A method for quickly predicting flow characteristics of a flow channel network by combining microfluidic experiments and numerical simulation

Abstract

The application discloses a kind of microfluidic experiment and numerical simulation flow channel network flow characteristic fast prediction method for combining, it is related to microfluidic chip field, by the axial parameterization processing to each target flow branch, the axial distance parameter of each center point on the target flow branch along the flow direction is determined, each geometric node in flow channel is given with the explicit physical flow direction measurement benchmark.This key step makes the originally discrete center point set have the continuous along-distance distance property, can accurately characterize the cumulative flow distance of fluid in complex topological path;In combination with deep operator learning architecture, this method uses the local feature set and global state vector containing the axial distance parameter, directly establishes the mapping relationship from geometric topology to flow field distribution, so as to realize the fast prediction of the flow characteristics in microfluidic chip flow channel network without the prerequisite of traditional grid iterative calculation.

Description

Technical Field

[0001] This invention relates to the field of microfluidic chips, and in particular to a method for rapid prediction of flow characteristics of flow channels by combining microfluidic experiments and numerical simulation. Background Technology

[0002] Precise manipulation of fluids within the microscale channels of microfluidic chips has wide applications in fields such as bioanalysis, chemical reactions, and microscale transport process research. In these systems, the design of the channel network structure directly affects the flow distribution, pressure distribution, and transport efficiency of the fluid in each branch.

[0003] In the design and optimization of microfluidic chips, it is often necessary to rapidly evaluate different flow channel structures to determine their flow performance. However, current technologies primarily rely on computational fluid dynamics (CFD) simulation methods to analyze the flow characteristics within the channels. While this method offers high accuracy, it requires complex mesh generation and time-consuming numerical solutions, resulting in high computational costs and long cycles, making it difficult to meet the design requirements of rapid iteration across multiple schemes.

[0004] On the other hand, with the development of microfluidic experimental techniques, local flow information inside the flow channel can be obtained through methods such as microscopic imaging and particle tracing, thereby providing experimental data support for flow characteristic analysis. However, the cost of acquiring this type of experimental data is high, the data scale is limited, and it is difficult to directly use it for rapid prediction of complex flow channel networks.

[0005] In recent years, data-driven deep learning-based flow field prediction methods have provided a new technical path for achieving rapid evaluation. However, existing methods still have significant shortcomings when dealing with complex three-dimensional multi-branch flow channel networks: Most existing methods primarily use the spatial coordinates of the flow path as input features, lacking information that can clearly characterize the position of a sampling point along the flow path and the branch it belongs to. Specifically, existing models struggle to accurately identify the preceding and following relationships of sampling points within the flow path from complex geometries, and also find it difficult to determine the specific flow path branch to which the sampling point belongs, leading to the following problems: Lack of positional awareness along the flow path: Due to the lack of features representing the cumulative distance of sampling points along the flow direction, the model cannot effectively distinguish the front and rear positional relationships of sampling points in the flow channel, and it is difficult to learn the laws of fluid parameters (such as pressure and velocity) as they change with the flow path, resulting in biases in the prediction of pressure loss and velocity distribution along the flow path. Branch attribution confusion: In complex flow network with multiple branches, especially in bifurcation and merging areas, the model has difficulty in accurately identifying the branch to which the sampling point belongs, which easily leads to confusion of samples from different branches. This affects the learning of the flow patterns of each branch and results in large prediction errors in local areas.

[0006] The aforementioned problems limit the prediction accuracy and generalization ability of existing deep learning methods in complex flow channel networks. Therefore, there is an urgent need for a flow characteristic prediction method that can combine flow channel geometric features with experimental data information and enhance the model's ability to identify the location of sampling points along the flow path and the branch affiliation, so as to achieve efficient and accurate prediction of the flow characteristics of complex microfluidic flow channel networks. Summary of the Invention

[0007] To improve the prediction accuracy of flow characteristics in complex microfluidic channel networks and achieve efficient utilization of experimental and simulation data, this invention proposes a rapid prediction method for flow characteristics of channel networks combining microfluidic experiments and numerical simulations, including: A three-dimensional structural model of the flow channel network in the microfluidic chip is obtained, and the inner wall surface of the flow channel is sampled to generate wall point cloud data. The wall point cloud data is spatially aggregated to obtain a set of center points, and in the set of center points, the point pairs with Euclidean distance less than a preset distance threshold are connected to construct a centerline network. In the centerline network, the starting node and the ending node corresponding to the inlet and outlet of the flow channel are determined respectively. Multiple discretely distributed center points are selected as candidate seed points in the middle area between the starting node and the ending node. Multiple target flow branches are obtained based on each candidate seed point. The axial parameterization of each target flow branch is performed to determine the axial distance parameter of each center point along the flow direction on the target flow branch. The set of center points is sampled to generate a set of sampling points; the axial distance parameters of each target flow branch and its center points along the flow direction are used to obtain the local feature set corresponding to each sampling point in the set of sampling points. A global state vector is constructed, and the global state vector and the local feature set corresponding to each sampling point in the sampling point set are input into the target prediction model to obtain the predicted flow characteristics of each sampling point; the global state vector represents the operating condition control parameters of the microfluidic chip during operation; the target prediction model is obtained by training an initial model based on a deep operator learning architecture.

[0008] Furthermore, the intermediate region includes: a region located between the starting node and the ending node, and covering the area where the flow channel geometry undergoes bifurcation, merging, or bending changes; The process of obtaining multiple target flow branches based on each candidate seed point specifically includes: For each candidate seed point, a first shortest path connecting the candidate seed point to the starting node and a second shortest path connecting the candidate seed point to the ending node are searched in the centerline network. The first shortest path and the second shortest path are then combined at the candidate seed point to form a candidate flow branch. All candidate flow branches are aggregated to form a candidate flow branch set; The candidate flow branches in the candidate flow branch set are deduplicated to obtain multiple independent flow branches; each independent flow branch is smoothed to obtain multiple target flow branches.

[0009] Furthermore, the process of deduplicating candidate flow branches in the candidate flow branch set to obtain multiple independent flow branches specifically involves: Select any candidate flow branch from the candidate flow branch set as an independent flow branch and add it to the preset independent flow branch library; The remaining candidate flow branches in the candidate flow branch set are compared with the existing independent flow branches in the independent flow branch library in turn. In each comparison, the shortest Euclidean distance from multiple center points on the candidate flow branch being compared to the independent flow branch being compared is calculated, and the percentage of distance values ​​less than the preset deduplication threshold is counted. If the proportion of the number is greater than the preset overlap ratio threshold, the subsequent comparison of the candidate flow branch will be terminated immediately. If the candidate flow branch being compared has completed the comparison with all independent flow branches in the independent flow branch library, and the proportion of the number of comparisons is less than or equal to the preset overlap ratio threshold, then it is added to the independent flow branch library as an independent flow branch. All paths in the independent flow branch database are considered as the multiple independent flow branches.

[0010] Furthermore, determining the axial distance parameter along the flow direction of each center point on the target flow branch specifically includes: Set the starting node as the zero point for the flow distance calculation; For each target flow branch, for each center point on that branch, calculate the cumulative path length from the zero point to the center point along the target flow branch; The cumulative path length is used as the axial distance parameter of the center point.

[0011] Furthermore, the acquisition of the local feature set corresponding to each sampling point in the sampling point set specifically includes: For each sampling point in the set of sampling points: Calculate the shortest vertical distance from it to the wall point cloud data; Determine the target flow branch to which the sampling point belongs, and determine its projection position on the target flow branch to which it belongs; Based on the projection position, obtain the axial distance parameter corresponding to the sampling point, and obtain the branch identifier of the target flow branch to which it belongs; obtain the three-dimensional spatial coordinates of the sampling point, and combine the three-dimensional spatial coordinates, the shortest vertical distance, the branch identifier and the axial distance parameter to construct a local feature set; The step of obtaining the axial distance parameter corresponding to the sampling point based on the projection position includes: If the projected position coincides with any center point on the target flow branch, then the axial distance parameter of that center point is directly used as the axial distance parameter corresponding to the sampling point. If the projected position does not coincide with any center point, the axial distance parameter corresponding to the sampling point is obtained by linear interpolation of the axial distance parameter of the center point on the target flow branch.

[0012] Further, determining the target flow branch to which the sampling point belongs specifically includes: projecting the sampling point to the target flow branch with the closest spatial distance, and determining that branch as the target flow branch to which it belongs; or: Set the preset core radius; Iterate through the set of sampling points, and for each sampling point: Calculate the vertical distance from each of the target flow branches, and mark the branch corresponding to the minimum vertical distance as the candidate nearest branch; If the minimum vertical distance is less than the preset core radius, the sampling point is defined as the safe core point, and the corresponding candidate nearest branch is determined as its target flow branch. If the minimum vertical distance is greater than or equal to the preset core radius, the sampling point is defined as a non-core point. For each non-core point, among all the determined safe core points, the safe core point with the closest spatial Euclidean distance to the non-core point is found, and the target flow branch to which the safe core point belongs is inherited as the target flow branch to which the current non-core point belongs.

[0013] Furthermore, the global state vector includes fluid property parameters and microfluidic chip inlet flow rate and / or operating condition control parameters; the operating condition control parameters include at least one of inlet pressure, applied electric field strength, or rotational angular velocity; The initial model includes: A branch network is used to receive the global state vector and encode it to obtain a global operating condition feature embedding. The backbone network is used to receive the local feature set and encode it to obtain local geometric feature embeddings; The fusion layer is used to fuse the global operating condition feature embedding with the local geometric feature embedding, and map and output the predicted flow characteristics of each sampling point; The predicted flow characteristics include at least one of three-dimensional pressure distribution values ​​and velocity scalar distribution values.

[0014] Furthermore, the training method for the target prediction model includes: A training dataset is constructed, which contains multiple training samples. Each training sample includes: a global state vector corresponding to a three-dimensional structural model of the flow channel network in a microfluidic chip, and a set of local features corresponding to each sampling point within the three-dimensional structural model; and a training label that matches each sampling point one by one. The training labels are derived from microfluidic experimental measurement data and / or numerical simulation data, and the training label corresponding to each sampling point includes the true pressure value and / or the true velocity scalar value of that point. The training samples are sequentially input into the initial model to obtain predicted values ​​of flow characteristics; Calculate the loss function value between the predicted flow characteristics and the training labels; Based on the loss function value, the model parameters of the initial model are iteratively updated until the loss function value converges to a preset threshold, thereby obtaining the target prediction model.

[0015] Furthermore, the rapid prediction method for flow characteristics also includes: If the predicted flow characteristics include three-dimensional pressure distribution values, then based on the three-dimensional pressure distribution values, the average pressure at the inlet and outlet sections of the flow channel is statistically analyzed to calculate the overall pressure drop of the flow channel; and using the branch identifiers in the local feature set, each sampling point is grouped and statistically analyzed, and combined with the three-dimensional pressure distribution values, the pressure distribution inside each target flow branch is obtained, and then the branch pressure drop of each target flow branch is calculated. If the predicted flow characteristics include velocity scalar distribution values, then based on the velocity scalar distribution values, sampling points located at specified cross sections are selected, and the velocity scalar distribution values ​​of sampling points with the same branch identifier are accumulated or integrated using the branch identifier in the local feature set to calculate the volumetric flow rate corresponding to each target flow branch; then, based on the volumetric flow rate of all target flow branches, the flow distribution ratio of each target flow branch is calculated.

[0016] Compared with the prior art, the present invention has at least the following beneficial effects: (1) This invention determines the axial distance parameter of each center point along the flow direction on each target flow branch by performing axial parameterization on each target flow branch, thus giving each geometric node in the flow channel a clear physical flow direction measurement benchmark. This key step enables the originally discrete set of center points to have continuous flow distance attributes, which can accurately characterize the cumulative flow distance of fluid in complex topological paths; combined with a deep operator learning architecture, this method uses the local feature set containing the axial distance parameter and the global state vector to directly establish a mapping relationship from geometric topology to flow field distribution, thereby realizing the rapid prediction of the internal flow characteristics of the microfluidic chip flow channel network without the need for traditional grid iterative calculation.

[0017] (2) In view of the actual distribution of sampling points in three-dimensional space, this invention adopts a strategy of projection positioning combined with linear interpolation to obtain the axial distance parameters of the sampling points. By mapping the sampling points to the projection position of their respective target flow branches, and using the existing axial distance parameters of the center points on the branches for interpolation calculation, the sampling points, which originally only had three-dimensional spatial coordinates, are supplemented with axial distance parameters that characterize their positions along the flow direction. This processing method establishes a numerical correspondence between the spatial sampling point positions and the flow path distance, so that the local feature set contains geometric information that can reflect the fluid's development state along the path, thereby supporting the model to deduce the flow characteristics based on the axial distance parameters.

[0018] (3) The branch attribution determination mechanism based on the preset core radius and the safety core point inheritance mechanism in this invention improves the sampling point attribution determination logic at the intersection of multiple branches, and supports the prediction of flow characteristics of microfluidic chip flow channel networks with complex structures such as bifurcation and merging. For non-core points that are far from the branches, the minimum vertical distance is no longer used for hard division, but the target flow branch to which it belongs is determined by inheriting the branch identifier of the neighboring safety core point. This logic enhances the robustness of the algorithm to geometrically ambiguous regions and reduces the risk of flow characteristic prediction distortion due to incorrect branch judgment, which is particularly suitable for scenarios with complex bifurcation channels.

[0019] (4) This invention achieves a complete closed loop from microscopic field distribution prediction to macroscopic engineering performance index calculation. Based on the three-dimensional pressure distribution value and velocity scalar distribution value output by the model, combined with the branch identifier in the local feature set, it can directly perform group statistics and integration operations on each sampling point, and then accurately calculate the overall pressure drop of the flow channel, the branch pressure drop of each target flow branch, the volumetric flow rate, and the flow distribution ratio. This capability enables the method to not only provide flow field details at the cloud map level, but also directly output key system-level indicators to guide engineering design, combining the efficiency of deep learning with the physical interpretability of traditional fluid dynamics analysis. Attached Figure Description

[0020] Figure 1 This is a flowchart of a method for rapid prediction of flow characteristics of a flow channel network that combines microfluidic experiments and numerical simulation, as described in an embodiment of the present invention. Detailed Implementation

[0021] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.

[0022] To achieve efficient and accurate prediction of the internal flow characteristics of microfluidic chip channel networks with complex bifurcation topologies without the need for traditional grid iterative calculations, such as... Figure 1 As shown, this invention proposes a rapid prediction method for the flow characteristics of flow channels by combining microfluidic experiments and numerical simulations, including: A three-dimensional structural model of the flow channel network in the microfluidic chip is obtained, and the inner wall surface of the flow channel is sampled to generate wall point cloud data. The wall point cloud data is spatially aggregated to obtain a set of center points, and in the set of center points, the point pairs with Euclidean distance less than a preset distance threshold are connected to construct a centerline network. In this embodiment, the specific implementation process of spatially aggregating the wall point cloud data to obtain the center point set is as follows: First, obtain discrete point cloud data of the inner wall surface of the flow channel, and initialize the state marker of all points as unprocessed.

[0023] Secondly, perform iterative spatial clustering: Seed point selection: Randomly select one point from the wall points that are currently unprocessed as the seed point; Neighborhood search: Using the seed point as the center and a preset aggregation radius as the threshold, search for all neighboring points within its spatial neighborhood; Constructing a local cluster: Merging the seed point and all its neighboring points to form a local point cloud cluster; State Update: Update the state label of all points contained in the local point cloud cluster to "processed" to ensure that these points are not selected as seed points or participate in clustering repeatedly in subsequent iterations; Centroid calculation: Calculate the arithmetic mean of the three-dimensional coordinates of all points within the local point cloud cluster, and define the spatial position corresponding to the average value as the center point of the region; Loop termination check: Repeat the above steps until the state of all wall points has changed to "processed".

[0024] Finally, all the center points generated during the iteration process are summarized to form the center point set.

[0025] Through the aforementioned non-overlapping iterative aggregation strategy, the originally high-density wall point cloud data is effectively compressed into a sparse set of center points. Each center point uniquely represents a non-overlapping local wall region, which eliminates data redundancy and ensures the uniform distribution of center points within the flow channel space.

[0026] In this embodiment, a connection relationship is established between point pairs whose Euclidean distance is less than a preset distance threshold, specifically as follows: First, based on the aforementioned set of center points, a spatial neighborhood search is performed on each center point in the set to find all neighboring center points within a preset search radius. Next, calculate the Euclidean distance between the current center point and each neighboring center point, and compare this distance with the preset distance threshold; wherein, the preset distance threshold is a physical constraint value set according to the minimum pipe spacing and wall thickness parameters of the flow channel, and its value is less than the minimum gap between adjacent flow channels to ensure that the connection only occurs within the same flow channel; Subsequently, for point pairs whose Euclidean distance is less than the preset distance threshold, a topological connection edge is established between them; for point pairs whose distance is greater than or equal to the threshold, no connection is established, thereby physically blocking the possibility of the centerline crossing the flow channel wall or mistakenly crossing to an adjacent pipe. Finally, after traversing all points in the set of center points, the centerline network, which reflects the true connectivity of the flow channels, is formed by all the established topological connection edges and center points.

[0027] In the centerline network, the starting node and the ending node corresponding to the inlet and outlet of the flow channel are determined respectively. Multiple discretely distributed center points are selected as candidate seed points in the middle area between the starting node and the ending node. Multiple target flow branches are obtained based on each candidate seed point. The axial parameterization of each target flow branch is performed to determine the axial distance parameter of each center point along the flow direction on the target flow branch. In this embodiment, the specific implementation process of determining the starting node and ending node corresponding to the flow channel inlet and outlet in the centerline network is as follows: First, the inlet and outlet regions are extracted from the three-dimensional structural model of the flow channel network in the microfluidic chip. The inlet region is a local tube segment entity that includes the inlet cross section of the flow channel and extends inward for a predetermined length, and the outlet region is a local tube segment entity that includes the outlet cross section of the flow channel and extends inward for a predetermined length. Secondly, the geometric centroids of the aforementioned inlet and outlet regions are calculated respectively to obtain the inlet centroid coordinates and the outlet centroid coordinates. Subsequently, all nodes in the constructed centerline network are traversed, and the Euclidean distance between the three-dimensional coordinates of each node and the aforementioned inlet centroid coordinates and outlet centroid coordinates is calculated. Finally, the node with the smallest distance from the inlet centroid is selected as the starting node, and the node with the smallest distance from the outlet centroid is selected as the ending node.

[0028] The intermediate region includes: the area located between the starting node and the ending node, and covering the area where the flow channel geometry undergoes bifurcation, merging, or bending changes; The process of obtaining multiple target flow branches based on each candidate seed point specifically includes: For each candidate seed point, a first shortest path connecting the candidate seed point to the starting node and a second shortest path connecting the candidate seed point to the ending node are searched in the centerline network. The first shortest path and the second shortest path are then combined at the candidate seed point to form a candidate flow branch. All candidate flow branches are aggregated to form a candidate flow branch set; The candidate flow branches in the candidate flow branch set are deduplicated to obtain multiple independent flow branches; each independent flow branch is smoothed to obtain multiple target flow branches.

[0029] In this embodiment, the specific implementation process of smoothing each independent flow branch to obtain multiple target flow branches is as follows: First, the positions of each selected independent flow branch are corrected: the local centroid of the inner wall surface points of the flow channel within a preset neighborhood around each node on the path is calculated, and the coordinates of the node are converged and adjusted towards the local centroid to eliminate discretization error and improve the coincidence of the branch with the actual flow channel axis. Subsequently, a curve smoothing algorithm (such as B-spline interpolation or moving least squares) is applied to the path after position correction to remove high-frequency jitter and bends in the path, generating a geometrically continuous and smooth curve. Finally, the processed continuous curve is output as the target flow branch to ensure that it can accurately and smoothly reflect the true flow direction characteristics of each branch.

[0030] The process of deduplicating candidate flow branches in the candidate flow branch set to obtain multiple independent flow branches is as follows: Select any candidate flow branch from the candidate flow branch set as an independent flow branch and add it to the preset independent flow branch library; The remaining candidate flow branches in the candidate flow branch set are compared with the existing independent flow branches in the independent flow branch library in turn. In each comparison, the shortest Euclidean distance from multiple center points on the candidate flow branch being compared to the independent flow branch being compared is calculated, and the percentage of distance values ​​less than the preset deduplication threshold is counted. If the proportion of the number is greater than the preset overlap ratio threshold, the subsequent comparison of the candidate flow branch will be terminated immediately. If the candidate flow branch being compared has completed the comparison with all independent flow branches in the independent flow branch library, and the proportion of the number of comparisons is less than or equal to the preset overlap ratio threshold, then it is added to the independent flow branch library as an independent flow branch. All paths in the independent flow branch database are considered as the multiple independent flow branches.

[0031] The determination of the axial distance parameters along the flow direction of each center point on the target flow branch specifically includes: Set the starting node as the zero point for the flow distance calculation; For each target flow branch, for each center point on that branch, calculate the cumulative path length from the zero point to the center point along the target flow branch; The cumulative path length is used as the axial distance parameter of the center point.

[0032] The set of center points is sampled to generate a set of sampling points; the axial distance parameters of each target flow branch and its center points along the flow direction are used to obtain the local feature set corresponding to each sampling point in the set of sampling points. The acquisition of the local feature set corresponding to each sampling point in the sampling point set specifically includes: For each sampling point in the set of sampling points: Calculate the shortest vertical distance from it to the wall point cloud data; Determine the target flow branch to which the sampling point belongs, and determine its projection position on the target flow branch to which it belongs; Determining the target flow branch to which the sampling point belongs specifically includes: The sampling point is projected onto the target flow branch with the closest spatial distance, and that branch is determined to be the target flow branch to which it belongs. or: Set the preset core radius; Iterate through the set of sampling points, and for each sampling point: Calculate the vertical distance from each of the target flow branches, and mark the branch corresponding to the minimum vertical distance as the candidate nearest branch; If the minimum vertical distance is less than the preset core radius, the sampling point is defined as the safe core point, and the corresponding candidate nearest branch is determined as its target flow branch. If the minimum vertical distance is greater than or equal to the preset core radius, the sampling point is defined as a non-core point. For each non-core point, among all the determined safe core points, the safe core point with the closest spatial Euclidean distance to the non-core point is found, and the target flow branch to which the safe core point belongs is inherited as the target flow branch to which the current non-core point belongs.

[0033] This invention improves the logic for determining the assignment of sampling points at the intersection of multiple branches by employing a branch assignment mechanism based on a preset core radius and a safety core point inheritance mechanism. This mechanism supports the prediction of flow characteristics in microfluidic chip channel networks with complex structures such as bifurcation and merging. For non-core points far from branches, the algorithm no longer relies solely on the minimum vertical distance for rigid division. Instead, it determines the target flow branch by inheriting the branch identifier of the nearest safety core point. This logic enhances the algorithm's robustness to geometrically ambiguous regions and reduces the risk of flow characteristic prediction distortion due to incorrect branch identification. It is particularly suitable for scenarios with complex bifurcation channels.

[0034] For the complex topology of the flow channel network of microfluidic chips, such as branching and merging, the minimum vertical distance calculated for the sampling point located in the branch boundary region can determine a candidate nearest branch. However, if the distance value is greater than or equal to the preset core radius, it indicates that the sampling point is in the fuzzy region of branch affiliation. Directly accepting the candidate nearest branch as the final result carries a high risk of misjudgment.

[0035] Therefore, this embodiment adopts a dual judgment mechanism of minimum vertical distance and preset core radius: First, traverse the set of sampling points, calculate the vertical distance from each sampling point to each target flow branch, and mark the branch corresponding to the minimum vertical distance as the candidate nearest branch; Secondly, a tiered determination is made based on the relationship between the minimum vertical distance and the preset core radius: If the minimum vertical distance is less than the preset core radius, the sampling point is defined as the safe core point, and its candidate nearest branch is directly determined as its target flow branch. If the minimum vertical distance is greater than or equal to the preset core radius, the sampling point is defined as a non-core point. In this case, its candidate nearest branch is not directly adopted. Instead, based on the spatial Euclidean distance, the node closest to the non-core point is found among all the determined safe core points, and the target flow branch of the nearest neighbor safe core point is inherited as the target flow branch of the current non-core point.

[0036] This processing method effectively identifies and isolates low-confidence boundary sampling points (i.e., non-core points) by presetting the core radius, and corrects their branch affiliation by using spatial proximity in an inheritance manner. It eliminates the jumps that may be caused by relying solely on a single minimum vertical distance for determination, and significantly improves the robustness of the determination of the target flow branch under complex flow channel structures.

[0037] Based on the projection position, obtain the axial distance parameter corresponding to the sampling point, and obtain the branch identifier of the target flow branch to which it belongs; obtain the three-dimensional spatial coordinates of the sampling point, and combine the three-dimensional spatial coordinates, the shortest vertical distance, the branch identifier and the axial distance parameter to construct a local feature set; The step of obtaining the axial distance parameter corresponding to the sampling point based on the projection position includes: If the projected position coincides with any center point on the target flow branch, then the axial distance parameter of that center point is directly used as the axial distance parameter corresponding to the sampling point. If the projected position does not coincide with any center point, the axial distance parameter corresponding to the sampling point is obtained by linear interpolation of the axial distance parameter of the center point on the target flow branch.

[0038] This invention addresses the actual distribution of sampling points in three-dimensional space by employing a strategy combining projection positioning and linear interpolation to obtain the axial distance parameters of these points. By mapping the sampling points to the projected positions of their respective target flow branches and using the existing axial distance parameters of the center points on those branches for interpolation calculations, the original sampling points, which only possessed three-dimensional spatial coordinates, are supplemented with axial distance parameters characterizing their positions along the flow direction. This approach establishes a numerical correspondence between the spatial sampling point positions and the flow path distance, enabling the local feature set to contain geometric information reflecting the fluid's development along the path, thus supporting the model's deduction of flow characteristics based on these axial distance parameters.

[0039] A global state vector is constructed, and the global state vector and the local feature set corresponding to each sampling point in the sampling point set are input into the target prediction model to obtain the predicted flow characteristics of each sampling point; the global state vector represents the operating condition control parameters of the microfluidic chip during operation; the target prediction model is obtained by training an initial model based on a deep operator learning architecture.

[0040] The term "operation" in this invention refers to the working state of a microfluidic chip where fluid flows within a microchannel under external driving conditions (such as inlet pressure, electric field, or rotation).

[0041] The global state vector includes fluid property parameters and microfluidic chip inlet flow rate (volume flow rate) and / or operating condition control parameters; the operating condition control parameters include at least one of inlet pressure, applied electric field strength, or rotational angular velocity; The fluid properties include fluid density and fluid dynamic viscosity.

[0042] The initial model includes: A branch network is used to receive the global state vector and encode it to obtain a global operating condition feature embedding. The backbone network is used to receive the local feature set and encode it to obtain local geometric feature embeddings; The fusion layer is used to fuse the global operating condition feature embedding with the local geometric feature embedding, and map and output the predicted flow characteristics of each sampling point; The predicted flow characteristics include at least one of three-dimensional pressure distribution values ​​and velocity scalar distribution values.

[0043] Specifically: In a preferred example of this embodiment, the target prediction model employs a deep operator learning architecture, specifically built upon a Deep Operator Network (DeepONet). DeepONet is a deep learning model for learning nonlinear operator mappings, and its architecture comprises two independent parts: a branch network and a backbone network. In this embodiment, the specific configuration and data flow of each part are as follows: Branch network: This network is used to process global information. Its input receives the global state vector, which includes fluid property parameters and the inlet flow rate of the flow channel. The branch network encodes the global state vector through a multi-layer fully connected neural network, outputting a high-dimensional global operating condition feature embedding. This embedding characterizes the overall flow state of the current flow channel under specific global operating conditions.

[0044] Backbone network: This network is used to process local geometric and spatial information. Its input receives the set of local features, and the backbone network also encodes the data in the set of local features through a multi-layer fully connected neural network, outputting a local geometric feature embedding that matches the output dimension of the branch networks. This embedding represents the response characteristics of the sampling point in the local geometric environment and spatial location.

[0045] Fusion layer: The target prediction model also includes a fusion layer for fusing the global operating condition feature embedding with the local geometric feature embedding. In specific implementations, the two feature embedding vectors can be fused using a dot product or by concatenating them and then passing them through a fully connected layer, and the predicted flow characteristics of each fluid sampling point can be directly mapped and output.

[0046] Through the above architecture, the target prediction model can effectively decouple the coupling influence of global operating conditions and local geometric and topological features on the flow field, and learn the nonlinear mapping relationship from the global state vector and local feature set to the distribution of physical quantities in the three-dimensional flow field; thus, for any new flow channel structure and operating conditions, the three-dimensional pressure and velocity distribution at the fluid sampling point can be directly predicted without re-grid generation and numerical iteration.

[0047] The training method for the target prediction model includes: A training dataset is constructed, which contains multiple training samples. Each training sample includes: a global state vector corresponding to a three-dimensional structural model of the flow channel network in a microfluidic chip, and a set of local features corresponding to each sampling point within the three-dimensional structural model; and a training label that matches each sampling point one by one. The training labels are derived from microfluidic experimental measurement data and / or numerical simulation data, and the training label corresponding to each sampling point includes the true pressure value and / or the true velocity scalar value of that point. The training samples are sequentially input into the initial model to obtain predicted values ​​of flow characteristics; Calculate the loss function value between the predicted flow characteristics and the training labels; In this embodiment, the loss function uses mean squared error (MSE) to minimize the deviation between the predicted pressure / velocity and the true value; Based on the loss function value, the model parameters of the initial model are iteratively updated until the loss function value converges to a preset threshold, thereby obtaining the target prediction model.

[0048] It should be noted that when constructing the training dataset, the three-dimensional structural models of the microfluidic chip channel networks corresponding to the multiple training samples are all different. Specifically, these three-dimensional structural models cover a variety of different channel topologies (e.g., serpentine, spiral, bifurcated, etc.), different cross-sectional shapes (e.g., circular, rectangular, trapezoidal, etc.), and different dimensional parameters (e.g., channel length, width, bending radius, etc.). By introducing diverse samples with significant geometric differences for training, the initial model can be forced to learn the deep mapping between fluid flow characteristics and complex geometric structures, thereby ensuring that the final target prediction model has good generalization ability, enabling it to be applied to the design of novel microfluidic chip channel networks that have never been seen before, rather than simply a memorization of known structures.

[0049] The methods for obtaining the true pressure values ​​and / or true velocity scalar values ​​in the training labels include, but are not limited to: using high-precision computational fluid dynamics (CFD) simulation software (such as ANSYS Fluent, COMSOL Multiphysics, OpenFOAM, etc.) to perform numerical simulation calculations on the corresponding three-dimensional structural models, or acquiring them in actual microfluidic chip systems through physical experimental measurement devices (such as particle image velocimeters (PIV), pressure sensor arrays, etc.).

[0050] In a preferred embodiment, high-precision CFD simulation results verified by mesh independence are used as the true values. The specific process is as follows: for each three-dimensional structural model, the corresponding boundary conditions and fluid property parameters (such as density, viscosity, etc.) are set, and the Navier-Stokes equations are solved. After the flow field calculation converges, the pressure distribution data and velocity distribution data of the entire field are obtained, and the true pressure value and / or true velocity scalar value corresponding to each sampling point are extracted from them. These values ​​are used as supervision signals (training labels) to guide the parameter update of the model.

[0051] In another preferred embodiment of the present invention, the process of constructing the global state vector further includes the step of introducing a global pressure drop reference value or an equivalent flow resistance coefficient.

[0052] Specifically, using well-known formulas for calculating friction loss and local resistance in fluid mechanics (such as the Darcy-Weisbach formula combined with the local resistance coefficient method), and based on the current microfluidic chip channel network's channel geometry characteristics (including total channel length, equivalent hydraulic diameter, number of bends, and branch connections) and operating parameters (including inlet velocity, fluid density, and dynamic viscosity), the method is applied. If a global pressure drop reference value is needed, the value representing the total pressure loss from the inlet to the outlet of the flow channel can be directly calculated. If an equivalent flow resistance coefficient is required, the dimensionless equivalent flow resistance coefficient, which characterizes the overall flow resistance of the microfluidic chip's channel network, can be calculated using the formula above, based on the well-known principle that total resistance equals the sum of friction resistance and local resistance.

[0053] Subsequently, the calculated global pressure drop reference value (or equivalent flow resistance coefficient) is directly added to the global state vector as a scalar value. At this point, the global state vector input to the target prediction model contains both operating parameters describing the basic properties of the fluid and global reference information characterizing the overall pressure loss magnitude or overall resistance characteristics of the flow channel.

[0054] During the prediction process, the target prediction model uses this global reference information as a numerical reference benchmark for the overall pressure distribution, assisting the model in constraining the distribution range of the pressure prediction values ​​at each sampling point, so that its overall pressure change trend matches the global reference information. By introducing this explicit overall magnitude information, it helps to reduce the pressure distribution deviation caused by the accumulation of point-by-point predictions in complex long channels or multi-branch channels, thereby improving the stability of the flow characteristic prediction results.

[0055] The rapid prediction method for flow properties also includes: If the predicted flow characteristics include three-dimensional pressure distribution values, then based on the three-dimensional pressure distribution values, the average pressure at the inlet and outlet sections of the flow channel is statistically analyzed to calculate the overall pressure drop of the flow channel; and using the branch identifiers in the local feature set, each sampling point is grouped and statistically analyzed, and combined with the three-dimensional pressure distribution values, the pressure distribution inside each target flow branch is obtained, and then the branch pressure drop of each target flow branch is calculated. If the predicted flow characteristics include velocity scalar distribution values, then based on the velocity scalar distribution values, sampling points located at specified cross sections are selected, and the velocity scalar distribution values ​​of sampling points with the same branch identifier are accumulated or integrated using the branch identifier in the local feature set to calculate the volumetric flow rate corresponding to each target flow branch; then, based on the volumetric flow rate of all target flow branches, the flow distribution ratio of each target flow branch is calculated.

[0056] This invention achieves a complete closed loop from microscopic field distribution prediction to macroscopic engineering performance index calculation. Based on the three-dimensional pressure distribution and velocity scalar distribution values ​​output by the model, combined with the branch identifiers in the local feature set, it can directly perform grouped statistics and integration operations on each sampling point, thereby accurately calculating the overall pressure drop of the flow channel, the branch pressure drop of each target flow branch, the volumetric flow rate, and the flow distribution ratio. This capability enables the method to not only provide flow field details at the cloud map level, but also directly output key system-level indicators to guide engineering design, combining the efficiency of deep learning with the physical interpretability of traditional fluid dynamics analysis.

[0057] This invention determines the axial distance parameter along the flow direction of each center point on each target flow branch by performing axial parameterization on the target flow branch, thus assigning a clear physical flow direction measurement benchmark to each geometric node within the flow channel. This key step endows the originally discrete set of center points with continuous friction distance attributes, accurately characterizing the cumulative flow distance of fluid in complex topological paths. Combined with a deep operator learning architecture, this method utilizes a local feature set containing this axial distance parameter and a global state vector to directly establish a mapping relationship from geometric topology to flow field distribution, thereby achieving rapid prediction of the internal flow characteristics of the microfluidic chip's flow channel network without the need for traditional grid iterative calculations.

[0058] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of the components in a specific posture (as shown in the attached figures). If the specific posture changes, the directional indication will also change accordingly. Furthermore, descriptions involving "first," "second," or "a" in the present invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. In the description of the present invention, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly defined. In the present invention, unless otherwise explicitly specified and defined, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can be a fixed connection, a detachable connection, or an integral part; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal communication of two components or the interaction between two components, unless otherwise explicitly defined. For those skilled in the art, the specific meanings of the above terms in this invention can be understood according to the specific circumstances. Furthermore, the technical solutions of the various embodiments of this invention can be combined with each other, but only on the basis that those skilled in the art can implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

Claims

1. A method for rapid prediction of flow characteristics in a flow channel network combining microfluidic experiments and numerical simulation, characterized in that, include: A three-dimensional structural model of the flow channel network in the microfluidic chip is obtained, and the inner wall surface of the flow channel is sampled to generate wall point cloud data. The wall point cloud data is spatially aggregated to obtain a set of center points, and in the set of center points, the point pairs with Euclidean distance less than a preset distance threshold are connected to construct a centerline network. In the centerline network, the starting node and the ending node corresponding to the inlet and outlet of the flow channel are determined respectively. Multiple discretely distributed center points are selected as candidate seed points in the middle area between the starting node and the ending node. Multiple target flow branches are obtained based on each candidate seed point. The axial parameterization of each target flow branch is performed to determine the axial distance parameter of each center point along the flow direction on the target flow branch. The set of center points is sampled to generate a set of sampling points; the axial distance parameters of each target flow branch and its center points along the flow direction are used to obtain the local feature set corresponding to each sampling point in the set of sampling points. A global state vector is constructed, and the global state vector and the local feature set corresponding to each sampling point in the sampling point set are input into the target prediction model to obtain the predicted flow characteristics of each sampling point; the global state vector represents the operating condition control parameters of the microfluidic chip during operation; the target prediction model is obtained by training an initial model based on a deep operator learning architecture.

2. The method for rapid prediction of flow characteristics of channel networks combining microfluidic experiments and numerical simulation as described in claim 1, characterized in that, The intermediate region includes: the area located between the starting node and the ending node, and covering the area where the flow channel geometry undergoes bifurcation, merging, or bending changes; The process of obtaining multiple target flow branches based on each candidate seed point specifically includes: For each candidate seed point, a first shortest path connecting the candidate seed point to the starting node and a second shortest path connecting the candidate seed point to the ending node are searched in the centerline network. The first shortest path and the second shortest path are then combined at the candidate seed point to form a candidate flow branch. All candidate flow branches are aggregated to form a candidate flow branch set; The candidate flow branches in the candidate flow branch set are deduplicated to obtain multiple independent flow branches; each independent flow branch is smoothed to obtain multiple target flow branches.

3. The method for rapid prediction of flow characteristics of channel networks combining microfluidic experiments and numerical simulation as described in claim 2, characterized in that, The process of deduplicating candidate flow branches in the candidate flow branch set to obtain multiple independent flow branches is as follows: Select any candidate flow branch from the candidate flow branch set as an independent flow branch and add it to the preset independent flow branch library; The remaining candidate flow branches in the candidate flow branch set are compared with the existing independent flow branches in the independent flow branch library in turn. In each comparison, the shortest Euclidean distance from multiple center points on the candidate flow branch being compared to the independent flow branch being compared is calculated, and the percentage of distance values ​​less than the preset deduplication threshold is counted. If the proportion of the number is greater than the preset overlap ratio threshold, the subsequent comparison of the candidate flow branch will be terminated immediately. If the candidate flow branch being compared has completed the comparison with all independent flow branches in the independent flow branch library, and the proportion of the number of comparisons in each comparison is less than or equal to the preset overlap ratio threshold, then it is added to the independent flow branch library as an independent flow branch. All paths in the independent flow branch database are considered as the multiple independent flow branches.

4. The method for rapid prediction of flow characteristics of channel networks combining microfluidic experiments and numerical simulation as described in claim 1, characterized in that, The determination of the axial distance parameters along the flow direction of each center point on the target flow branch specifically includes: Set the starting node as the zero point for the flow distance calculation; For each target flow branch, for each center point on that branch, calculate the cumulative path length from the zero point to the center point along the target flow branch; The cumulative path length is used as the axial distance parameter of the center point.

5. The method for rapid prediction of flow characteristics of channel networks combining microfluidic experiments and numerical simulation as described in claim 4, characterized in that, The acquisition of the local feature set corresponding to each sampling point in the sampling point set specifically includes: For each sampling point in the set of sampling points: Calculate the shortest vertical distance from it to the wall point cloud data; Determine the target flow branch to which the sampling point belongs, and determine its projection position on the target flow branch to which it belongs; Based on the projection position, obtain the axial distance parameter corresponding to the sampling point, and obtain the branch identifier of the target flow branch to which it belongs; obtain the three-dimensional spatial coordinates of the sampling point, and combine the three-dimensional spatial coordinates, the shortest vertical distance, the branch identifier and the axial distance parameter to construct a local feature set; The step of obtaining the axial distance parameter corresponding to the sampling point based on the projection position includes: If the projected position coincides with any center point on the target flow branch, then the axial distance parameter of that center point is directly used as the axial distance parameter corresponding to the sampling point. If the projected position does not coincide with any center point, the axial distance parameter corresponding to the sampling point is obtained by linear interpolation of the axial distance parameter of the center point on the target flow branch.

6. The method for rapid prediction of flow characteristics of channel networks combining microfluidic experiments and numerical simulation as described in claim 5, characterized in that, The step of determining the target flow branch to which the sampling point belongs specifically includes: projecting the sampling point to the target flow branch with the closest spatial distance, and determining that the branch is the target flow branch to which it belongs; or: Set the preset core radius; Iterate through the set of sampling points, and for each sampling point: Calculate the vertical distance from each of the target flow branches, and mark the branch corresponding to the minimum vertical distance as the candidate nearest branch; If the minimum vertical distance is less than the preset core radius, the sampling point is defined as the safe core point, and the corresponding candidate nearest branch is determined as its target flow branch. If the minimum vertical distance is greater than or equal to the preset core radius, the sampling point is defined as a non-core point. For each non-core point, among all the determined safe core points, the safe core point with the closest spatial Euclidean distance to the non-core point is found, and the target flow branch to which the safe core point belongs is inherited as the target flow branch to which the current non-core point belongs.

7. The method for rapid prediction of flow characteristics of channel networks combining microfluidic experiments and numerical simulation as described in claim 1, characterized in that, The global state vector includes fluid property parameters and microfluidic chip inlet flow rate and / or operating condition control parameters; the operating condition control parameters include at least one of inlet pressure, applied electric field strength, or rotational angular velocity. The initial model includes: A branch network is used to receive the global state vector and encode it to obtain a global operating condition feature embedding. The backbone network is used to receive the local feature set and encode it to obtain local geometric feature embeddings; The fusion layer is used to fuse the global operating condition feature embedding with the local geometric feature embedding, and map and output the predicted flow characteristics of each sampling point; The predicted flow characteristics include at least one of three-dimensional pressure distribution values ​​and velocity scalar distribution values.

8. The method for rapid prediction of flow characteristics of a flow channel network combining microfluidic experiments and numerical simulation as described in claim 7, characterized in that, The training method for the target prediction model includes: A training dataset is constructed, which contains multiple training samples. Each training sample includes: a global state vector corresponding to a three-dimensional structural model of the flow channel network in a microfluidic chip, and a set of local features corresponding to each sampling point within the three-dimensional structural model; and a training label that matches each sampling point one by one. The training labels are derived from microfluidic experimental measurement data and / or numerical simulation data, and the training label corresponding to each sampling point includes the true pressure value and / or the true velocity scalar value of that point. The training samples are sequentially input into the initial model to obtain predicted values ​​of flow characteristics; Calculate the loss function value between the predicted flow characteristics and the training labels; Based on the loss function value, the model parameters of the initial model are iteratively updated until the loss function value converges to a preset threshold, thereby obtaining the target prediction model.

9. The method for rapid prediction of flow characteristics of a flow channel network combining microfluidic experiments and numerical simulation as described in claim 7, characterized in that, The rapid prediction method for flow properties also includes: If the predicted flow characteristics include three-dimensional pressure distribution values, then based on the three-dimensional pressure distribution values, the average pressure at the inlet and outlet sections of the flow channel is statistically analyzed to calculate the overall pressure drop of the flow channel; and using the branch identifiers in the local feature set, each sampling point is grouped and statistically analyzed, and combined with the three-dimensional pressure distribution values, the pressure distribution inside each target flow branch is obtained, and then the branch pressure drop of each target flow branch is calculated. If the predicted flow characteristics include velocity scalar distribution values, then based on the velocity scalar distribution values, sampling points located at specified cross sections are selected, and the velocity scalar distribution values ​​of sampling points with the same branch identifier are accumulated or integrated using the branch identifier in the local feature set to calculate the volumetric flow rate corresponding to each target flow branch; then, based on the volumetric flow rate of all target flow branches, the flow distribution ratio of each target flow branch is calculated.

Images