Viscoelastic microfluidic particle sorting parameter optimization method and system

Abstract

The present application provides a kind of viscoelastic microfluidic particle sorting parameter optimization method and system, including the theoretical calculation for obtaining the theoretical results of various diameter particles in the transverse equilibrium position and the required channel length to reach equilibrium in straight channel section under different sheath liquid polyethylene oxide concentration, sample PEO concentration, total flow, sheath liquid and sample flow ratio.Utilize random forest machine learning algorithm to train the theoretical results, judge whether the particle can completely penetrate the sample-sheath liquid interface under each working condition parameter combination and predict the transverse equilibrium position of the particle and the required channel length to reach equilibrium.Determine the optimal sorting parameter of viscoelastic microfluidic.The present application can realize the design optimization of viscoelastic microfluidic chip from experience dependence to data-driven based on machine learning, and can significantly improve the scalability of viscoelastic microfluidic technology in different research fields by simplifying the working condition parameter iteration process.

Description

Technical Field

[0001] This invention relates to the technical field of the intersection of microfluidics and artificial intelligence, specifically to a method and system for optimizing the sorting parameters of viscoelastic microfluidic particles, and more particularly to a machine learning-driven method for optimizing the sorting parameters of viscoelastic microfluidic particles. Background Technology

[0002] Accurate and rapid separation of target biological particles from complex samples is fundamental to disease diagnosis, therapeutic intervention, and mechanistic research. However, existing separation technologies remain insufficient to provide robust support for clinical research due to several limitations. For example, separating circulating tumor cells (CTCs) from human blood typically requires lysing red blood cells followed by large-scale sample dilution, significantly increasing operational complexity, processing time, and the risk of sample contamination. Furthermore, personalized precision medicine places increasingly stringent demands on key separation performance indicators such as throughput, purity, and recovery rate. More importantly, the traditional "experiment-analysis-optimization" paradigm is becoming increasingly inadequate to meet the growing demands for higher sorting performance and diverse applications. In many application scenarios, achieving the desired sorting target often requires significant time and financial investment due to iterative trial-and-error processes, with results that may still be unsatisfactory. Moreover, adapting separation technologies to different research contexts often necessitates repeating the entire iterative optimization process, severely limiting their clinical application and flexibility in expanding to other research areas.

[0003] Particle sorting technologies can generally be divided into two categories: active and passive methods. Active methods utilize external multiphysics fields to sort particles, typically enabling precise and diverse manipulation. However, these systems are often sophisticated and expensive, posing a significant challenge to the development of further integrated systems. Furthermore, the high-precision manipulation of active technologies often comes at the cost of reduced sample throughput, greatly limiting their clinical applicability. In contrast, passive methods typically rely solely on fluid forces within channels to drive cells, offering advantages such as high throughput, small size, low cost, and simple structure. However, passive devices have lower separation accuracy and flexibility, making it difficult to adjust parameters to meet diverse separation needs like active systems. Moreover, due to high stress and the risk of clogging, passive methods may compromise the integrity and activity of biological particles.

[0004] Viscoelastic microfluidics is an advanced particle sorting technology that uses polymer solutions as a fluid medium to introduce additional elastic forces to achieve precise sorting of target particles. Compared with other mainstream separation technologies, viscoelastic microfluidics has unique advantages, including the ability to handle high-concentration particle samples and achieve high-resolution separation. For example, viscoelastic microfluidics can directly sort CTCs from undiluted human blood and separate exosomes from extracellular vesicles. However, the widespread application of viscoelastic microfluidics in biology, chemistry, and medicine still faces challenges. For specific sorting targets and needs, it is usually necessary to first conduct particle sorting experiments under multiple sets of operating conditions, then investigate the influence of various parameters on particle motion based on a large number of experimental phenomena, and optimize the operating parameters. Repeating these steps multiple times is necessary to finally iterate and arrive at a relatively ideal parameter combination. Furthermore, if the sorting targets and needs change, the above experimental iteration steps often need to be repeated. Therefore, existing viscoelastic microfluidics technologies often require a significant investment of unnecessary time and money in their use. Developing a rapid optimization strategy for viscoelastic microfluidic particle sorting parameters can significantly improve the application value of viscoelastic microfluidics in the biomedical field, thereby contributing to research on disease mechanisms and drug development. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for optimizing the sorting parameters of viscoelastic microfluidic particles.

[0006] According to the present invention, a method for optimizing the sorting parameters of viscoelastic microfluidic particles includes the following steps: Step S1: Based on the total flow in the straight channel segment The flow rate ratio of sheath fluid to sample fluid Distribution function of PEO concentration along the channel width Horizontal coordinates of particles Particle diameter Particles in x velocity in direction Particles in y velocity in direction Channel height Width of straight passage section Fluid density PEO molecular weight PEO concentration in sheath fluid PEO concentration of the sample Solvent viscosity of PEO solution shear rate Avogadro's constant Boltzmann constant and temperature Constructing the inertial lift force on the particles Elastic force Viscous drag force and virtual mass force The expression; obtained through calculation and Based on the distribution along the channel width, a sample-sheath fluid interface penetration criterion is constructed; , , and A particle dynamics model is constructed to calculate the trajectory of the particle after it has fully entered the sheath fluid, and the lateral equilibrium position of the particle in the straight channel section is given. y equ And the channel length required to achieve balance L equ ; Step S2: Set particle diameter PEO concentration in sheath fluid PEO concentration of the sample Total flow in straight sections The flow rate ratio of sheath fluid to sample fluid The range of values ​​is determined by using Latin sampling to obtain sample points and calculating the range of values ​​for each sample point. y equ and L equ ; Step S3: Train three random forest sub-models using the above sample points to predict particle diameters in different environments. , , , The penetration ability of the sample-sheath fluid interface and y equ and L equ ; Step S4: Generate sample points in the parameter space using Latin sampling, and determine the optimal sorting parameters for viscoelastic microfluidics based on hard constraints, Pareto optimization results, fitting error, similarity between solutions, stability, and the importance of each performance index.

[0007] Preferably, the height of the viscoelastic microfluidic chip channel in step S1 is... h The width of the straight channel section is 50μm. w The molecular weight of PEO is 120 μm. The distribution function of PEO concentration along the channel width at 600 kDa. Expressed using equation (1): (1) Where w1 is the width of the sheath fluid in the straight channel segment, which can be solved from equation (2): (2) Inertial lift force on the particle Elastic force Viscous drag force and virtual mass force The expressions for are shown in equation (3): (3) in and Let be the unit vectors pointing from the main flow direction and the channel center to the wall, respectively; the sample-sheath interface penetration criterion is: if and The curve of the resultant force is in If a negative slope zero point exists within the interval, it is determined that the particle cannot completely penetrate the sample-sheath fluid interface; otherwise, it is determined that the particle can completely penetrate the sample-sheath fluid interface. The particle dynamics model is characterized by equation (4): (4).

[0008] Preferably, the particle diameter in step S2 is... PEO concentration in sheath fluid PEO concentration of the sample Total flow in straight sections The flow rate ratio of sheath fluid to sample fluid The range of values ​​for: , , , , The number of Latin samples is greater than or equal to 45,000.

[0009] Preferably, the three random forest sub-models in step S3 are penetration classification models used to determine whether particles can completely penetrate the sample-sheath fluid interface. y equ Fitting a random forest regression model, L equ Fitting a conditional random forest regression model; where the penetration classification model and y equ The fitted random forest regression model is trained on the entire dataset. L equ The fitted conditional random forest regression model is trained only on data samples where particles can completely penetrate the sample-sheath fluid interface.

[0010] Preferably, in step S4, the initial Latin sample count is greater than or equal to 10,000, and hard constraints are used to remove sample points from the initial sample that do not meet the hard constraint conditions; wherein, hard constraints refer to... and In this context, subscript 1 indicates smaller particles in the mixture, and subscript 2 indicates larger particles in the mixture. The Pareto optimization objectives in step S4 are threefold: maximizing... y equ1 , minimize y equ2 and maximizing sample flow rate Q s = Q / (1+ Q r These correspond to maximizing the purity of large-diameter particles, maximizing the recovery rate of large-diameter particles, and maximizing the sorting throughput, respectively. Based on this, the initial Pareto front solution set is obtained. In step S4 y equ and L equ The fitting error tolerances are 3% and 10%, respectively. Sample points that do not meet the fitting error tolerance in the initial Pareto front solution set are removed. In step S4, ±0.8μm, ±0.8μm, ±0.2mL·h are used. -1 Each is used to determine the relationship between two samples. y equ1 , y equ2 and Q s Whether the two sample points are similar to a threshold; if the two sample points are at... y equ1 , y equ2 , Q s If two or three samples simultaneously exhibit similarity, the sample point with the lower weighted score is removed; the weighted scoring formula is: 0.45*( y equ1 / 60)-0.45*( y equ1 / 60)+0.1*( Q s / 10); In step S4, ±1 mL·h -1 ±0.4%, ±0.01%, and ±0.01% were respectively calculated. Q , Q r , cc , c s The tolerance range; for each sample point, 35 test points are generated within this tolerance range. If more than 80% of the test points can satisfy the hard constraints, the sample point is considered stable; otherwise, the sample point is considered unstable and is removed from the solution set. In step S4, the initial Pareto front solution set is processed by removing fitting errors, similarity between solutions, and stability to obtain the final Pareto front solution set; the optimal sorting parameters for viscoelastic microfluidics are determined based on purity, recovery rate, and flux.

[0011] The present invention also provides a viscoelastic microfluidic particle sorting parameter optimization system, the system comprising the following modules: Module M1: Based on the total flow in the straight section The flow rate ratio of sheath fluid to sample fluid Distribution function of PEO concentration along the channel width Horizontal coordinates of particles Particle diameter Particles in x velocity in direction Particles in y velocity in direction Channel height Width of straight passage section Fluid density PEO molecular weight PEO concentration in sheath fluid PEO concentration of the sample Solvent viscosity of PEO solution shear rate Avogadro's constant Boltzmann constant and temperature Constructing the inertial lift force on the particles Elastic force Viscous drag force and virtual mass force The expression; obtained through calculation and Based on the distribution along the channel width, a sample-sheath fluid interface penetration criterion is constructed; , , and A particle dynamics model is constructed to calculate the trajectory of the particle after it has fully entered the sheath fluid, and the lateral equilibrium position of the particle in the straight channel section is given. y equ And the channel length required to achieve balanceL equ ; Module M2: Set particle diameter PEO concentration in sheath fluid PEO concentration of the sample Total flow in straight sections The flow rate ratio of sheath fluid to sample fluid The range of values ​​is determined by using Latin sampling to obtain sample points and calculating the range of values ​​for each sample point. y equ and L equ ; Module M3: Trains three random forest sub-models using the above sample points to predict particle diameters at different times. , , , The penetration ability of the sample-sheath fluid interface and y equ and L equ ; Module M4: Generates sample points in the parameter space using Latin sampling, and determines the optimal sorting parameters for viscoelastic microfluidics based on hard constraints, Pareto optimization results, fitting error, similarity between solutions, stability, and the importance of each performance index.

[0012] Preferably, the viscoelastic microfluidic chip channel height in module M1 h The width of the straight channel section is 50μm. w The molecular weight of PEO is 120 μm. The distribution function of PEO concentration along the channel width at 600 kDa. Expressed using equation (1): (1) Where w1 is the width of the sheath fluid in the straight channel segment, which can be solved from equation (2): (2) Inertial lift force on the particle Elastic force Viscous drag force and virtual mass force The expressions for are shown in equation (3): (3) in and Let be the unit vectors pointing from the main flow direction and the channel center to the wall, respectively; the sample-sheath interface penetration criterion is: if and The curve of the resultant force is in If a negative slope zero point exists within the interval, it is determined that the particle cannot completely penetrate the sample-sheath fluid interface; otherwise, it is determined that the particle can completely penetrate the sample-sheath fluid interface. The particle dynamics model is characterized by equation (4): (4).

[0013] Preferably, the particle diameter in module M2 is... PEO concentration in sheath fluid PEO concentration of the sample Total flow in straight sections The flow rate ratio of sheath fluid to sample fluid The range of values ​​for: , , , , The number of Latin samples is greater than or equal to 45,000.

[0014] Preferably, the three random forest sub-models in module M3 are penetration classification models used to determine whether particles can completely penetrate the sample-sheath fluid interface. y equ Fitting a random forest regression model, L equ Fitting a conditional random forest regression model; where the penetration classification model and y equ The fitted random forest regression model is trained on the entire dataset. L equ The fitted conditional random forest regression model is trained only on data samples where particles can completely penetrate the sample-sheath fluid interface.

[0015] Preferably, the initial Latin sample count in module M4 is greater than or equal to 10,000, and hard constraints are used to remove sample points in the initial sample that do not meet the hard constraint conditions; where hard constraints refer to... and In this context, subscript 1 indicates smaller particles in the mixture, and subscript 2 indicates larger particles in the mixture. The Pareto optimization objectives in module M4 are three: maximizing y equ1 , minimize y equ2 And maximizing sample flow rate Q s = Q / (1+ Q rThese correspond to maximizing the purity of large-diameter particles, maximizing the recovery rate of large-diameter particles, and maximizing the sorting throughput, respectively. Based on this, the initial Pareto front solution set is obtained. In module M4 y equ and L equ The fitting error tolerances are 3% and 10%, respectively. Sample points that do not meet the fitting error tolerance in the initial Pareto front solution set are removed. The module M4 will contain ±0.8μm, ±0.8μm, and ±0.2mL·h. -1 Each is used to determine the relationship between two samples. y equ1 , y equ2 and Q s Whether the two sample points are similar to a threshold; if the two sample points are at... y equ1 , y equ2 , Q s If two or three samples simultaneously exhibit similarity, the sample point with the lower weighted score is removed; the weighted scoring formula is: 0.45*( y equ1 / 60)-0.45*( y equ1 / 60)+0.1*( Q s / 10); The module M4 will contain ±1 mL·h -1 ±0.4%, ±0.01%, and ±0.01% were respectively calculated. Q , Q r , c c , c s The tolerance range; for each sample point, 35 test points are generated within this tolerance range. If more than 80% of the test points can satisfy the hard constraints, the sample point is considered stable; otherwise, the sample point is considered unstable and is removed from the solution set. In module M4, the initial Pareto front solution set is processed by removing fitting errors, similarity between solutions, and stability to obtain the final Pareto front solution set; the optimal sorting parameters for viscoelastic microfluidics are determined based on purity, recovery rate, and flux.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention transforms the optimization process from experience-driven to data-driven by constructing a framework that combines theoretical calculation models with machine learning prediction models. It can quickly train high-precision prediction models using a large amount of theoretical calculation data generated in the early stage, thereby enabling efficient exploration and evaluation within the parameter space. This greatly shortens the optimization cycle, reduces unnecessary experimental iterations, significantly improves R&D efficiency, and reduces R&D costs. 2. This invention constructs a computational model based on rigorous fluid mechanics and particle dynamics theories, ensuring the scientific nature and accuracy of the basic data; it uses robust machine learning algorithms such as random forests to train massive theoretical results, and the established sub-models all exhibit excellent predictive performance; this makes the prediction of particle behavior more accurate and reliable, providing a solid data foundation for subsequent optimization and avoiding the uncertainty and blindness of traditional trial-and-error methods. 3. This invention innovatively introduces a multi-objective Pareto optimization framework; it can automatically find a series of non-dominated solutions that achieve the best trade-off among multiple objectives in a huge parameter space; it can further screen out a robust set of optimal solutions from the Pareto front that not only meets the physical requirements of sorting, has small prediction errors, obvious differentiation between schemes, but also resists small parameter fluctuations; it provides scientific, systematic and flexible decision support for researchers to quickly determine the final operating parameters according to specific application priorities.

[0017] 4. By introducing machine learning and multi-objective optimization techniques, this invention has completely changed the optimization mode of viscoelastic microfluidic particle sorting parameters, realizing efficient, accurate and automated parameter optimization. It has brought about fundamental improvements in terms of improving sorting performance, reducing R&D costs and enhancing the universality of the technology, and has important theoretical significance and wide practical value. Attached Figure Description

[0018] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram illustrating the working principle of viscoelastic microfluidics in an embodiment of the present invention; Figure 2 This is a schematic diagram of particle sorting in a viscoelastic microfluidic channel in an embodiment of the present invention; Figure 3 This is a schematic diagram of the movement of particles in the straight channel section of a viscoelastic microfluidic chip in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the machine learning-driven viscoelastic microfluidic particle sorting parameter optimization process in an embodiment of the present invention. Figure 5 For the 10 μm and 15 μm particles in the embodiments of the present invention, cc =0.13%, c s =0.05%, Q =20 mL·h -1 , Q r Lateral force diagram under the condition of =10; Figure 6 For the 15 μm particles in the embodiments of the present invention c c =0.13%, c s =0.05%, Q =20 mL·h -1 , Q r Motion trajectory diagram under the condition of =10; Figure 7 This is a regression performance graph of yequ and Lequ in an embodiment of the present invention; Figure 8 This is a performance comparison chart of Lequ direct regression and conditional regression in an embodiment of the present invention; Figure 9 This is a three-dimensional Pareto front diagram of 10 μm / 20 μm mixed particle sorting in an embodiment of the present invention; Figure 10 This is a two-dimensional Pareto front diagram of the sorting of 10 μm / 20 μm mixed particles in an embodiment of the present invention.

[0019] in: Detailed Implementation

[0020] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0021] Example 1: According to the present invention, a method for optimizing the sorting parameters of viscoelastic microfluidic particles includes the following steps: Step S1: Based on the total flow in the straight channel segment The flow rate ratio of sheath fluid to sample fluid Distribution function of PEO concentration along the channel width Horizontal coordinates of particles Particle diameter Particles inx velocity in direction Particles in y velocity in direction Channel height Width of straight passage section Fluid density PEO molecular weight PEO concentration in sheath fluid PEO concentration of the sample Solvent viscosity of PEO solution shear rate Avogadro's constant Boltzmann constant and temperature Constructing the inertial lift force on the particles Elastic force Viscous drag force and virtual mass force The expression; obtained through calculation and Based on the distribution along the channel width, a sample-sheath fluid interface penetration criterion is constructed; , , and A particle dynamics model is constructed to calculate the trajectory of the particle after it has fully entered the sheath fluid, and the lateral equilibrium position of the particle in the straight channel section is given. y equ And the channel length required to achieve balance L equ ; Viscoelastic microfluidic chip channel height h The width of the straight channel section is 50μm. w The molecular weight of PEO is 120 μm. The distribution function of PEO concentration along the channel width at 600 kDa. Expressed using equation (1): (1) Where w1 is the width of the sheath fluid in the straight channel segment, which can be solved from equation (2): (2) Inertial lift force on the particle Elastic force Viscous drag force and virtual mass force The expressions for are shown in equation (3): (3) in and Let be the unit vectors pointing from the main flow direction and the channel center to the wall, respectively; the sample-sheath interface penetration criterion is: if and The curve of the resultant force is in If a negative slope zero point exists within the interval, it is determined that the particle cannot completely penetrate the sample-sheath fluid interface; otherwise, it is determined that the particle can completely penetrate the sample-sheath fluid interface. The particle dynamics model is characterized by equation (4): (4).

[0022] Step S2: Set particle diameter PEO concentration in sheath fluid PEO concentration of the sample Total flow in straight sections The flow rate ratio of sheath fluid to sample fluid The range of values ​​is determined by using Latin sampling to obtain sample points and calculating the range of values ​​for each sample point. y equ and L equ ; Particle diameter PEO concentration in sheath fluid PEO concentration of the sample Total flow in straight sections The flow rate ratio of sheath fluid to sample fluid The range of values ​​for: , , , , The number of Latin samples is greater than or equal to 45,000.

[0023] Step S3: Train three random forest sub-models using the above sample points to predict particle diameters in different environments. , , , The penetration ability of the sample-sheath fluid interface and y equ and L equ ; The three random forest sub-models are penetration classification models used to determine whether particles can completely penetrate the sample-sheath fluid interface. y equ Fitting a random forest regression model, L equ Fitting a conditional random forest regression model; where the penetration classification model and y equThe fitted random forest regression model is trained on the entire dataset. L equ The fitted conditional random forest regression model is trained only on data samples where particles can completely penetrate the sample-sheath fluid interface.

[0024] Step S4: Generate sample points in the parameter space using Latin sampling, and determine the optimal sorting parameters for viscoelastic microfluidics based on hard constraints, Pareto optimization results, fitting error, similarity between solutions, stability, and the importance of each performance index.

[0025] The initial Latin sample size is greater than or equal to 10,000. Hard constraints are used to remove sample points from the initial sample that do not meet the hard constraint conditions; where hard constraints refer to… and In this context, subscript 1 indicates smaller particles in the mixture, and subscript 2 indicates larger particles in the mixture. Pareto optimization has three objectives: maximizing y equ1 , minimize y equ2 And maximizing sample flow rate Q s = Q / (1+ Q r These correspond to maximizing the purity of large-diameter particles, maximizing the recovery rate of large-diameter particles, and maximizing the sorting throughput, respectively. Based on this, the initial Pareto front solution set is obtained. y equ and L equ The fitting error tolerances are 3% and 10%, respectively. Sample points that do not meet the fitting error tolerance in the initial Pareto front solution set are removed. Put ±0.8μm, ±0.8μm, ±0.2mL·h -1 Each is used to determine the relationship between two samples. y equ1 , y equ2 and Q s Whether the two sample points are similar to a threshold; if the two sample points are at... y equ1 , y equ2 , Q s If two or three samples simultaneously exhibit similarity, the sample point with the lower weighted score is removed; the weighted scoring formula is: 0.45*( y equ1 / 60)-0.45*( y equ1 / 60)+0.1*( Q s / 10); ±1 mL·h -1 ±0.4%, ±0.01%, and ±0.01% were respectively calculated. Q , Q r , c c , c s The tolerance range; for each sample point, 35 test points are generated within this tolerance range. If more than 80% of the test points can satisfy the hard constraints, the sample point is considered stable; otherwise, the sample point is considered unstable and is removed from the solution set. The initial Pareto front solution set was processed by removing fitting errors, similarity between solutions, and stability to obtain the final Pareto front solution set. The optimal sorting parameters for viscoelastic microfluidics were determined based on purity, recovery rate, and flux.

[0026] The present invention also provides a viscoelastic microfluidic particle sorting parameter optimization system, which can be implemented by executing the process steps of the viscoelastic microfluidic particle sorting parameter optimization method. That is, those skilled in the art can understand the viscoelastic microfluidic particle sorting parameter optimization method as a preferred embodiment of the viscoelastic microfluidic particle sorting parameter optimization system.

[0027] Example 2: Those skilled in the art can understand this embodiment as a more specific description of Embodiment 1.

[0028] The present invention also provides a viscoelastic microfluidic particle sorting parameter optimization system, the system comprising the following modules: Module M1: Based on the total flow in the straight section The flow rate ratio of sheath fluid to sample fluid Distribution function of PEO concentration along the channel width Horizontal coordinates of particles Particle diameter Particles in x velocity in direction Particles in y velocity in direction Channel height Width of straight passage section Fluid density PEO molecular weight PEO concentration in sheath fluid PEO concentration of the sample Solvent viscosity of PEO solution shear rate Avogadro's constant Boltzmann constant and temperature Constructing the inertial lift force on the particles Elastic force Viscous drag force and virtual mass force The expression; obtained through calculation and Based on the distribution along the channel width, a sample-sheath fluid interface penetration criterion is constructed; , , and A particle dynamics model is constructed to calculate the trajectory of the particle after it has fully entered the sheath fluid, and the lateral equilibrium position of the particle in the straight channel section is given. y equ And the channel length required to achieve balance L equ ; Viscoelastic microfluidic chip channel height h The width of the straight channel section is 50μm. w The molecular weight of PEO is 120 μm. The distribution function of PEO concentration along the channel width at 600 kDa. Expressed using equation (1): (1) Where w1 is the width of the sheath fluid in the straight channel segment, which can be solved from equation (2): (2) Inertial lift force on the particle Elastic force Viscous drag force and virtual mass force The expressions for are shown in equation (3): (3) in and Let be the unit vectors pointing from the main flow direction and the channel center to the wall, respectively; the sample-sheath interface penetration criterion is: if and The curve of the resultant force is in If a negative slope zero point exists within the interval, it is determined that the particle cannot completely penetrate the sample-sheath fluid interface; otherwise, it is determined that the particle can completely penetrate the sample-sheath fluid interface. The particle dynamics model is characterized by equation (4): (4).

[0029] Module M2: Set particle diameter PEO concentration in sheath fluid PEO concentration of the sample Total flow in straight sections The flow rate ratio of sheath fluid to sample fluid The range of values ​​is determined by using Latin sampling to obtain sample points and calculating the range of values ​​for each sample point. y equ and L equ ; Particle diameter PEO concentration in sheath fluid PEO concentration of the sample Total flow in straight sections The flow rate ratio of sheath fluid to sample fluid The range of values ​​for: , , , , The number of Latin samples is greater than or equal to 45,000.

[0030] Module M3: Trains three random forest sub-models using the above sample points to predict particle diameters at different times. , , , The penetration ability of the sample-sheath fluid interface and y equ and L equ ; The three random forest sub-models are penetration classification models used to determine whether particles can completely penetrate the sample-sheath fluid interface. y equ Fitting a random forest regression model, L equ Fitting a conditional random forest regression model; where the penetration classification model and y equ The fitted random forest regression model is trained on the entire dataset. L equ The fitted conditional random forest regression model is trained only on data samples where particles can completely penetrate the sample-sheath fluid interface.

[0031] Module M4: Generates sample points in the parameter space using Latin sampling, and determines the optimal sorting parameters for viscoelastic microfluidics based on hard constraints, Pareto optimization results, fitting error, similarity between solutions, stability, and the importance of each performance index.

[0032] The initial Latin sample size is greater than or equal to 10,000. Hard constraints are used to remove sample points from the initial sample that do not meet the hard constraint conditions; where hard constraints refer to… and In this context, subscript 1 indicates smaller particles in the mixture, and subscript 2 indicates larger particles in the mixture. Pareto optimization has three objectives: maximizing y equ1 , minimize y equ2 And maximizing sample flow rate Q s = Q / (1+ Q r These correspond to maximizing the purity of large-diameter particles, maximizing the recovery rate of large-diameter particles, and maximizing the sorting throughput, respectively. Based on this, the initial Pareto front solution set is obtained. y equ and L equ The fitting error tolerances are 3% and 10%, respectively. Sample points that do not meet the fitting error tolerance in the initial Pareto front solution set are removed. Put ±0.8μm, ±0.8μm, ±0.2mL·h -1 Each is used to determine the relationship between two samples. y equ1 , y equ2 and Q s Whether the two sample points are similar to a threshold; if the two sample points are at... y equ1 , y equ2 , Q s If two or three samples simultaneously exhibit similarity, the sample point with the lower weighted score is removed; the weighted scoring formula is: 0.45*( y equ1 / 60)-0.45*( y equ1 / 60)+0.1*( Q s / 10); ±1 mL·h-1 ±0.4%, ±0.01%, and ±0.01% were respectively calculated. Q , Q r , c c , c s The tolerance range; for each sample point, 35 test points are generated within this tolerance range. If more than 80% of the test points can satisfy the hard constraints, the sample point is considered stable; otherwise, the sample point is considered unstable and is removed from the solution set. The initial Pareto front solution set was processed by removing fitting errors, similarity between solutions, and stability to obtain the final Pareto front solution set. The optimal sorting parameters for viscoelastic microfluidics were determined based on purity, recovery rate, and flux.

[0033] Example 3: Those skilled in the art can understand this embodiment as a more specific description of Embodiment 1.

[0034] like Figure 1 As shown, the viscoelastic microfluidic chip mainly consists of a first inlet 1, a second inlet 2, a circular bifurcation branch 3, a straight channel section 4, an amplification section 5, a first outlet 6, and a second outlet 7. Particle samples and sheath fluid are injected into the viscoelastic microfluidic chip through the first inlet 1 and the second inlet 2, respectively. After flowing through the circular bifurcation branch 3, the particle samples flow into the straight channel section 4 together with the sheath fluid. Small-diameter and large-diameter particles flow out of the microfluidic chip through the amplification section 5 through the first outlet 6 and the second outlet 7, respectively.

[0035] like Figure 2 As shown, after entering the straight channel section 4, the mixed particulate solution is squeezed to both sides of the channel by the sheath fluid. Under inertial lift... elastic force viscous drag force and virtual mass force Under the combined effect of these factors, large-diameter and small-diameter particles will form different trajectories, causing the small-diameter and large-diameter particles to reach equilibrium near the sample-sheath fluid interface and the channel centerline, respectively, when they reach the end of the straight channel section 4. When the particles enter the amplification section 5, the small-diameter particles will continue to flow out from the first outlet 6 along the sample-sheath fluid interface, while the large-diameter particles will flow out from the second outlet 7.

[0036] like Figure 3 As shown, under different operating conditions, the particles will be in equilibrium at different lateral positions in the straight channel section 4. y equ And the channel length required to reach equilibrium. L equ They are different.

[0037] like Figure 4 As shown, a machine learning-driven framework for optimizing viscoelastic microfluidic particle sorting parameters is presented. , , , , These five parameters serve as input. y equ , L equ The ability of particles to penetrate the interface is used as the output. Three random forest sub-models are trained using 45,000 theoretical calculation results: a penetration classification model to determine whether particles can completely penetrate the sample-sheath fluid interface; and a penetration classification model to determine whether particles can completely penetrate the sample-sheath fluid interface. y equ Fitting a random forest regression model, L equ Fitting a conditional random forest regression model. This includes a penetration classification model and... y equ The fitted random forest regression model is trained on the entire dataset. L equ The fitted conditional random forest regression model is trained only on data samples where particles can completely penetrate the sample-sheath fluid interface. During multi-objective optimization, the initial Latin sample size is greater than or equal to 10,000, and then hard constraints are used to remove sample points from the initial sample that do not meet the hard constraints. The hard constraints refer to… and Here, subscript 1 represents the smaller particle in the mixture, and subscript 2 represents the larger particle. Pareto optimization has three objectives: maximizing... y equ1 , minimize y equ2 And maximizing sample flow rate Q s = Q / (1+ Q r These correspond to maximizing the purity of large-diameter particles, maximizing the recovery rate of large-diameter particles, and maximizing the sorting flux, respectively, based on which the initial Pareto front solution set is obtained. (Setting...) y equ and L equ The fitting error tolerances were 3% and 10%, respectively. Sample points that did not meet the fitting error tolerances in the initial Pareto front solution set were removed. ±0.8 μm, ±0.8 μm, ±0.2 mL·h -1 Each is used to determine the relationship between two samples. y equ1 ,y equ2 and Q s The threshold for similarity. If two sample points are at... y equ1 , y equ2 , Q s If two or three samples show similarity simultaneously, the sample point with the lower weighted score is removed. The weighted score formula is 0.45*( y equ1 / 60)-0.45*( y equ1 / 60)+0.1*( Q s / 10). Add 1 mL·h -1 ±0.4, ±0.01%, and ±0.01% are respectively used as... Q , Q r , c c , c s The tolerance range is defined. For each sample point, 35 test points are generated within this tolerance range. If more than 80% of the test points can satisfy the hard constraints, the sample point is considered stable and will not have a significant positional deviation due to slight fluctuations in parameters. Otherwise, the sample point is considered unstable and needs to be removed from the solution set. After removing fitting errors, similarity between solutions, and stability from the initial Pareto front solution set, the final Pareto front solution set can be obtained. For specific application needs, researchers can determine the optimal sorting parameters for viscoelastic microfluidics based on key indicators such as purity, recovery rate, and flux.

[0038] like Figure 5 As shown, in c c =0.13%, c s =0.05%, Q =20 mL·h -1 , Q r Under the condition of 10, the lateral equilibrium positions of 10 μm and 15 μm particles are respectively y equ1 =47.26 μm and y equ2 =19.80 μm, which means that the 10 μm and 15 μm particles are in equilibrium near the sample-sheath fluid interface and the channel midline, respectively.

[0039] like Figure 6 As shown, inc c =0.13%, c s =0.05%, Q =20 mL·h -1 , Q r Under operating conditions of 10, 15 μm particles, after completely penetrating the sample-sheath fluid interface, will continue to migrate towards the channel centerline. The channel length required to reach equilibrium is [value missing]. L equ =20.88 mm.

[0040] like Figure 7 As shown, the random forest model proposed in this embodiment of the invention... y equ and L equ It has good prediction accuracy. R 2 They can achieve 0.984 and 0.998 respectively.

[0041] like Figure 8 As shown, compared to L equ Direct regression analysis is proposed in this embodiment of the invention. L equ Conditional regression models have higher prediction accuracy.

[0042] like Figures 9-10 As shown, taking the sorting of 10 μm / 20 μm mixed particles as an example, a machine learning-driven viscoelastic microfluidic particle sorting parameter optimization method proposed in this invention can quickly obtain the three-dimensional and two-dimensional Pareto fronts.

[0043] The working principle of this invention is a machine learning-driven method for optimizing viscoelastic microfluidic particle sorting parameters, mainly comprising three parts: theoretical calculation, machine learning training and prediction, and multi-objective optimization. The theoretical calculation is used to rapidly and extensively obtain theoretical results on the lateral equilibrium position of particles of various diameters in a straight channel section and the channel length required to reach equilibrium under different sheath fluid concentrations, PEO concentrations in the sample, total flow rates, and sheath fluid to sample flow rate ratios. Subsequently, a random forest machine learning algorithm is used to train on a large number of theoretical results to accurately determine whether particles can completely penetrate the sample-sheath fluid interface under various combinations of operating parameters and to predict the lateral equilibrium position of the particles and the channel length required to reach equilibrium. Finally, for specific sorting requirements, the optimal sorting parameters of the viscoelastic microfluidic system are determined based on hard constraints, Pareto optimization results, fitting error, solution similarity, stability, and the importance of each performance index.

[0044] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0045] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for optimizing the sorting parameters of viscoelastic microfluidic particles, characterized in that, The method includes the following steps: Step S1: Based on the total flow in the straight channel segment The flow rate ratio of sheath fluid to sample fluid Distribution function of PEO concentration along the channel width Horizontal coordinates of particles Particle diameter Particles in x velocity in direction Particles in y velocity in direction Channel height Width of straight passage section Fluid density PEO molecular weight PEO concentration in sheath fluid PEO concentration of the sample Solvent viscosity of PEO solution shear rate Avogadro's constant Boltzmann constant and temperature Constructing the inertial lift force on the particles Elastic force Viscous drag force and virtual mass force The expression; obtained through calculation and Based on the distribution along the channel width, a sample-sheath fluid interface penetration criterion is constructed; , , and A particle dynamics model is constructed to calculate the trajectory of the particle after it has fully entered the sheath fluid, and the lateral equilibrium position of the particle in the straight channel section is given. y equ And the channel length required to achieve balance L equ ; Step S2: Set particle diameter PEO concentration in sheath fluid PEO concentration of the sample Total flow in straight sections The flow rate ratio of sheath fluid to sample fluid The range of values ​​is determined by using Latin sampling to obtain sample points and calculating the range of values ​​for each sample point. y equ and L equ ; Step S3: Train three random forest sub-models using the above sample points to predict particle diameters in different environments. , , , The penetration ability of the sample-sheath fluid interface and y equ and L equ ; Step S4: Generate sample points in the parameter space using Latin sampling, and determine the optimal sorting parameters for viscoelastic microfluidics based on hard constraints, Pareto optimization results, fitting error, similarity between solutions, stability, and the importance of each performance index.

2. The method for optimizing viscoelastic microfluidic particle sorting parameters according to claim 1, characterized in that, The viscoelastic microfluidic chip channel height in step S1 h The width of the straight channel section is 50μm. w The molecular weight of PEO is 120 μm. The distribution function of PEO concentration along the channel width at 600 kDa. Expressed using equation (1): （1） Where w1 is the width of the sheath fluid in the straight channel segment, which can be solved from equation (2): （2） Inertial lift force on the particle Elastic force Viscous drag force and virtual mass force The expressions for are shown in equation (3): （3） in and Let be the unit vectors pointing from the main flow direction and the channel center to the wall, respectively; the sample-sheath interface penetration criterion is: if and The curve of the resultant force is in If a negative slope zero point exists within the interval, it is determined that the particle cannot completely penetrate the sample-sheath fluid interface; otherwise, it is determined that the particle can completely penetrate the sample-sheath fluid interface. The particle dynamics model is characterized by equation (4): （4）。 3. The method for optimizing viscoelastic microfluidic particle sorting parameters according to claim 1, characterized in that, In step S2, the particle diameter PEO concentration in sheath fluid PEO concentration of the sample Total flow in straight sections The flow rate ratio of sheath fluid to sample fluid The range of values ​​for: , , , , The number of Latin samples is greater than or equal to 45,000.

4. The method for optimizing viscoelastic microfluidic particle sorting parameters according to claim 1, characterized in that, The three random forest sub-models in step S3 are penetration classification models used to determine whether particles can completely penetrate the sample-sheath fluid interface. y equ Fitting a random forest regression model, L equ Fitting a conditional random forest regression model; Among them, the penetration classification model and y equ The fitted random forest regression model is trained on the entire dataset. L equ The fitted conditional random forest regression model is trained only on data samples where particles can completely penetrate the sample-sheath fluid interface.

5. The method for optimizing viscoelastic microfluidic particle sorting parameters according to claim 1, characterized in that, In step S4, the initial Latin sample count is greater than or equal to 10,000. Hard constraints are used to remove sample points from the initial sample that do not meet the hard constraint conditions. Here, hard constraints refer to… and In this context, subscript 1 indicates smaller particles in the mixture, and subscript 2 indicates larger particles in the mixture. The Pareto optimization objectives in step S4 are threefold: maximizing... y equ1 , minimize y equ2 And maximizing sample flow rate Q s = Q / (1+ Q r These correspond to maximizing the purity of large-diameter particles, maximizing the recovery rate of large-diameter particles, and maximizing the sorting throughput, respectively. Based on this, the initial Pareto front solution set is obtained. In step S4 y equ and L equ The fitting error tolerances are 3% and 10%, respectively. Sample points that do not meet the fitting error tolerance in the initial Pareto front solution set are removed. In step S4, ±0.8μm, ±0.8μm, ±0.2mL·h are used. -1 Each is used to determine the relationship between two samples. y equ1 , y equ2 and Q s Whether the two sample points are similar to a threshold; if the two sample points are at... y equ1 , y equ2 , Q s If two or three samples simultaneously exhibit similarity, the sample point with the lower weighted score is removed; the weighted scoring formula is: 0.45*( y equ1 / 60)-0.45*( y equ1 / 60)+0.1*( Q s / 10)； In step S4, ±1 mL·h -1 ±0.4%, ±0.01%, and ±0.01% were respectively calculated. Q , Q r , c c , c s The tolerance range; for each sample point, 35 test points are generated within this tolerance range. If more than 80% of the test points can satisfy the hard constraints, the sample point is considered stable; otherwise, the sample point is considered unstable and is removed from the solution set. In step S4, the initial Pareto front solution set is processed by removing fitting errors, similarity between solutions, and stability to obtain the final Pareto front solution set; the optimal sorting parameters for viscoelastic microfluidics are determined based on purity, recovery rate, and flux.

6. A viscoelastic microfluidic particle sorting parameter optimization system, characterized in that, The system includes the following modules: Module M1: Based on the total flow in the straight section The flow rate ratio of sheath fluid to sample fluid Distribution function of PEO concentration along the channel width Horizontal coordinates of particles Particle diameter Particles in x velocity in direction Particles in y velocity in direction Channel height Width of straight passage section Fluid density PEO molecular weight PEO concentration in sheath fluid PEO concentration of the sample Solvent viscosity of PEO solution shear rate Avogadro's constant Boltzmann constant and temperature Constructing the inertial lift force on the particles Elastic force Viscous drag force and virtual mass force The expression; obtained through calculation and Based on the distribution along the channel width, a sample-sheath fluid interface penetration criterion is constructed; , , and A particle dynamics model is constructed to calculate the trajectory of the particle after it has fully entered the sheath fluid, and the lateral equilibrium position of the particle in the straight channel section is given. y equ And the channel length required to achieve balance L equ ; Module M2: Set particle diameter PEO concentration in sheath fluid PEO concentration of the sample Total flow in straight sections The flow rate ratio of sheath fluid to sample fluid The range of values ​​is determined by using Latin sampling to obtain sample points and calculating the range of values ​​for each sample point. y equ and L equ ; Module M3: Trains three random forest sub-models using the above sample points to predict particle diameters at different times. , , , The penetration ability of the sample-sheath fluid interface and y equ and L equ ; Module M4: Generates sample points in the parameter space using Latin sampling, and determines the optimal sorting parameters for viscoelastic microfluidics based on hard constraints, Pareto optimization results, fitting error, similarity between solutions, stability, and the importance of each performance index.

7. The viscoelastic microfluidic particle sorting parameter optimization system according to claim 6, characterized in that, The viscoelastic microfluidic chip channel height in module M1 h The width of the straight channel section is 50μm. w The molecular weight of PEO is 120 μm. The distribution function of PEO concentration along the channel width at 600 kDa. Expressed using equation (1): （1） Where w1 is the width of the sheath fluid in the straight channel segment, which can be solved from equation (2): （2） Inertial lift force on the particle Elastic force Viscous drag force and virtual mass force The expressions for are shown in equation (3): （3） in and Let be the unit vectors pointing from the main flow direction and the channel center to the wall, respectively; the sample-sheath interface penetration criterion is: if and The curve of the resultant force is in If a negative slope zero point exists within the interval, it is determined that the particle cannot completely penetrate the sample-sheath fluid interface; otherwise, it is determined that the particle can completely penetrate the sample-sheath fluid interface. The particle dynamics model is characterized by equation (4): （4）。 8. The viscoelastic microfluidic particle sorting parameter optimization system according to claim 6, characterized in that, Particle diameter in module M2 PEO concentration in sheath fluid PEO concentration of the sample Total flow in straight sections The flow rate ratio of sheath fluid to sample fluid The range of values ​​for: , , , , The number of Latin samples is greater than or equal to 45,000.

9. The viscoelastic microfluidic particle sorting parameter optimization system according to claim 6, characterized in that, The three random forest sub-models in module M3 are penetration classification models used to determine whether particles can completely penetrate the sample-sheath fluid interface. y equ Fitting a random forest regression model, L equ Fitting a conditional random forest regression model; Among them, the penetration classification model and y equ The fitted random forest regression model is trained on the entire dataset. L equ The fitted conditional random forest regression model is trained only on data samples where particles can completely penetrate the sample-sheath fluid interface.

10. The viscoelastic microfluidic particle sorting parameter optimization system according to claim 6, characterized in that, In module M4, the initial Latin sample count is greater than or equal to 10,000. Hard constraints are used to remove sample points from the initial sample that do not meet the hard constraint conditions; where hard constraints refer to… and In this context, subscript 1 indicates smaller particles in the mixture, and subscript 2 indicates larger particles in the mixture. The Pareto optimization objectives in module M4 are three: maximizing y equ1 , minimize y equ2 And maximizing sample flow rate Q s = Q / (1+ Q r These correspond to maximizing the purity of large-diameter particles, maximizing the recovery rate of large-diameter particles, and maximizing the sorting throughput, respectively. Based on this, the initial Pareto front solution set is obtained. In module M4 y equ and L equ The fitting error tolerances are 3% and 10%, respectively. Sample points that do not meet the fitting error tolerance in the initial Pareto front solution set are removed. The module M4 will contain ±0.8μm, ±0.8μm, and ±0.2mL·h. -1 Each is used to determine the relationship between two samples. y equ1 , y equ2 and Q s Whether the two sample points are similar to a threshold; if the two sample points are at... y equ1 , y equ2 , Q s If two or three samples simultaneously exhibit similarity, the sample point with the lower weighted score is removed; the weighted scoring formula is: 0.45*( y equ1 / 60)-0.45*( y equ1 / 60)+0.1*( Q s / 10)； The module M4 will contain ±1 mL·h -1 ±0.4%, ±0.01%, and ±0.01% were respectively calculated. Q , Q r , c c , c s The tolerance range; for each sample point, 35 test points are generated within this tolerance range. If more than 80% of the test points can satisfy the hard constraints, the sample point is considered stable; otherwise, the sample point is considered unstable and is removed from the solution set. In module M4, the initial Pareto front solution set is processed by removing fitting errors, similarity between solutions, and stability to obtain the final Pareto front solution set; the optimal sorting parameters for viscoelastic microfluidics are determined based on purity, recovery rate, and flux.

Images