Laser detection system

Abstract

The invention provides a laser detection system. The laser detection system comprises a laser emitting end, a condensing objective lens, a semi-transparent and semi-reflecting mirror, a sample table, a detection end, an imaging module and a regulation and control module, the laser transmitting end comprises a laser, a beam splitter, a plurality of phase shifters and a QBH array; the transflective mirror divides the flat-topped light beam into a detection sub-light beam and a monitoring sub-light beam, and projects the detection sub-light beam to the surface of a to-be-detected sample loaded on the sample table, so that the to-be-detected sample reflects the detection sub-light beam to a detection end, and projects the monitoring sub-light beam to the imaging module; the regulation and control module is electrically connected with the imaging module, the phase shifter and the QBH array, and the regulation and control module is used for regulating and controlling the phase shifter and / or the QBH array according to the imaging result of the imaging module so as to change the state of the sub laser beam and optimize the light condensing objective lens to output the flat-topped beam. According to the invention, wafer-level defect detection can be completed efficiently and accurately.

Description

Technical Field

[0001] This invention relates to the field of laser beam shaping technology, and more particularly to a laser detection system. Background Technology

[0002] As the feature size of integrated circuits continues to shrink, problems caused by defects in wafers are becoming increasingly apparent. Among these, wafer interface defects have become an important research direction in the field of wafer defect detection.

[0003] Traditional methods for detecting wafer interface defects typically involve electrical detection methods or deep-level transient spectroscopy (DLTS). However, electrical detection methods have certain limitations. For example, capacitance-voltage (CV) methods require sample preparation and are destructive to the wafer; while DLTS is sensitive, its temperature cycling testing process is cumbersome and inefficient.

[0004] Therefore, how to quickly and non-destructively perform wafer-level defect detection has become an urgent problem to be solved. Summary of the Invention

[0005] To address the aforementioned issues, the laser inspection system provided by this invention, by incorporating a beam splitter, multiple phase shifters, a QBH array, an imaging module, and a control module, can efficiently and accurately perform wafer-level defect detection.

[0006] This invention provides a laser detection system, which includes: a laser emitting end, a condenser lens, a semi-transparent and semi-reflective mirror, a sample stage, a detection end, an imaging module, and a control module; The laser emitter includes: a laser, a beam splitter, multiple phase shifters, and a QBH array; The laser is used to generate a laser beam, which is then passed sequentially through a beam splitter, a phase shifter, a QBH array, a condenser lens, and a semi-transparent mirror. A beam splitter is used to split a laser beam into multiple sub-laser beams for projection onto different phase shifters; each phase shifter is used to adjust the phase of the corresponding sub-laser beam; a QBH array is used to project the multiple sub-laser beams onto a condenser lens; A condenser lens is used to focus multiple laser beams so that the focused multiple laser beams form a flat-top beam; The semi-transparent and semi-reflective mirror splits the flat-top beam into a detection sub-beam and a monitoring sub-beam, and projects the detection sub-beam onto the surface of the sample to be tested mounted on the sample stage, so that the sample to be tested reflects the detection sub-beam to the detection end, and projects the monitoring sub-beam onto the imaging module. The control module is electrically connected to the imaging module, the phase shifter, and the QBH array. The control module is used to adjust the phase shifter and / or the QBH array according to the imaging results of the imaging module to change the state of the sub-laser beam and optimize the output flat-top beam of the condenser objective.

[0007] Optionally, the laser detection system may also include a polarizer; The polarizer is located between the laser emitter and the condenser lens. It is used to adjust the polarization state of the multiple sub-laser beams emitted by the QBH array so that the detection sub-beam reflected by the sample under test is projected onto the detector.

[0008] Optionally, the laser detection system may also include a polarizer; The analyzer is located between the detector and the sample stage and is used to project a beam of a specified polarization state from the detection sub-beam reflected by the sample to the detector.

[0009] Optionally, the laser detection system may also include a collimating objective lens; The collimating objective is located between the sample stage and the analyzer. It is used to collimate the detection sub-beam reflected by the sample under test so that it can be projected onto the analyzer.

[0010] Optionally, the laser detection system may also include a filter; The filter is located between the analyzer and the detector and is used to filter out the fundamental frequency light in the reflected detector sub-beam.

[0011] Optionally, the laser emitter includes: multiple amplifiers; Multiple amplifiers are located between the phase shifter and the QBH array, and correspond one-to-one with the phase shifter, to amplify the corresponding sub-laser beams and project them onto the QBH array.

[0012] Optionally, the QBHs in the QBH array are formed by a hexagonal close-packed structure.

[0013] Optionally, the laser beam generated by the laser is a Gaussian beam.

[0014] Optionally, the shape of the flat-top beam can be circular, rectangular, triangular, or a combination of at least one of the above.

[0015] Optionally, the control module is also used to evaluate the imaging results of the imaging module through a system performance evaluation function, obtain the evaluation results, and adjust the phase shifter and / or QBH array according to the evaluation results so that the condenser objective outputs a flat-top beam.

[0016] The laser inspection system provided in this invention, by setting a beam splitter, multiple phase shifters, a QBH array, an imaging module, and a control module in the laser inspection system, enables the beam splitter to divide the laser beam generated by the laser into multiple sub-laser beams, and the control module to dynamically adjust the phase and propagation direction of each sub-laser beam according to the imaging results of the imaging module, thereby enabling efficient and accurate wafer-level defect detection. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic partial structural diagram of a laser detection system according to an embodiment of this application; Figure 2 This is a schematic partial structural diagram of a laser detection system according to an embodiment of this application; Figure 3 This is a schematic structural diagram of a QBH array according to an embodiment of this application; Figure 4 This is a diagram illustrating the interaction between an intelligent agent and its environment according to an embodiment of this application. Figure 5 This is a schematic flowchart of an embodiment of the SPGD algorithm of this application.

[0019] Figure label: 1. Laser emitter; 11. Laser; 12. Beam splitter; 13. Phase shifter; 14. QBH; 15. Amplifier; 2. Condenser objective; 3. Semi-transparent mirror; 4. Sample stage; 5. Detector; 6. Imaging module; 7. Polarizer; 8. Analyzer; 9. Collimating objective; 10. Filter; 100: Control module. Detailed Implementation

[0020] To facilitate understanding of this application, a more complete description will be provided below with reference to the accompanying drawings, which illustrate embodiments of the present application. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of this application will be thorough and complete.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0022] Spatial relation terms such as "below," "under," "below," "below," "above," and "above" are used here to describe the relationship between one element or feature shown in the figure and other elements or features. Similarly, "directly above" can be used here to describe an element or feature shown in the figure that coincides in a vertical straight line direction, which may be partial or complete, depending on the actual situation or the content of the illustration. It should be understood that, in addition to the orientation shown in the figure, spatial relation terms also include different orientations of the device in use and operation. For example, if the device in the figure is flipped, the element or feature described as "below," "below," or "below" of other elements will be oriented "above" other elements or features. Therefore, the exemplary terms "below" and "below" can include both upper and lower orientations. Furthermore, the device may also include other orientations (e.g., rotated 90 degrees or other orientations), and the spatial descriptive terms used herein will be interpreted accordingly.

[0023] It should be noted that when an element is referred to as "fixedly connected" to another element, it can be directly on the other element or there may be an intervening element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there may be an intervening element. Conversely, when an element is referred to as being "directly on" another element, there is no intervening element. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0024] When used herein, the singular forms of “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising,” “including,” or “having,” etc., specify the presence of the stated feature, whole, step, operation, component, part, or combination thereof, but do not preclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof.

[0025] In a first aspect, one embodiment of the present invention provides a laser detection system, combined with Figure 1 and Figure 2 The laser detection system includes: a laser emitter 1, a condenser lens 2, a semi-transparent mirror 3, a sample stage 4, a detector 5, an imaging module 6, and a control module 100.

[0026] The laser emitter 1 includes: a laser 11, a beam splitter 12, multiple phase shifters 13, and a QBH14 (Quartz BlockHead, laser output head) array. The laser 11 is used to generate a laser beam, which is then passed sequentially through the beam splitter 12, the phase shifters 13, the QBH array, the condenser lens 2, and the semi-transparent mirror 3.

[0027] Beam splitter 12 is used to split the laser beam into multiple sub-laser beams for projection onto different phase shifters 13; each phase shifter 13 is used to adjust the phase of the corresponding sub-laser beam; QBH array is used to project the multiple sub-laser beams onto condenser lens 2; The condenser objective 2 is used to focus multiple laser beams so that the focused multiple laser beams form a flat-top beam; The semi-transparent and semi-reflective mirror 3 splits the flat-top beam into a detection sub-beam and a monitoring sub-beam, and projects the detection sub-beam onto the surface of the sample to be tested loaded on the sample stage 4, so that the sample to be tested reflects the detection sub-beam to the detector end 5 and projects the monitoring sub-beam onto the imaging module 6. The control module 100 is electrically connected to the imaging module 6, the phase shifter 13 and the QBH array. The control module 100 is used to control the phase shifter 13 and / or the QBH array according to the imaging result of the imaging module 6, so as to change the state of the sub-laser beam and optimize the output flat-top beam of the condenser objective 2.

[0028] Understandably, taking a wafer as an example, after the detection sub-beam is directed at the wafer on the sample stage 4, the detection sub-beam will interact with the defects in the wafer to generate a laser signal carrying defect information. The detection end 5 obtains the laser signal carrying defect information based on the received detection sub-beam in order to determine whether there are defects on the wafer surface.

[0029] Specifically, if there are defects such as lattice mismatch on the wafer surface that cause changes in the symmetry of the wafer surface, it will affect the second-order nonlinear polarizability χ(2) at the wafer surface. Therefore, the interface state density of the wafer can be non-destructively characterized by the second harmonic to detect whether there are defects on the wafer surface. This embodiment will not elaborate on this.

[0030] Furthermore, the shape of the flat-top beam can be circular, rectangular, triangular, or a combination of at least one of the above; combined with Figure 3 The QBH array contains multiple QBH14s, each of which corresponds to a phase shifter 13 and a QBH14 to adjust the phase and propagation direction of the sub-laser beam; the semi-transparent mirror 3 divides the flat-top beam into a detection sub-beam and a monitoring sub-beam according to a transmittance to reflectance ratio of 1:(0.5~5).

[0031] In this embodiment, the control module 100 is a PC (Personal Computer); the laser 11 is the main oscillating laser 11; the shape of the flat-top beam is square; the QBH14 in the QBH array adopts a hexagonal close-packed structure to form the QBH array, so as to achieve the QBH array with the highest duty cycle, combined with Figure 3 The laser beam generated by laser 11 is a Gaussian beam; the transmittance to reflectance ratio of the semi-transparent mirror 3 is 1:1, and the beam transmitted through the semi-transparent mirror 3 is the monitoring sub-beam, while the reflected beam is the detection sub-beam. This embodiment does not limit the specific number of QBH14 and phase shifter 13.

[0032] The laser detection system provided in this embodiment of the invention, by setting up a beam splitter 12, multiple phase shifters 13, a QBH array, an imaging module 6, and a control module 100 in the laser detection system, enables the beam splitter 12 to divide the laser beam generated by the laser 11 into multiple sub-laser beams, and the control module 100 to dynamically adjust the phase and propagation direction of each sub-laser beam according to the imaging results of the imaging module 6, thereby enabling efficient and accurate wafer-level defect detection.

[0033] In a further optional embodiment of this embodiment, the laser detection system further includes a polarizer 7 and an analyzer 8.

[0034] The polarizer 7 is located between the laser emitting end 1 and the condenser objective lens 2. It is used to adjust the polarization state of the multiple sub-laser beams emitted by the QBH array so that the detection sub-beam reflected by the sample under test is projected onto the detector end 5.

[0035] The analyzer 8 is located between the detector end 5 and the sample stage 4, and is used to project the beam of a specified polarization state in the detection sub-beam reflected by the sample to be tested onto the detector end 5.

[0036] In this embodiment, both the polarizer 7 and the analyzer 8 are polarizers used to adjust the polarization state of the corresponding beam so that the polarization state of the corresponding beam reaches the optimal level, thereby ensuring the accuracy of wafer defect detection. The optimal polarization state of the beam is determined by the performance of the system in processing the beam, and this embodiment does not limit this.

[0037] In a further optional embodiment of this invention, the laser detection system further includes a collimating objective lens 9. The collimating objective lens 9 is located between the sample stage 4 and the analyzer 8, and is used to collimate the detection sub-beam reflected by the sample to be tested, so as to project it onto the analyzer 8.

[0038] In a further optional embodiment of this invention, the laser detection system further includes a filter 10. The filter 10 is located between the analyzer 8 and the detector end 5 and is used to filter out the fundamental frequency light in the reflected detection sub-beam.

[0039] It is understandable that the fundamental frequency light refers to the beam directly emitted by laser 11, while the frequency-doubled light refers to the second harmonic of the fundamental frequency light generated within the wafer due to nonlinear effects. The two beams are emitted coaxially and mixed together. Because the intensity of the fundamental frequency light is very high, while the intensity of the second harmonic carrying defect information, i.e., the frequency-doubled light, is very weak, filtering out the fundamental frequency light from the reflected detection sub-beam using filter 10 can effectively improve the accuracy of the laser detection system.

[0040] In a further optional embodiment of this embodiment, the laser emitting end 1 includes: a plurality of amplifiers 15; the plurality of amplifiers 15 are located between the phase shifter 13 and the QBH array, and correspond one-to-one with the phase shifter 13, for amplifying the corresponding sub-laser beams and projecting them into the corresponding QBH14 in the QBH array.

[0041] The laser inspection system provided in this invention can convert a Gaussian beam into a flat-top beam, resulting in a uniform square distribution of the light spot when incident on the wafer. This facilitates precise scanning point stitching and significantly improves inspection accuracy. Simultaneously, the system achieves coherent combining of multiple beams through a beam splitter 12, multiple phase shifters 13, a QBH array, and a condenser lens 2. This ensures uniform beam energy distribution and consistency of the laser signal (i.e., the detection sub-beam) on the wafer surface, improving the stitching accuracy of mapping scanning points. It is suitable for high-power laser sources and the inspection needs of large-size wafers, providing strong support for integrated circuit manufacturing process optimization. Furthermore, it overcomes the limitations of traditional shaping modules on laser power, supports beam size expansion, and further enhances the system's adaptability.

[0042] In a further optional embodiment of this embodiment, the control module 100 is also used to evaluate the imaging result of the imaging module 6 through a system performance evaluation function, obtain the evaluation result, and adjust the phase shifter 13 and / or the QBH array according to the evaluation result so that the condenser lens 2 outputs a flat-top beam.

[0043] The system performance evaluation function uses iterative algorithms, such as the Stochastic Parallel Gradient Descent (SPGD) algorithm and the Deep Deterministic Policy Gradient (DDPG) algorithm, to perform phase compensation of the beam in order to obtain a high-quality flat-top beam.

[0044] Secondly, one embodiment of the present invention provides a control method for the laser detection system as described in the first aspect. The method is applied to the control module of a laser detection system, and employs a stochastic parallel gradient descent algorithm to adjust the phase shifter and / or QBH array based on evaluation results, thereby enabling the condenser objective to output a flat-top beam. The method includes: Determine whether the similarity between the imaging result and the preset target result reaches a threshold; If not, apply a random perturbation to the system performance evaluation function in the stochastic parallel gradient descent (SPGD) algorithm, use the stochastic gradient descent algorithm with the applied random perturbation to update the control parameters of multiple phase shifters and / or QBH, and perform the step of judging whether the similarity between the imaging result and the preset target result reaches the threshold.

[0045] The control parameters include: horizontal control parameters of the QBH used to control the propagation path of the sub-laser beam in the horizontal direction. The vertical control parameters of QBH used to control the propagation path of the sub-laser beam in the vertical direction. And the phase control parameters of the phase shifter used to control the phase of the sub-laser beam. The random perturbation follows a Bernoulli distribution with a mean of 0.

[0046] In this laser detection system, the control module regulates the phase shifter and QBH array by outputting control signals. The iterative formula of the SPGD algorithm can be expressed as: In the formula U n (ν x ν y ν p ) represents the control parameter for the nth iteration. These represent the horizontal control voltages respectively. Vertical control voltage Phase control voltage ;ν x1 ν y1 ν p1 Let ν represent the horizontal control voltage, the vertical control voltage, and the phase control voltage corresponding to the QBH through which the first sub-laser beam passes, respectively. x2 ν y2 ν p2 Let ν represent the horizontal control voltage, vertical control voltage, and phase control voltage corresponding to the QBH through which the second sub-laser beam passes, respectively, and the phase control voltage corresponding to the phase shifter through which it passes, ..., ν xN ν yN ν pN These represent the horizontal control voltage, vertical control voltage, and phase control voltage of the QBH through which the Nth sub-laser beam passes, respectively. This is the iteration step size; This represents the change in the performance evaluation function before and after the disturbance; The system performance evaluation function values ​​before and after the disturbance; For statistically independent random disturbances.

[0047] Meanwhile, the variance of the Bernoulli distribution followed by the random perturbation is fixed, and the amplitude of the random perturbation is expressed as the standard deviation of the Bernoulli distribution. The specific value of the variance of the Bernoulli distribution depends on the actual settings of the laser detection system and factors such as the focal length of each lens in the system; this embodiment does not impose specific limitations on this. Similarly, the amplitude of the random perturbation is in a black box state, and this embodiment will not elaborate on this aspect.

[0048] In a further optional embodiment of this example, the step of applying a random perturbation to the system performance evaluation function in the stochastic gradient descent algorithm and updating the control parameters of multiple phase shifters using the stochastic gradient descent algorithm after applying the random perturbation includes: The sum of the random disturbance parameters and the control parameters of the current iteration is multiplied by the system performance evaluation function value of the current iteration to obtain the evaluation function value of the positive disturbance. The difference between the control parameter and the random disturbance parameter of the current iteration is multiplied by the system performance evaluation function value of the current iteration to obtain the evaluation function value of the negative disturbance. The difference between the evaluation function value of the positive perturbation and the evaluation function value of the negative perturbation is used to obtain the change in the system performance evaluation function for the current iteration. The control parameters for the next iteration are obtained by multiplying the preset iteration step size by the change in the system performance evaluation function of the current iteration and the random disturbance parameter of the current iteration, and then adding the result to the control parameters of the current iteration. Configure the control parameters for the next iteration into the corresponding phase shifter and QBH array to update the control parameters; The random perturbation parameter for the current iteration is randomly selected within a preset random value range; the system performance evaluation function value for the current iteration is obtained by evaluating the current laser detection system using the system performance evaluation function.

[0049] Specifically, the evaluation function value of the positive perturbation Evaluation function value of negative perturbation ; Calculate the change in the evaluation function for the nth iteration based on the evaluation function values ​​of the positive and negative perturbations. According to the formula Update control parameters to allow the laser detection system to proceed to the next iteration; the phase shifter and QBH array are based on U... (n+1) The corresponding pair of laser beams are controlled.

[0050] In a further optional embodiment of this example, the system performance evaluation function value J of the current iteration is calculated using Formula 1; Formula 1 is: ; Among them, A mn This is the actual distribution array of the beam intensity at the focal point of the current imaging result acquired by the imaging module in m rows and n columns; For A mn The conjugate matrix of B; mn The target distribution array is formed by the beam intensity corresponding to the focal point of the target result at m rows and n columns; For B mn The conjugate matrix.

[0051] In this system, the correlation coefficient between the target distribution matrix and the actual distribution matrix is ​​used as the reward value, which is used to calculate the measured light intensity value A of each beam. mn and target value B mn The correlation coefficient is the closest the imaging result is to the target result. When the correlation coefficient is 1, the actual imaging result is exactly the same as the target result.

[0052] The control method of the laser detection system provided in this embodiment introduces an optimization method based on the SPGD algorithm into laser coherent combining technology. This algorithm is simple to implement and has significant advantages in convergence speed and performance. Through coherent combining technology, not only does it overcome the power limitations of traditional lasers, but it also enables beam expansion, effectively eliminating the problem of uneven light intensity within the beam spot caused by the incident Gaussian beam, thereby improving beam quality and uniformity. Simultaneously, this control method further optimizes the stitching accuracy of wafer mapping scan data, providing a reliable guarantee for high-precision detection, and also has broad applicability and expansion potential.

[0053] Thirdly, one embodiment of the present invention provides a control method for a laser detection system as described in the first aspect.

[0054] The method is applied to the control module of a laser detection system and employs a depth-deterministic strategy gradient algorithm to control the phase shifter and / or QBH array based on the current state output by the control module, thereby changing the state of the sub-laser beam and optimizing the flat-top beam output by the condenser objective.

[0055] Specifically, the control module is configured with a DDPG policy network based on the DDPG algorithm, which includes an Actor network and a Critic network.

[0056] The Actor network is used to output actions based on the state parameters of the laser detection system in its current state, thereby controlling the phase shifter and / or the QBH array. The state parameters in the current state... That is, the state space includes: the phase of all sub-laser beams in the current state. Control parameters of the phase shifter and QBH array in the current state .

[0057] Among them, the phase of all sub-laser beams Also available Indicates; control parameters Includes: horizontal control parameters for controlling the propagation path of the sub-laser beam in the horizontal direction. Vertical control parameters used to control the propagation path of the sub-laser beam in the vertical direction. and phase control parameters used to control the phase of the sub-laser beam. .

[0058] Based on the content of the embodiments in the second aspect, ν x1 ν y1 ν p1 These represent the horizontal and vertical control voltages corresponding to the QBH through which the first sub-laser beam passes, and the phase control voltage corresponding to the phase shifter it passes through. Additionally, Ф 1 indicates the phase of the first sub-laser beam. Ф 2 represents the phase of the second sub-laser beam, ..., Ф N This indicates the phase of the Nth sub-laser beam.

[0059] The Critic network is used during the training phase of the DDPG policy network to optimize the actions output by the Actor network based on the imaging results in the current state and the actions output by the Actor network in the current state.

[0060] The method includes: The state parameters of the current state are input into the Actor network, so that the Actor network can output the action of the current state. ; Based on the action in the current state Adjust the phase shifter and / or QBH array to optimize the flat-top beam output from the condenser objective.

[0061] Understandable It indicates the current state inside the laser emitter, including the phase of the multi-channel laser beam, the phase shifter, and various control parameters of the QBH; Indicates the current state The action to be executed can be represented as: That is, the control parameters Adjustments.

[0062] By controlling parameters Adjustments can be made to control multiple phase shifters and QBH arrays. Specifically, the control parameters can be adjusted within a certain range. Adjustments are made by applying the control action to the drivers that regulate each control parameter, so that the corresponding drivers output the corresponding analog voltage signals to regulate the control parameters of the phase shifter and QBH.

[0063] In this embodiment, the current state can be understood as follows: The decision-making actions (actions) generated through the Actor network After the adjustment, The included parameters are processed through this action. After adjustment, the laser beams are closer to the desired state. The master oscillator laser beams are obtained with more closely spaced phases through fiber beam splitters, phase shifters, QBH arrays, and condenser lenses.

[0064] This embodiment uses a DDPG strategy network to dynamically control the phase shifter and / or QBH array in the laser detection system, enabling the flat-top beam to detect wafer-level defects with high quality, thereby improving the efficiency and accuracy of defect detection.

[0065] It is understandable that the DDPG strategy network used in the laser detection system for defect detection is a pre-trained network structure. During the training process, convergence conditions can be set, and training of the DDPG strategy network can be terminated when these conditions are met. Specific convergence conditions can be set by limiting the number of training iterations of the DDPG strategy network or based on the similarity between the imaging result and the target result.

[0066] In a further optional embodiment of this example, the DDPG policy network further includes: a target Actor network and a target Critic network; the Actor network has the same structure as the target Actor network, and the Critic network has the same structure as the target Critic network, combined with... Figure 4 ,exist Figure 4 In this context, the decision network is the Actor network, and the target network integrated within the Actor box is the target Actor network; the evaluation network is the Critic network, and the target network integrated within the Critic box is the target Critic network; the system integrated within the environment box represents some components of the laser detection system. See the attached reference for details. Figure 2 .

[0067] During the training phase of the DDPG policy network, the network parameters of the target Actor network are... Based on soft update coefficient And the network parameters of the Actor network in the current state Perform a soft update on the network parameters of the target Critic network. Based on soft update coefficient And the network parameters of the Critic network in the current state Perform a soft update; Critic networks are used to determine the state parameters based on the current state. Actions in the current state output by the Actor network Output the precision parameter Q in the current state. Specifically, the precision parameter Q is represented as... Understandably, in the current state... The more precise the actions generated by the Actor network, the better the performance of the Critic network. The larger the value.

[0068] The target actor network is used to output the next state based on the state parameters of the next state in the current state. The following action That is, the target action. Specifically, the target Actor network outputs the following in its current state: Represented as .

[0069] The target Critic network is used to output the target precision parameter Q' in the current state, based on the state parameters of the next state and the action of the next state output by the Actor network. Here, the target precision parameter Q' in the current state is denoted as... .

[0070] During the training phase of the DDPG policy network, the DDPG policy network maximizes the accuracy parameter Q, which affects the network parameters of the Actor network. Update the network parameters of the updated Actor network. Network parameters of the corresponding target Actor network The difference becomes smaller.

[0071] in, The better the settings, the more accurate the actions generated by the Actor network, enabling the laser coherence in the system to converge faster.

[0072] Understandably, during the training phase of the DDPG policy network, initialization of the DDPG policy network is required first. Specifically, the process of initializing the Actor network, Critic network, target Actor network, and target Critic network involves randomly initializing the Actor network... and Critic network and make , Initialize the attenuation factor Soft update coefficient This embodiment does not impose specific limitations on the initial values. Among them, the attenuation factor... Used to solve for the objective Q value; soft update coefficients Used to update the target network and .

[0073] The DDPG policy network provided in this embodiment includes two working modes: a training phase and an application phase. In the training phase, the DDPG policy network is repeatedly trained to converge as quickly as possible, learning to shape the laser beam generated by the laser into a flat-top beam, and using a reward value... The training terminates when the convergence reaches a preset convergence threshold of 0.99. During the application phase, the trained network parameters (i.e., network weights) are frozen, and only the Actor network is used for real-time inference. The flat-top beam output is maintained long-term to ensure that the laser detection system can quickly recover if it deviates from the ideal state. This is controlled by a system operation switch, which controls the laser detection system's on / off state.

[0074] Combination Figure 4 and Figure 5 The training process of the DDPG policy network includes the following steps: Step 1: Set the state parameters in the current state. Inputting the data into the Actor network yields the actions in the current state. .

[0075] Step 2: Perform the action , obtain the state parameters of the next state and reward value This allows us to obtain the experience tuple in the current state. Among them, the reward value It is calculated by the reward function. In this embodiment, the formula for calculating the reward function, i.e., the formula for calculating the correlation coefficient between the actual distribution array and the target distribution array, is as follows, but is not limited to this: .

[0076] Among them, A mn This is the actual distribution array of the beam intensity at the focal point of the current imaging result acquired by the imaging module in m rows and n columns; For A mn The conjugate matrix of B; mn The target distribution array is formed by the beam intensity corresponding to the focal point of the target result at m rows and n columns; For B mn The conjugate matrix.

[0077] In this system, the correlation coefficient between the target distribution matrix and the actual distribution matrix is ​​used as the reward value, which is used to calculate the measured light intensity value A of each beam. mn and target value B mn The correlation coefficient is the closest the imaging result is to the target result. When the correlation coefficient is 1, the actual imaging result is exactly the same as the target result.

[0078] Step 3: Obtain the Stored in the experience replay pool Replay Buffer.

[0079] Step 4: When the number of samples in the experience replay pool exceeds the minimum value At that time, the network parameters of the Actor network were started. An update is being performed. This is understandable, during the update... Updated simultaneously during the process , , The specific process is as follows: Step 401: Randomly sample from the Replay Buffer M training samples are used to calculate M target Q values ​​from the output Q' of the target Critic network. The formula for calculating the target Q value is as follows: .

[0080] Understandably, for There are 10 training samples, and each training sample contains a state parameter in its current state. Actions in the current state Reward value in the current state and the state parameters in the next state ; This is the value output by the target Critic network in the current state.

[0081] Step 402: Obtain the Q-value in the current state using the Critic network. The Critic network uses the mean squared error loss function as its loss function. The Critic network's loss function is as follows: .

[0082] in, The result is the output of the target Critic network; This represents the Q-value obtained from the Critic network in the current state.

[0083] Step 403: Minimize using gradient descent algorithm To update the network parameters of the Citic network .in, The loss function of the Citic network is given. Step 403 involves the following steps: First, the gradient of the Citic network's loss function with respect to the network parameters is calculated. Then, the network parameters of the Citic network are decreased along the negative gradient direction. This reduces the loss by modifying the network parameters of the Citic network, thus achieving the desired reduction in loss. Optimization is then performed. It should be noted that the gradient descent algorithm minimizes... This can be achieved using existing technologies, which will not be elaborated upon in this embodiment.

[0084] Step 404: Based on the gradient formula of the Actor network, optimize the network parameters of the Actor network by maximizing the Q-value. .

[0085] The specific implementation step 404 includes the following: First, the gradient of the Q-value output by the Critic network with respect to the network parameters of the Actor network is calculated (chain rule). Then, the network parameters of the Actor network are updated along the positive gradient direction using gradient ascent, thereby maximizing the Q-value output by the Critic network. It should be noted that maximizing the Q-value optimizes the network parameters of the Actor network. This can be achieved using existing technologies, which will not be elaborated upon in this embodiment.

[0086] The gradient formula for the Actor network is as follows: .

[0087] In the formula It is the result of calculating the gradient of the Actor network's output with respect to its network parameters, as shown in the formula. It is the result of calculating the gradient of the input action by the output of the Critic network.

[0088] Step 405: Soft update the network parameters by soft updating the target network formula. and .

[0089] The formula for soft-update target network is as follows: .

[0090] Then, repeat steps 401 to 405 above until the reward value is reached. The threshold of 0.99 is reached and convergence occurs, or the system operation switch is turned off.

[0091] In a further optional embodiment of this example, during the training phase of the DDPG policy network, the Actor network is used to output the reward value of the current state based on the imaging result of the current state and the preset target result.

[0092] The method also includes the following steps during the training phase of the DDPG policy network: After the Actor network outputs the action in the current state, it collects the experience tuples in the current state. The empirical tuple includes: state parameters in the current state. Actions in the current state Reward value in the current state and the state parameters in the next state ; The experience tuples in the current state are put into the experience replay buffer for use as training samples; Determine whether the number of training samples in the experience replay pool exceeds the training threshold; in this embodiment, the training threshold is the same as the number of training samples sampled from the experience replay pool when the number of training samples in the experience replay pool reaches or exceeds the training threshold, but it is not limited to this. When the number of training samples in the experience replay pool reaches or exceeds the training threshold, M training samples are randomly sampled from the experience replay pool. The state parameters corresponding to the current state in the M training samples. and actions in the current state The parameters are sequentially input into the Critic network to obtain M precision parameters Q; Based on M precision parameters Q and the corresponding M target precision parameters output by the target Critic network. The loss value of the Critic network is calculated using its loss function. Loss ; Based on the loss value of the Critic network Loss Update the network parameters of the Critic network. ; Based on the M precision parameters Q, calculate the state parameters of the Critic network in the corresponding current state of the M training samples. Evaluation gradient ; According to the evaluation gradient And the gradient formula for the Actor network, optimizing the network parameters of the Actor network. ; Network parameters based on the optimized Actor network Soft update the network parameters of the target Actor network .

[0093] In a further optional embodiment of this example, the method further includes: During the training phase of the DDPG policy network, the actual distribution array A is formed by the beam intensity corresponding to the focal point at m rows and n columns based on the imaging results currently acquired by the imaging module. mn The target distribution array B, consisting of the beam intensity corresponding to the focal point of the target result at m rows and n columns. mn Calculate the reward value in the current state. Among them, the reward value With actual distributed array A mn and target distribution array B mn The similarity is positively correlated; the actual distribution array A mn and target distribution array B mn The higher the similarity, the higher the reward value. The larger; in the actual distributed array A mn and target distribution array B mn When they are the same, the reward value The value is 1.

[0094] Determine whether the reward value in the current state has reached the convergence threshold. If yes, end the training of the DDPG policy network; otherwise, input the state parameters of the next state into the Actor network to continue optimizing the network parameters of the Actor network.

[0095] In a further optional embodiment of this example, the method further includes: during the training phase of the DDPG policy network, outputting the action in the current state to the Actor network. Apply preset current state noise .

[0096] Based on the applied current state noise The actions taken in the current state after the phase shifter and / or QBH array are adjusted to make the condenser objective output a flat-top beam.

[0097] Specifically, current state noise The noise is generated through a noise function. In this embodiment, the noise function is the Ornstein-Uhlenbeck noise function. During the training phase of the DDPG policy network, a preset current-state noise is applied to the actions in the current state output by the Actor network. This allows the DDPG policy network to introduce a degree of exploration during the learning process, using noise to explore the action space. The noise varies in different training phases, and its determination can be achieved through the following noise function: .

[0098] in, It is the noise added when the Actor network outputs the action in the next state. Noise added when the Actor network outputs actions in the current state The differences are all around the noise mean. fluctuation, Control the magnitude of the noise variation toward the mean. It is random noise, noise standard deviation Control the amplitude of random noise.

[0099] During actual training, the coefficients of the random noise function are adjusted based on intermediate experimental results. By rigorously adjusting the random noise function according to the convergence of the reward curve during actual training, the agent (DDPG policy network) can fully explore the action space, allowing for thorough exploration in the early stages of training and timely utilization of previously gained experience after training. The values ​​of the coefficients in the random noise function are determined based on the control parameters. The different determinations. It should be noted that, The three types of control parameters correspond to different noise units and have different value ranges. This embodiment does not make specific limitations on this.

[0100] In this embodiment, parameters or coefficients that are not specifically disclosed or explained are either determined by the system's performance or generated randomly, and this embodiment does not impose specific limitations on them.

[0101] The control method of the laser detection system provided in this embodiment, based on the coherent combining technology of the DDPG algorithm, is applicable to tasks involving continuous action spaces and high-dimensional state spaces. It boasts excellent sampling efficiency and strategy search capabilities, as well as relatively efficient utilization of computational resources, and exhibits excellent convergence speed and performance. Furthermore, this embodiment overcomes the limitation of laser power through coherent combining technology, enabling beam expansion and eliminating the problem of uneven light intensity within the beam spot caused by an incident Gaussian beam, which prevents precise stitching of the entire wafer mapping scan data points, thereby improving the accuracy of laser defect detection.

[0102] In the description of this specification, the references to terms such as "some embodiments," "other embodiments," "ideal embodiments," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example that are included in at least one embodiment or example of this application. In this specification, the illustrative descriptions of the above terms do not necessarily refer to the same embodiments or examples.

[0103] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0104] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A laser detection system, characterized by, The laser detection system comprises a laser emission end, a condenser objective, a half-transmission half-reflection mirror, a sample stage, a detection end, an imaging module and a control module. The laser emission end comprises a laser, a beam splitter, a plurality of phase shifters and a QBH array. The laser is configured to generate a laser beam, and sequentially pass the laser beam through the beam splitter, the phase shifters, the QBH array, the condenser objective and the half-transmission half-reflection mirror. The beam splitter is configured to split the laser beam into a plurality of sub-laser beams to be projected to different phase shifters; each phase shifter is configured to adjust the phase of the corresponding sub-laser beam; and the QBH array is configured to project the plurality of sub-laser beams to the condenser objective. The condenser objective is configured to condense the plurality of sub-laser beams to form a flat-top beam. The half-transmission half-reflection mirror splits the flat-top beam into a detection sub-beam and a monitoring sub-beam, and projects the detection sub-beam onto the surface of a sample to be measured loaded on the sample stage, so that the sample to be measured reflects the detection sub-beam to the detection end, and projects the monitoring sub-beam to the imaging module. The control module is electrically connected to the imaging module, the phase shifters and the QBH array, and is configured to control the phase shifters and / or the QBH array according to the imaging result of the imaging module, to change the state of the sub-laser beams and optimize the flat-top beam output by the condenser objective.

2. The laser detection system of claim 1, wherein, The laser detection system further comprises a polarizer. The polarizer is located between the laser emission end and the condenser objective, and is configured to adjust the polarization state of the plurality of sub-laser beams emitted by the QBH array, so that the detection sub-beam reflected by the sample to be measured is projected to the detection end.

3. The laser detection system of claim 1, wherein, The laser detection system further comprises an analyzer. The analyzer is located between the detection end and the sample stage, and is configured to project the light beam of a specified polarization state in the detection sub-beam reflected by the sample to be measured to the detection end.

4. The laser detection system of claim 3, wherein, The laser detection system further comprises a collimator. The collimator is located between the sample stage and the analyzer, and is configured to collimate the detection sub-beam reflected by the sample to be measured to be projected to the analyzer.

5. The laser detection system of claim 3, wherein, The laser detection system further comprises a filter. The filter is located between the analyzer and the detection end, and is configured to filter out the fundamental frequency light in the reflected detection sub-beam.

6. The laser detection system of claim 1, wherein, The laser emission end comprises a plurality of amplifiers. The plurality of amplifiers are located between the phase shifters and the QBH array, and correspond to the phase shifters one by one, and are configured to amplify the corresponding sub-laser beams and project them to the QBH array.

7. The laser detection system of any one of claims 1 to 6, wherein, The QBHs in the QBH array adopt a hexagonal close-packed structure to form the QBH array.

8. The laser detection system of any one of claims 1 to 6, wherein, The laser beam generated by the laser is a Gaussian beam.

9. The laser detection system of any one of claims 1 to 6, wherein, The flat-top beam has a circular, rectangular, triangular or combined shape of at least one of the above.

10. The laser detection system of any one of claims 1 to 6, wherein, The control module is further configured to evaluate the imaging result of the imaging module by a system performance evaluation function to obtain an evaluation result, and control the phase shifters and / or the QBH array according to the evaluation result to change the state of the sub-laser beams and optimize the flat-top beam output by the condenser objective.

Images