Manipulation system and fluid chip

Abstract

[Problem] To provide technology for facilitating control of the position of a manipulation object. [Solution] Provided is a manipulation system for arranging and manipulating a manipulation object at a prescribed position in a manipulation region containing a liquid. The manipulation system comprises: the manipulation region; first flow paths that are a plurality of flow paths connected to the manipulation region and containing the liquid; a liquid control unit that moves the liquid inside the first flow paths; and second flow paths that are a series of flow paths containing the liquid and that connect to the respective plurality of first flow paths.

Description

Manipulation system and fluidic chip

[0001] The present invention relates to a manipulation system and a fluidic chip.

[0002] In recent years, there has been active development of a technology called Micro Total Analysis Systems (MicroTAS), which uses minute flow channels and the like to perform biochemical processes. MicroTAS is used, for example, for analyzing cell-cell interactions and constructing systems in which multiple cells are arranged in the field of regenerative medicine. Conventionally, a technology has been known in which an object to be manipulated within a liquid is moved to a target position by moving the liquid within a flow channel containing the liquid (see, for example, Patent Document 1).

[0003] Japanese Patent Application Laid-Open No. 2021-185782

[0004] In general, controlling the movement of an object in a liquid to a desired position by moving the liquid is difficult. One of the reasons for the difficulty of control is that the object is moved indirectly using the liquid. Because the shape of the liquid is not fixed, simply repeating the action of applying a predetermined pressure to the liquid makes it difficult to predict the behavior of the liquid and the object around it, making it difficult to control the position of the object.

[0005] The present invention has been made in consideration of the above-mentioned problems, and has an object to provide a technique for facilitating position control of an operation target object.

[0006] In order to achieve the above-mentioned object, the manipulation system of the present invention comprises an operation area containing a liquid and an object to be manipulated whose position within the liquid is manipulated, a first flow path which is a plurality of flow paths connected to the operation area and which contains the liquid, a liquid control unit which moves the liquid within the first flow path, and a second flow path which is a series of flow paths containing the liquid and which is connected to each of the plurality of first flow paths.

[0007] That is, the manipulation area contains a liquid and a manipulation object contained in the liquid, and a first flow path connected to the manipulation area can move the liquid in the first flow path. Multiple first flow paths are provided, and as the liquid in the first flow paths moves, the liquid flows into the manipulation area from the first flow paths and flows out from the manipulation area to the first flow paths. Therefore, the liquid in the manipulation area moves in conjunction with the movement of the liquid in the first flow paths, resulting in the movement of the manipulation object within the manipulation area. In this configuration, when the liquid in the first flow path moves, local pressure fluctuations occur, and the pressure eventually returns to a steady state where it is uniform. If such pressure fluctuations occur locally or remain for a long period of time, it becomes difficult to control the repeated movement of the liquid within the first fluid.

[0008] Therefore, the manipulation system is provided with a second flow path. The second flow path is a series of flow paths containing liquid and is connected to each of the multiple first flow paths. Therefore, the second flow path forms a flow path connecting the multiple first flow paths to each other. As a result, the pressure in the multiple first flow paths is equalized, making it difficult for local pressure differences to occur. Furthermore, even if pressure fluctuates due to the movement of liquid in a specific first flow path, the pressure fluctuation is quickly equalized. Therefore, when manipulating the position of a manipulation object by repeatedly moving the liquid, it is possible to easily control the position of the manipulation object.

[0009] FIG. 1A is a diagram showing a schematic configuration of a manipulation system, and FIG. 1B is a diagram showing the main parts of the manipulation system. FIG. 2A is a diagram showing an example of an image captured by a microscope, FIG. 2B is a diagram showing the shape of the manipulation area, and FIGS. 2C to 2E are diagrams showing the structure of a switching unit. FIG. 4A is a diagram showing a simulation model, and FIG. 4B is a diagram showing a machine learning model. FIG. 4A is a diagram explaining reinforcement learning. FIG. 4B is a flowchart of reinforcement learning processing. FIG. 4C is a flowchart of manipulation processing. FIG. 8A to 8C are diagrams showing an example of moving a manipulation object to a target position. FIG. 4C is a diagram showing an example of a configuration for introducing a manipulation object into a manipulation area. FIG. 4A is a diagram showing a board realizing a manipulation system having multiple manipulation areas. FIG. 4C is an enlarged view of a board realizing a manipulation system having multiple manipulation areas.

[0010] Here, the embodiments of the present invention will be described in the following order: (1) Configuration of the manipulation system: (2) Machine learning processing using A.I. (artificial intelligence): (3) Manipulation processing: (4) Other embodiments:

[0011] (1) Configuration of the manipulation system:

[0012] 1A is a diagram showing a schematic configuration of a manipulation system, and FIG. 1B is a diagram showing the main parts of a manipulation system 10. The manipulation system 10 according to this embodiment includes a computer 11, a microscope 12, a signal conversion interface 13, an electromagnetic valve 14, a pressure supply pipe 15, and a microfluidic chip 16.

[0013] In this embodiment, the computer 11 is connected to the microscope 12 and the signal conversion interface 13. The computer 11 acquires image data output by the microscope 12. The computer 11 also outputs a digital signal, which is input to the signal conversion interface 13. Details will be described later. The microscope 12 is equipped with an image sensor capable of capturing an image included in its field of view. In this embodiment, the microscope 12 is capable of capturing an image with the field of view being an operation area 60 of the microfluidic chip 16, which will be described later. Image data representing the captured image is transferred to the computer 11.

[0014] The signal conversion interface 13 is connected to the computer 11 and the solenoid valve 14 and outputs a predetermined current to the solenoid valve 14 in response to a digital signal output by the computer 11. A compressor (not shown) and a pressure supply pipe 15 are connected to the solenoid valve 14, and the connection between the compressor and the pressure supply pipe 15 is switched by opening and closing the solenoid valve 14. This allows the solenoid valve 14 to switch the internal pressure of the pressure supply pipe 15 between a high state and a low state. That is, the solenoid valve 14 adjusts the internal pressure of the pressure supply pipe 15 in response to the current supplied from the signal conversion interface 13. In this embodiment, the solenoid valve 14 keeps the internal pressure of the pressure supply pipe 15 high when the current output from the signal conversion interface 13 is high, and keeps the internal pressure of the pressure supply pipe 15 low when the current output from the signal conversion interface 13 is low.

[0015] The pressure supply pipe 15 is connected to a microfluidic chip 16. An air pressure adjustment unit 70 (described later) is formed in the microfluidic chip 16. Therefore, the solenoid valve 14 can switch the internal pressure of the air pressure adjustment unit 70, to which the pressure supply pipe 15 is connected, between a high state and a low state. In this embodiment, six solenoid valves 14 and six pressure supply pipes 15 are provided as shown in FIG. 1A, but only one of them is indicated by a reference numeral in FIG. 1A.

[0016] The microfluidic chip 16 is provided with a liquid containing portion that contains a liquid and a structure that contains a gas. Fig. 1B shows the liquid containing portion and structure provided in the microfluidic chip 16. The microfluidic chip 16 is a rectangular parallelepiped, i.e., a plate, with two sides larger than the other four sides, and the liquid containing portion and structure are formed on the side with the larger area. Fig. 1B shows the microfluidic chip 16 as viewed from a direction perpendicular to the side with the larger area.

[0017] The microfluidic chip 16 is provided with a first flow path 40, a second flow path 50, an operation area 60, and an air pressure adjustment unit 70. In this embodiment, the liquid is water and the gas is air. Note that the operation object is introduced into the first flow path 40 or the operation area 60, but the configuration for introducing the operation object is omitted in FIG. 1B.

[0018] The first flow path 40, the second flow path 50, and the operation area 60 are flow paths and areas that contain liquid. The first flow path 40 is a linear hollow space, and one end of the first flow path 40 is connected to the operation area 60. The other end of the first flow path 40 is connected to the second flow path 50. In this embodiment, three first flow paths 40 are provided. In this embodiment, the first flow paths 40 are arranged at equal intervals within the widest plane of the microfluidic chip 16 (a plane parallel to the paper surface of FIG. 1B). Therefore, in FIG. 1B, the angle between adjacent first flow paths 40 is 120°.

[0019] The operation area 60 contains a liquid and is an area for placing and operating an operation object at a predetermined position, and has an internal space for moving the operation object in a plane. In this embodiment, the operation area 60 and the like are formed on the largest surface of the microfluidic chip 16, and the internal space of the operation area 60 has a shape that allows the operation object to be moved in a plane in a direction parallel to the largest surface of the microfluidic chip 16. In other words, the operation area 60 has an internal space that extends to a certain depth (e.g., a depth of 30 μm or less) in the largest surface of the microfluidic chip 16 in a direction parallel to the surface.

[0020] Furthermore, the internal space of the operation area 60 has a triangular shape when viewed from a direction perpendicular to the largest surface of the microfluidic chip 16. Figure 2A shows an example of an image including the operation area 60 captured by the microscope 12. The operation area 60 is a space surrounded by three sides. If we imagine vertices V1 to V3 where the three sides are extended and intersect as shown by the dashed lines in Figure 2B, the operation area 60 has a triangular internal space with vertices V1 to V3. That is, in this embodiment, if the inner wall of the internal space of the operation area 60 is extended, it becomes a triangle.

[0021] The first flow paths 40 are connected to the vertices of a triangle formed by the operation area 60. In this embodiment, three first flow paths 40 are formed in a direction substantially parallel to a line extending from the center of gravity of the triangle formed by the operation area 60 to the vertex. In this embodiment, the first flow paths 40 have the same depth as the operation area 60 and a width smaller than one side of the triangle formed by the operation area 60. In this embodiment, the first flow paths 40 are linearly extending flow paths. In this embodiment, the first flow paths 40 are hollow spaces with rectangular cross sections. One end of each of the first flow paths is connected to the operation area 60 as shown in FIG. 2A and the other end is connected to the second flow path 50 as shown in FIG. 1B. In this specification, the depth refers to the length perpendicular to the largest surface of the microfluidic chip 16, and the width refers to the direction perpendicular to the depth direction and the short direction.

[0022] As shown in FIG. 1B , the second flow path 50 includes a storage portion 51 and a connecting portion 52. In this embodiment, the storage portion 51 has a circular shape when viewed from a direction perpendicular to the largest surface of the microfluidic chip 16. The storage portion 51 is connected to the first flow path 40 and has a depth greater than that of the first flow path 40 (e.g., 3 mm to 5 mm). The diameter of the circle formed by the storage portion 51 is greater than the width of the first flow path 40. In this embodiment, three storage portions 51 are provided, all of which have the same depth and shape. In this embodiment, the rectangular parallelepiped microfluidic chip 16 is placed on a horizontal surface so that the liquid levels in the three storage portions 51 are at the same height.

[0023] The connecting portion 52 is formed between two adjacent accumulation portions 51 so as to connect these accumulation portions 51. In this embodiment, the connecting portion 52 is a hollow space with a rectangular cross section. The depth of the connecting portion 52 is the same as that of the first flow path 40. The width is smaller than the diameter of the circle formed by the accumulation portions 51. Therefore, the accumulation portions 51 are larger than the connecting portion 52. Note that in this embodiment, the accumulation portions 51 and the connecting portion 52 exist around a triangle formed by the operation area 60, with the accumulation portions 51 corresponding to vertices of the triangle larger than the operation area 60 and the connecting portion 52 corresponding to sides of the triangle larger than the operation area 60. Therefore, the second flow path 50 can also be considered as an annular flow path formed around the operation area 60 so as to connect the accumulation portions 51.

[0024] The air pressure adjusting unit 70 is a structure including a passage containing gas. The air pressure adjusting unit 70 includes a switching unit 71, a gas passage 72, and a connecting unit 73. The connecting unit 73 is a portion to which the pressure supply pipe 15 is connected. In this embodiment, six connecting units 73 are formed, and the six pressure supply pipes 15 are connected to the connecting units 73, respectively. The gas passage 72 is connected to the connecting units 73 and is a hollow space that communicates with the space directly below the switching unit 71 formed directly below the first flow path 40 and the connecting units 73. Note that six air pressure adjusting units 70 are provided as shown in FIG. 1B, but only one of them is indicated by a reference numeral in FIG. 1B.

[0025] The switching unit 71 includes a mechanism that can switch the first flow paths 40 between a closed state and an open state. In this embodiment, two switching units 71 are provided for each of the three first flow paths 40. FIGS. 2C to 2E are diagrams for explaining the structure of the switching unit 71 and the adjustment of the internal pressure of the first flow paths 40 by the switching unit 71. FIG. 2C shows the switching unit 71 as viewed from a direction perpendicular to the largest surface of the microfluidic chip 16. FIGS. 2D and 2E show the switching unit 71 as viewed from a direction parallel to the largest surface of the microfluidic chip 16.

[0026] In this embodiment, the switching unit 71 includes a diaphragm 71 a, and a space 71 b is provided directly below the diaphragm 71 a. The first flow path 40 and the space 71 b are separated by a flexible thin film 40 a, and the portion of the thin film 40 a where the first flow path 40 and the space 71 b overlap is the diaphragm 71 a.

[0027] 2C to 2E is the diaphragm 71a, and the light gray shown in FIGS. 2D and 2E is the liquid contained in the first flow path 40. The space 71b is a rectangular parallelepiped cavity that is connected to the gas passage 72. The material of the diaphragm 71a is not limited, but in this embodiment it is made of resin.

[0028] As described above, the diaphragm 71 a is the portion where the first flow path 40 and the space 71 b overlap, and is rectangular when viewed in a direction perpendicular to the largest surface of the microfluidic chip 16 .

[0029] When the internal pressure of the pressure supply pipe 15 becomes high or low, the internal pressure of the space 71b also becomes high or low. Because the diaphragm 71a is a flexible membrane, its shape changes depending on the internal pressure of the space 71b. FIG. 2D shows the shape when the internal pressure of the space 71b is low. FIG. 2E shows the shape when the internal pressure of the space 71b is high. As shown in FIG. 2E, when the internal pressure of the space 71b becomes high, the diaphragm 71a expands toward the first flow path 40, closing the first flow path 40. In other words, the first flow path 40 is closed. As shown in FIG. 2D, when the internal pressure of the space 71b becomes low, the diaphragm 71a becomes flat, allowing liquid to flow through the first flow path 40. In other words, the first flow path 40 is open.

[0030] In this way, when the first flow path 40 is switched between a closed state and an open state, the internal pressure of the first flow path 40 fluctuates during this change, causing the liquid to flow. For example, when the state shown in FIG. 2D is switched to the state shown in FIG. 2E , the liquid flows through the first flow path 40 toward the left and right. In this embodiment, two switching units 71 are provided for one first flow path 40. Therefore, by combining the operations of the two switching units 71, the liquid can be moved from the first flow path 40 to the operation area 60 and from the operation area 60 to the first flow path 40.

[0031] In this embodiment, one switching unit 71 can realize two states: a state in which the first flow path 40 is closed, and a state in which the first flow path 40 is not closed. Furthermore, in this embodiment, two switching units 71 are provided for each of the three first flow paths 40, so a total of six switching units 71 are provided. If a 1 corresponds to a state in which the first flow path 40 is closed and a 0 corresponds to a state in which the first flow path 40 is not closed, the states realized by the total of six switching units 71 can be expressed by a 6-bit numerical value. Hereinafter, the states realized by the total of six switching units 71 and expressed by a 6-bit numerical value will be referred to as the states of the switching units 71.

[0032] In this embodiment, the above devices are controlled by a computer 11, thereby allowing the operation target object to be moved to a predetermined target position in the operation area 60. FIG. 3 is a diagram illustrating the configuration of the computer 11. The computer 11 includes a control unit 20 and a storage medium 30. The control unit 20 includes a CPU, RAM, ROM, and GPU (not shown), and is capable of executing various programs stored in the storage medium 30, etc. The control unit 20 and the storage medium 30 may be configured as an integrated computer, or at least a portion of them may be separate devices connected by various cables, etc. Of course, the computer 11 may also include an output unit that outputs images, sounds, etc., and an input unit for inputting user instructions, etc. The output unit and the input unit may be connectable via various interfaces.

[0033] In this embodiment, the control unit 20 can execute a manipulation program 21 and a machine learning program 22. The manipulation program 21 is a program that causes the control unit 20 to execute a function of moving an operation target object to a target position based on an image of the operation area 60 captured by the microscope 12. When the manipulation program 21 is executed, the control unit 20 functions as an imaging unit 21a and a position control unit 21b.

[0034] In this embodiment, the movement of the operation target is performed based on the results of machine learning that has been performed in advance. The machine learning program 22 is a program that causes the control unit 20 to perform a function of machine learning a model for determining a control target for the switching unit 71, based on, for example, an image of the operation area 60 captured by the microscope 12. When the machine learning program 22 is executed, the control unit 20 functions as a machine learning unit 22a.

[0035] (2) Machine learning processing using A.I. (artificial intelligence):

[0036] In this embodiment, a model for determining a control target for the switching unit 71 is learned by machine learning, and a model for simulating the movement of an operation object by a liquid is also learned by machine learning to support the machine learning. Here, the former is simply referred to as machine learning model 30b, and the latter is referred to as simulation model 30c.

[0037] In this embodiment, the simulation model 30c receives input data indicating the position of the operation target object and the current and next states of the switching unit 71, and outputs output data indicating the next position of the operation target object. FIG. 4A is a diagram schematically illustrating the simulation model 30c. In FIG. 4A, input data is shown on the left and output data on the right, and the input data is input to the simulation model 30c, and the output data is output from the right. In FIG. 4A, Xm and Ym indicate the position (coordinates) of the operation target object. V1 to V6 indicate the respective states (1 or 0) of the six switching units 71. When V1 to V6 are arranged, a 6-bit numerical value indicating the state of the switching unit 71 is obtained.

[0038] The symbol t is added to the symbols indicating the position of the operation object and the state of the switching unit 71. The symbol t is a symbol for identifying the number of operations performed by the switching unit 71. Therefore, as shown in FIG. 4A , the position of the operation object after the tth operation is input to the simulation model 30c. Furthermore, the state of the switching unit 71 after the tth operation and the state of the switching unit 71 after the t+1th operation are input to the simulation model 30c. The position of the operation object after the t+1th operation is then output. Therefore, according to the simulation model 30c, when the position of the operation object is at a specific position and the switching unit 71 is in a specific state after the tth operation, it is possible to simulate the position to which the operation object will be displaced by next changing the switching unit 71 to the specific state.

[0039] Such a simulation model 30c can be generated by machine learning using various techniques. For example, the simulation model 30c can be generated by optimizing a model constructed by a neural network through machine learning. Specifically, when the state of the switching unit 71 is in a specific state, the microscope 12 in the manipulation system 10 captures an image and identifies the position (coordinates) of the manipulation target object. In this state, the computer 11 outputs a signal to the signal conversion interface 13 to change the state of the switching unit 71.

[0040] Then, an image is captured again using the microscope 12, and the position of the operation object corresponding to the change in state is identified. By collecting information regarding a series of operations, the position of the operation object after the state change can be associated with the current position of the operation object and information indicating the current state and next state of the switching unit 71. This set is used as training data, and a sufficient number of training data samples are collected for training the neural network. Once the training data has been collected, the control unit 20 executes a known machine learning process using the function of the machine learning unit 22a to optimize the neural network. The model obtained in this manner becomes the simulation model 30c and is recorded in the storage medium 30.

[0041] Machine learning of the machine learning model 30b in the environment is performed after the simulation model 30c has already been created. Machine learning of the machine learning model 30b may also be performed using various techniques. Here, an example of machine learning of the machine learning model 30b using reinforcement learning is described. FIG. 5 is a diagram illustrating learning of the machine learning model 30b using a reinforcement learning model consisting of an agent and an environment. The agent shown in FIG. 5 corresponds to a function that selects an action a that can be taken in the current state s. The environment is used to determine the next state s' based on the action a selected by the agent and the current state s, and to determine a reward r' based on the action a, the state s, and the state s'. The environment may be defined so that the next state s' can be determined from the action a and the state s, and the reward r' can be determined. Therefore, the state may be actually observed, or the state may be determined in a virtual environment. In this embodiment, the virtual environment is defined, for example, by virtually reproducing the operation area 60 in a three-dimensional space virtually created in the computer 11. Then, the position of the manipulation target object in the virtual three-dimensional space, etc., are input to the simulation model 30c, thereby simulating changes in the state in the virtual environment. In this way, when reinforcement learning is performed in a virtual environment using the simulation model 30c, optimization of the machine learning model 30b can be achieved more quickly than in an environment requiring actual observation. Note that after the machine learning model 30b is optimized in a virtual environment and operation of the manipulation system 10 using the machine learning model 30b is started, information observed in a real environment may be used to further perform machine learning of the machine learning model 30b.

[0042] Various methods can be adopted as the reinforcement learning algorithm, and Q-learning will be described here as an example. In Q-learning, an action a is selected according to the current state s, and then an action value function Q(s, a) is assumed that indicates the expected profit when action a is selected according to a predetermined policy. The optimal policy is obtained when the action value function Q(s, a) is optimized. The policy may be defined in various ways. For example, a greedy policy is a policy that selects an action that maximizes the action value function Q(s, a). When the action value function Q(s, a) is optimized in such a state, the optimal policy is a policy that selects an action a such that the action value function Q(s, a) is maximized in state s. The machine learning model 30b may be defined in various ways, but here it is defined so that it includes the action value function Q(s, a) and the data indicating the action a that maximizes the action value function Q(s, a) is the output data.

[0043] In this embodiment, state s is the position of the operation object, the state of the switching unit 71, and the target position of the operation object. Action a is the operation of the switching unit 71 required to bring the operation object closer to the target position, i.e., the next state of the switching unit 71. For this reason, the machine learning model 30b is defined so as to receive information indicating the position of the operation object, the state of the switching unit 71, and the target position of the operation object as input data, and to output information indicating the operation of the switching unit 71 required to bring the operation object closer to the target position as output data.

[0044] FIG. 4B is a schematic diagram illustrating the machine learning model 30b. The notation of symbols is the same as in FIG. 4A . Xg and Yg are the positions (coordinates) of the target position of the operation object. In a virtual environment, these positions are coordinates in a virtually constructed environment, and in a real environment, they are positions in an image captured by the microscope 12. As shown in FIG. 4B , the machine learning model 30b receives the position of the operation object at the tth time, the state of the switching unit 71 after the tth operation, and the target position of the operation object. These positions and states input to the machine learning model 30b correspond to the current state s shown in FIG. 5 . The output from the machine learning model 30b is the state of the switching unit 71 after the t+1th operation, which corresponds to action a shown in FIG. 5 . Therefore, assuming that the tth time is the present, the machine learning model 30b outputs the next action a (the next state of the switching unit 71) based on the current state s (the current state and target position of the operation object and the switching unit 71).

[0045] FIG. 6 is a flowchart illustrating the main parts of the reinforcement learning process for generating the machine learning model 30b described above. FIG. 6 illustrates the process of machine learning the machine learning model 30b in a virtual environment. The reinforcement learning process illustrated in FIG. 6 is performed in a state in which the optimized simulation model 30c, the unoptimized machine learning model 30b, and information indicating the virtually constructed environment are recorded in the storage medium 30. When the reinforcement learning process is started, the control unit 20 initializes the virtual environment constructed for learning using the function of the machine learning unit 22a (step S100). That is, the control unit 20 sets values ​​of variables and the like used in the reinforcement learning process to initial values. Specifically, the control unit 20 sets the position of the control object and the initial state of the switching unit 71, and further sets a virtual target position. The initial value of the position of the virtual control object is the initial position during learning, and a virtually predetermined position becomes the initial position. There may be multiple initial positions depending on the nature of the task, and processing is performed for each initial position in each episode repeated in the learning loop process of steps S100 to S130.

[0046] The virtual target position can be set within the virtually reproduced operation area 60, and any position within the operation area 60 can be set as the virtual target position. In each episode repeated in the loop processing of steps S100 to S130, the target position may be changed or fixed. In the manipulation system 10, if the operation target object is moved toward a specific target position, it is preferable that the target position be fixed to the specific target position within the same episode (S100 to S125). If the target position is not limited to a specific target position and multiple positions can be the target position, it is preferable that machine learning proceed with multiple positions as initial values ​​of the target position in the reinforcement learning processing. During the process of repeating the loop processing of steps S100 to S130, the initial value of the state of the switching unit 71 may be fixed to a specific state (e.g., 000000) or may be variable.

[0047] Next, the control unit 20 inputs the state s to the machine learning model 30b using the function of the machine learning unit 22a (step S105). That is, the control unit 20 inputs the current state (positions Xmt, Ymt of the operation target object, states V1t to V6t of the switching unit 71, and target positions Xg, Yg) to the machine learning model 30b currently being trained. As a result, states V1t+1 to V6t+1 of the switching unit 71 are output as the next action. Note that in the process of machine learning, the next states, states V1t+1 to V6t+1 of the switching unit 71, are not necessarily appropriate, but are gradually optimized as the machine learning progresses.

[0048] Next, the control unit 20 uses the function of the machine learning unit 22a to simulate the position of the operation object according to the action (step S110). That is, the control unit 20 inputs the operation object positions Xmt, Ymt, the current states V1t to V6t of the switching unit 71, and the next states V1t+1 to V6t+1 of the switching unit 71 obtained in step S105 to the simulation model 30c recorded in the storage medium 30. As a result, the next operation object positions Xmt+1, Ymt+1 are output. This makes it possible to reproduce an equivalent movement phenomenon of the operation object in a virtual environment without using a real environment.

[0049] Next, the control unit 20 performs empirical observation using the function of the machine learning unit 22a (step S115). That is, the control unit 20 identifies a state and a reward corresponding to the action taken in steps S105 and S110. Specifically, the control unit 20 identifies a reward based on the target positions Xg, Yg and the positions Xmt+1, Ymt+1 of the next control object. The reward may be determined using various methods. Specifically, the reward is defined so that a positive reward is obtained when the distance between the target positions Xg, Yg and the positions Xmt+1, Ymt+1 of the next control object is within a predetermined distance. However, other methods may be used, for example, a larger reward may be provided as the distance between the target positions Xg, Yg and the positions Xmt+1, Ymt+1 of the next control object is smaller, or a negative reward whose absolute value increases as the distance is greater. In any case, once the reward is defined, the action value function Q(s, a) can be defined, for example, using a known definition.

[0050] Next, the control unit 20 updates the machine learning model 30b using the function of the machine learning unit 22a (step S120). The machine learning model 30b includes an action value function Q(s, a), and reinforcement learning aims to optimize the action value function Q(s, a), which is not initially optimized. In this embodiment, the action value function Q(s, a) is defined by a multilayer neural network, and the multilayer neural network is optimized during the reinforcement learning process. Optimization of the multilayer neural network may be performed using various known algorithms. For example, a configuration in which the multilayer neural network is optimized using an objective function that minimizes the TD error (Temporal Difference) can be employed.

[0051] After updating the action value function Q(s, a) according to a known algorithm, i.e., updating the machine learning model 30b, the control unit 20 determines whether processing of the current episode has ended using the function of the machine learning unit 22a (step S125). In this embodiment, one episode is defined as the period from the first observation of the position of the control object until the control object reaches the target position Xg, Yg, or until the control object fails to reach the target position Xg, Yg. The failure condition may be determined in advance. For example, a configuration may be employed in which the failure condition is considered to be satisfied when the state s changes a predetermined number of times or more without the control object reaching the target position Xg, Yg. Alternatively, a configuration may be employed in which the failure condition is considered to be satisfied when the control object reaches outside the operation area 60.

[0052] If it is determined in step S125 that the processing of the current episode has not ended, the control unit 20 repeats the processing from step S105 onward. If it is determined in step S125 that the processing of the current episode has ended, the control unit 20 determines whether the machine learning has ended (step S130). The completion condition for the machine learning may be determined in various ways. For example, a configuration may be adopted in which the completion condition for the machine learning is deemed to be satisfied when the processing of steps S105 to S125 is completed for all of a plurality of environments prepared in advance. Furthermore, satisfaction of the completion condition may be determined by an index for evaluating whether the machine learning model 30b has been sufficiently optimized, such as the task completion rate.

[0053] If it is determined in step S130 that the machine learning is not complete, the control unit 20 repeats the processes from step S100 onward using the function of the machine learning unit 22a. In this embodiment, the environment is initialized at this time so that the initial state is different from that of the previous episode. If it is determined in step S130 that the machine learning is complete, the control unit 20 terminates the reinforcement learning process. As a result, an optimized machine learning model 30b is obtained.

[0054] (3) Manipulation process:

[0055] Next, a manipulation process for manipulating and placing the manipulation target object at a target position will be described. FIG. 7 is a flowchart showing the manipulation process. The manipulation process is realized by the control unit 20 executing the manipulation program 21 while the optimized machine learning model 30b is recorded in the storage medium 30. When the manipulation program 21 is executed, the control unit 20 functions as an imaging unit 21a and a position control unit 21b. The imaging unit 21a has a function of outputting a control signal to the microscope 12 and acquiring image data representing an image within the field of view of the microscope 12. The position control unit 21b has a function of outputting a control signal to the signal conversion interface 13 and causing the signal conversion interface 13 to output a current.

[0056] The user issues a command to start the manipulation process to the manipulation system 10 via an input unit (not shown) or the like. When the manipulation process starts, the control unit 20 accepts the setting of a target position (step S200). That is, the control unit 20 acquires the coordinates input by the user via an input unit (not shown) or the like, and regards them as the target position within the operation area 60.

[0057] Next, the control unit 20 introduces the operation object using the function of the position control unit 21b (step S205). That is, the user introduces the operation object into the first flow path 40, the second flow path 50, or the operation area 60 by performing an operation on a component (not shown). The control unit 20 controls the signal conversion interface 13 using the function of the position control unit 21b, and operates the switching unit 71 to perform processing for moving the operation object into the operation area 60. The processing for moving the operation object into the operation area 60 is, for example, processing for moving the operation object in the first flow path 40 toward the operation area 60, and the operation required by the switching unit 71 is a predetermined simple operation.

[0058] Next, the control unit 20 determines whether or not the operation object is present within the operation area using the function of the imaging unit 21a (step S210). That is, the control unit 20 controls the microscope 12 to acquire image data and extracts the operation object from the image data. The control unit 20 then determines whether or not the position of the operation object is within the operation area 60. If it is not determined that the operation object is present within the operation area, the control unit 20 repeats the processes from step S205 onwards. However, when step S205 is repeated, a new operation object is not introduced, but rather a process is performed to move the operation object relative to the operation area 60.

[0059] If it is determined in step S210 that the operation object exists within the operation area, the control unit 20 performs the processes in step S215 and subsequent steps. The state in which it is determined that the operation object exists within the operation area is, for example, the state shown in Fig. 2A. In Fig. 2A, a rectangular frame is drawn alongside the image of the operation object, and a rectangular frame is also shown at the target position, but these rectangular frames are superimposed on the image and are not actual objects that exist within the operation area 60.

[0060] In step S215, the control unit 20 performs status observation. That is, the control unit 20 controls the microscope 12 using the function of the imaging unit 21a to acquire image data, and acquires the current position (Xmt, Ymt) of the operation target object from the image data. The control unit 20 also outputs a control signal to the signal conversion interface 13 to acquire the current status (V1t to V6t) of the switching unit 71. Furthermore, the control unit 20 acquires the target position (Xg, Yg) set in step S200.

[0061] Next, the control unit 20 determines whether the operation object has reached the target position (step S220). That is, the control unit 20 determines that the operation object has reached the target position when the current position (Xmt, Ymt) of the operation object acquired in step S215 is within a predetermined distance from the target position (Xg, Yg) set in step S200. If it is determined in step S220 that the operation object has reached the target position, the control unit 20 ends the manipulation process.

[0062] If it is not determined in step S220 that the operation object has reached the target position, the control unit 20, using the function of the position control unit 21b, inputs the position (Xmt, Ymt) of the operation object, the current state (V1t to V6t) of the switching unit 71, and the target position (Xg, Yg) of the operation object to the machine learning model 30b (step S225). That is, the control unit 20 inputs the current state to the machine learning model 30b. As a result, the machine learning model 30b outputs the next action, that is, the state (V1t+1 to V6t+1) of the switching unit 71 to be set next.

[0063] Based on this output, the control unit 20 controls the switching unit 71 using the function of the position control unit 21b (step S230). That is, the output of the machine learning model 30b is a 6-bit numerical value indicating the next state of the switching unit 71 to be set. Therefore, the control unit 20 controls the signal conversion interface 13 so that current is supplied to the solenoid valves 14 in response to the numerical value 1 and current supply to the solenoid valves 14 is stopped in response to the numerical value 0. As a result, the six solenoid valves 14 adjust the air pressure, and the diaphragm 71a of the switching unit 71 operates via the pressure supply pipe 15, changing the state of the first flow path 40. As a result, the liquid in the first flow path 40 moves, and the movement of the liquid moves the operation target object in the operation area 60. As described above, in this embodiment, the air pressure adjustment unit 70, the imaging unit 21a, the position control unit 21b, the signal conversion interface 13, the solenoid valves 14, and the pressure supply pipe 15 constitute the liquid control unit.

[0064] When step S230 is executed, the control unit 20 repeats the processing from step S215 onward. According to the above processing, the control unit 20 can move the operation object to the target position by repeatedly observing the state and inputting the state to the machine learning model 30b. The machine learning model 30b inputs the current state and outputs the next state of the switching unit 71 required to move the operation object to the target position. By repeating this control of the operation object's position, the operation object reaches the target position. If the machine learning model 30b has been properly trained, this processing can be achieved by the simple processing of observing and inputting the state, and the operation object can be accurately moved to the target position. Therefore, the generally very difficult task of moving the operation object by switching the state of the switching unit 71 can be achieved very simply and accurately.

[0065] Because the first flow path 40 is provided with a total of six switching units 71, the state that can be realized by the switching units 71 is one of 64 possible states that can be expressed using 6 bits. In the machine learning model 30b, the next state is selected from the 64 possible states depending on the current state, and the switching units 71 are repeatedly switched to achieve the next state, thereby moving the operation target object to the target position within the operation area 60. In each of the 64 possible states, the following may occur in each of the first flow path 40: movement of liquid from the first flow path 40 to the operation area 60, movement of liquid in the opposite direction to that movement, movement of liquid from the first flow path 40 to the second flow path 50, and movement of liquid in the opposite direction to that movement. Immediately after a change in state of the switching units 71, movement of liquid occurs, causing a pressure fluctuation. However, if the pressure fluctuation remains until the next state change, the movement of liquid due to the next state change will differ from the originally expected movement. Therefore, in the process of repeating the state change of the switching unit 71, the direction and amount of movement of the liquid differ from the original assumption, and as a result, it becomes difficult to control the position of the operation object.

[0066] However, in this embodiment, as shown in FIG. 1B , multiple first flow paths 40 are connected to the operation area 60, and each first flow path 40 is connected to a second flow path 50. In this embodiment, the second flow paths 50 form a series of flow paths surrounding the first flow paths 40. That is, the multiple first flow paths 40 are connected to each other via the second flow paths 50. With this configuration, in this embodiment, the pressure in the first flow paths 40 is easily equalized via the second flow paths 50. Therefore, local pressure differences are unlikely to occur within the first flow paths 40 or the operation area 60. Furthermore, even if pressure fluctuates due to the movement of liquid within a specific first flow path 40, the pressure fluctuations are quickly equalized. Therefore, when repeatedly moving liquid to manipulate the position of a manipulation object, it is easy to control the movement of liquid to manipulate the manipulation object. Furthermore, in this embodiment, the switching unit 71 is provided in the first flow path 40 but not in the second flow path 50. Therefore, the operation of the switching unit 71 required to operate the position of the operation target and the equalization of the pressure in the first flow path 40 can be performed at different locations. Therefore, the operation of the switching unit 71 does not hinder the equalization of the pressure.

[0067] Furthermore, in this embodiment, the first flow path 40 is narrower than the second flow path 50. Specifically, the diameter of the circle formed by the accumulation portion 51 constituting the second flow path 50 is at least 10 times larger than the width of the first flow path 40 and at least several times larger than the width of the connecting portion 52. Furthermore, the width of the connecting portion 52 is several times larger than the width of the first flow path 40.

[0068] Therefore, the accumulation section 51 is significantly larger than the first flow paths 40 and the connecting section 52. Therefore, the accumulation section 51 can effectively absorb pressure fluctuations occurring in the first flow paths 40 and quickly equalize the pressure in the first flow paths 40. Furthermore, because the accumulation sections 51 are connected to each other by the connecting section 52, even if different pressure fluctuations occur in each of the multiple first flow paths 40, the accumulation sections 51 can quickly and effectively absorb the pressure fluctuations and quickly equalize the pressure in the first flow paths 40. In this way, the second flow path 50 functions as a highly effective buffer against pressure fluctuations and liquid movement occurring in the first flow paths 40. Therefore, the second flow path 50 quickly eliminates and equalizes pressure fluctuations in the first flow paths 40 and the operation region 60, quickly reaching an equilibrium state. Therefore, the machine learning model 30b can quickly transition to the next state output. Furthermore, it is possible to easily control the movement of liquid when transitioning to the next state.

[0069] Furthermore, the second flow path 50 is composed of a connecting portion 52 and an accumulation portion 51 having a diameter larger than the width of the connecting portion 52. The accumulation portion 51 is also connected to the first flow path 40. Therefore, the accumulation portion 51, which is larger than the first flow path 40 and the connecting portion 52, effectively functions as a buffer for quickly equalizing pressure fluctuations that occur in the first flow path 40 and the connecting portion 52. Furthermore, by connecting the accumulation portions 51 to each other by the connecting portion 52, pressure differences that occur between multiple first flow paths 40 are also quickly equalized. Note that the connecting portion 52 has a linear structure that connects the accumulation portions 51. If the connecting portion 52 is linear, the accumulation portions 51 can be connected to each other over as short a distance as possible, making it possible to efficiently equalize pressure differences between the accumulation portions 51.

[0070] Furthermore, the operation area 60 has an internal space that allows the operation object to be moved in a planar manner in a direction parallel to the largest surface of the microfluidic chip 16. Therefore, in this embodiment, the movement direction of the operation object can be assumed to be a substantially two-dimensional direction, making it easy to identify the position of the operation object and to control it. Furthermore, the shape of the internal space of the operation area 60 when viewed from a direction perpendicular to the largest surface of the microfluidic chip 16 is triangular. Therefore, the internal space has an extremely simple shape, making it easy to form the operation area 60.

[0071] Furthermore, the first flow path 40 is connected to the vertex of a triangle formed by the operation area 60. Therefore, when liquid flows into the operation area 60 from the first flow path 40, or when liquid flows out of the operation area 60 from the operation area 60 to the first flow path 40, stagnation or vortexes of the liquid are unlikely to occur within the operation area 60. On the other hand, if the first flow path 40 were connected to a side of a triangle rather than a vertex, stagnation or vortexes of the liquid would be more likely to occur within the operation area 60. Therefore, in this embodiment, it is easier to control the position of the operation target object compared to a configuration in which the first flow path 40 is connected to the side of a triangle.

[0072] Furthermore, in the above-described embodiment, two switching units 71 are provided for one first flow path 40. Therefore, it is possible to cause the liquid to flow in and out between the first flow paths 40 and the operation area 60 independently in each of the three first flow paths 40 or by combining the three first flow paths 40. This makes it possible to easily control the position of the operation object within the operation area 60.

[0073] (4) Other embodiments:

[0074] The above embodiment is merely an example for implementing the present invention, and various other embodiments are possible. For example, the shape and configuration of the second flow path 50 are not limited to those of the above embodiment. Therefore, the second flow path 50 may be formed by forming a flow path of a constant width around the operation area 60 and the first flow path 40, the flow path being similar to the shape of the operation area 60 (e.g., polygonal) but larger than the shape of the operation area 60. Furthermore, the second flow path may be formed by forming a single large accumulation portion around the operation area 60 and the first flow path 40, and various other embodiments may be adopted. Furthermore, when the second flow path 50 is formed by the accumulation portion 51 and the connecting portion 52, the accumulation portion 51 may have a shape other than circular.

[0075] Furthermore, the simulation of the position of the operation target object according to the action in step S110 may be realized in a manner that does not use the simulation model 30c. For example, simulations using other methods, such as finite element analysis, various fluid simulations, and simulations using molecular dynamics, may be used.

[0076] Furthermore, the input and output modes of the machine learning model 30b are not limited to the above examples. For example, the state of the switching unit 71 is not limited to being expressed in 6 bits, but may be expressed as a single-digit decimal number (e.g., any number from 0 to 5). Furthermore, the position of the operation target object at the t+1th time output by the simulation model 30c may be expressed not by coordinates but by a displacement vector from the current position, or the like. Furthermore, the target position may be defined by a region with a certain extent, rather than by a single point.

[0077] Furthermore, the state of one switching unit 71 may not be two states, but may be multiple states, or may change continuously. Furthermore, the machine learning model 30b may be updated during operation. For example, the simulation model 30c may be updated based on data accumulated during operation, and reinforcement learning may be performed using the updated simulation model 30c, so that the accuracy increases as operation progresses. In this case, the environment for identifying the state may be a real environment, and in this case, the coordinates of the operation target in the real environment are observed. Furthermore, the movement direction of the operation target is not limited to two-dimensional directions, and the position may be controlled in three-dimensional directions. In this case, imaging by the microscope 12 is performed from multiple different directions.

[0078] Furthermore, the reinforcement learning method is not limited to the above-described Q-learning, and methods such as SARSA may be used. Furthermore, a method in which a policy model and an action value function model are modeled separately, such as the Actor-Critic algorithm, may be used. Furthermore, the machine learning model that outputs the next state of the switching unit 71 is not limited to a configuration generated by reinforcement learning, but may be a configuration generated by a neural network or the like. The mode of machine learning is not limited to the above-described embodiment, and machine learning may be performed by appropriately selecting various factors, such as the number of layers and nodes constituting the model, the type of activation function, the type of loss function, the type of gradient descent, the type of optimization algorithm for gradient descent, the presence or absence of mini-batch learning and the number of batches, the learning rate, the discount rate, the initial value, the type or presence of an overlearning suppression method, the presence or absence of a convolutional layer, the filter size in the convolution operation, the type of filter, the type of padding and stride, the type or presence or absence of a pooling layer, the presence or absence of a fully connected layer, and the presence or absence of a recursive structure. Of course, learning may also be performed using other machine learning methods, such as a support vector machine or clustering. Furthermore, machine learning may be performed in which the structure of the model (e.g., the number of layers, the number of nodes per layer, etc.) is automatically optimized.

[0079] The manipulation system may be realized by multiple devices, or may be a system in which machine learning is performed by a server and position control is performed by a client. Furthermore, at least some of the imaging unit 21 a, position control unit 21 b, and machine learning unit 22 a may be separated and present in multiple devices. Furthermore, some of the configurations of the above-described embodiments may be omitted, and the order of processing may be changed or omitted.

[0080] The operation area may be any area capable of containing a liquid and a manipulation object whose position within the liquid is manipulated. That is, the operation area may be a liquid accumulation area, and the position of the manipulation object may be variable within the operation area. Because the position of the manipulation object is manipulated by being carried by the liquid, liquid is present at least in the target position of the manipulation object or in the path leading to the target position. Because the operation area contains movable liquid, it has a three-dimensional structure capable of accumulating liquid, but the actual movement direction of the manipulation object may or may not be three-dimensional. That is, the manipulation object may move planarly along a predetermined plane, as in the above-described embodiment, or may move one-dimensionally.

[0081] The liquid is not particularly limited, and examples thereof include water, physiological saline, blood, serum, phosphate buffer, Good's buffer such as HEPES, oil, magnetic fluid, polymer solution, photocurable resin, etc. The liquid may also be a combination of these. The water is not particularly limited, and can be appropriately selected depending on the purpose, and examples thereof include industrial water, tap water, ion-exchanged water, ultrafiltered water, reverse osmosis water, distilled water (pure water), and ultrapure water.

[0082] The manipulation object may be any object that can move through a liquid and can be selected appropriately. For example, any particle such as latex particles, silica particles, magnetic particles, cells, or liposomes can be used as the manipulation object. Other examples include beads, microparticles modified with receptors such as antibodies, or nanocolloids. Moving objects such as sperm or microorganisms are also possible.

[0083] The operation object may be operated so as to be placed at a predetermined position within the liquid present in the operation area, and the predetermined position may be a predetermined target position, a predetermined trajectory position, etc. In either case, it is sufficient that the position where the operation object should be placed before movement is determined as the predetermined position.

[0084] The liquid and the manipulation object only need to be contained within the manipulation area when the manipulation system manipulates the position of the manipulation object, and the liquid and the manipulation object may be located outside the manipulation area before the manipulation begins. Furthermore, the state in which the liquid and the manipulation object are contained may be any state in which they exist within the area. For example, in the case of a liquid, the manipulation area and each flow path may be filled with the liquid, or some air or the like may be contained.

[0085] The first flow paths are connected to the operation area and are multiple flow paths containing liquid. That is, at least two first flow paths containing liquid are connected to the operation area. As a result, an inflow path for liquid to the operation area and an outflow path for liquid from the operation area are secured, making it possible to operate the operation object by moving the liquid. The first flow paths are not limited to any particular shape or cross-sectional shape as long as liquid can move through them. To prevent the pressure distribution from becoming complicated or stagnation from occurring, a simple shape, such as a linear flow path, is preferable. A simple cross-sectional shape is also preferable; for example, a linear flow path with a circular or rectangular cross section and a constant diameter can be used. The number of first flow paths is not limited to three as in the above embodiment, and various numbers may be used.

[0086] The liquid control unit may be configured to move the liquid in the first flow path. Various configurations may be employed for moving the liquid in the first flow path. The configuration for moving the liquid by directly applying pressure to the liquid is not limited to the configuration for switching the first flow path between a closed state and an open state, as described above. For example, various configurations for moving the liquid in the first flow path may be employed. The configuration for applying a force to the liquid is also not limited to a configuration for changing the diaphragm using air pressure, and may be configured to apply a force to the liquid using various actuators such as piezoelectric elements. Furthermore, when applying a force to the liquid, a configuration for applying a force indirectly may be employed, such as a configuration in which a force is applied to a deformable first flow path to move the liquid in the first flow path.

[0087] The second flow path is a series of flow paths containing liquid, and may be connected to each of the multiple first flow paths. That is, the multiple first flow paths may be connected to each other via the second flow path. The second flow path may be configured so that the liquid in the second flow path and the liquid in the first flow path are connected by being connected to each of the multiple first flow paths, thereby equalizing the pressure difference between the first flow paths. Therefore, the shape of the second flow path is not limited. For example, in the above-described embodiment, three connecting portions 52 are provided, and the second flow path 50 forms a triangle. However, this configuration is not limited to this, and the connecting portions 52 may have a shape that only spans two sides of the triangle. Furthermore, the accumulation portion and the connecting portions 52 do not need to be clearly distinguishable. For example, the connecting portions 52 may be formed to connect the accumulation portions 51 with each other, with a thickness equal to the diameter of the accumulation portions 51.

[0088] The second flow paths are a series of flow paths, and the second flow paths themselves may be connected. To equalize the pressure, it is preferable that the second flow paths form a closed loop as in the above-described embodiment, but the second flow paths may also be a series of flow paths having ends. In any case, it is sufficient that the pressure in the first flow paths is equalized by connecting multiple first flow paths separately connected to the operation region to a second flow path formed as a continuous flow path.

[0089] The shape of the internal space of the operation area when viewed from a direction perpendicular to the plane on which the operation object moves may be any polygon. Therefore, the shape of the operation area is not limited to a triangle as in the above-described embodiment, but may be a rectangle or other polygon with more vertices. Furthermore, the polygonal shape may be any internal space, i.e., a space in which the operation object can move. Therefore, when the outer shape of the operation area and the internal space have different shapes, the internal space may have a polygonal shape. Furthermore, the internal space may be any space in which the operation object can exist within the operation area. Therefore, the interior of the operation area may be partitioned by a fence-like or lattice-like structure that allows liquid to pass through but not the operation object.

[0090] Furthermore, in the above-described embodiment, two switching units 71 are provided per first flow path 40, but the number of first flow paths 40 and switching units 71 is not limited to this example. For example, three or more switching units 71 may be provided per first flow path 40, or the number of switching units 71 provided per each of the multiple first flow paths 40 may vary for each first flow path 40. If one or more switching units 71 are provided per each first flow path 40, for a total of (the number of first flow paths 40 + 1) or more switching units 71, the inflow of liquid into and outflow of liquid from the operation area 60 can be achieved by combining the states of the switching units 71. By configuring a total of (the number of first flow paths 40 + 1) switching units 71, it is possible to control the position of the operation target within the operation area 60 with a small number of units. Furthermore, by configuring a total of (the number of first flow paths 40 + 1) or more switching units 71, more precise position control can be achieved as the number of switching units 71 increases.

[0091] Furthermore, there may be multiple target positions, and there may be multiple operation objects. FIG. 8A shows an image captured by the microscope 12 in an example where there are two operation objects and one target position. Even in this case, by introducing two operation objects into the operation area 60, setting one target position, and generating the machine learning model 30b through machine learning, it becomes possible to manipulate the position of the operation object and control it to reach the target position. In FIG. 8A, the image on the left shows the initial state immediately after the operation object is introduced, and the image in the center shows the state in the middle of moving to the target position. Furthermore, the image on the right shows the state where the operation object has reached the target position. In this way, it has been confirmed that it is possible to move two operation objects to one target position through machine learning.

[0092] 8B and 8C show images captured by the microscope 12 in an example in which there are two control objects and two target positions. Even in this configuration, if two control objects are introduced into the operation area 60, two target positions are set, and the machine learning model 30b is generated by machine learning, it becomes possible to manipulate the position of the control object and control it to reach the target position. In FIGS. 8B and 8C, the left image shows the initial state immediately after the control object is introduced, while the center image shows the state in the middle of moving to the target position. Furthermore, the right image shows the state in which the control object has reached the target position. As shown in these figures, it has been confirmed that machine learning can move two control objects to two target positions. Furthermore, as shown in FIGS. 8A to 8C, various combinations of the number of control objects, the number of target positions, and the initial positions may be adopted.

[0093] Fig. 9 is a diagram showing an example of a configuration for introducing a manipulation object into the first flow path 40 or the manipulation region 60. In the example shown in Fig. 9, the configuration for introducing a manipulation object includes a third flow path 80, a second accumulation unit 81, a second switching unit 82, a second gas passage 83, and a second connection unit 84. In the example shown in Fig. 9, three sets of the third flow path 80, the second accumulation unit 81, the second switching unit 82, the second gas passage 83, and the second connection unit 84 are formed, but for simplicity of the drawing, only one set is shown with a reference numeral.

[0094] The third flow path 80 is connected to a side of the triangle formed by the operation area 60 and extends in the extension direction of the first flow path 40. The third flow path 80 contains the same liquid as the liquid in the operation area 60 and may have the same depth, width, etc. as the first flow path 40. The second accumulation section 81 is connected to one of the two ends of the third flow path 80 that is not connected to the operation area 60. The diameter of the second accumulation section 81 is several times larger than the width of the third flow path 80 and contains the same liquid as the liquid in the operation area 60, in which the operation object is accumulated.

[0095] The second switching unit 82, the second gas passage 83, and the second connecting unit 84 have the same configuration as the switching unit 71, the gas passage 72, and the connecting unit 73. However, their installation locations are different. That is, the second switching unit 82 is formed directly below the third flow path 80, and the second gas passage 83 is a passage connecting the location directly below the second switching unit 82 with the second connecting unit 84. The second connecting unit 84 is connected to the solenoid valves 14. That is, in the example shown in FIG. 9 , there are a total of nine solenoid valves 14. The object to be operated is introduced by selecting one of the nine solenoid valves 14.

[0096] That is, the solenoid valve 14 switches the internal pressure of the second gas passage 83 and the gas passage 72 between a high state and a low state. As a result, the third flow path 80 and the first flow path 40 are switched between a closed state and an open state. By using this state switching, it is possible to move the manipulation object stored in one of the three second storage units 81 to the manipulation area 60. For example, only one of the three third flow paths 80 is opened and the other two third flow paths 80 are closed. In this state, one of the switching units 71 is operated to form a liquid flow from the manipulation area 60 to the first flow path 40, thereby introducing the manipulation object stored in the second storage unit 81 into the manipulation area 60. The above configuration makes it possible to easily introduce the manipulation object into the manipulation area 60.

[0097] In the configuration shown in FIG. 9 , the third flow path 80, the second storage unit 81, and the like are located more inward than the second flow path 50, allowing manipulation objects accumulated near the manipulation area 60 to be introduced into the manipulation area 60. This shortens the distance the manipulation object must travel for introduction, making it easy to introduce the manipulation object into the manipulation area 60. Furthermore, in the configuration shown in FIG. 9 , multiple second storage units 81 are provided. Therefore, different manipulation objects can be stored in different second storage units 81, allowing for switching between the manipulation objects to be introduced. Specifically, one second storage unit 81 can store manipulation objects for machine learning, such as beads, and another second storage unit 81 can store manipulation objects for research, such as cells. A second storage unit 81 can be selected depending on the situation, such as machine learning or research, and the desired manipulation object can be introduced into the manipulation area 60.

[0098] Furthermore, the operation area is not limited to one location, and a configuration may be adopted in which a placement operation of an operation target object can be performed using multiple operation areas. Figures 10 and 11 are diagrams for explaining a manipulation system having multiple operation areas. Figure 10 is a diagram showing a substrate 100 for realizing the manipulation system, and Figure 11 is an enlarged view of the periphery of the operation area provided on the substrate 100.

[0099] The manipulation system is realized by a substrate 100 having a plurality of units Ut. In FIG. 10 , the boundaries of the units Ut are indicated by dashed lines. A plurality of solenoid valves 140 are connected to the substrate 100. The solenoid valves 140 are connected to a signal conversion interface (not shown). The signal conversion interface can be realized by a circuit having the same function as the signal conversion interface shown in FIG. 1. As in the embodiment shown in FIG. 1 , a computer (not shown) is connected to the signal conversion interface, and a microscope (not shown) is connected to the computer. These computers and microscopes may have the same functions as the computer 11 and microscope 12.

[0100] That is, a machine learning model that has been trained in advance on an operation area or the like provided on the substrate 100 is recorded in a storage medium (not shown) of the computer. The computer also acquires an image of the operation area (described later) of each unit Ut using a microscope, outputs a signal to the signal conversion interface based on the image, and controls the solenoid valve 140 to perform a placement operation of the operation target within the operation area.

[0101] A compressor (not shown) and a pressure supply pipe 150 are connected to the solenoid valve 140, and the connection between the compressor and the pressure supply pipe 150 is switched by opening and closing the solenoid valve 140. Therefore, the solenoid valve 140 can switch the internal pressure of the pressure supply pipe 150 between a high state and a low state.

[0102] The pressure supply pipe 150 is connected to an air pressure adjustment unit 700 (described later) formed on the substrate 100. Therefore, the solenoid valve 140 can switch the internal pressure of the air pressure adjustment unit 700, to which the pressure supply pipe 150 is connected, between a high state and a low state.

[0103] Each of the multiple units Ut provided on the substrate 100 is provided with a liquid-containing structure that contains a liquid and a structure that contains a gas. Specifically, each unit Ut includes a first flow path 400, a second flow path 500, an operation area 600, and an air pressure adjustment unit 700. In this embodiment, the liquid is water and the gas is air. Note that, although the manipulation object is introduced into the first flow path 400 or the operation area 600, the configuration for introducing the manipulation object is omitted in FIGS. 10 and 11 .

[0104] The first flow path 400, the second flow path 500, and the operation area 600 are flow paths and areas containing a liquid. The operation area 600 contains a liquid, is an area in which an operation object is placed and operated, and has an internal space for moving the operation object in a planar manner. In this embodiment, the operation area 600 and the like are formed on the largest surface of the substrate 100, and the internal space of the operation area 600 has a shape that allows the operation object to be moved in a planar manner in a direction parallel to the largest surface of the substrate 100. That is, the operation area 600 has an internal space that extends in a direction parallel to the surface to a certain depth (e.g., a depth of 30 μm or less) on the largest surface of the substrate 100. Furthermore, the internal space of the operation area 600 has a rectangular shape when viewed from a direction perpendicular to the largest surface of the substrate 100.

[0105] The first flow path 400 is a linear hollow space, and one end thereof is connected to a vertex of the rectangle formed by the operation area 600. In this embodiment, the first flow path 400 is a flow path extending in a direction parallel to a diagonal of the rectangle formed by the operation area 600. The other end of the first flow path 400 is connected to a flow path 510 that is part of the second flow path 500. In this embodiment, the depth of the first flow path 400 is the same as the depth of the operation area 600, and the width of the first flow path 400 is smaller than the width of the second flow path 500.

[0106] The second flow path 500 includes a flow path 510, a flow path 520, a flow path 530, and an accumulation section 540. The flow path 510 is a square frame-shaped flow path formed around the operation region 600. The first flow path 400 is connected to the inner periphery of the vertex of the square of the flow path 510. The flow paths 530 are flow paths that extend in the direction in which the units Ut are arranged, and are flow paths formed at two locations on either side of the flow path 510.

[0107] The flow path 510 is a series of flow paths that connect the four first flow paths 400, and the flow path 520 is a flow path that connects the flow path 510 and the flow path 530. The depths of the flow paths 510, 520, and 530 are the same as that of the first flow path 400. The widths of the flow paths 510, 520, and 530 are greater than that of the first flow path 400.

[0108] The accumulation portion 540 is a circular portion and is connected to an end of the flow path 530. The depth of the accumulation portion 540 is deeper (for example, 3 mm to 5 mm) than the first flow path 400. The diameter of the circle formed by the accumulation portion 540 is greater than the width of the flow path 530.

[0109] The air pressure adjusting unit 700 is a structure including a passage containing gas. In this embodiment, the air pressure adjusting unit 700 includes switching units 710 and 711 (see FIG. 11 ) and gas passages 720 and 721. Of course, a portion such as the connecting unit 73 shown in FIG. 1B may be provided at the connection between the pressure supply pipe 150 and the gas passage 720.

[0110] The switching unit 710 includes a mechanism capable of switching the first flow path 400 between a closed state and an open state. In this embodiment, one or more switching units 710 are provided for each of the four first flow paths 400. The total number of switching units 710 is preferably the number of first flow paths 400 plus one. In the example shown in FIGS. 10 and 11, two switching units 710 are provided for the first flow path 400 located in the upper left of the operation area 600, and one switching unit 710 is provided for each of the other first flow paths 400. Therefore, the total number of switching units 710 is five. The switching units 710 and the first flow paths 400 also have the structures shown in FIGS. 2C to 2E, and the internal pressure of the first flow path 400 can be adjusted by the switching unit 710.

[0111] The gas passages 720 are gas passages extending from each of the switching units 710, and in each unit Ut, they extend from the switching unit 710 side to the solenoid valve 140 side and further extend in the direction in which the units Ut are lined up, connecting the gas passages 720 of the units Ut to each other. In this embodiment, among the switching units 710 present in each unit Ut, the gas passages 720 extending from the switching units 710 that have the same relative positional relationship with respect to the operation area 600 are connected to each other. For example, in Figures 10 and 11, the gas passages 720 extending from the switching units 710 present in the upper right of the operation area 600 are connected to each other.

[0112] In the substrate 100, one pressure supply pipe 150 extending from one solenoid valve 140 is connected to one gas passage 720. Therefore, when the internal pressure of the pressure supply pipe 150 becomes high or low under the control of the solenoid valve 140, the internal pressure of the gas passage 720 also becomes high or low. As a result, it is possible to switch between a state in which liquid cannot flow within the first flow path 400 and a state in which liquid can flow within the first flow path 400.

[0113] As described above, in this embodiment, the air pressure adjusting unit 700 constitutes a structure that applies pressure to the liquid in the first flow path to move the liquid. Furthermore, the solenoid valve 140 and the pressure supply pipe 150 constitute a pressure supply unit that supplies pressure to the structure provided in the multiple units Ut.

[0114] In this embodiment, a switching unit 711 is also formed in the flow path 520, and a gas path 721 is connected to the switching unit 711. The structures of the switching unit 711 and the gas path 721 are similar to the structures of the switching unit 710 and the gas path 720. Therefore, when the internal pressure of the pressure supply pipe 150 becomes high or low under the control of the solenoid valve 140, the internal pressure of the gas path 721 also becomes high or low. As a result, it is possible to switch between a state in which liquid cannot flow through the flow path 520 and a state in which liquid can flow through the flow path 520.

[0115] If the flow path 520 is in a state where liquid cannot flow, the flow of liquid between the flow path 510 and the flow path 530 is blocked, so that the pressure in the flow path 510 provided in each unit Ut can be prevented from being affected by pressure fluctuations due to operations, etc. in the other units Ut. If the flow path 520 is in a state where liquid can flow, the flow of liquid can be allowed between the flow path 510 and the flow path 530, so that the pressure in each of the flow paths 510 provided in the multiple units Ut can be made uniform.

[0116] In this embodiment, the substrate 100 further includes supply control units 910a to 910d (see FIG. 10 ) that switch between a state in which pressure can be supplied to the air pressure adjustment unit 700 and a state in which pressure cannot be supplied to the air pressure adjustment unit 700 for each of the multiple units Ut. That is, the supply control units 910a to 910d can switch between a state in which gas can flow through the intersecting gas passages 720 and 721 and a state in which gas cannot flow through the intersecting gas passages 720 and 721.

[0117] The supply control units 910a to 910d can be realized by various configurations, and for example, they may apply pressure to a diaphragm or the like as shown in Figures 2C to 2E using a solenoid valve and a pressure supply pipe to block or open the gas passages 720 and 721. Alternatively, the gas passages 720 and 721 may be directly blocked or opened by a solenoid valve.

[0118] Each of the supply control units 910a to 910d simultaneously switches whether to supply pressure to multiple intersecting gas passages 720, 721. In the example shown in FIG. 10 , each of the supply control units 910a to 910d simultaneously switches whether to supply pressure to 14 intersecting gas passages. For example, the supply control unit 910a simultaneously switches whether to supply pressure to the five gas passages 720 and two gas passages 721 in the unit Ut shown in the top row in FIG. 10 and the five gas passages 720 and two gas passages 721 in the unit Ut shown in the third row from the top. Meanwhile, the supply control unit 910c simultaneously switches whether to supply pressure to the five gas passages 720 and two gas passages 721 in the unit Ut shown in the top row in FIG. 10 and the five gas passages 720 and two gas passages 721 in the unit Ut shown in the second row from the top.

[0119] The supply control unit 910 is connected to a signal conversion interface (not shown), and the computer controls the operation of the supply control units 910a to 910d via the signal conversion interface. That is, the computer can select any unit Ut and enable or disable the adjustment of air pressure by the solenoid valve 140 by switching the combination of blocking or opening the gas passages by the supply control units 910a to 910d.

[0120] For example, assume that the supply control units 910a and 910c disable pressure supply and the supply control units 910b and 910d enable pressure supply. In this case, the supply control unit 910a disables pressure supply to the top and third-highest units Ut shown in FIG. 10 and disables them. The supply control unit 910c disables pressure supply to the second-highest unit Ut and disables it. Meanwhile, the supply control units 910b and 910d enable pressure supply to the bottom-lowest unit Ut shown in FIG. 10 and enable it.

[0121] Pressure can be supplied to the activated unit Ut by the solenoid valve 140, and switching is performed by the switching units 710, 711 via the gas passages 720, 721. That is, the pressure on the liquid in the first flow path 400 is adjusted and the flow path 520 is opened and closed. With the above configuration, in this embodiment, an arbitrary unit Ut can be selected under computer control, and an operation to place an operation target object can be performed within the operation area 600 of the selected unit Ut.

[0122] Furthermore, in a configuration in which the position of an object to be manipulated within an operation area is controlled by the movement of liquid from multiple first flow paths, a technique in which a second flow path is connected to each of the first flow paths can also be realized as a method invention. Furthermore, the above-described system and method can be realized as a single device or multiple devices, and include various aspects. For example, it is possible to provide a microscope equipped with the above-described means, or an attachment for mounting on an existing microscope. Furthermore, the invention can also be realized as a recording medium for a program executed by the control unit. Of course, the software recording medium may be a magnetic recording medium or a semiconductor memory, and any recording medium developed in the future can be considered in the same way.

[0123] 10... Manipulation system, 11... Computer, 12... Microscope, 13... Signal conversion interface, 14... Solenoid valve, 15... Pressure supply pipe, 16... Microfluidic chip, 20... Control unit, 21... Manipulation program, 21a... Photography unit, 21b... Position control unit, 22... Machine learning program, 22a... Machine learning unit, 30... Storage medium, 30b... Machine learning model, 30c... Simulation model, 40... First flow path, 50... Second flow path, 51... Storage unit, 52... Connection unit, 60... Operation area, 70... Air pressure adjustment unit, 71... Switching unit, 71a... Diaphragm, 71b... Space, 72... Gas passage, 73... Connection unit, 80... Third flow path, 81... Second storage unit, 82... Second switching unit, 83... Second gas passage, 84... Second connection unit

Claims

1. A manipulation system for arranging and manipulating an operation target object at a predetermined position in an operation area containing a liquid, comprising: The operation area has an internal space for moving the operation object in a plane; A first flow path is a plurality of flow paths connected to the operation area and containing the liquid; A liquid control unit that moves the liquid in the first flow path; a second flow path that is a series of flow paths containing the liquid and is connected to each of the first flow paths; A manipulation system comprising:

2. The shape of the internal space as viewed in a direction perpendicular to the plane is polygonal. The manipulation system of claim 1 .

3. The first flow path is connected to a vertex of the polygon. The manipulation system according to claim 2 .

4. The first flow path is narrower than the second flow path. A manipulation system according to claim 1 or 2.

5. The liquid control unit is a switching unit that switches the first flow path between a closed state and an open state to move the liquid, One or more switching units are provided for each of the first flow paths, A total of (the number of the first flow paths + 1) or more are provided. The manipulation system of claim 1 .

6. The liquid control unit is a switching unit that switches the first flow path between a closed state and an open state to move the liquid, The switching unit is Two of the first flow paths are provided for each of the first flow paths. The manipulation system of claim 1 .

7. The liquid control unit is A machine learning model that is trained by machine learning to input the position of the operation object, the state of the switching unit, and a target position of the operation object and output an operation of the switching unit required to bring the operation object closer to the target position is a position of the operation object, a state of the switching unit, and a target position of the operation object are input, and the switching unit is operated based on an obtained output; A manipulation system according to claim 5 or claim 6.

8. The second flow path is a connecting portion and a liquid accumulation portion larger than the connecting portion, The storage unit is connected to the first flow path, The connecting portion is The storage portions are formed between the plurality of storage portions so as to connect the storage portions, A manipulation system according to claim 1 or 2.

9. an operation area containing a liquid, in which an operation object is placed and operated, and having an internal space in which the operation object is moved in a plane; A first flow path is a plurality of flow paths connected to the operation area and containing the liquid; a structure that applies pressure to the liquid to move the liquid in the first flow path; a second flow path that is a series of flow paths containing the liquid and is connected to each of the first flow paths; A fluid chip comprising:

10. an operation area containing a liquid, in which an operation object is placed and operated, and having an internal space in which the operation object is moved in a plane; A first flow path is a plurality of flow paths connected to the operation area and containing the liquid; a structure that applies pressure to the liquid to move the liquid in the first flow path; a plurality of units each having a series of flow paths containing the liquid, and a second flow path connected to each of the plurality of first flow paths; The second flow paths of the plurality of units are connected to each other, a pressure supply unit that supplies the pressure to the structures provided in the plurality of units; a supply control unit that switches between a state in which the pressure can be supplied from the pressure supply unit to the structure and a state in which the pressure cannot be supplied, for each of the plurality of units; A manipulation system comprising: