micro-helical flow, with a deviation range of +1% to -6% and a MAPE of 0.7%. MAPE value and R 2 The values ​​all showed excellent performance, indicating that the pitch, diameter, and amplitude prediction model trained based on the XGBoost algorithm can accurately predict the pitch, diameter, and amplitude of the micro-spiral flow according to the operating conditions.

[0141] (6) Analysis of size control rules

[0142] The SHAP model was used to analyze the pitch, diameter, and amplitude prediction models trained based on the XGBoost algorithm. At the same time, distributed computation was performed on the pitch, diameter, and amplitude prediction models trained based on the XGBoost algorithm to solve the contour map data values, and then the data were summarized to draw the two-phase contour map.

[0143] SHAP analysis was performed on the pitch, diameter, and amplitude prediction models trained based on the XGBoost algorithm to investigate the influence of operating conditions on the structural parameters of the micro-helical flow. Specifically, the external phase fluid velocity (u... o The screw pitch (P) has a significant impact, followed by the internal phase fluid velocity (u). i ), Inner diameter of receiving tube (D) o ), Inner diameter of the injection tube conical opening (D) i ), internal phase fluid viscosity (μ) i ), external phase fluid viscosity (μ) o Its average SHAP value is approximately 120, indicating that the pitch of the micro-helical flow is mainly affected by kinetic factors, exhibiting a flexible and adjustable morphological structure in two-phase shearing, such as... Figure 12 Figures a to b show that the inner diameter of the injection tube cone has a significant impact on the diameter, followed by the viscosity of the inner phase fluid, the velocity of the outer phase fluid, the inner diameter of the receiving tube, the velocity of the inner phase fluid, and the viscosity of the outer phase fluid. The average SHAP value is 28, indicating that the diameter of the micro-spiral flow is mainly affected by the size effect. The Barus effect at the cone of the injection tube in high-viscosity flow determines the transverse dimension of the micro-spiral flow, and it is no longer significantly affected by the two-phase shearing action. Figure 12Figures c to d show that the inner diameter of the injection tube cone has a significant impact on the amplitude, followed by the inner diameter of the receiving tube, the viscosity of the inner phase fluid, the flow velocity of the outer phase fluid, and the flow velocity of the inner phase fluid. The average SHAP value is 95, indicating that the amplitude of the micro-spiral flow is mainly affected by the size effect. When the inner diameter of the injection tube cone is large, the resulting large-diameter microfluidic fluid has a large bending moment under pressure bending and is not easy to form a tightly contracted structure. Consequently, it curls up significantly in the space inside the receiving tube to form a micro-spiral flow with a large amplitude, such as... Figure 12 The diagram from e to f is shown.

[0144] Further analysis of the control mechanisms of pitch, diameter, and amplitude in micro-spiral flow revealed that the internal and external fluid velocities have a comprehensive impact on pitch, exhibiting negative and positive correlations, respectively. Therefore, the pitch can be adjusted by precisely controlling the two-phase fluid velocities, such as... Figure 13 As shown in Figure a; simultaneously, the combined effects of the internal phase fluid velocity and the inner diameter of the receiving tube on the pitch are as follows: Figure 13 As shown in Figure b, a large adjustable pitch range is also observed. In the control of diameter by the inner diameter of the injection tube cone and the inner diameter of the receiving tube, the inner diameter of the injection tube cone exhibits a significant positive correlation with the diameter of the micro-spiral flow, demonstrating continuous adjustability. Figure 13 As shown in Figure c; in the amplitude control of the inner diameter of the injection tube cone and the inner diameter of the receiving tube, the inner diameter of the injection tube cone also shows a significant positive correlation with the amplitude, but it is easily deformed by the friction of the inner wall of the receiving tube. Therefore, the micro-spiral should be designed as a free-flow preparation state, such as... Figure 13 As shown in Figure d.

[0145] Example 6

[0146] In this embodiment, a reverse design method for micro-helices driven by visual recognition and machine learning is provided, with the following steps:

[0147] S611 sets the target size and target size range of the micro spiral according to user requirements, and sets the boundary conditions of the operating conditions to standardize the operating conditions. When setting the target size and target size range, it specifically includes setting the pitch, diameter and amplitude of the micro spiral, as well as the pitch range, diameter range and amplitude range. When standardizing the operating conditions, it is necessary to set the mean and standard deviation of the operating conditions.

[0148] S612, using the pitch prediction model, diameter prediction model, and amplitude prediction model trained by the XGBoost algorithm in Example 5 as surrogate models, the Pareto algorithm is used to search in the parameter space of the operating conditions to back-calculate the optimal solution set of the operating conditions that can achieve the target size. The operating conditions in the optimal solution set are then put into the flow prediction model constructed in Example 4 for flow verification.

[0149] S62, if any operating condition passes the flow regime verification, it is used as a candidate operating condition; if no operating condition passes the flow regime verification, the boundary conditions of the operating condition are modified, the operating condition is standardized again, and step S61 is repeated until a candidate operating condition that can pass the flow regime verification is obtained. During the flow regime verification, the back-calculated operating condition is fed into the flow regime prediction model for fluid prediction. If the flow regime prediction result is a spiral, then the operating condition passes the flow regime verification.

[0150] S63, the candidate operating conditions are input into the pitch, diameter, and amplitude prediction models of the micro-spiral flow trained using the XGBoost algorithm in Example 5 for size prediction. The mean absolute percentage error (MASE) between the predicted pitch and target pitch of the micro-spiral flow is calculated, as are the mean absolute percentage errors between the predicted diameter and target diameter, and the predicted amplitude and target amplitude. These three MAS are summed to obtain the total MAS. Candidate operating conditions with a total MAS of less than 10% are selected as feasible operating conditions.

[0151] Since Example 5 performed normalization preprocessing on each pitch, diameter, and amplitude data bar in the pitch, diameter, and amplitude database before training the size prediction model, in step S63, after obtaining the feasible operating conditions, it is necessary to perform inverse normalization processing on the feasible operating conditions to obtain the final feasible operating conditions.

[0152] The reverse design of a target microhelix will be used as an example for illustration:

[0153] In step S611, the target size of the micro-helix and the pitch P are set. set = 600 μm, diameter D set = 210 μm, amplitude A set = 900 μm, set the target size range: pitch range P_target_range = (590, 610), diameter range D_target_range = (205, 215), amplitude range A_target_range = (885, 915), in μm. Set the operating boundary conditions, specifically including setting the internal phase fluid viscosity μ. i _values ​​= [108, 168,247, 345, 462, 597, 751], external phase fluid viscosity μ o _values ​​= [0.87], Inner diameter D of the injection tube tapered orifice i_values ​​= [70, 80, 90, 100, 110, 120, 130], inner diameter D of the receiving tube. o _values ​​= [700, 1000,1200], the minimum and maximum flow velocity u of the internal phase fluid i _min, u i _max = 0.023, 0.442, minimum and maximum flow velocities u of the external phase fluid. o _min, u o `_max = 0.002, 0.025`, where the viscosity of the internal phase fluid, the viscosity of the external phase fluid, the inner diameter of the injection tube cone, and the inner diameter of the receiving tube are set as discrete values, and the flow rates of the internal phase fluid and the external phase fluid are set as continuous values. The operating conditions are standardized by setting the mean `MEAN = np.array([463, 2, 97, 1014, 0.1999, 0.0091])` and the standard deviation `STD = np.array([189, 1, 22, 158, 0.0833, 0.0041])`, with the mean and standard deviation parameters in the following order: internal phase fluid viscosity, external phase fluid viscosity, inner diameter of the injection tube cone, inner diameter of the receiving tube, internal phase fluid flow rate, and external phase fluid flow rate. In this step, not only is it necessary to set the target size, but also to set the target size range for the solution, which is different from the traditional fixed value setting method. At the same time, it is necessary to constrain the boundary conditions of the operating conditions so that the optimal solution set obtained in step S612 is feasible in actual operation.

[0154] Based on the conditions in step S611, and following the operations in step S612 above, the optimal solution set of operating conditions for achieving the target size is obtained, and a Pareto front plot is generated to visualize the results. Then, following the operations in step S62 above, flow regime verification is performed, and it is found that all operating conditions in the optimal solution set can pass the flow regime verification. All operating conditions that pass the flow regime verification are selected as candidate operating conditions. Then, following the operations in step S63 above, size prediction is performed, and feasible operating conditions are selected. The feasible operating conditions are sorted in ascending order according to the total average absolute percentage error.

[0155] Figure 14 Figure a shows the distribution of predicted pitch and diameter values ​​for each candidate operating condition, as well as the total average absolute percentage error (Sum of MAPE(%)) for each candidate operating condition. Figure 14 Figure b shows the distribution of predicted pitch and amplitude values ​​for each candidate operating condition, as well as the total mean absolute percentage error (Sum of MAPE (%)) for each candidate operating condition. The data points in the figure represent different candidate operating conditions; the darker the color of the data point, the smaller the total mean absolute percentage error, and the better the performance of the corresponding operating condition. Figure 14The candidate working condition with the smallest total average absolute percentage error is selected as the optimal working condition.

[0156] Example 7

[0157] In this embodiment, a precise preparation method for microhelices driven by visual recognition and machine learning is provided, and the steps are as follows:

[0158] (1) Ten sets of micro-spirals were designed according to the actual application requirements. The target dimensions of the ten sets of micro-spirals are shown in Table 1, where P tgt D represents the target pitch of the micro-helix. tgt A represents the target diameter of the microhelix. tgt The target amplitude of the micro-helix is ​​represented by P in Table 1. tgt D tgt A tgt Based on this, a target size range is set, that is, according to the actual application needs, within P tgt D tgt A tgt Based on this, it can fluctuate up or down by a certain value. For example, the pitch range can be (P... tgt -30, P tgt The selection range is +30), and the diameter range can be within (D). tgt -30, D tgt The amplitude range can be selected within the range of +30, and the amplitude range can be selected within (A). tgt -60, A tgt Select within the range of -60). At the same time, set the boundary conditions of the operating conditions and standardize the operating conditions. During the standardization process, it is necessary to set the mean and standard deviation of the operating conditions.

[0159] Table 1. Detailed list of target and fabrication dimensions of microspirals.

[0160]

[0161] (2) Based on the conditions of step (1), the optimal operating conditions corresponding to each target size are obtained by following the steps S612~S63 of Example 6.

[0162] (3) Prepare internal phase fluid and external phase fluid

[0163] Preparation of the internal phase fluid: Dissolve the monomer polyethylene glycol diacrylate (PEGDA) and the photoinitiator 2-hydroxy-2-methyl-1-phenyl-1-propanone (HMPP) in deionized water, then add surfactant F127 and stir for 30 min, and then use an emulsifier to demulsify for 5 min to obtain a prepolymer solution; in the prepolymer solution, the mass concentrations of PEGDA, HMPP and F127 are 15%, 1.5% and 1%, respectively; add NaAlg to the prepolymer solution, then stir in a water bath at 30 ℃ for 60 min, filter with a water-based filter head to obtain the internal phase fluid, and the concentration of NaAlg is determined according to the viscosity of the internal phase fluid in the optimal operating conditions obtained in step (2).

[0164] Preparation of external phase fluid: Deionized water is used as the external phase fluid, or glycerol is dissolved in deionized water to form an external phase fluid. The concentration of glycerol is determined according to the viscosity of the external phase fluid in the optimal operating conditions obtained in step (2).

[0165] (4) Preparation of microhelices

[0166] Microhelices were prepared using a single-stage microfluidic device with the structure described above. The single-stage microfluidic device was arranged vertically, with the injection tube positioned above the receiving tube. A receiving container filled with receiving liquid was placed below the receiving tube, with the outlet end of the receiving tube below the liquid surface in the receiving container. An ultraviolet point light source was positioned midway down the receiving tube to apply ultraviolet light to irradiate the receiving tube, initiating a photopolymerization reaction of the PEGDA monomers in the microhelical flow within the receiving tube to form microhelices.

[0167] The inner and outer phase fluids are continuously injected into the injection tube and receiving tube respectively using a constant flow syringe pump. The inner phase fluid jets and expands into the receiving tube from the tapered opening of the injection tube. After entering the receiving tube, the inner phase fluid expands and decelerates, forming a coaxial laminar flow with the outer phase fluid. Under the action of the fluid rope effect, the inner phase fluid becomes unstable and coils due to buckling instability, forming a continuous and stable micro-spiral flow in the receiving tube. Ultraviolet light is applied to the receiving tube using an ultraviolet point light source to induce the photopolymerization reaction of PEGDA monomers in the micro-spiral flow within the receiving tube, forming microspirals. The formed microspirals are naturally discharged from the outlet end of the receiving tube.

[0168] In this step, the inner diameter of the injection tube cone and the inner diameter of the receiving tube of the microfluidic device, as well as the flow velocity of the inner phase fluid and the flow velocity of the outer phase fluid, are determined according to the optimal operating conditions obtained in step (2).

[0169] The microspirals prepared in step (4) were observed and photographed using an industrial microscope, as shown in the following figures. Figure 15 As shown in Figure b, the prepared microhelices exhibit excellent helical morphology. The pitch, diameter, and amplitude of the microhelices prepared in step (4) were measured, and the results are shown in Table 1, where P... exp Dexp A exp The values ​​represent the pitch, diameter, and amplitude of the experimentally prepared microhelices, respectively. Using the dimensions measured in Table 1 as experimental values ​​and the target dimensions in Table 1 as target values, the deviations between the experimental and target values ​​are analyzed, and the results are as follows: Figure 15 As shown in Figures c to e, the fabrication deviations for pitch were +6% to -3%, diameter was +2% to -9%, and amplitude was +8% to -5%, all within a range of less than 10%. Furthermore, a uniformity analysis was performed on one group of microspirals, revealing an average pitch of 553 μm, an average diameter of 188 μm, and an average amplitude of 757 μm. The coefficients of variation (CV) for pitch, diameter, and amplitude were 3.6%, 2.0%, and 0.9%, respectively, indicating uniform morphology.

Claims

1. A reverse design method for micro-helices driven by visual recognition and machine learning, characterized in that, Includes the following steps: S1, Microfluidic image acquisition The microfluidic device is operated under different operating conditions to construct microfluidics with different flow regimes and acquire optical images of the microfluidics; the operating conditions include the viscosity of the inner phase fluid, the viscosity of the outer phase fluid, the inner diameter of the injection tube cone, the inner diameter of the receiving tube, the flow velocity of the inner phase fluid, and the flow velocity of the outer phase fluid; the flow regimes include blockage, spiral, and jet. S2, Construction of the Flow Recognition Model Images are collected according to the method in step S1 and then fed into CVAT software for flow regime annotation. The annotated images are then fed into the YOLO object detection model for training to obtain a flow regime recognition model. The flow regime recognition model is then used to perform flow regime recognition on the images collected in step S1, and images with a spiral flow regime are selected. S3, Construction of the Dimension Inspection Model The images of the spiral microfluidic selected in step S2 are input into CVAT software for size annotation. The annotated images are then input into the YOLO object detection model for training to obtain the size detection model. S4, Construction of the Flow Prediction Model The flow pattern recognition model is used to identify the flow pattern of the images collected according to the method in step S1, and different flow pattern values ​​are assigned to the blockage, spiral and jet flow patterns. The operating conditions and the corresponding flow pattern values ​​are constructed into flow pattern data bars. All flow pattern data bars are fed into the SVM algorithm for training to obtain the flow pattern prediction model. S5, Construction of Size Prediction Model The images of the helical microfluidic flow selected in step S2 are input into the size detection model for size detection to obtain the pitch, diameter and amplitude data of the micro-helical flow. The operating conditions are respectively constructed into pitch, diameter and amplitude data bars with the corresponding pitch, diameter and amplitude size data. All pitch data bars, diameter data bars and amplitude data bars are respectively input into the XGBoost algorithm for training to obtain pitch prediction model, diameter prediction model and amplitude prediction model. S6, reverse design of micro-helices S61, set the target size and target size range of the micro-spiral, set the boundary conditions of the operating conditions, and standardize the operating conditions; based on the pitch prediction model, diameter prediction model and amplitude prediction model, combined with the multi-objective optimization algorithm, back-calculate the operating conditions that can achieve the target size, and put the back-calculated operating conditions into the flow prediction model for flow verification. S62, if an operating condition passes the flow regime verification, then the operating condition that passes the flow regime verification shall be used as the candidate operating condition. If no operating condition passes the flow verification, modify the boundary conditions of the operating condition, re-standardize the operating condition, and repeat step S61 until a candidate operating condition that can pass the flow verification is obtained. S63, input the candidate operating conditions into the size prediction model to predict the size, calculate the total average absolute percentage error between the predicted size and the target size, and take the candidate operating conditions with the total average absolute percentage error less than the set threshold as the feasible operating conditions.

2. The reverse design method for micro-helices driven by visual recognition and machine learning according to claim 1, characterized in that, Step S61 includes the following steps: S611 sets the target size and target size range of the micro-spiral according to user needs, sets the boundary conditions of the operating conditions, and standardizes the operating conditions. S612 uses pitch prediction model, diameter prediction model and amplitude prediction model as surrogate models. It uses Pareto algorithm to search in the parameter space of operating conditions and back-calculates the optimal solution set of operating conditions that can achieve the target size. The operating conditions in the optimal solution set are then put into the flow prediction model for flow verification.

3. The reverse design method for micro-helices driven by visual recognition and machine learning according to claim 1, characterized in that, The target dimensions of the micro-spiral include pitch, diameter, and amplitude, and the target size range includes pitch range, diameter range, and amplitude range.

4. The reverse design method for micro-helices driven by visual recognition and machine learning according to claim 1, characterized in that, In step S63, the set threshold refers to the total average absolute percentage error being 10%.

5. The reverse design method for micro-helices driven by visual recognition and machine learning according to claim 4, characterized in that, In step S63, the feasible working condition that has the total average absolute percentage error less than a set threshold and the total average absolute percentage error is the smallest is taken as the optimal working condition.

6. The reverse design method for microhelices driven by visual recognition and machine learning according to any one of claims 1 to 5, characterized in that, In steps S2-S3, when training the flow recognition model and the size detection model, the number of images required for training is at least 1000; in steps S4-S5, when training the flow prediction model, pitch prediction model, diameter prediction model and amplitude prediction model, the number of flow data bars, pitch data bars, diameter data bars and amplitude data bars required for training is at least 1000.

7. The reverse design method for a micro-helix driven by visual recognition and machine learning according to any one of claims 1 to 5, characterized in that, Step S1 involves setting up orthogonal experiments under different operating conditions when constructing microfluidics with different flow regimes using a microfluidic device.

8. A method for precise fabrication of microhelices driven by visual recognition and machine learning, characterized in that, Includes the following steps: A feasible operating condition is obtained according to the method described in any one of claims 1 to 7. The microfluidic device is operated under the feasible operating condition to construct a microspiral flow. Ultraviolet light is applied to a position where the morphology of the microspiral flow in the microfluidic device has reached a stable state to induce the photopolymerization reaction of the photopolymerizable polymer monomers in the microspiral flow, thereby obtaining a microspiral. The operating conditions include the viscosity of the inner phase fluid, the viscosity of the outer phase fluid, the inner diameter of the injection tube cone, the inner diameter of the receiving tube, the flow rate of the inner phase fluid, and the flow rate of the outer phase fluid; the inner phase fluid is an aqueous solution containing photopolymers, photoinitiators, and polymers, and the outer phase fluid is an aqueous solution of water or small molecule substances.

9. The precise fabrication method for microhelices driven by visual recognition and machine learning according to claim 8, characterized in that, The viscosity of the inner phase fluid is greater than that of the outer phase fluid. The viscosity of the inner phase fluid is adjusted by regulating the concentration of polymers and photopolymers in the inner phase fluid, while the viscosity of the outer phase fluid is adjusted by regulating the concentration of small molecules in the outer phase fluid.

10. The precise fabrication method of microhelices driven by visual recognition and machine learning according to claim 8 or 9, characterized in that, The polymer includes sodium alginate, sodium carboxymethyl cellulose, or polyvinyl alcohol; the small molecule includes ethylene glycol or glycerol.

Images