Method for predicting fluid field of low reynolds number microfluidic chip based on artificial neural network

Abstract

The present application relates to a method for predicting fluid field of low Reynolds number microfluidic chip based on artificial neural network, which can quickly predict the simulation microfluidic performance before the microfluidic chip is manufactured. A nine-square model is proposed, and 15 kinds of artificial neural network models are summarized. Under the premise of only boundary conditions and geometric shape, the fluid field of the whole microfluidic mixing area is predicted. Starting from the upper left corner of the whole microfluidic mixing area, the left boundary and the upper boundary are given, and the fluid field information of the whole area is predicted by using the 15 kinds of artificial neural network models. From the mathematical point of view, any shape can be infinitely approached by the nine-square model, and it can be deduced that theoretically, the nine-square model can be used to predict the fluid information of the microfluidic mixer of any shape.

Description

TECHNICAL FIELD

[0001] The present application relates to a method for predicting a fluid field of a low Reynolds number microfluidic chip based on an artificial neural network, and belongs to the technical field of automated design of biochips and fluid mechanics. BACKGROUND

[0002] With the progress of science and technology, microfluidic technology shows great potential in the fields of biology, chemistry, medicine, etc. Microfluidic chips, also known as lab-on-a-chip, transfer the experimental platform from traditional large laboratories to micro portable chips. These chips are usually a few square centimeters or smaller. Due to their ability to integrate multiple independent units on a micro platform, their small size, low energy consumption, and fast reaction speed, they are widely used in medical diagnosis, cell sorting, single-cell RNA sequencing or droplet digital PCR, gene protein analysis, drug discovery, etc.

[0003] Microfluidic experimental platforms are used by researchers in various disciplines, and are mostly custom-made equipment based on their experimental backgrounds. They cannot be widely used. In the design process of previous microfluidic chips, researchers first design on a computer according to the expected function, then produce and test the actual chip function. However, due to the complexity of custom equipment design, manufacturing and operation, these complexities limit their reproducibility and universality. Although manufacturing can be outsourced (although at a high cost), microfluidic design and operation may require several months to years of "design-manufacture-test" iterations to optimize performance. Not all researchers are familiar with the design and manufacture of microfluidic chips, and they only need to use microfluidic chips for their research. This requires a lot of time to design the chip, which greatly inconveniences their research.

[0004] Machine learning is the use of trainable statistical models to identify patterns and predict future behavior, and has been widely applied in various fields in recent years. Artificial neural networks are a type of machine learning that abstracts the neural network of the human brain from the perspective of information processing, establishes a certain simple model, and forms different networks according to different connection methods. Applying artificial neural networks to microfluidic chip design can narrow the gap between experts and end users, and achieve automation of microfluidic design and operation.

[0005] Currently, many microfluidic technologies are too complex to be accurately modeled. In the microfluidic chip design process, researchers first need to design it on a computer based on the intended function, and then produce and test the actual chip function. However, it is limited by the complexity of custom equipment design, manufacturing and operation, which limits its reproducibility and versatility. If a microfluidic chip with a different structure needs to repeat the above steps, it is time-consuming and labor-intensive. Usually, not all researchers are familiar with the design and manufacture of microfluidic chips. They just need to use microfluidic chips for their research content, which requires a lot of time to design the chip, which brings great inconvenience to their research. Although fluid dynamics computational modeling is fast, it is easy to make mistakes in analyzing fluid properties.

[0006] With the rapid development of artificial neural networks, people try to use artificial neural networks to solve problems in other fields across disciplines. The present invention is applied to the fields of biochip automation design and fluid mechanics technology, and is a method for predicting the fluid field of a general microfluidic chip under low Reynolds number based on artificial neural networks.

[0007] This paper designs a method for predicting the fluid field of a universal microfluidic chip at low Reynolds numbers based on an artificial neural network. By establishing a large number of random microfluidic chips to obtain a fluid velocity field dataset, an artificial neural network model is synthesized based on a nine-square grid model. The artificial neural network model is trained and then used to predict the fluid field of a universal microfluidic chip at low Reynolds numbers. Finally, the structure similarity index (SSIM) is used to measure the image similarity based on brightness, contrast, and structure. Summary of the Invention

[0008] To overcome the shortcomings of existing research, the present invention provides a technology based on artificial neural networks to predict the fluid field of a universal microfluidic chip, converting the fluid mechanics problem into a nonlinear problem based on artificial neural networks to explore the intrinsic relationship of data.

[0009] The specific steps of the method for predicting the fluid field of low Reynolds number microfluidic chip based on artificial neural network are as follows:

[0010] S1, Create microfluidic chip model. The data of the present application is derived from a finite element simulation software model. Random microfluidic chip design scheme, because the possible chip design types are basically unlimited, therefore some restrictions are added in the design process. Each microfluidic chip has two inlets, two outlets and a design area of 500um·500um, and a random mixing structure is located in the area. The design domain is divided into 100·100 grids with a size of 5um·5um, and we use a 100·100 01 matrix to represent whether to create a grid, where 1 represents to create a matrix and 0 represents not to create a matrix. The randomness of the microfluidic chip is controlled by controlling the probability of 01, so as to realize the random design of the mixing domain, to balance the calculation resources in the 500um·500um design domain as much as possible to design different chip design schemes. Each grid structure acts as a building block of mixing features in each design domain. We generated 10000 different design schemes accordingly.

[0011] Random microfluidic chip performance simulation. Finite element analysis software is used to design and simulate chip performance. Set up a laminar flow physical module and a dilute substance transfer physical module, and two fixed solvers to calculate. Among them, in the laminar flow physical module, the boundary conditions of the two inlets are defined as 1mm·s -1 normal inflow velocity, the boundary conditions of each outlet are defined as 0Pa pressure, and the remaining boundary is no-slip boundary condition, and the material filled in the channel is incompressible flow type water. In the dilute substance transfer physical module, the boundary condition of inlet 1 is 1mmol·L -1 , the boundary condition of inlet 2 is 0mmol·L -1 , and the two outlets are defined as outflow. The solute diffusion coefficient of fluorescein is 4.25·10 -10 m 2 ·s -1 . Thus the fluid field of the microfluidic chip can be obtained. After simulation calculation, the fluid field information of the chip design can be obtained, two inlets, two outlets, chip design domain, and the flow velocity is represented by color legend, wherein the bottom deep blue represents the minimum value 0m·s -1 , and the top deep red represents the maximum value 20·10 -4 m·s -1 . Using this method, 10000 different microfluidic chip fluid performances are simulated on the computer.

[0012] S2, an artificial neural network model, predicts the fluid field of a microfluidic chip. S1 is a traditional FEA simulation of the fluid field of a microfluidic chip. We now propose a novel, completely data-based method for predicting the fluid field of a microfluidic chip using an artificial neural network model. Data is extracted from 10,000 different random microfluidic chip models in S1, meshed, boundary conditions set, and the artificial neural network model is summarized. Finally, a customized algorithm is used to call a pre-trained artificial neural network model to predict the fluid field of the entire microfluidic chip. This method avoids the complex and time-consuming geometric modeling process while achieving results in a short period of time.

[0013] S3. Extract data based on the low Reynolds number general microfluidic chip model. Extract data: Export the 100·100 01 matrix 301 that controls the random design of the microfluidic chip from the above model in .txt format. A microfluidic chip model has more than 70,000 nodes. Export the horizontal component and vertical component of the velocity of each node and its (x, y) coordinate value in the coordinate system. Through the exported data, divide the horizontal component and vertical component of the fluid velocity in the design domain into 100·100 matrices according to the (x, y) coordinate values. Each design corresponds to three 100·100 matrices, namely the 01 matrix, the horizontal component matrix of the velocity, and the vertical component matrix. The horizontal velocity component matrix is ​​denoted as V x , the vertical component velocity matrix is ​​recorded as V y .

[0014] S4. Generate a data set. Now define a nine-square grid 302, a matrix with 3 rows and 3 columns, numbered from left to right and from top to bottom, the first row is numbered: 1, 2, 3, the second row is numbered: 4, 5, 6, and the third row is numbered: 7, 8, 9. Using the numpy library of python, the three matrices 301, 401, 501 are sliced ​​with a size of 3·3 and a step length of (3,3) from the positions 303, 403, 503 using 302, 402, 502. After one slice, three matrices with 3 rows and 3 columns can be obtained, namely: whether to create a grid 01 matrix, denoted as [[P1, P2, P3], [P4, P5, P6], [P7, P8, P9]]; the horizontal velocity component matrix, denoted as [[V x,1 ,V x,2 ,V x,3 ],[V x,4 ,V x,5 ,V x,6 ],[V x,7 ,V x,8 ,V x,9 ]]; vertical velocity component matrix, denoted as [[V y,1 ,V y,2 ,Vy,3 ],[V y,4 ,V y,5 ,V y,6 ],[V y,7 ,V y,8 ,V y,9 ]]. Expand the cut data and splice them into a row to get a matrix with 1 row and 27 columns, that is, [P1,…,P9,V x,1 ,…,V x,9 ,V y,1 ,…,V y,9 ], after processing, one microfluidic chip model can generate 33·33 rows of data, and the 10,000 models generated previously will generate 10,000·33·33 pieces of data as a data set.

[0015] S5. Initialize V x , V y Matrix. In the 01 matrix of S3's 100·100, all 1 corresponding positions V x , V y Initialized to 0, all V corresponding to 0 x , V y Initialized to -1. This is because in the 01 matrix, 1 represents the creation of a grid, and the corresponding speed is 0; the position where the value is 0 in the 01 matrix is ​​the area to be tested. So we only need to predict V x , V y The position with the value -1 in the matrix is ​​sufficient.

[0016] S6, add boundary conditions. In S3, we get a 01 matrix 100·100, and in S5, we get two initialized V x , V y Now add a 1-row, 100-column matrix above each of these three matrices as the upper boundary condition, and add a 101-row, 1-column matrix on the far left as the left boundary condition. The upper boundary value is extracted from the 0 <x<500,500<y<505这一第一范围,由于这一范围为实体,所以对应01矩阵上边界,这一行所有值为0,对应V x , V y The upper boundary is extracted from the 0 of the microfluidic chip model <x<500,500<y<505这一范围,左边界值提取自微流体混合器的-5<x<0,0<y<505这一范围,由于该范围内全部被创建网格,所以对应01矩阵左边界所有值为1;对应V x , V yThe left boundary is extracted from the microfluidic mixer in the range of -5 < x < 0, 0 < y < 505, and all values in this column are 0. After the above operation, we obtain three 101·101 matrices, and the first row and the first column are the known boundary conditions, V x , V y The area with a median value of -1 represents the to-be-measured area, and the area with a value of 0 represents the created grid area without fluid passing through.

[0017] S7, induction artificial neural network model. If the V x , V y matrix in S7 is sliced using the nine-square grid idea, the first nine-square grid is predicted from the top left corner, and the 1st, 2nd, 3rd, 4th, and 7th positions are known boundary conditions, and only the 5th, 6th, 8th, and 9th positions are uncertain. All possible cases of the first nine-square grid are induced, and if the values of the 5th, 6th, 8th, and 9th positions are {V x，5 , V x，6 , V x，8 , V x，9 , V y，5 , V y，6 , V y，8 , V y，9 =-1, 0, 0, 0, -1, 0, 0, 0}, it represents that only the 5th position is the to-be-measured area, and the other positions are the created grid with a speed of 0. We record this model as ANN_5, and similarly, ANN_6, ANN_8, and ANN_9 can be deduced. Similarly, {V x，5 , V x，6 , V x，8 , V x，9 , V y，5 , V y，6 , V y，8 , V y，9 =-1, -1, 0, 0, -1, -1, 0, 0} represents that the 5th and 6th positions are to-be-measured areas, and the 8th and 9th positions are created grids, recorded as ANN_56, and the following models can be deduced: ANN_58, ANN_59, ANN_68, ANN_69, and ANN_89. Similarly, {V x，5 , V x，6 , V x，8 , V x，9 , V y，5 , V y，6 , V y，8 , V y，9 =-1, -1, -1, 0, -1, -1, -1, 0} represents that the 5th, 6th, and 8th positions are to-be-measured areas, and the 9th position is a created grid, recorded as ANN_568, and the following models can be deduced: ANN_569, ANN_589, and ANN_689. {V x，5 , Vx，6 ,V x，8 ,V x，9 ,V y，5 ,V y，6 ,V y，8 ,V y，9 =-1,-1,-1,-1,-1,-1,-1,-1} means that positions 5, 6, 8, and 9 are the areas to be measured, which are recorded as ANN_5689.

[0018] S8. Train the artificial neural network model. Specific parameters for the artificial neural network include the number of layers, type, number of neurons, and activation function type. Training results include the name of each artificial neural network model, the predicted output location, the accuracy of the training set, the accuracy of the test set, and the loss value of the training set.

[0019] S9. Use the trained artificial neural network model to predict the fluid field of the entire mixing domain.

[0020] Using the above 15 ANN models to predict the fluid field of the entire microfluidic mixing domain from the upper left corner, the 3·3 slice in 701 is used to obtain the first nine-square grid, and the values ​​of positions 8 and 9 are -1. The ANN_89 model is used to predict V x,8 , V y,8 ,V x,9 , V y,9 and update it to the value of V x , V y Then, we move the 3.3 slice one square to the right to get the second nine-square grid. The 1st, 2nd, and 3rd positions of the second nine-square grid are known boundary conditions. The values ​​of the 4th, 5th, 7th, and 8th positions are the values ​​predicted by the ANN model above for the first nine-square grid. Therefore, only the values ​​of the 6th and 9th positions are -1. The ANN_69 model predicts V x,6 , V y,6 ,V x,9 , V y,9 The value of . And so on, until the last column; when the last nine-square grid 704 is predicted, the slice starts to predict from position 705. Positions 1, 4, and 7 are known boundary conditions. Positions 2, 3, 5, and 6 are obtained from the prediction of the first nine-square grid. The value of position 8 is -1, and the value of position 9 is 0, so ANN_8 is used for prediction. And so on, these 15 ANN models can predict all V x , V y .

[0021] S10. Use the SSIM algorithm to compare the similarity between the predicted results and the actual simulation results. The structural similarity index (SSIM index) is an indicator used to measure the similarity between two digital images. It measures the brightness, contrast, and structure of the image. The index ranges from 0 to 1. The SSIM algorithm scores 1 for two identical images. The closer the score is to 1, the more similar the two images are and the more accurate the prediction. To explain the prediction results, the SSIM algorithm was used to calculate the similarity between the original image and the image rotated 30°, with a score of 0.21. The SSIM algorithm was used to calculate the similarity between the original image and the image translated 30px to the lower right, with a score of 0.20. We also calculated the similarity between the original image and the image predicted by the artificial neural network using the SSIM algorithm, with a score of 0.40. To test the effectiveness of the artificial neural network model, 500 microfluidic chips were redesigned, and the SSIM algorithm was used to calculate the similarity between the results simulated on the computer and the results predicted by the artificial neural network. The results were presented in a histogram.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] The present invention avoids the steps of designing chips on computers, producing and testing actual chip functions. It is no longer limited by the complexity of custom equipment design, manufacturing and operation. The fluid mechanics problem is converted into a nonlinear regression problem based on artificial neural networks to explore the intrinsic relationships of big data. From a mathematical point of view, any regular and irregular shape can be gridded, so in theory, the nine-square grid model can predict the fluid field of microfluidic chips of any shape. Combining machine learning with microfluidic design and testing will help eliminate many obstacles in research and development, improve the success rate, and accelerate the commercialization process of microfluidic platforms.

[0024] The present invention is mainly designed to be a universal microfluidic chip under low Reynolds number laminar flow. In fluid mechanics, the Reynolds number is a dimensionless number that can be used to characterize the flow conditions of a fluid. When the Reynolds number is small, the effect of viscous force on the flow field is usually greater than inertia, and the disturbance of the flow velocity in the flow field will be attenuated by the viscous force, and the fluid flow is stable. Machine learning is very suitable for predicting such fluid behavior. By predicting the fluid field of a microfluidic chip based on an artificial neural network, the desired chip can be designed in a relatively short time and the expected fluid performance can be seen. The fluid mechanics problem is converted into a nonlinear regression problem based on an artificial neural network to explore the intrinsic relationship of big data.

[0025] The present invention uses a nine-grid model to grid the area to be predicted, simplifies the boundary conditions of the microfluidic chip, trains an artificial neural network model based on an existing data set, uses the artificial neural network model to predict the fluid field of a low-Reynolds number universal microfluidic chip, starts the prediction from the upper left corner of the area to be measured, calls the pre-trained artificial neural network model to predict the fluid field of the microfluidic chip, and uses the SSIM algorithm to measure the similarity between the target result and the predicted result in terms of brightness, contrast and structure.

[0026] The present invention can quickly predict and simulate microfluidic performance before the microfluidic chip is manufactured. A nine-square grid model is proposed, and 15 artificial neural network models are summarized. Under the premise of only boundary conditions and geometric shapes, the fluid field of the entire microfluidic mixing area is predicted. The prediction starts from the upper left corner of the fluid field of the entire microfluidic mixing domain. Given the upper boundary of the left boundary, 15 artificial neural network models are used to predict the fluid field information of the entire area. From a mathematical point of view, any shape can be infinitely approximated by the nine-square grid. It can be deduced that, in theory, the nine-square grid model can be used to predict the fluid information of a microfluidic mixer of any shape. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0028] Figure 1 It is a schematic diagram of the geometric model of the universal chip of the present invention;

[0029] Figure 2 It is a schematic diagram of the velocity field of a universal chip of the present invention;

[0030] Figure 3 This is a schematic diagram of the 100·100 01 matrix for controlling randomness of the hybrid domain of a universal chip according to the present invention;

[0031] Figure 4 The V of 100·100 obtained from the general chip model of the present invention is x Matrix diagram;

[0032] Figure 5 The V of 100·100 obtained from the general chip model of the present invention is y Matrix diagram;

[0033] Figure 6 101·101 is a schematic diagram of the 01 matrix with boundary conditions added to the present invention;

[0034] Figure 7The present invention adds boundary conditions to the V of 101·101 x Matrix diagram;

[0035] Figure 8 The present invention adds boundary conditions to the V of 101·101 y Matrix diagram;

[0036] Figure 9 Schematic diagram of the artificial neural network model of the present invention;

[0037] Figure 10 It is the overall flow chart of the present invention;

[0038] Figure 11 This is the prediction result of 500 microfluidic chips evaluated by SSIM of the present invention;

[0039] Figure 12 It is the training result of the ANN_5 model in the present invention;

[0040] Figure 13 It is the training result of the ANN_6 model in the present invention.

[0041] In the figure, 101-first entrance; 102-second entrance; 103-square chip design domain; 104-first exit; 105-second exit; 201-top of the color legend; 202-bottom of the color legend; 203-third entrance; 204-fourth entrance; 205-square area; 206-third exit; 207-fourth exit; 208-color legend; 209-upper boundary; 210-left boundary; 301-100·100 01 matrix; 302-illustration of the nine-square grid position; 303-the first position of the nine-square grid traversal in the 01 matrix; 401-100·100 V x Matrix; 402-Horizontal velocity component matrix; 403-V x The first position of the nine-square grid traversal in the matrix; V of 501-100·100 y Matrix; 502-vertical velocity component matrix; 503-V y The first position of the nine-square grid in the matrix; the 01 matrix of 601-101·101; the first nine-square grid content in the 01 matrix of 602-101·101; the first position of the nine-square grid in the 01 matrix of 603-101·01; the V of 701-101·101 x Matrix; 702-101·101 V x The first nine-square grid in the matrix; 703-101·101 V x The first position of the first row of the nine-square grid in the matrix; V of 704-101·101 xThe last position of the first row in the matrix is traversed; 705-101·101 V x The first position of the second row in the matrix is traversed; 801-101·101 V y Matrix; 802-101·101 V y The content of the first nine-square matrix in the matrix; 803-101·101 V y The first position of the first row in the matrix is traversed; 901-9-square matrix position diagram; 902-ANN_5; 903-ANN_6; 904-ANN_8; 905-ANN_9; 906-ANN_56; 907-ANN_58; 908-ANN_59; 909-ANN_68; 910-ANN_69; 911-ANN_89; 912-ANN_568; 913-ANN_569; 914-ANN_589; 915-ANN_689; 916-ANN_5689. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0043] The specific steps of the method for predicting a fluid field of a low Reynolds number microfluidic chip based on an artificial neural network are as follows:

[0044] S1, creating a microfluidic chip model. The data of the present application is derived from a finite element simulation software model. In the process of designing, some restrictions are added because the possible types of chip design are basically unlimited. Each microfluidic chip has two inlets (a first inlet 101 and a second inlet 102), two outlets (a first outlet 104 and a second outlet 105), and a design area of 500um·500um (a square chip design domain 103), and a random mixing structure is located in the area. The design domain is divided into 100·100 grids with a size of 5um·5um, and a 100·100 01 matrix is used to represent whether to create a grid, wherein 1 represents to create a matrix and 0 represents not to create a matrix. The randomness of the microfluidic chip is controlled by controlling the probability of 01, so as to realize the random design of the mixing domain, balance the calculation resources in the 500um·500um design domain as much as possible to design different chip design schemes, and each grid structure acts as a construction unit of the mixing feature in each design domain. Accordingly, 10000 different design schemes are generated.

[0045] Stochastic microfluidic chip performance simulation. Finite element analysis software was used to design and simulate the chip performance. The laminar flow physics module and the dilute species transport physics module were set up with two fixed solvers for calculation. In the laminar flow physics module, the boundary conditions of the two inlets (the third inlet 203 and the fourth inlet 204) were defined as 1 mm·s -1 The normal inflow velocity is 0 Pa, the boundary condition of each outlet defines the pressure, the other boundaries are no-slip boundary conditions, and the material filling the channel is water of incompressible flow type. In the Diluted Species Transport Physics Module, the boundary condition of the inlet is 1 mmol·L -1 , the boundary condition at the inlet is 0mmol·L -1 , two outlets (third outlet 206, fourth outlet 207) are defined as outflow. The solute diffusion coefficient of fluorescein is 4.25·10 -10 m 2 ·s -1 . Thus, the fluid field of the microfluidic chip can be obtained. After simulation calculation, the fluid field information of the chip design can be obtained as follows Figure 2 As shown, there are two inlets, two outlets, and a chip design domain (square area 205). The flow velocity is represented by a color legend 208, where the darkest color at the bottom (bottom 202 of the color legend) represents the minimum value of 0 m·s -1 , the lightest color at the top (201 at the top of the color legend) represents the maximum value of 20·10 -4 m·s -1 This approach was used to simulate the fluidic properties of 10,000 different microfluidic chips on a computer.

[0046] S2, an artificial neural network model, predicts the fluid field of a microfluidic chip. S1 is a traditional FEA simulation of the fluid field of a microfluidic chip. We now propose a novel, completely data-based method for predicting the fluid field of a microfluidic chip using an artificial neural network model. Data is extracted from 10,000 different random microfluidic chip models in S1, meshed, boundary conditions set, and the artificial neural network model is summarized. Finally, a customized algorithm is used to call a pre-trained artificial neural network model to predict the fluid field of the entire microfluidic chip. This method avoids the complex and time-consuming geometric modeling process while achieving results in a short period of time.

[0047] S3. Extract data based on the low Reynolds number general microfluidic chip model. Extract data: Export the 100·100 01 matrix 301 that controls the random design of the microfluidic chip from the above model in .txt format. A microfluidic chip model has more than 70,000 nodes. Export the horizontal component and vertical component of the velocity of each node and its (x, y) coordinate value in the coordinate system. Through the exported data, divide the horizontal component and vertical component of the fluid velocity in the design domain into 100·100 matrices according to the (x, y) coordinate values. Each design corresponds to three 100·100 matrices, namely the 01 matrix, the horizontal component matrix of the velocity and the vertical component matrix. Let the horizontal velocity component matrix be Vx(Vx of 100·100) x Matrix 401), the vertical component velocity matrix is ​​recorded as Vy (V y Matrix 501).

[0048] S4. Generate a data set. Now define a nine-square grid (the first nine-square grid 302), a matrix with 3 rows and 3 columns, numbered from left to right and from top to bottom. The first row is numbered: 1, 2, 3, the second row is numbered: 4, 5, 6, and the third row is numbered: 7, 8, 9. Using the Python numpy library, three matrices (100·100 01 matrix 301), (100·100 V x Matrix 401),(100·100 of V y Matrix 501), using (first nine-square grid 302), (horizontal velocity component matrix 402), (vertical velocity component matrix 502), from (the first position 303 of the nine-square grid traversal in the 01 matrix), (V x The first position of the nine-square grid traversal in the matrix is ​​403), (V y The first position 503 of the nine-square grid traversal in the matrix is ​​sliced ​​with a size of 3·3 and a step size of (3,3). After one slice, three matrices with 3 rows and 3 columns can be obtained, namely: the 01 matrix of whether to create a grid, recorded as [[P1,P2,P3],[P4,P5,P6],[P7,P8,P9]]; the horizontal velocity component matrix 402, recorded as [[V x,1 ,V x,2 ,V x,3 ],[V x,4 ,V x,5 ,V x,6 ],[V x,7 ,V x,8 ,V x,9 ]]; vertical velocity component matrix 502, denoted as [[V y,1 ,V y,2 ,V y,3 ],[Vy,4 ,V y,5 ,V y,6 ],[V y ,7,V y,8 ,V y,9 ]]. Expand the cut data and splice them into a row to get a matrix with 1 row and 27 columns, that is, [P1,…,P9,V x,1 ,…,V x,9 ,V y,1 ,…,V y,9 ], after processing, one microfluidic chip model can generate 33·33 rows of data, and the 10,000 models generated previously will generate 10,000·33·33 pieces of data as a data set.

[0049] S5. Initialize V x , V y Matrix. In the 01 matrix of S3's 100·100, all 1 corresponding positions V x , V y Initialized to 0, all V corresponding to 0 x , V y Initialized to -1. This is because in the 01 matrix, 1 represents the creation of a grid, and the corresponding speed is 0; the position where the value is 0 in the 01 matrix is ​​the area to be tested. So we only need to predict V x , V y The position with the value -1 in the matrix is ​​sufficient.

[0050] S6, add boundary conditions. In S3, we get a 01 matrix 100·100, and in S5, we get two initialized V x , V y Now add a 1-row, 100-column matrix above each of these three matrices as the upper boundary condition, and add a 101-row, 1-column matrix on the far left as the left boundary condition. The upper boundary value is extracted from the 0 <x<500,500<y<505这一范围(上边界209),由于这一范围为实体,所以对应01矩阵上边界(101·101的01矩阵601中第一行),这一行所有值为0,对应V x , V y Upper boundary (V of 101·101 x V of matrix 701,101·101 yThe first row in matrix 801 is extracted from the 0 of the microfluidic chip model <x<500,500<y<505这一范围(上边界209),左边界值提取自微流体混合器的-5<x<0,0<y<505这一范围(左边界210),由于该范围内全部被创建网格,所以对应01矩阵左边界(101·101的01矩阵601中第一列)所有值为1;对应V x , V y Left border (101·101 of V x V of matrix 701,101·101 y The first column in matrix 801) is extracted from the microfluidic mixer-5 <x<0,0<y<505这一范围(左边界210),这一列所有值为0。经过上述操作,我们就得到了三个101·101的矩阵且第一行第一列是已知的边界条件,V x , V y The area with a median value of -1 indicates the area to be tested, and the area with a median value of 0 indicates the area where the mesh is created and no fluid passes through.

[0051] S7, summarize the artificial neural network model. If V in S7 x , V y Matrix (101·101 of V x V of matrix 701,101·101 y Matrix 801) uses the nine-square grid idea to slice and predict from the upper left corner (V of 101·101 x The V of the first position 703,101·101 in the first row of the nine-square grid in the matrix y The first row and first position of the nine-square grid in the matrix are traversed (803), the first nine-square grid (101·101 V x The first nine-square grid in the matrix has 702, 101, and 101 V y The first nine-square grid (content 802) in the matrix has known boundary conditions for positions 1, 2, 3, 4, and 7. Only positions 5, 6, 8, and 9 are uncertain. Summarize all possible situations of the first nine-square grid (second nine-square grid 901). If the values ​​of positions 5, 6, 8, and 9 are {V x，5 ,V x，6 ,V x ,8,V x，9 ,V y，5 ,V y，6 ,V y，8 ,V y，9=-1,0,0,0,-1,0,0,0}, which means that only position 5 is the test area, and all other positions are grids with a speed of 0. We record this model as ANN_5(902), and similarly we can deduce ANN_6(903), ANN_8(904), and ANN_9(905); Similarly, we can deduce {V x，5 ,V x，6 ,V x，8 ,V x，9 ,V y，5 ,V y，6 ,V y，8 ,V y，9 =-1,-1,0,0,-1,-1,0,0} means that positions 5 and 6 are the areas to be measured, and grids are created at positions 8 and 9, which are recorded as ANN_56(906). Then the following models can be derived: ANN_58(907), ANN_59(908), ANN_68(909), ANN_69(910), ANN_89(911); Similarly, {V x，5 ,V x，6 ,V x，8 ,V x，9 ,V y，5 ,V y，6 ,V y，8 ,V y，9 =-1,-1,-1,0,-1,-1,-1,0} means that positions 5, 6, and 8 are the areas to be measured, and a grid is created at position 9, which is recorded as ANN_568(912). Then the following models can be derived: ANN_569(913), ANN_589(914), ANN_689(915). {V x，5 ,V x，6 ,V x，8 ,V x，9 ,V y，5 ,V y，6 ,V y，8 ,V y，9 =-1,-1,-1,-1,-1,-1,-1,-1} indicates that positions 5, 6, 8, and 9 are the areas to be measured, recorded as ANN_5689(916).

[0052] S8. Train the artificial neural network model. The specific parameters of the artificial neural network are shown in Table 1, including the number of layers, type, number of neuron nodes, and activation function type of the artificial neural network.

[0053] Table 1

[0054]

[0055] The training results are shown in Table 2, including the name of each artificial neural network model, the output position to be predicted, the accuracy of the training set, the accuracy of the test set, and the loss value of the training set.

[0056] Table 2

[0057]

[0058] S9. Use the trained artificial neural network model to predict the fluid field of the entire mixing domain.

[0059] The above 15 ANN models are used to predict the fluid field of the entire microfluidic mixing domain starting from the upper left corner. The first nine-square grid 602 is obtained by traversing the first position 603 in the 01 matrix 601 of 101·101 using the nine-square grid. The V x V from 101·101 in matrix 701 x The nine-square grid in the matrix traverses the first row and the first position 703, starting with 3·3 slices to get the first nine-square grid (V of 101·101 x The first nine-square grid in the matrix contains 702), 101·101 V y V from 101·101 in matrix 801 y The nine-square grid in the matrix traverses the first row and the first position 803, starting with 3·3 slices to get the first nine-square grid (V of 101·101 y The first nine-square grid in the matrix has a value of -1 at position 8 and position 9. The ANN_89 model is used to input data (input data = {P1, P2, P3, P4, P5, P6, P7, P8, P9, V x，1 ,V x，2 ,V x，3 ,V x，4 ,V x，5 ,V x，6 ,V x，7 ,V y，1 ,V y，2 ,V y，3 ,V y，4 ,V y，5 ,V y，6 ,V y，7}), predict {output data = V x,8 , V x,9 ,V y,8 , V y,9} and update it to V x , V yThen, we move the 3.3 slice one square to the right to get the second nine-square grid. The 1st, 2nd, and 3rd positions of the second nine-square grid are known boundary conditions, and the values ​​of the 4th, 5th, 7th, and 8th positions are the values ​​predicted by the ANN model of the first nine-square grid. Therefore, only the values ​​of the 6th and 9th positions are -1. Using the ANN_69 model input data (input data = {P1, P2, P3, P4, P5, P6, P7, P8, P9, V x,1 ,V x,2 ,V x,3 ,V x，4 ,V x，5 ,V x，7 ,V x，8 ,V y，1 ,V y，2 ,V y，3 ,V y，4 ,V y，5 ,V y，7 ,V y，8}), predict {output data = V x,6 , V x,9 ,V y,6 , V y,9}. And so on, until the last column; when the last nine-square grid (101·101 of V x After the prediction of the last position 704 in the first row of the nine-square grid in the matrix, the slice starts from V of 101·101 x The nine-square grid in the matrix traverses the second row and starts prediction at the first position 705. Positions 1, 4, and 7 are known boundary conditions. Positions 2, 3, 5, and 6 are predicted by the first nine-square grid. The value of position 8 is -1, and the value of position 9 is 0, so ANN_8 is used for prediction. By analogy, these 15 ANN models can predict all V x , V y . Now only the training results of ANN_5 and ANN_6 models are shown ( Figure 12 , Figure 13 ), where 1201 is V x,5 Training curve diagram, 1202 is V y,5 The training curve diagram, 1203 is the V in the test set x,5 , V y,5 The absolute error histogram of V x,6 Training curve diagram, 1302 is V y,6 The training curve diagram, 1303 is the V in the test set x,6 , V y,6 The absolute error histogram of all ANN model training results is shown in Table 2.

[0060] S10. Use the SSIM algorithm to compare the similarity between the predicted results and the actual simulation results. The structural similarity index (SSIM index) is an indicator used to measure the similarity between two digital images. It is measured from three aspects: image brightness, contrast, and structure. The index range is 0 to 1. The SSIM algorithm scores 1 for two identical images. The closer the score is to 1, the more similar the two images are, and the more accurate the prediction is. In order to explain the prediction results, the SSIM algorithm is used to calculate the similarity between the original image and the image rotated 30°, with a score of 0.21. The SSIM algorithm is used to calculate the similarity between the original image and the image translated 30px to the lower right, with a score of 0.20. We use the SSIM algorithm to calculate the similarity between the original image and the image predicted by the artificial neural network, and the score is 0.40. In order to test the effect of the artificial neural network model, 500 microfluidic chips were redesigned, and the results of the simulation on the computer and the results predicted by the artificial neural network were calculated using the SSIM algorithm. The results are presented in a histogram. Figure 11 .

[0061] Table 3

[0062] Original image Rotate 30° Offset 30 pixels to the lower right Artificial Neural Networks SSIM algorithm score 1.0 0.21 0.20 0.40

[0063] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. It is apparent to those skilled in the art that various changes, modifications, substitutions, and variations to these embodiments may be made without departing from the principles and spirit of the present invention, and these changes and modifications still fall within the scope of protection of the present invention.

Claims

1. A method for predicting the fluid field of a low-Reynolds-number microfluidic chip based on an artificial neural network, characterized by: The following steps are involved: S1. Create a microfluidic chip model: S2, artificial neural network model predicts the fluid field of the microfluidic chip: extract data from the random microfluidic chip model in S1, divide the grid, set boundary conditions, summarize the artificial neural network model, and finally use a customized algorithm to call the pre-trained artificial neural network model to predict the fluid field of the entire microfluidic chip; S3. Data extraction based on a general microfluidic chip model with low Reynolds number: Export a 100·100 01 matrix for controlling the random design of the microfluidic chip from the above model in .txt format. A microfluidic chip model has more than 70,000 nodes. Export the horizontal component and vertical component of the velocity of each node and their (x, y) coordinate values ​​in the coordinate system. Through the exported data, divide the horizontal component and vertical component of the fluid velocity in the design domain into 100·100 matrices according to the (x, y) coordinate values. Each design corresponds to three 100·100 matrices, namely the 01 matrix, the horizontal component matrix of velocity and the vertical component matrix. The horizontal velocity component matrix is ​​denoted as V x , the vertical component velocity matrix is ​​recorded as V y ; S4. Generate data set: Now define a nine-square grid, a matrix with 3 rows and 3 columns, numbered from left to right and from top to bottom, the first row is numbered: 1, 2, 3, the second row is numbered: 4, 5, 6, the third row is numbered: 7, 8, 9, and use the numpy library of python to slice the three matrices at the same time from the three matrix positions with a size of 3·3 and a step size of (3,3). After one slicing, three matrices with 3 rows and 3 columns can be obtained, namely: whether to create a grid 01 matrix, recorded as [[P1, P2, P3], [P4, P5, P6], [P7, P8, P9]]; the horizontal velocity component matrix, recorded as [[V x,1 ,V x,2 ,V x,3 ],[V x,4 ,V x,5 ,V x,6 ],[V x,7 ,V x,8 ,V x,9 ]]; vertical velocity component matrix, denoted as [[V y,1 ,V y,2 ,V y,3 ],[V y,4 ,V y,5 ,V y,6 ],[V y,7 ,V y,8 ,V y,9 ]], expand the cut data and splice them into a row to get a matrix with 1 row and 27 columns, that is, [P1,…,P9,V x,1 ,…,V x,9 ,V y,1 ,…,V y,9 ], after processing, one microfluidic chip model can generate 33·33 rows of data, and the 10,000 models generated previously will generate 10,000·33·33 pieces of data as a data set; S5. Initialize V x , V y Matrix: In the 100·100 01 matrix of S3, all 1 corresponding positions V x , V y Initialized to 0, all V corresponding to 0 x , V y Initialized to -1, 1 in the 01 matrix represents the creation of a grid, and the corresponding speed is 0; the position where the value is 0 in the 01 matrix is ​​the area to be measured; S6. Add boundary conditions: Obtain a 01 matrix of 100×100 in S3 and two initialized Vs in S5. x , V y . Add a 1×100 matrix above these three matrices as the upper boundary condition and a 101×1 matrix on the leftmost side as the left boundary condition. The upper boundary values are extracted from the range 0 < x < 500, 500 < y < 505 of the microfluidic chip model. Since this range is a solid, it corresponds to the upper boundary of the 01 matrix, and all values in this row are 0, corresponding to V x , V y . The upper boundary is extracted from the range 0 < x < 500, 500 < y < 505 of the microfluidic chip model, and the left boundary values are extracted from the range -5 < x < 0, 0 < y < 505 of the microfluidic mixer. Since the grid is created in this range, all values corresponding to the left boundary of the 01 matrix are 1; corresponding to V x , V y . The left boundary is extracted from the range -5 < x < 0, 0 < y < 505 of the microfluidic mixer, and all values in this column are 0. After the above operations, three 101×101 matrices are obtained, and the first row and the first column are known boundary conditions, V x , V y . The area with a value of -1 represents the待测区域 (to-be-measured area), and the area with a value of 0 represents the area where the grid is created and no fluid passes through. S7, Inductive Artificial Neural Network Model: V in S7 x , V y The matrix is ​​sliced ​​using the nine-square grid idea, starting from the upper left corner. The first nine-square grid positions 1, 2, 3, 4, and 7 are known boundary conditions. Only positions 5, 6, 8, and 9 are uncertain. Summarize all possible situations of the first nine-square grid. If the values ​​of positions 5, 6, 8, and 9 are {V x，5 ,V x，6 ,V x，8 ,V x，9 ,V y，5 ,V y，6 ,V y，8 ,V y，9 =-1,0,0,0,-1,0,0,0}, which means that only position 5 is the test area, and the other positions are grids created with a speed of 0. This model is recorded as ANN_5. Similarly, ANN_6, ANN_8, and ANN_9 can be inferred. Similarly, {V x，5 ,V x，6 ,V x，8 ,V x，9 ,V y，5 ,V y，6 ,V y，8 ,V y，9 =-1,-1,0,0,-1,-1,0,0} means that positions 5 and 6 are the test areas, and grids are created at positions 8 and 9, which are recorded as ANN_56. Then the following models can be derived: ANN_58, ANN_59, ANN_68, ANN_69, ANN_89; Similarly, {V x，5 ,V x，6 ,V x，8 ,V x，9 ,V y，5 ,V y，6 ,V y，8 ,V y，9 =-1,-1,-1,0,-1,-1,-1,0} means that positions 5, 6, and 8 are the areas to be measured, and a grid is created at position 9, which is recorded as ANN_568. Then the following models can be derived: ANN_569, ANN_589, ANN_689, {V x，5 ,V x，6 ,V x，8 ,V x，9 ,V y，5 ,V y，6 ,V y，8 ,V y，9 =-1,-1,-1,-1,-1,-1,-1,-1} means positions 5, 6, 8, and 9 are the areas to be measured, recorded as ANN_5689; S8. Training artificial neural network model: The specific parameters of the artificial neural network include the number of layers, type, number of neuron nodes, activation function type of the artificial neural network. The training results include the name of each artificial neural network model, the output position to be predicted, the accuracy of the training set, the accuracy of the test set, and the loss value of the training set; S9. Use the trained artificial neural network model to predict the fluid field in the entire mixing domain: Using the above ANN model to predict the fluid field of the entire microfluidic mixing domain from the upper left corner, the first nine-square grid is obtained by slicing. The values ​​of positions 8 and 9 are -1. The ANN_89 model is used to predict V x,8 , V y,8 ,V x,9 , V y,9 and update it to the value of V x , V y In the matrix, the 3·3 slice is then moved one grid to the right to obtain the second nine-square grid. The 1st, 2nd, and 3rd positions of the second nine-square grid are known boundary conditions. The values ​​of the 4th, 5th, 7th, and 8th positions are predicted by the above ANN model for the first nine-square grid. Only the values ​​of the 6th and 9th positions are -1. The ANN_69 model is used to predict V x,6 , V y,6 ,V x,9 , V y,9 The value of , and so on, until the last column; after the last nine-square grid is predicted, the slice starts to predict from position 705, positions 1, 4, and 7 are known boundary conditions, positions 2, 3, 5, and 6 are obtained by the first nine-square grid prediction, position 8 has a value of -1, and position 9 has a value of 0, so ANN_8 is used for prediction, and so on. These 15 ANN models can predict all V x , V y ; S10. Use the SSIM algorithm to compare the similarity between the predicted results and the actual simulation results: measure the image brightness, contrast, and structure in three aspects, with an index range of 0 to 1. The SSIM algorithm score of two identical images is 1. The closer the score is to 1, the more similar the two images are, and the more accurate the prediction is. The results of the simulation on the computer and the results of the artificial neural network prediction are calculated using the SSIM algorithm, and the results are displayed in a histogram.

2. The method for predicting the fluid field of a low Reynolds number microfluidic chip based on an artificial neural network according to claim 1, characterized in that: The step S1 specifically includes: Each microfluidic chip has two inlets, two outlets, and a design area where the random mixing structure is located. The design domain is divided into 100·100 grids of 5μm·5μm in size. A 100·100 01 matrix is ​​used to indicate whether a grid is created, where 1 represents the creation of a matrix and 0 represents the non-creation of a matrix. By controlling the probability of 01, the randomness of the microfluidic chip is controlled, thereby achieving random design of the mixing domain. Stochastic microfluidic chip performance simulation uses finite element analysis software to design and simulate chip performance. Laminar flow physics module and dilute species transport physics module are set up with two fixed solvers for calculation. In the laminar flow physics module, the boundary conditions of the two inlets are defined as 1 mm·s -1 The normal inflow velocity is 0 Pa. The boundary condition of each outlet defines a pressure of 0 Pa. The remaining boundaries are no-slip boundary conditions. The material filling the channel is water of incompressible flow type. In the Diluted Species Transport Physics Module, the boundary condition of the first inlet is 1 mmol·L -1 , the boundary condition of the second inlet is 0mmol·L -1 , the two outlets are defined as outflow, and the fluid field information of the microfluidic chip design can be obtained through simulation calculation.

Images