A laser focusing spot position positioning and monitoring device

Abstract

The present invention discloses a device for positioning and monitoring the position of a laser focused spot, which designs the structure of an optical system, including the structure design of a laser focusing optical path and an equivalent feedback optical path, and uses a first CCD component and a second CCD component in cooperation. The wafer surface graphic information collected by the first CCD component extracts the wafer contour and surface microstructure features through image processing, providing high-precision input for subsequent wafer positioning and alignment. The second CCD component is located on a one-dimensional moving platform with a moving accuracy in the micron order. By moving the second CCD component left and right, the positions of multiple focal points in the sample are judged. The laser spot image information in the feedback optical path collected by the second CCD component extracts the center coordinates and radius of the laser spot through image processing, and reversely calibrates the focal depth inside the wafer, thereby indirectly monitoring the laser spot inside the wafer. In addition, based on deep reinforcement learning, the position of the second CCD component is dynamically adjusted to achieve automatic focusing.

Description

Technical Field

[0001] The present invention relates to the technical field of high-precision laser cutting processing, and particularly relates to a device for positioning and monitoring the position of a laser focusing spot. Background Art

[0002] Laser stealth cutting technology is a revolutionary process in the field of semiconductor wafer precision processing. Its core principle is to focus an ultrashort pulse laser inside the wafer, form a controllable modified layer on the subsurface of the material, and then achieve wafer separation through mechanical or thermal stress expansion. The traditional diamond blade cutting technology has inherent defects such as large cut width, low processing efficiency, significant edge chipping, and serious cutting debris pollution. To solve these problems, laser ablation cutting technology emerged. It realizes cutting separation by focusing a high-energy density laser beam on the wafer surface to cause phase change processes such as melting and vaporization of the material. However, this technology still has process defects such as slag residue on the cut surface, insufficient control of the cut width, and too wide heat-affected zone, which severely restrict its application in the field of precision semiconductor processing. Compared with traditional cutting methods, laser stealth cutting technology has significant advantages such as non-contact processing, no debris generation, narrow cutting lanes, and suitability for ultra-thin wafers, and has become a key technology for advanced packaging and compound semiconductor processing.

[0003] During the laser stealth cutting process, the focus of the laser spot is inside the wafer. Due to surface reflection interference and penetration depth limitations, traditional optical detection means (CCD imaging, confocal measurement) cannot achieve focus positioning and monitoring. Existing monitoring means all adopt an offline post-test detection scheme. After cutting is completed, a scanning electron microscope or an optical microscope is used to observe the focus position, multi-focus interval, and stealth cutting quality. By comparing the actual focus position with the theoretical focus position, the offset amount of the actual coordinates of the modified layer focus from the theoretical coordinates, the consistency of the multi-focus spacing, and the cross-sectional morphology are evaluated.

[0004] The prior art proposed a multi-focus laser stealth cutting algorithm based on a large numerical aperture. Multiple foci are generated through a spatial light modulator (SLM) to achieve multi-layer cutting inside a silicon wafer. First, the multi-focus positions inside the silicon wafer are obtained through simulation, and then through experiments, the cross-section of the silicon wafer is observed under an optical microscope. The centers of the three scratches are respectively located at positions about 35.0μm, 105.2μm, and 176.0μm from the upper surface of the silicon wafer, which are basically consistent with the initial designed focus positions and the simulation results.

[0005] The prior art also proposed a research on the cutting process of silicon wafers. The control of the focus position involves the optimization of parameters such as laser power, frequency, and cutting speed to ensure the formation of the modified layer and the cutting quality. By comparing the theoretical focus position with the actual focus position, the focus offset law under different process parameters is tested.

[0006] In the above-mentioned technology, the monitoring of the internal focal position of the silicon wafer adopts an off-line post-test detection scheme, that is, after the cutting is completed, the actual multi-focal positions are obtained by observing the morphology of the cross-section of the cut surface. The prior art cannot accurately and dynamically track the focal point in real time during the laser stealth cutting process through traditional optical detection means, and this monitoring blind area directly affects the precision control and process stability of the stealth cutting. Summary of the Invention

[0007] In order to solve the problem that the prior art cannot accurately and dynamically track the focal point in real time during the laser stealth cutting process through traditional optical detection means, and this monitoring blind area directly affects the precision control and process stability of the stealth cutting, the present invention provides a device for positioning and monitoring the position of a laser focusing spot, the structural design of its optical system, including the structural design of a laser focusing optical path and an equivalent feedback optical path, and the dual CCD components cooperate to achieve high-precision positioning of the surface morphology and internal spots of the wafer, determine the depth of the stealth cutting spot through the focal position of the feedback optical path, and use the SAC algorithm to replace the traditional manual focusing, and perform automatic focusing through computer control to improve the focusing efficiency.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] The present invention provides a device for positioning and monitoring the position of a laser focusing spot, including a laser, a first optical path structure, a spatial light modulator, a first beam splitter, a first focusing objective lens, a second focusing objective lens, a first lens, a first CCD component, a second CCD component, a first moving platform, a wafer, a second moving platform, and a control system;

[0010] After the laser emits a laser beam, it is deflected and modulated by the first optical path structure and then incident on the spatial light modulator. After passing through the spatial light modulator, the laser beam generates a multi-focal spot, and then is split into a focusing optical path and a feedback optical path by the first beam splitter. The focusing optical path is focused by the first focusing objective lens and then enters the wafer placed on the first moving platform for laser stealth cutting operation. The first CCD component is used to collect the image information on the surface of the wafer and position the wafer;

[0011] After the feedback optical path is focused by the second focusing objective lens, it passes through the first lens and finally enters the second CCD component. The second moving platform drives the second CCD component to move left and right along the optical axis direction. The second CCD component synchronously collects the spot images corresponding to the focal points in the feedback optical path, and extracts the center coordinates and radius of the laser spot circle;

[0012] The control system includes an image processing module, an algorithm processing module, and a signal control module. The image processing module executes the image processing process in parallel. According to the wafer surface image information collected by the first CCD component and the laser spot image information in the feedback optical path collected by the second CCD component, the wafer contour, surface microstructure features, the center coordinates and radius of the laser spot circle are obtained respectively. The algorithm processing module integrates the wafer positioning and alignment algorithm and the deep reinforcement learning focusing algorithm. The signal control module controls the positioning and alignment of the wafer based on the algorithm results of the algorithm processing module and controls the lateral movement of the second CCD component for automatic focusing based on the algorithm training model. The second CCD component observes the laser spot at the focal position in real time and obtains the cutting depth of the laser stealth cutting of the wafer, so as to locate and monitor the stealth cutting spot in real time.

[0013] Adopting the above technical solution:

[0014] In the present invention, the structural design of the optical system includes the structural design of the laser focusing optical path and the equivalent feedback optical path, and the first CCD component and the second CCD component are used in cooperation. The wafer surface graphic information collected by the first CCD component extracts the wafer contour and surface microstructure features through image processing, providing high-precision input for the subsequent positioning and alignment of the wafer. The second CCD component is located on a one-dimensional moving platform with a moving precision in the micron level. By moving the second CCD component left and right to judge the positions of multiple focal points in the sample, the laser spot image information in the feedback optical path collected by the second CCD component extracts the center coordinates and radius of the laser spot through image processing, and reversely calibrates the focal depth inside the wafer, thereby indirectly monitoring the laser spot inside the wafer. In addition, based on deep reinforcement learning (SAC algorithm), the position of the second CCD component is dynamically adjusted to perform automatic focusing, improving the focusing efficiency.

[0015] The first CCD component obtains the wafer surface topography information through image acquisition and realizes feature extraction by adopting a three-stage processing process. Specifically, it includes median filtering of the image: constructing an N×N pixel sliding window to scan line by line, calculating the median value of the image in the window coverage area, effectively suppressing pulse noise and retaining edge details; image threshold segmentation: traversing the gray value histogram of the preprocessed image based on the Otsu algorithm, and taking the maximum value of the between-class variance result of the target area and the background area as the threshold to be set to achieve high-fidelity feature separation; edge detection: performing edge detection based on the Canny operator to accurately extract the wafer contour and surface microstructure features, providing high-precision input for the subsequent positioning and alignment of the wafer.

[0016] The lasers used in wafer stealth dicing generally are infrared ultrashort pulse lasers with a central wavelength of 1064 nm, and the power and single pulse energy of such lasers do not need to be too high to avoid damaging the wafer surface. The main function of the laser is to form multiple foci inside the wafer through a spatial light modulator, and the foci are focused to form a modified layer inside the material.

[0017] Further, the first focusing objective lens and the second focusing objective lens have the same model, and their distances from the first beam splitter are the same. The first lens has the same refractive index as the wafer, and the distance from the first lens to the second CCD assembly is the axial position of the focus inside the wafer.

[0018] Specifically, on the feedback optical path, a laser energy attenuator is passed through first to avoid damage to the second CCD assembly due to excessive energy. The distances from the first focusing objective lens and the second focusing objective lens to the first beam splitter are equal. The first lens is used to equivalently refract the light entering the wafer. The aperture stop is located behind the first lens and at the center of the optical path to limit the imaging of stray light. The second CCD assembly is on a one-dimensional moving platform with a moving accuracy in the micron range. By moving the second CCD assembly left and right, the positions of the multiple foci in the sample can be judged. The second CCD assembly synchronously acquires the laser spot images in the feedback optical path, and a three-level processing process is adopted to realize the extraction of the center coordinates and radius of the spot circle. Specifically, it includes wavelet threshold filtering: by decomposing the image into multiple scales of wavelets, using the difference in energy distribution of the signal and noise in the wavelet domain to divide the signal and noise in the image, and removing the noise by setting the coefficients to zero to obtain a denoised image; edge detection: extracting the edge of the laser spot based on the Canny operator; least squares circle fitting: constructing an objective function based on the edge point set, and calculating the center coordinates (x, y) and radius r of the spot through iterative optimization. When the laser is in an ideal focusing state, the spot diameter reaches the minimum value, and thus the focal depth inside the wafer can be inversely calibrated. According to the principle of optical path symmetry, the distance from the first lens to the second CCD assembly is equivalent to the axial position of the focus inside the wafer.

[0019] Further, the spacing between the multiple focus spots generated by the spatial light modulator is Δd. The wafer is driven by the first moving platform to make a lateral step displacement that is an integer multiple of Δd, and the second CCD assembly can continuously capture the spot images corresponding to each focus.

[0020] Specifically, the laser modulated by the spatial light modulator forms an equally spaced multiple focus array inside the wafer, and the spacing is uniformly Δd. The number and positions of the foci are determined by the spatial light modulator. The wafer is driven by a precision displacement stage to make a lateral step displacement that is an integer multiple of Δd, and the second CCD assembly can continuously capture the spot images corresponding to each focus, and the consistency of the focus spacing is monitored in real time and the modulation accuracy of the spatial light modulator is verified.

[0021] Furthermore, the deep reinforcement learning focusing algorithm adopts the SAC algorithm. During the training process, the reward function of the laser spot radius is adjusted to converge to zero, and a scaling factor of 10 is added to increase the training speed.

[0022] The wafer positioning and alignment algorithm mainly obtains the detection image of the wafer through a CCD camera, and calculates the deviation of the coordinates and rotation angle by comparing with a pre-prepared aligned template image, so as to realize wafer alignment.

[0023] Furthermore, the first optical path structure includes a double mirror assembly, which is used to adjust the transmission path of the laser beam emitted by the laser. The adjusted laser beam enters the center of the subsequent optical device and is incident on the spatial light modulator at an incident angle less than 10°.

[0024] The spatial light modulator modulates the laser beam by loading a hologram to generate multiple foci. When the incident angle is too large, the adjustment efficiency will decrease, resulting in uneven energy distribution of the generated multiple foci. At the same time, a smaller incident angle can avoid distortion when the laser beam is refracted or reflected on the surface of the spatial light modulator, ensuring good transmission quality of the beam in the optical path.

[0025] Furthermore, the first optical path structure further includes a laser beam expander, a half-wave plate, a polarization beam splitter, and a first doublet lens assembly. The laser beam adjusted by the double mirror assembly sequentially passes through the laser beam expander, the half-wave plate, the polarization beam splitter, and the first doublet lens assembly and then enters the spatial light modulator.

[0026] The laser beam expander is used to compress the laser divergence angle and improve the beam quality. The polarization direction of the incident laser after passing through the half-wave plate and the polarization beam splitter is consistent with the response direction of the spatial light modulator. The two lenses in the first doublet lens assembly form a 4f system, which is used to modulate the beam. The laser generates multiple focus spots after passing through the spatial light modulator.

[0027] Furthermore, a second doublet lens assembly is provided between the spatial light modulator and the first beam splitter.

[0028] Furthermore, the first doublet lens assembly and the second doublet lens assembly respectively form a 4f system for modulating the laser beam.

[0029] Furthermore, an illumination light source is coaxially provided directly above the focusing optical path. The illumination light source is divided into two illumination optical paths by a second beam splitter and enters the first CCD assembly and the second CCD assembly respectively.

[0030] The illumination beam emitted by the illumination light source enters the first CCD assembly and the second CCD assembly for supplementary illumination, improving the imaging quality and detection accuracy of the CCD camera.

[0031] Further, the image processing module adopts a three - level processing flow of image median filtering, image threshold segmentation, and edge detection to process the wafer surface image information collected by the first CCD component to obtain the wafer contour and surface microstructure features; the image processing module adopts a three - level processing flow of wavelet threshold filtering, edge detection, and least - squares circle fitting to process the laser spot image information in the feedback optical path collected by the second CCD component to obtain the center coordinates and radius of the laser spot circle.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] In the present invention, the structural design of the optical system includes the structural design of the laser focusing optical path and the equivalent feedback optical path, and the first CCD component and the second CCD component are used in cooperation. The wafer surface graphic information collected by the first CCD component is processed by image processing to extract the wafer contour and surface microstructure features, providing high - precision input for the subsequent positioning and alignment of the wafer. The second CCD component is located on a one - dimensional moving platform with a moving precision in the micron level. By moving the second CCD component left and right, the positions of multiple focal points in the sample are judged. The laser spot image information in the feedback optical path collected by the second CCD component is processed by image processing to extract the center coordinates and radius of the laser spot, and the depth of the internal focal point of the wafer is inversely calibrated, thereby indirectly monitoring the internal laser spot of the wafer. In addition, based on deep reinforcement learning (SAC algorithm), the position of the second CCD component is dynamically adjusted, combined with a reward function based on the laser spot radius, and automatic focusing is carried out through computer control to improve the focusing efficiency. Description of the Drawings

[0034] The following further elaborates on the present invention in detail in conjunction with the drawings and specific embodiments.

[0035] Figure 1 It is a schematic structural diagram of the laser focusing spot position positioning and monitoring device in the present invention;

[0036] Figure 2 It is a schematic diagram of the multi - focal spot positions generated by the spatial light modulator in the present invention;

[0037] Figure 3 It is a schematic structural diagram of the control system in the present invention;

[0038] Figure 4 It is a flow chart of the training process of the deep reinforcement learning focusing algorithm in the present invention;

[0039] Among them, the specific reference numerals are:

[0040] Laser 1, double mirror assembly 2, laser beam expander 3, half-wave plate 4, polarization beam splitter 5, first doublet lens assembly 6, spatial light modulator 7, second doublet lens assembly 8, first beam splitter 9, first focusing objective lens 10, wafer 11, first motion platform 12, first CCD assembly 13, laser energy attenuator 14, second focusing objective lens 15, first lens 16, aperture 17, second CCD assembly 18, second motion platform 19, illumination light source 20, second beam splitter 21, control system 22, image processing module 23, algorithm processing module 24, signal control module 25. Detailed implementation manners

[0041] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0042] An embodiment of the present invention discloses a device for positioning and monitoring the position of a laser focused spot, as Figure 1 shown, including a laser 1, a first optical path structure, a spatial light modulator 7, a first beam splitter 9, a first focusing objective lens 10, a second focusing objective lens 15, a first lens 16, a first CCD assembly 13, a second CCD assembly 18, a first motion platform 12, a wafer 11, a second motion platform 19 and a control system 22;

[0043] After the laser 1 emits a laser beam, it is deflected and modulated by the first optical path structure and then incident on the spatial light modulator 7. After passing through the spatial light modulator 7, the laser beam generates a multi-focus spot, and then is split into a focusing optical path and a feedback optical path by the first beam splitter 9. The focusing optical path is focused by the first focusing objective lens 10 and then enters the wafer 11 placed on the first motion platform 12 for laser stealth cutting operation. The first CCD assembly 13 is used to collect the image information on the surface of the wafer 11 and locate the position of the wafer 11;

[0044] After the feedback optical path is focused by the second focusing objective lens 15, it passes through the first lens 16 and finally enters the second CCD assembly 18. The second motion platform 19 drives the second CCD assembly 18 to move left and right along the optical axis direction. The second CCD assembly 18 synchronously collects the spot images corresponding to each focus in the feedback optical path and extracts the center coordinates and radius of the laser spot circle;

[0045] As Figure 3As shown in the figure, the control system 22 includes an image processing module 23, an algorithm processing module 24, and a signal control module 25. The image processing module 23 executes the image processing process in parallel. According to the surface image information of the wafer 11 collected by the first CCD component 13 and the laser spot image information in the feedback optical path collected by the second CCD component 18, the contour and surface microstructure features of the wafer 11, and the center coordinates and radius of the laser spot circle are obtained respectively. The algorithm processing module 24 integrates the wafer 11 positioning and alignment algorithm and the deep reinforcement learning focusing algorithm. The signal control module 25 controls the positioning and alignment of the wafer 11 based on the algorithm results of the algorithm processing module 24 and controls the lateral movement of the second CCD component 18 for automatic focusing based on the algorithm training model. The second CCD component 18 observes the laser spot at the focal position in real time and obtains the cutting depth of the laser stealth cutting of the wafer 11, so as to locate and monitor the stealth cutting spot in real time.

[0046] In the present invention, the structural design of the optical system includes the structural design of the laser focusing optical path and the equivalent feedback optical path, and the first CCD component 13 and the second CCD component 18 are used in cooperation. The surface pattern information of the wafer 11 collected by the first CCD component 13 is used to extract the contour and surface microstructure features of the wafer 11 through image processing, providing high-precision input for the subsequent positioning and alignment of the wafer 11. The second CCD component 18 is located on a one-dimensional moving platform with a moving accuracy in the micron range. By moving the second CCD component 18 left and right, the positions of multiple focal points in the sample are judged. The laser spot image information in the feedback optical path collected by the second CCD component 18 is used to extract the center coordinates and radius of the laser spot through image processing, and the focal depth inside the wafer 11 is inversely calibrated, thereby indirectly monitoring the laser spot inside the wafer 11. In addition, based on deep reinforcement learning (SAC algorithm), the position of the second CCD component 18 is dynamically adjusted to perform automatic focusing and improve the focusing efficiency.

[0047] Specifically, the first CCD component 13 obtains the surface topography information of the wafer 11 through image acquisition and realizes feature extraction through a three-level processing process, which specifically includes: image median filtering: constructing an N×N pixel sliding window to scan line by line, calculating the median value of the image in the window coverage area, effectively suppressing pulse noise and retaining edge details; image threshold segmentation: based on the Otsu algorithm, traversing the gray value histogram of the preprocessed image, and taking the maximum value of the between-class variance result of the target area and the background area as the threshold to be set to achieve high-fidelity feature separation; edge detection: based on the Canny operator for edge detection, accurately extracting the contour and surface microstructure features of the wafer 11, providing high-precision input for the subsequent positioning and alignment of the wafer 11.

[0048] Among them, the first focusing objective lens 10 and the second focusing objective lens 15 have the same model and the same distance from the first beam splitter 9. The first lens 16 has the same refractive index as the wafer 11. The distance from the first lens 16 to the second CCD assembly 18 is the axial position where the focus is inside the wafer 11. Specifically, on the feedback optical path, a laser energy attenuator 14 is passed through first to avoid damage to the second CCD assembly 18 due to excessive energy. The distances from the first focusing objective lens 10 and the second focusing objective lens 15 to the first beam splitter 9 are equal. The first lens 16 is used to equivalently refract the light entering the wafer 11. The aperture stop 17 is located behind the first lens 16 and at the center of the optical path, and is used to limit the imaging of stray light. The second CCD assembly 18 is on a one-dimensional moving platform with a moving accuracy in the micron order. By moving the second CCD assembly 18 left and right, the position of the multi-focus in the sample can be judged. The second CCD assembly 18 synchronously collects the laser spot images in the feedback optical path, and adopts a three-stage processing flow to realize the extraction of the circle center coordinates and radius of the spot. Specifically, it includes wavelet threshold filtering: by performing multi-scale wavelet decomposition on the image, using the difference in the energy distribution of the signal and noise in the wavelet domain to divide the signal and noise in the image, and removing the noise by setting the coefficients to zero to obtain a denoised image; edge detection: based on the Canny operator to extract the edge of the laser spot; least squares circle fitting: based on the edge point set to construct an objective function, and through iterative optimization to calculate the spot center coordinates (x, y) and radius r. When the laser is in the ideal focusing state, the spot diameter reaches the minimum value, and thus the focal depth inside the wafer 11 can be inversely calibrated. According to the principle of optical path symmetry, the distance from the first lens 16 to the second CCD assembly 18 is equivalent to the axial position where the focus is inside the wafer 11.

[0049] Among them, as Figure 2 shown, the spacing between the multi-focus spots generated by the spatial light modulator 7 is Δd. The wafer 11 is driven by the three-dimensionally movable first moving platform 12 to laterally step by an integer multiple of Δd, and the second CCD assembly 18 can continuously capture the spot images corresponding to each focus. Specifically, the laser modulated by the spatial light modulator 7 forms an equidistant multi-focus array inside the wafer 11, with a uniform spacing of Δd, and the number and position of the foci are determined by the spatial light modulator 7. By driving the wafer 11 to laterally step by an integer multiple of Δd through a precision displacement stage, the second CCD assembly 18 can continuously capture the spot images corresponding to each focus, and can monitor the consistency of the focus spacing in real time and verify the modulation accuracy of the spatial light modulator 7.

[0050] Among them, the depth reinforcement learning focusing algorithm adopts the SAC algorithm. During the training process, the reward function of the laser spot radius is adjusted to converge to zero, and at the same time, a scaling factor of 10 is added to increase the training speed.

[0051] Specifically, the training process of the depth reinforcement learning focusing algorithm, as Figure 4 shown, includes the following steps:

[0052] Step S1: First, obtain the initial horizontal position of the second CCD component 18 of the camera. Subsequently, initialize the algorithm, set the reward function, hyperparameters, and the environment. The present invention adopts the SAC algorithm in deep reinforcement learning. The SAC algorithm is based on the theory of maximum entropy reinforcement learning. Entropy is used here as a measure of the uncertainty of a random variable, and the entropy H is defined as:

[0053] ;

[0054] X is a random variable with a probability density function p(x), represents the degree of randomness of the policy π in state s. In maximum entropy reinforcement learning, entropy is introduced as a regularization term in the objective function of reinforcement learning, defined as:

[0055] ;

[0056] where α is the regularization coefficient, controlling the exploration of the policy. Reasonable selection of α helps to accelerate policy learning and reduce the risk of falling into a poor local optimum. and represent the action and state at time t, respectively.

[0057] In the SAC algorithm, based on the idea of the double Q network, the network with the smaller Q value is selected each time for update to alleviate the problem of overestimation of Q values. At the same time, based on the reparameterization trick, the loss function of the policy can be expressed as:

[0058] ;

[0059] is a noise random variable.

[0060] In the SAC algorithm, the algorithm can automatically adjust the entropy regularization term. After mathematical simplification, the loss function of α can be expressed as:

[0061] ;

[0062] When the entropy of the policy is lower than the target value H 0 , the training objective will increase the value of α; while when the entropy of the policy is higher than the target value H 0 , the training objective will decrease the value of α.

[0063] Step S2: Obtain the laser spot image through the second CCD component 18. The extraction of the center coordinates and radius of the spot circle is realized by adopting a three-level processing flow. Specifically, it includes wavelet threshold filtering: by performing multi-scale wavelet decomposition on the image, using the difference in the energy distribution of the signal and noise in the wavelet domain to divide the signal and noise in the image, and removing the noise by setting the coefficients to zero to obtain a denoised image. Edge detection: Extract the edge of the laser spot based on the Canny operator. Least squares circle fitting: Construct an objective function based on the edge point set, and calculate the center coordinates (x, y) and radius r of the spot through iterative optimization. In order to more effectively punish invalid behaviors and reward valid behaviors during the training process, the reward function of the laser spot radius in the feedback optical path is adjusted to converge to zero. At the same time, a scaling factor of 10 times is added to increase the training speed. Thus, the reward function is set as:

[0064] ;

[0065] The smaller the laser spot radius in the feedback optical path, the closer it is to the focal position, and the larger the reward function; conversely, the larger the laser spot radius in the feedback optical path, the farther it is from the focal position, and the smaller the reward function.

[0066] Step S3: The environment of the SAC algorithm consists of a state space and an action space. The state space is a one-dimensional space, representing the lateral position x1 of the second CCD component 18. The action space is also a one-dimensional space [-1, 1], indicating that the range of movement of the second CCD component 18 in each step is [-1, 1] × search step size, and the minimum value of the search step size is the displacement accuracy δ of the motor.

[0067] The process of the SAC algorithm is as follows. First, randomly initialize the current value network , and the policy network , and initialize the same parameters for the target network; then, initialize the experience replay pool R; for each round of training, first obtain the initial state of the environment ; in each step of training, select an action according to the current policy and execute it to obtain a reward , update the environmental state to , and then store the quadruple in the experience replay pool. Sample N quadruples from the replay pool, and calculate the target value using the target network, and then minimize the loss function to update the two value networks. Sample actions through the reparameterization technique, and update the current policy network with the loss function . At the same time, update the entropy regularization term coefficient α and the target network.

[0068] Thus, the algorithm outputs the corresponding action according to the current state space and the reward function, that is, adjust the axial position of the first lens 16.

[0069] Step S4: According to the current training progress, if the current round of training is not completed, i.e., the number of steps then return to Step S2. To avoid the problem of recognizing other focal positions due to excessive number of steps and the displacement accuracy δ of the motor, the displacement accuracy of the motor must be on the order of μm and δ×N < Δd / 2. If the current round of training is completed but the entire training process is not completed, i.e., the number of steps , the number of training rounds , calculate the total reward for the current round and reset the second CCD component 18 of the camera to the initial position. If the entire training process is completed, then reach Step S5.

[0070] Step S5: After the training is completed, the above-mentioned trained model information will be stored so that the SAC algorithm can load this trained model for the autofocus process. Finally, the second CCD component 18 can observe the laser spot at the focal position and obtain the cutting depth of the laser wafer 11 for hidden cutting, thereby real-time positioning and monitoring the hidden cutting spot.

[0071] Among them, the first optical path structure includes a double mirror assembly 2. The double mirror assembly 2 is used to adjust the transmission path of the laser beam emitted by the laser 1. The adjusted laser beam enters the center of the subsequent optical device and is incident on the spatial light modulator 7 at an incident angle less than 10°. The first optical path structure also includes a laser beam expander 3, a half-wave plate 4, a polarization beam splitter 5, and a first doublet lens assembly 6. The laser beam adjusted by the double mirror assembly 2 sequentially passes through the laser beam expander 3, the half-wave plate 4, the polarization beam splitter 5, and the first doublet lens assembly 6 and then enters the spatial light modulator 7. The laser beam expander 3 is used to compress the laser divergence angle and improve the beam quality. The polarization direction of the incident laser after passing through the half-wave plate 4 and the polarization beam splitter 5 is consistent with the response direction of the spatial light modulator 7. The two lenses in the first doublet lens assembly 6 form a 4f system for modulating the beam. The laser generates a multi-focal spot after passing through the spatial light modulator 7.

[0072] Among them, a second doublet lens assembly 8 is provided between the spatial light modulator 7 and the first beam splitter 9. The first doublet lens assembly 6 and the second doublet lens assembly 8 respectively form 4f systems for modulating the laser beam.

[0073] Among them, an illumination light source 20 is coaxially provided directly above the focusing optical path. The illumination light source 20 is divided into two illumination optical paths by the second beam splitter 21 and respectively enter the first CCD component 13 and the second CCD component 18. The illumination beam emitted by the illumination light source enters the first CCD component 13 and the second CCD component 18 for supplementary illumination to improve the imaging quality and detection accuracy of the CCD camera.

[0074] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A laser focus spot position positioning and monitoring device, characterized in that: It includes a laser, a first optical path structure, a spatial light modulator, a first beam splitter, a first focusing lens, a second focusing lens, a first lens, a first CCD component, a second CCD component, a first motion platform, a wafer, a second motion platform and a control system; After the laser emits a laser beam, it is deflected and modulated by the first optical path structure and then incident on the spatial light modulator. After the laser beam passes through the spatial light modulator, a multi-focal spot is generated, and then it is divided into a focusing optical path and a feedback optical path by the first beam splitter. After the focusing optical path is focused by the first focusing objective lens, it enters the interior of the wafer placed on the first motion platform for laser invisible cutting operation. The first CCD component is used to collect wafer surface image information and locate the wafer position. After the feedback light path is focused by the second focusing objective lens, it passes through the first lens and finally enters the second CCD component. The second motion platform drives the second CCD component to move left and right along the optical axis. The second CCD component synchronously collects the spot image corresponding to each focus in the feedback light path and extracts the center coordinates and radius of the laser spot circle. The control system includes an image processing module, an algorithm processing module and a signal control module. The image processing module executes the image processing process in parallel, and obtains the wafer contour and surface microstructure characteristics and the center coordinates and radius of the laser spot circle according to the wafer surface image information collected by the first CCD component and the laser spot image information in the feedback light path collected by the second CCD component. The algorithm processing module integrates the wafer positioning and correction algorithm and the deep reinforcement learning focus search algorithm. The signal control module controls the positioning and correction of the wafer based on the algorithm results of the algorithm processing module and controls the lateral movement of the second CCD component for automatic focus search based on the algorithm training model. The second CCD component observes the laser spot at the focal position in real time and obtains the cutting depth of the wafer laser hidden cutting, thereby locating and monitoring the hidden cutting spot in real time.

2. The laser focus spot position positioning and monitoring device according to claim 1, characterized in that: The first focusing lens and the second focusing lens are of the same model and are at the same distance from the first beam splitter. The first lens has the same refractive index as the wafer. The distance from the first lens to the second CCD component is the axial position of the focus inside the wafer.

3. The laser focus spot position positioning and monitoring device according to claim 1, characterized in that: The spacing between the multi-focal light spots generated by the spatial light modulator is Δd. The first motion platform drives the wafer to move laterally by an integer multiple of Δd, and the second CCD component can continuously capture the light spot image corresponding to each focus.

4. The laser focus spot position positioning and monitoring device according to claim 1, characterized in that: The deep reinforcement learning focus finding algorithm adopts the SAC algorithm. During the training process, the reward function of the laser spot radius is adjusted to converge to zero, and a 10-fold scaling factor is added to increase the training speed.

5. The laser focus spot position positioning and monitoring device according to claim 1, characterized in that: The first optical path structure includes a double reflector assembly, which is used to adjust the transmission path of the laser beam emitted by the laser. The adjusted laser beam enters the center of the subsequent optical device and is incident on the spatial light modulator at an incident angle of less than 10°.

6. The laser focus spot position positioning and monitoring device according to claim 5, characterized in that: The first optical path structure also includes a laser beam expander, a half-wave plate, a polarization beam splitter and a first double lens assembly. The laser beam adjusted by the double reflector assembly passes through the laser beam expander, the half-wave plate, the polarization beam splitter and the first double lens assembly in sequence and then enters the spatial light modulator.

7. The laser focus spot position positioning and monitoring device according to claim 6, characterized in that: A second double lens assembly is provided between the spatial light modulator and the first beam splitter.

8. The laser focus spot position positioning and monitoring device according to claim 7, characterized in that: The first double lens assembly and the second double lens assembly respectively constitute a 4f system for modulating the laser beam.

9. The laser focus spot position positioning and monitoring device according to claim 1, characterized in that: An illumination light source is coaxially arranged just above the focusing light path. The illumination light source is divided into two illumination light paths by a second beam splitter and enters the first CCD component and the second CCD component respectively.

10. The laser focus spot position positioning and monitoring device according to claim 1, characterized in that: The image processing module adopts a three-level processing flow of image median filtering, image threshold segmentation, and edge detection to process the wafer surface image information collected by the first CCD component to obtain the wafer contour and surface microstructure characteristics; the image processing module adopts a three-level processing flow of wavelet threshold filtering, edge detection, and least squares circle fitting to process the laser spot image information in the feedback light path collected by the second CCD component to obtain the center coordinates and radius of the laser spot circle.

Images