Processing method and processing device

Abstract

Provided is a processing method for performing more accurate processing on target coordinates. More specifically, the processing method involves performing processing on a target coordinate Q. The processing method includes: a learning step in which experimental data D relating to an instruction value qa for specifying a target coordinate Q and a measured value qb of the target coordinate Q measured at the time of the instruction value qa is used as training data to train a machine learning model M; and an execution step for executing the trained machine learning model M. In the learning step, the relationship m between the instruction value qa and the measured value qb in the experimental data D is learned. In the execution step, the instruction value qa is input to the trained machine learning model M, and the machine learning model M outputs a corrected instruction value qa' obtained by correcting the instruction value qa on the basis of the relationship m.

Description

Processing method and processing apparatus

[0001] The present disclosure relates to a processing method and a processing apparatus.

[0002] The mounting method disclosed in Patent Document 1 includes: a first transfer step of transferring a plurality of semiconductor chips formed on a carrier substrate to a first transfer substrate; an inspection step of inspecting the state of the semiconductor chips transferred onto the first transfer substrate; a second transfer step of transferring only the semiconductor chips judged as normal through the inspection step from the first transfer substrate to a second transfer substrate; and a mounting step of mounting the semiconductor chips transferred onto the second transfer substrate onto a circuit substrate.

[0003] Japanese Patent Application Laid-Open No. 2021-150614

[0004] In this type of mounting method, chips are mounted at specific coordinates on a substrate. When coordinates are specified, chips are mounted at the specified coordinates on the substrate.

[0005] However, chips may not be accurately mounted at the specified coordinates on the substrate. Such mounting defects lead to defective products and reduced yield, and are therefore undesirable.

[0006] This problem also applies to other processing methods that perform processing other than mounting on target coordinates of a substrate or an article other than a substrate.

[0007] An object of the present disclosure is to provide a processing method for performing processing more accurately on target coordinates.

[0008] The processing method according to the present disclosure is a processing method for performing processing on target coordinates, and includes: a learning step of causing a machine learning model to learn, as teacher data, experimental data related to an indicated value for indicating the target coordinates and a measured value of the target coordinates measured at the indicated value; and an execution step of executing the machine learning model after learning. In the learning step, a relationship between the indicated value and the measured value in the experimental data is learned, and in the execution step, by inputting the indicated value into the machine learning model after learning, a corrected indicated value obtained by correcting the indicated value based on the relationship is output from the machine learning model.

[0009] In the learning phase, the machine learning model is trained using the indicated values ​​and measured values ​​related to the target coordinates in the experimental data as training data, thereby learning the relationship between indicated values ​​and measured values ​​in the experimental data. In the execution phase, when the indicated values ​​are input to the trained machine learning model, the machine learning model outputs corrected indicated values, which are adjusted based on the relationship between indicated values ​​and measured values ​​in the experimental data. Processing can then be performed on the target coordinates based on the corrected indicated values.

[0010] This provides a processing method for performing more accurate operations on target coordinates.

[0011] In one embodiment, the processing method includes a preprocessing step performed before the learning step, in which the linear component in which the relationship between the indicated value and the measured value is linear is removed.

[0012] When the relationship between the indicated value and the measured value is linear, the error between the indicated value and the measured value can be determined by other methods without using a machine learning model. Machine learning models can be avoided for simple processes such as suppressing errors in the linear component.

[0013] In one embodiment, the processing method involves irradiating the target coordinates with a laser through a lens, and in the pre-processing step, the linear component and the known component related to the lens, for which the relationship between the indicated value and the measured value is known, are removed.

[0014] If the relationship between the indicated value and the measured value is known, the error between the indicated value and the measured value can be determined by other methods without using a machine learning model. Machine learning models can be avoided for simple processes such as suppressing errors in the linear and known components mentioned above.

[0015] The processing apparatus according to this disclosure is a processing apparatus that performs processing on target coordinates, comprising: a learning unit that trains a machine learning model using experimental data relating to an instruction value for indicating the target coordinates and a measured value of the target coordinates measured at the time of the instruction value as training data; an execution unit that executes the machine learning model after training; and a control unit that controls the processing on the target coordinates, wherein the learning unit learns the relationship between the instruction value and the measured value in the experimental data; the execution unit inputs the instruction value to the machine learning model after training, outputs a corrected instruction value from the machine learning model that corrects the instruction value based on the relationship, instructs the control unit to use the outputted corrected instruction value; and the control unit performs processing on the target coordinates based on the instructed corrected instruction value.

[0016] This allows us to provide a processing device that performs more accurate processing on target coordinates.

[0017] This disclosure provides a processing method for performing more accurate processing on target coordinates.

[0018] Figure 1 shows the laser irradiation device. Figure 2 shows the discrepancy between the indicated value and the measured value related to the target coordinates. Figure 3 shows the nonlinear distortion in the Fθ lens. Figure 4 shows the learning and execution using a machine learning model. Figure 5 shows the preprocessing.

[0019] Embodiments of the present disclosure will be described in detail below with reference to the drawings. The following description of preferred embodiments is illustrative in nature and is not intended to limit the present disclosure, its applications, or its uses in any way.

[0020] Figure 1 shows a laser irradiation device 1. The laser irradiation device 1 is an example of a processing device. The laser irradiation device 1 performs processing on a target coordinate Q on the workpiece W. Specifically, the laser irradiation device 1 performs laser irradiation processing on a target coordinate Q on the workpiece W. The workpiece W is, for example, a substrate.

[0021] The target coordinate Q is composed of the x and y coordinates in the XY plane of the Cartesian coordinate system. In the Cartesian coordinate system, the X, Y, and Z axes are orthogonal to each other.

[0022] The laser irradiation device 1 comprises a laser light source 10, a first galvanometer mirror 21, a second galvanometer mirror 22, an Fθ lens 23, a control device 30, and a learning execution device 40.

[0023] The laser light source 10 is, for example, a laser oscillator. The laser light source 10 emits a laser L. Between the laser light source 10 and the workpiece W, a first galvanometer mirror 21, a second galvanometer mirror 22, and an Fθ lens 23 are interposed in order from the upstream side (laser light source 10 side) to the downstream side (workpiece W side). The Fθ lens 23 is an example of a lens.

[0024] The laser L emitted from the laser light source 10 is reflected by the first galvanometer mirror 21, then reflected again by the second galvanometer mirror 22, then passes through the Fθ lens 23, and finally irradiates the target coordinate Q on the workpiece W. The laser irradiation device 1 irradiates the target coordinate Q on the workpiece W with the laser L by passing through the Fθ lens 23.

[0025] The first galvanometer mirror 21 and the second galvanometer mirror 22 adjust the incident angle of the laser L to the Fθ lens 23. The first galvanometer mirror 21 is tilted around the Z axis by an actuator (not shown). The first galvanometer mirror 21 adjusts the incident angle of the laser L to the Fθ lens 23 in the X axis direction. The second galvanometer mirror 22 is tilted around the X axis by an actuator (not shown). The second galvanometer mirror 22 adjusts the incident angle of the laser L to the Fθ lens 23 in the Y axis direction.

[0026] The irradiation position (irradiation coordinate) of the laser L on the workpiece W is proportional to the incident angle of the laser L with respect to the Fθ lens 23. By adjusting the incident angle of the laser L with respect to the Fθ lens 23 using the first galvanometer mirror 21 and the second galvanometer mirror 22, the irradiation position (irradiation coordinate) of the laser L on the workpiece W can be adjusted.

[0027] The control device 30 is composed of, for example, a computer. The control device 30 has a processor mounted on a circuit board and a memory that stores software for operating the processor.

[0028] The control device 30 of the laser irradiation device 1 includes a control unit 31. The control unit 31 controls the processing of the target coordinate Q on the workpiece W. Specifically, the control unit 31 controls the irradiation process of the laser L on the target coordinate Q on the workpiece W.

[0029] The control unit 31 is connected to the laser light source 10, the first galvanometer mirror 21 (and its actuator), and the second galvanometer mirror 22 (and its actuator) by wire or wireless means. The control unit 31 controls the emission of the laser L from the laser light source 10, the tilt of the first galvanometer mirror 21, and the tilt of the second galvanometer mirror 22.

[0030] The control unit 31 controls the irradiation position (irradiation coordinates) of the laser L on the workpiece W by controlling the tilt of the first galvanometer mirror 21 and the tilt of the second galvanometer mirror 22.

[0031] The user inputs an instruction value qa related to the target coordinate Q to the control unit 31. The instruction value qa consists of an instruction value xa related to the coordinate x in the XY plane of the Cartesian coordinate system and an instruction value ya related to the coordinate y.

[0032] The control unit 31 controls the irradiation position (irradiation coordinate, processing position) of the laser L on the workpiece W by controlling the tilt of the first galvanometer mirror 21 and the tilt of the second galvanometer mirror 22 based on the input instruction value qa.

[0033] The learning execution device 40 will be described later.

[0034] (Discrepancy between indicated value and measured value related to target coordinates) The discrepancy between the indicated value qa and the measured value qb related to the target coordinate Q will be explained. Figure 2 shows the discrepancy between the indicated value qa and the measured value qb. Figure 3 shows the nonlinear distortion in an Fθ lens.

[0035] The measured value qb is the actual measured value related to the target coordinate Q. The measured value qb consists of the measured value xb related to the coordinate x in the XY plane of the Cartesian coordinate system and the measured value yb related to the coordinate y. In this example, the measured value qb is obtained based on the image captured by the camera.

[0036] As shown in Figure 2, the measured value qb should ideally coincide with the indicated value qa, but in reality, it deviates from the indicated value qa. In Figure 2, the starting point of the arrow indicates the indicated value qa, and the ending point of the arrow indicates the measured value qb.

[0037] There are various reasons for the discrepancy between the two. In particular, (1) to (3) below have a significant impact on the discrepancy between the indicated value qa and the measured value qb.

[0038] (1) Linear distortion caused by non-horizontal and non-orthogonal movements in the X-axis guide and Y-axis guide, which are equipped with optical systems such as the first galvanometer mirror 21 and the second galvanometer mirror 22.

[0039] (2) Distortion of the Fθ lens 23 itself.

[0040] (3) Location-dependent nonlinear distortion caused by uneven polishing or refractive index in the Fθ lens 23.

[0041] Regarding (1). The optical system, including the first galvanometer mirror 21 and the second galvanometer mirror 22, is mounted on an X-axis guide and a Y-axis guide (not shown). The X-axis guide guides the movement or rotation of the optical system in the X-axis direction. The Y-axis guide guides the movement or rotation of the optical system in the Y-axis direction. The X-axis guide and the Y-axis guide are calibrated to be horizontal and orthogonal to each other. However, the X-axis guide and the Y-axis guide inevitably become non-horizontal due to deviation from horizontal, or non-orthogonal due to deviation from mutual orthogonality. When the X-axis guide and the Y-axis guide become non-horizontal and non-orthogonal, the indicated value qa and the measured value qb will be out of sync. The deviation between the indicated value qa and the measured value qb in this case has a linear relationship (linear distortion), and can be specifically represented by the projection transformation matrix described later.

[0042] Regarding (2), the indicated value qa and the measured value qb are misaligned due to the distortion of the Fθ lens 23 itself. This misalignment between the indicated value qa and the measured value qb is not linear (nonlinear distortion) and therefore cannot be expressed using the projection transformation matrix described later, but it can be expressed using the equation for Brown's lens, a known model.

[0043] Regarding (3). As shown in Fig. 3, the indicated value qa and the measured value qb deviate from each other due to causes such as polishing unevenness and refractive index unevenness in the Fθ lens 23. These unevenness vary depending on the location in the Fθ lens 23 and are location-dependent. In Fig. 2, the solid line indicates the laser L whose traveling direction is deviated due to unevenness, and the two-dot chain line indicates the laser L in the ideal traveling direction assuming that there is no unevenness. Since the deviation between the indicated value qa and the measured value qb at this time does not have a linear relationship (non-linear distortion), it cannot be represented not only by the projection transformation matrix described later, but also cannot be represented by a known model.

[0044] (Learning Execution Apparatus) The learning execution apparatus 40 is constituted by, for example, a computer. The learning execution apparatus 40 includes a processor mounted on a base board and a memory storing software for operating the processor. The learning execution apparatus 40 may be integrated with the control apparatus 30, or may be independent of the control apparatus 30. The learning execution apparatus 40 may be an external server.

[0045] The learning execution apparatus 40 of the laser irradiation apparatus 1 includes a preprocessing unit 41, a learning unit 42, and an execution unit 43. The preprocessing unit 41, the learning unit 42, and the execution unit 43 may be integrated with each other, or may be independent of each other.

[0046] Fig. 4 shows learning and execution using the machine learning model M. The machine learning model M is a supervised learning model. Specific examples of the machine learning model M include ridge regression, GDBT (Gradient Boosting Decision Tree), multilayer perceptron (MLP), and the like. The machine learning model M is stored, for example, in the memory of the learning execution apparatus 40.

[0047] The learning unit 42 performs a learning process. The learning unit 42 causes the machine learning model M to learn the experimental data D related to the indicated value qa and the measured value qb at the target coordinates Q as teacher data. The indicated value qa is for indicating the target coordinates Q.

[0048] The indicated value qa may be a value input by the user to the control unit 31. The measured value qb is measured when the indicated value is qa. For example, the measured value qb is actually measured by a sensor (e.g., a camera) when the user inputs the indicated value qa to the control unit 31.

[0049] The experimental data D is composed of sets of indicated values qa and measured values qb that correspond to each other. There are a plurality of pieces of experimental data D. It is preferable that there are a large number of pieces of experimental data D.

[0050] The learning unit 42 learns the relationship m between the indicated value qa and the measured value qb in the experimental data D. The relationship m is, for example, an error G between the indicated value qa and the measured value qb. The relationship m includes a linear component m1 according to (1), a known component m2 according to (2), and a nonlinear component m3 according to (3). The relationship m is complex.

[0051] The execution unit 43 performs an execution step. The execution unit 43 executes the machine learning model M after learning. By inputting the indicated value qa into the machine learning model M after learning, the execution unit 43 outputs a corrected indicated value qa' obtained by correcting the indicated value qa based on the relationship m from the machine learning model M, and instructs the corrected indicated value qa' to the control unit 31 of the control device 30.

[0052] The corrected indicated value qa' is composed of a corrected indicated value xa' related to the coordinate x and a corrected indicated value ya' related to the coordinate y on the XY plane of an orthogonal coordinate system.

[0053] For example, the corrected indicated value qa' deviates from the indicated value qa by an amount corresponding to the error G between the indicated value qa and the measured value qb. Due to the nonlinear component m3 described later, the error G is not fixed to a constant value, nor is it derived by a linear model based on the indicated value qa and the measured value qb, and it is complex, for example, it takes different values depending on the location in the Fθ lens 23.

[0054] The execution unit 43 inputs the correct value (position) related to the target coordinate Q as an instruction value qa to the machine learning model M, and then, based on the error G, calculates inversely and outputs a corrected instruction value qa' from the machine learning model M. The execution unit 43 then instructs the control unit 31 to use the outputted corrected instruction value qa'. Alternatively, the user may also instruct the control unit 31 to use the corrected instruction value qa' output by the execution unit 43.

[0055] The control unit 31 of the control device 30 performs processing on the target coordinate Q in the workpiece W based on the instructed correction instruction value qa'. The control unit 31 irradiates the target coordinate Q in the workpiece W with the laser L based on the instructed correction instruction value qa'. The control unit 31 controls the irradiation position (irradiation coordinate, processing position) of the laser L in the workpiece W by controlling the tilt of the first galvanometer mirror 21 and the tilt of the second galvanometer mirror 22 based on the instructed correction instruction value qa'.

[0056] (Preprocessing) Figure 5 shows the preprocessing. The preprocessing unit 41 performs the preprocessing steps. The preprocessing steps are performed before the learning steps.

[0057] As described above, the relationship m between the indicated value qa and the measured value qb in experimental data D includes a linear component m1 related to (1), a known component m2 related to (2), and a nonlinear component m3 related to (3).

[0058] The linear component m1 is the relationship m between the indicated value qa and the measured value qb, which is linear. The linear component m1 relates to non-horizontal and non-orthogonal aspects in the X-axis guide and Y-axis guide. The linear component m1 is represented by a projection transformation matrix (homography transformation matrix). The linear component m1 is expressed by equation [Equation 1].

[0059]

[0060] The nine unknowns a-h and s can be obtained by substituting nine or more sets of experimental data D (sets of indicated value qa and measured value qb) into a system of equations and solving them. Specific numerical values ​​are given for a-h and s. Specifically, equation [Equation 1] contains "noise," and after substituting nine or more sets of experimental data D (sets of indicated value qa and measured value qb) into equation [Equation 1], the nine unknowns a-h and s are calculated using the least squares method to minimize the "noise."

[0061] The known component m2 is related to the Fθ lens 23, where the relationship m between the indicated value qa and the measured value qb is known. The known component m2 is expressed by the Brown lens equation, which is known. The known component m2 is represented by equation [Equation 2].

[0062]

[0063] a1, a2 and K1 to K4 are known and determined by the specifications of the Fθ lens 23.

[0064] The preprocessing unit 41 removes the linear component m1 and the known component m2. The preprocessing does not change the indicated value qa. The preprocessing changes (corrects) the measured value qb, taking into account the removal of the linear component m1 and the known component m2, and treats this as a new measured value qb (for training the machine learning model M).

[0065] The only remaining relation m is the nonlinear component m3. The nonlinear component m3 is related to the nonlinear relationship m between the indicated value qa and the measured value qb. The nonlinear component m3 relates to location-dependent components such as polishing irregularities and refractive index irregularities in the Fθ lens 23. The nonlinear component m3 cannot be expressed by linear models such as projection transformation matrices or known models such as the equation for Brown's lens. The nonlinear component m3 is complex. The nonlinear component m3 can be expressed, for example, as a function f(x, y).

[0066] Therefore, in the subsequent learning process, the learning unit 42 causes the machine learning model M to learn the nonlinear component m3.

[0067] The processing method according to this embodiment performs processing (laser irradiation processing) on ​​a target coordinate Q in the workpiece W. This processing method comprises a learning step in which experimental data D of an instruction value qa for indicating the target coordinate Q in the workpiece W and the measured value qb of the target coordinate Q measured at the time of instruction value qa is used as training data to train a machine learning model M; an execution step in which the trained machine learning model M is executed; and a control step in which the processing (laser irradiation processing) on ​​the target coordinate Q in the workpiece W is controlled.

[0068] In the learning phase, a machine learning model M is used to learn the relationship m between the command value qa and the measured value qb in the experimental data D. In the execution phase, the command value qa is input to the learned machine learning model M, and a corrected command value qa' is output from the machine learning model M, correcting the command value qa based on the relationship m. In the control phase, processing (laser irradiation processing) is performed on the target coordinate Q in the workpiece W based on the commanded corrected command value qa'.

[0069] In this processing method, a laser L is irradiated onto the target coordinate Q in the workpiece W, passing through the Fθ lens 23. This processing method includes a preprocessing step that is performed before the learning step. In the preprocessing step, a linear component m1, in which the relationship m between the indicated value qa and the measured value qb is linear, and a known component m2 related to the Fθ lens 23, in which the relationship between the indicated value qa and the measured value qb is known, are removed. The nonlinear component m3 remaining as the relationship m is then used to train the machine learning model M in the learning step.

[0070] (Effects) In the learning process, the machine learning model M is trained using the instruction value qa and measured value qb related to the target coordinate Q of the workpiece W in the experimental data D as training data. As a result, the relationship m between the instruction value qa and the measured value qb in the experimental data D is learned by the machine learning model M.

[0071] In the execution phase, the instruction value qa is input to the trained machine learning model M. Based on the relationship m between the instruction value qa and the measured value qb in the experimental data D, the machine learning model M outputs a corrected instruction value qa', which is a corrected version of the instruction value qa. In the control phase, processing (laser irradiation processing) is performed on the target coordinate Q in the workpiece W based on the instructed corrected instruction value qa'.

[0072] This invention provides a processing method (processing method, processing device) for performing more accurate processing (laser irradiation processing) on ​​the target coordinate Q in the workpiece W.

[0073] In the case of a linear component m1 where the relationship m between the indicated value qa and the measured value qb is linear, the error between the indicated value qa and the measured value qb can be determined by other methods (e.g., projection transformation matrices) without using a machine learning model M. In the case of a known component m2 where the relationship m between the indicated value qa and the measured value qb is known, the error between the indicated value qa and the measured value qb can be determined by other methods (e.g., the equation for Brown's lens) without using a machine learning model M.

[0074] The machine learning model M can be excluded from simple processing such as error suppression in the linear component m1 and the known component m2. Instead, the machine learning model M can be dedicated to more complex processing such as error suppression in the nonlinear component m3.

[0075] (Other Embodiments) Although the present disclosure has been described above with reference to preferred embodiments, this description is not limiting, and of course, various modifications, substitutions, or combinations are possible.

[0076] The processing device is not limited to a laser irradiation device. The processing device may be, for example, a mounting device. The mounting device comprises a header and a stage. The header holds the chip. The header moves while being guided by X-axis guides and Y-axis guides. A substrate is placed on the stage. The stage moves while being guided by X-axis guides and Y-axis guides (separate from those for the header). In the mounting device, the header mounts the held chip to the target coordinates on the substrate placed on the stage. In the mounting device, there may be a discrepancy between the indicated value and the measured value related to the target coordinates. Causes of the error between the two include, for example, deviation from orthogonality between the X-axis guide and the Y-axis guide (axis tilt, linear distortion), location-dependent nonlinear distortion due to deflection of the axis guides, and minute nonlinear distortion due to uneven polishing of the axis guides. It is advisable to train a machine learning model on the relationship (error) between the indicated value and the measured value to obtain a corrected indicated value.

[0077] The processing device may be, for example, an inkjet printer or a laser trimming device.

[0078] This disclosure is extremely useful and highly industrially applicable, as it can be applied to processing methods and apparatus.

[0079] 1 Laser irradiation device (processing device) 10 Laser light source 21 First galvanometer mirror 22 Second galvanometer mirror 23 Fθ lens (lens) 30 Control device 31 Control unit 40 Learning execution device 41 Preprocessing unit 42 Learning unit 43 Execution unit W Work L Laser M Machine learning model D Experimental data m Relationship G Error m1 Linear component m2 Known component m3 Nonlinear component X X axis Y Y axis Z Z axis Q Target coordinates x coordinate y coordinate qa Indicated value xa Indicated value ya Indicated value qb Measured value xb Measured value yb Measured value qa' Correction indicated value xa' Correction indicated value ya' Correction indicated value

Claims

1. A processing method for performing processing on target coordinates, comprising: a learning step of training a machine learning model using experimental data relating to an instruction value for indicating the target coordinates and a measured value of the target coordinates measured at the time of the instruction value as training data; and an execution step of executing the machine learning model after training, wherein the learning step learns the relationship between the instruction value and the measured value in the experimental data, and the execution step inputs the instruction value to the machine learning model after training, thereby outputting a corrected instruction value from the machine learning model that corrects the instruction value based on the relationship.

2. The processing method of claim 1, comprising a preprocessing step performed prior to the learning step, wherein the preprocessing step removes a linear component in which the relationship between the indicated value and the measured value is linear.

3. The processing method of claim 2, wherein a laser is irradiated onto the target coordinates through a lens, and in the pre-processing step, the linear component and the known component related to the lens, for which the relationship between the indicated value and the measured value is known, are removed.

4. A processing device for processing target coordinates, comprising: a learning unit that trains a machine learning model using experimental data relating to an instruction value for indicating the target coordinates and a measured value of the target coordinates measured at the time of the instruction value as training data; an execution unit that executes the machine learning model after training; and a control unit that controls the processing of the target coordinates, wherein the learning unit learns the relationship between the instruction value and the measured value in the experimental data; the execution unit inputs the instruction value to the machine learning model after training, outputs a corrected instruction value from the machine learning model that corrects the instruction value based on the relationship, instructs the control unit to use the outputted corrected instruction value; and the control unit processes the target coordinates based on the instructed corrected instruction value.

Images