Van der Waals heterojunction automatic preparation instrument based on machine vision

Abstract

The invention provides a Van der Waals heterojunction automatic preparation instrument based on machine vision, which is characterized in that a precise hardware system integrating an optical microscope, a sample displacement table, a rotating table, a glass slide displacement table, a focusing lifting table, a CCD (Charge Coupled Device) camera and a temperature regulator is constructed; and the controller executes a computer program to solve the manual operation bottleneck in Van der Waals heterojunction preparation. A controller firstly obtains a microscope view image collected by a CCD camera, analyzes the image by using a YOLOv11 instance segmentation model, identifies contour and position information of a two-dimensional material and a Newton ring, and realizes automatic detection and positioning of the material; according to the information, the sample displacement table is coordinated to control the silicon wafer to move, the rotating table adjusts the angle, the glass slide displacement table executes the picking action, a temperature regulator is combined to assist material transfer, and picking, accurate positioning and automatic stacking of two-dimensional materials are achieved.

Description

Technical Field

[0001] This invention relates to the field of laboratory instruments, and in particular to an automated instrument for the preparation of van der Waals heterojunctions based on machine vision. Background Technology

[0002] Van der Waals heterostructures are artificial heterostructures formed by vertically stacking different types of two-dimensional materials through van der Waals forces. Because the layers are held together only by weak van der Waals forces, lattice mismatch problems are avoided, allowing for arbitrary combinations of different materials. By precisely controlling the stacking order, rotation, and alignment of the layers, the electronic band structure of the heterostructure can be tuned, producing novel physical phenomena such as Mohr's superlattices and strongly correlated electronic states. Van der Waals heterostructures are considered an important pathway to extending Moore's Law and driving the development of next-generation electronic devices.

[0003] Currently, the preparation of high-quality van der Waals heterostructures mainly relies on dry transfer printing technology. This technology uses polydimethylsiloxane (PDMS) and polycarbonate (PC) or polypropylene carbonate (PPC) as transfer media, and with the assistance of an optical microscope, mechanically exfoliated two-dimensional materials are transferred and stacked layer by layer. Dry transfer printing technology can ensure interface cleanliness and avoid solvent contamination, making it the preferred method for preparing high-quality heterostructures.

[0004] However, in existing van der Waals heterojunction fabrication processes, the two-dimensional material obtained through mechanical exfoliation is randomly distributed on the silicon wafer, and finding suitable materials is time-consuming. The entire transfer process is highly dependent on manual operation, and fabricating a single heterojunction device typically takes several hours or even days. Furthermore, manual operation makes it difficult to guarantee the consistency of accuracy and quality for each fabrication, affecting the repeatability of device performance. In addition, errors exist in manual alignment and angle control, making it difficult to achieve sub-micron level precise positioning and sub-degree level angle control.

[0005] In view of the above, this application is hereby submitted. Summary of the Invention

[0006] This invention discloses an automated van der Waals heterojunction fabrication instrument based on machine vision, aiming to solve the problems of low efficiency, insufficient repeatability, and inadequate precision in existing manual dry transfer techniques. This invention provides an automated van der Waals heterojunction fabrication instrument based on machine vision, comprising: an optical microscope, a sample displacement stage disposed below the optical microscope, a rotary stage mounted on the sample displacement stage, a slide displacement stage disposed in the optical observation path of the optical microscope, a focusing lifting stage connected to the optical microscope, a CCD camera mounted on the imaging end of the optical microscope, and a temperature regulator thermally connected to the slide displacement stage. The controller is electrically connected to the control terminal of the sample displacement stage, the control terminal of the rotary stage, the control terminal of the slide displacement stage, the control terminal of the focusing lifting stage, and the control terminal of the temperature regulator; and the CCD camera is electrically connected to the input terminal of the controller. The controller is configured to execute a computer program stored therein to perform the following steps: The microscope field-of-view image acquired by the CCD camera is obtained, and the image is analyzed based on the YOLOv11 instance segmentation model to identify the contour and position information of the two-dimensional material and Newton's rings. Based on the contour and position information, the sample displacement stage, rotary stage, and glass slide displacement stage are controlled to move in coordination to achieve the picking, positioning, and automated stacking of two-dimensional materials.

[0007] Preferably, before acquiring the microscope field-of-view image captured by the CCD camera, the method further includes: Acquire multiple frames of images continuously captured by the CCD camera, and calculate the sharpness value of each frame, wherein the sharpness value is obtained by the DWT sharpness measurement algorithm; Based on the position corresponding to the image with the highest sharpness value, the focusing platform is moved from the adaptive lowest point to that position to complete automatic focusing.

[0008] Preferably, the step of controlling the sample displacement stage, rotary stage, and glass slide displacement stage to move in coordination based on the contour and position information to achieve the picking, positioning, and automated stacking of two-dimensional materials specifically involves: The sample displacement stage is controlled to move along an S-shaped path, traversing the lower and upper halves of the silicon wafer in sequence, while acquiring and recording the position coordinates of all detected two-dimensional materials. Based on the target material location selected by the user, the sample displacement stage is controlled to move the target material to the center of the field of view; The slide displacement stage is controlled to descend. When Newton's rings are detected, the slide displacement stage is controlled to remove the surrounding debris from the four directions of left, right, up, and down from the target material according to the relative position of Newton's rings and the target material.

[0009] Preferably, after controlling the slide displacement stage to remove debris around the target material from four directions, the method further includes: The optical microscope magnification is reduced, and the CCD camera exposure value is adjusted. According to the preset stacking design parameters, the rotary table is controlled to rotate to a specified angle. Based on the coordinate transformation after rotation, the sample displacement stage is controlled to move and the target material is positioned at the design location. The slide displacement stage is controlled to descend along the Z-axis, and the expansion state of the Newton's rings is detected in real time.

[0010] Preferably, after controlling the Z-axis descent of the slide displacement stage and detecting the expansion state of Newton's rings in real time, the method further includes: When the edge of the Newton's rings is detected to be close to the outline of the target material, the temperature regulator is controlled to heat up to the preset temperature range, causing the stamp material to thermally expand and the Newton's rings to spread until they completely cover the target material. After maintaining the temperature for a preset time, the temperature regulator is controlled to cool down to below room temperature, causing the stamp material to shrink due to cold. When the Newton's rings are detected to have contracted and separated from the target material contour, the Z-axis of the slide displacement stage is slowly raised to complete the material transfer.

[0011] Preferably, it further includes: switching to a corner stacking mode when two materials are detected to be imported consecutively, wherein the execution process of the corner stacking mode is as follows: Control the slide displacement stage to move along the Y-axis to the target position of 1 mm, and control the focusing lifting stage to refocus on the stamp surface; Obtain a material image of the stamp surface, determine the actual position and angle of the transferred material using a contour matching algorithm, and calculate the position compensation amount based on the deviation between the matching result and the design parameters. By controlling the X and Y axis movement of the slide displacement stage with corresponding compensation amounts, the next layer of material can be transferred after alignment.

[0012] Preferably, the step of acquiring the microscope field-of-view image captured by the CCD camera and analyzing the image based on the YOLOv11 instance segmentation model to identify the contour and position information of the two-dimensional material and Newton's rings specifically involves: The YOLOv11 instance segmentation model is trained and optimized based on a self-built materials and Newton's rings dataset to perform instance segmentation on the input image and generate the initial segmentation mask for materials and Newton's rings. When the segmentation contour is determined to be inaccurate, the initial segmentation mask is optimized through contour editing and repair. Contour information is extracted from the optimized segmentation mask, and the center coordinates, area, perimeter, and roundness parameters of the contour are calculated. For Newton's ring detection results, when the roundness, size, and color meet the preset thresholds, it is determined to be a valid Newton's ring.

[0013] This invention provides an automated van der Waals heterostructure fabrication instrument based on machine vision. It constructs a sophisticated hardware system integrating an optical microscope, sample displacement stage, rotary stage, slide displacement stage, focusing stage, CCD camera, and temperature regulator. A controller executes a computer program to overcome the bottleneck of manual operation in van der Waals heterostructure fabrication. The controller first acquires the microscope field-of-view image from the CCD camera, analyzes the image using the YOLOv11 instance segmentation model, and identifies the contours and positions of the two-dimensional material and Newton's rings, achieving automatic material detection and positioning. Based on this information, it coordinates the sample displacement stage to control the silicon wafer movement, the rotary stage to adjust its angle, and the slide displacement stage to perform pick-up actions. Combined with the temperature regulator to assist in material transfer, it achieves the picking up, precise positioning, and automated stacking of two-dimensional materials. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of an automated van der Waals heterojunction fabrication instrument based on machine vision provided by the present invention; Figure 2 This is a schematic diagram of the controller execution steps provided by the present invention; Figure 3 This is a diagram illustrating the segmentation and positioning effects of YOLOv11 provided by this invention; Figure 4 This is a diagram illustrating the segmentation and positioning effect of Newton's rings provided by the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0017] This invention discloses an automated van der Waals heterojunction fabrication instrument based on machine vision, aiming to solve the problems of low efficiency, insufficient repeatability, and inadequate precision in existing manual dry transfer techniques. Please see Figure 1This invention provides an automated preparation instrument for van der Waals heterojunctions based on machine vision, comprising: an optical microscope, a sample displacement stage disposed below the optical microscope, a rotary stage mounted on the sample displacement stage, a slide displacement stage disposed in the observation optical path of the optical microscope, a focusing lifting stage connected to the optical microscope, a CCD camera mounted on the imaging end of the optical microscope, and a temperature regulator thermally connected to the slide displacement stage. The controller is electrically connected to the control terminal of the sample displacement stage, the control terminal of the rotary stage, the control terminal of the slide displacement stage, the control terminal of the focusing lifting stage, and the control terminal of the temperature regulator; and the CCD camera is electrically connected to the input terminal of the controller. Specifically, in this embodiment, an optical microscope is vertically positioned above the work platform to magnify and observe micron-sized two-dimensional materials. The microscope is equipped with 20x and 50x magnification lenses, which can be switched according to different operational needs. The 50x lens is mainly used for precise material positioning and observation, while the 20x lens is used for operations with a larger field of view, such as debris removal and material stacking. A sample displacement stage is positioned directly below the optical microscope. This stage has two degrees of freedom of movement in the X and Y axes, used to precisely control the position of the silicon wafer carrying the two-dimensional material in the horizontal plane. Its movement accuracy reaches the micron level, and its movement range covers the entire silicon wafer area, allowing the microscope to observe material at any position on the silicon wafer. A rotary stage is mounted on the upper surface of the sample displacement stage, forming a stacked structure. The silicon wafer is placed directly on the rotary stage. This rotary stage can rotate 360 ​​degrees around its vertical axis with a rotation accuracy of 0.1 degrees, used to achieve precise angle control when stacking two-dimensional materials, playing a crucial role, especially in the preparation of heterostructures such as twisted graphene that require precise control of interlayer rotation angles.

[0018] The slide stage is positioned in the observation optical path of the optical microscope, above the sample stage. This stage has three degrees of freedom of movement in the X, Y, and Z directions. The Z-axis controls the raising and lowering of the stamp, enabling contact and separation between the stamp and the silicon wafer surface. The X and Y axes adjust the stamp's position in the horizontal plane. The stamp is fixed to the stage by a glass slide. The stamp material uses PDMS as a substrate, coated with PC or PPC material, utilizing the transparency and adhesion properties of these materials to achieve the pickup and transfer of two-dimensional materials. The focusing stage is mechanically connected to the microscope barrel, using a precision screw drive mechanism to move the microscope up and down along the Z-axis. The movement range is 0-10 mm, with a positioning accuracy of 0.1 μm. This is used to adjust the distance between the microscope and the object being observed to achieve precise focusing and ensure clear microscopic images.

[0019] The CCD camera is mounted on the imaging end of the optical microscope and connects to the microscope via a standard C-interface. This camera features high resolution and a high frame rate, enabling real-time image acquisition of the microscope's field of view. The acquired images are transmitted to the computer control system via a USB 3.0 interface, providing raw data for subsequent image processing and machine vision analysis. The temperature regulator is thermally connected to the slide stage via a thermally conductive material. This regulator incorporates heating and cooling elements, allowing for temperature control from room temperature to 60°C and from room temperature to 0°C, with a temperature control accuracy of ±0.5°C. Precise temperature control adjusts the thermal expansion and contraction characteristics of the stamp material, achieving dynamic control of the contact area between the stamp and the two-dimensional material.

[0020] The controller, as the core of the entire instrument, establishes electrical connections via RS232 serial ports with the control terminals of the sample displacement stage, rotary stage, slide displacement stage, focusing stage, and temperature regulator, enabling precise control of each actuator. Simultaneously, the CCD camera connects to the controller's input via a USB 3.0 interface, transmitting acquired image data to the control system for processing in real time. This control system integrates control software developed using a hybrid C++ and Python programming framework. The software embeds a trained YOLOv11 instance segmentation model and related image processing algorithms, enabling real-time processing of images acquired by the CCD camera, identification of the contours of two-dimensional materials and Newton's rings, and coordinated control of each actuator based on the identification results, achieving a fully automated preparation process for two-dimensional material heterostructures.

[0021] Please see Figure 2 The controller is configured to execute a computer program stored therein to perform the following steps: S101, acquire the microscope field-of-view image captured by the CCD camera, and analyze the image based on the YOLOv11 instance segmentation model to identify the contour and position information of the two-dimensional material and Newton's rings. In this embodiment, the controller can acquire microscope field-of-view images captured by the CCD camera in real time via the USB 3.0 interface. The image resolution is 2048×2048 pixels and the frame rate can reach 30fps. The acquired image data first enters the preprocessing module for format conversion and normalization, and then is input into the YOLOv11-based instance segmentation model for intelligent recognition and analysis.

[0022] It should be noted that the training of this YOLOv11 instance segmentation model is based on a self-built professional dataset. The dataset construction adopted a combination of pre-trained model-assisted annotation and precise manual annotation. The materials dataset contains microscopic images of two types of two-dimensional materials, graphene and hexagonal boron nitride (BN), under different lighting conditions, thicknesses, and morphologies. Each material has more than 1,000 annotated images. The Newton's rings dataset contains more than 800 images of Newton's rings generated by a stamp and a silicon wafer under different contact states and pressure conditions. All images have been precisely annotated at the pixel level, including the target category, contour boundary, and instance ID. Data augmentation techniques, including random rotation, brightness adjustment, and contrast variation, were used during model training to improve the model's generalization ability. After multiple rounds of iterative training and hyperparameter tuning, the final model achieved an mAP of 0.92 on the test set, accurately identifying two-dimensional materials and Newton's rings in various complex scenes.

[0023] Please combine Figure 3 When an input image enters the YOLOv11 model, the model's backbone network first extracts multi-scale features from the image, then performs feature fusion through a feature pyramid network, and finally, the detection head outputs the class probability and instance attribution for each pixel, generating an initial segmentation mask for the material and Newton's rings. Due to the complexity of real-world application scenarios, such as blurred material edges, background interference, and uneven lighting, the initial segmentation mask may have inaccurate boundaries or local missing parts. In such cases, the system automatically judges the segmentation quality. When the completeness of the segmentation contour is detected to be lower than a set threshold, or when the user manually determines that the contour is inaccurate through the interface, the system will initiate a contour optimization process.

[0024] Contour optimization employs an interactive editing and repair mechanism. Users can adjust the initial segmentation results using the software's editing tools, including adding missing parts, deleting erroneous regions, and smoothing contour boundaries. The system automatically optimizes the contour based on user editing and original image features, using image processing algorithms such as morphological operations, edge detection, and region growing to ensure accurate material boundaries. After optimization, the system extracts complete contour information from the corrected segmentation mask, obtains the contour point set using the `findContours` function from the OpenCV image processing library, and then calculates various geometric parameters, including calculating the contour center coordinates (cx, cy) = (M10 / M00, M01 / M00) using image moments, calculating the contour area using the `contourArea` function, calculating the contour perimeter using the `arcLength` function, and roundness using the formula 4π × area / perimeter. 2 The calculation yielded the result.

[0025] Please combine Figure 4The detection of Newton's rings has its own unique characteristics. Newton's rings are concentric circular patterns formed by light interference when a stamp contacts a silicon wafer. Their shape and color features are relatively obvious but easily affected by environmental interference. After obtaining the initial detection result, the system performs multiple verifications to ensure accuracy. First, it checks the roundness parameter; only when the roundness is greater than 0.8 is it considered a valid Newton's ring, as a true Newton's ring should exhibit a relatively regular circle. Second, it checks the area size. Depending on the stamp size and microscope magnification, the area of ​​a valid Newton's ring should be within the range of 200-5000 square pixels. Too small an area may indicate noise, while too large an area may indicate other interference. Finally, it checks the color characteristics. A true Newton's ring exhibits a specific interference color pattern under a microscope. The system verifies this by analyzing the hue and saturation distribution in the HSV color space. Only when all three conditions are met is the system considered a valid Newton's ring, and its center coordinates, radius, and other information are transmitted to the motion control module for subsequent automated operation control.

[0026] S102, based on the contour and position information, control the sample displacement stage, rotary stage and glass slide displacement stage to move in coordination to realize the picking, positioning and automated stacking of two-dimensional materials.

[0027] In this embodiment, after the system completes initialization and loading of the material detection model, it begins to execute the automated material search and stacking process. The entire process is achieved by coordinating the precise movement of each execution component through a computer control system.

[0028] In the material search phase, the control system first moves the sample displacement stage to the preset starting position of the silicon wafer, namely the lower left corner of the silicon wafer at coordinates (0,0). Then, it controls the focusing stage to automatically focus, so that the microscope is clearly focused on the surface of the silicon wafer. Next, the sample displacement stage begins to systematically scan according to the preset S-shaped path. The specific path is planned as follows: first scan the lower half of the silicon wafer, the sample displacement stage moves to the right from the lower left corner, and each movement step is 80% of the microscope's field of view width to ensure sufficient overlap and avoid omissions. When the detection reaches the right boundary of the silicon wafer, the displacement stage moves down one field of view height and then moves to the left. This is repeated until the lower half of the silicon wafer is scanned. Then, the upper half is scanned in a similar way, that is, starting from the middle of the silicon wafer, moving to the left, reaching the left boundary, moving up and then to the right, forming a continuous S-shaped scanning trajectory. At each scanning position, the CCD camera acquires the current field of view image, and the YOLOv11 model analyzes the two-dimensional materials in the image in real time. The contour information and center coordinates of each material are converted into the global coordinate system of the silicon wafer and stored in the database. At the same time, the material distribution is displayed in real time on the silicon wafer map on the software interface. The entire silicon wafer scanning process takes about 15-20 minutes and can detect and record hundreds of two-dimensional material layers on the silicon wafer.

[0029] After completing the full-wave scan, the user can view all detected materials through the software interface, select the target material that meets the experimental requirements, and the system will obtain the precise position of the target material in the silicon wafer coordinate system. Then, it will calculate the X and Y displacement required to move from the current position to the target position and send control commands to drive the sample displacement stage to move. The movement process uses a trapezoidal velocity curve to ensure positioning accuracy. When the target material enters the microscope field of view, the system will make fine adjustments and provide real-time feedback on the material position through image recognition. The fine-tuning of the sample displacement stage will make the target material accurately positioned in the center of the field of view, with a positioning accuracy of ±2 micrometers.

[0030] After the target material is located, the debris removal stage begins, as unwanted material fragments often exist around the target material, affecting subsequent stacking. The system first switches the microscope from 50x to 20x to obtain a wider field of view, while simultaneously adjusting the CCD camera's exposure parameters to adapt to the magnification change. Then, the slide stage is rapidly lowered along the Z-axis from a safe height to approximately 3.5mm above the target material. At this point, the Newton's rings detection thread is activated, and the system continues to slowly lower the slide stage at a rate of 0.01mm / s, while simultaneously analyzing the image in real-time at 30fps. When the first Newton's rings are detected, it indicates that the stamp has approached or contacted the silicon wafer surface, and the system immediately records the current Z-axis position and the coordinates of the Newton's rings center.

[0031] The waste removal process employs a four-directional sequential operation strategy. First, the system calculates the relative position between the center of Newton's rings and the center of the target material. Then, it controls the X-axis movement of the slide displacement stage to position the stamp center 0.15 mm to the left of the target material. This optimized distance avoids contact with the target material while effectively covering the waste area. Next, the Z-axis is controlled to descend at an extremely low speed of 0.005 mm / s, continuously monitoring the morphological changes of Newton's rings. When the system detects that the edge of Newton's rings extends to within 0.04 mm of the waste material, it determines that the stamp has made sufficient contact with the waste material. At this point, the descent is stopped, and the position is maintained for 3-5 seconds to allow the PC material to fully adhere to the waste. Then, the Z-axis is slowly raised at a speed of 0.01 mm / s, using the adhesiveness of the PC material to peel the waste material off the silicon wafer surface. The raising process continues until Newton's rings are no longer detectable in the image, indicating that the stamp has completely left the silicon wafer surface.

[0032] After the debris removal on the left side is completed, the system moves the center of the stamp to the right, top, and bottom of the target material by 0.15mm each. The process of pressing, adhering, and lifting is repeated at each position. The removal operation in four directions ensures that a clean ring-shaped area is formed around the target material, creating ideal conditions for subsequent precise stacking. The entire debris removal process is controlled in a closed loop through visual feedback. If incomplete removal is detected, the system will automatically adjust the parameters and repeat the removal action until the preset cleanliness standard is reached.

[0033] After clearing debris from the target material, the system enters the material stacking stage. At this point, the system automatically switches the magnification of the optical microscope from 50x to 20x. This magnification reduction expands the field of view by 2.5 times, allowing the operator to observe a wider range of material distribution, which is beneficial for overall control during the subsequent stacking process. Simultaneously, the system automatically adjusts the exposure parameters of the CCD camera according to the magnification change, adjusting the exposure time from 20ms at 50x magnification to 8ms at 20x magnification, and reducing the ISO value from 800 to 400, ensuring that images with moderate brightness and good contrast can be obtained at different magnifications.

[0034] After the magnification and exposure are adjusted, the system reads the stacking parameters that the user has set in the material design interface. The stacking parameters include the final position coordinates of the target material, the rotation angle, and the interlayer alignment requirements. According to the design requirements, the control system sends precise angle control commands to the rotary table. The rotary table achieves angle positioning through a built-in stepper motor and a high-precision encoder. For example, when preparing corner graphene, the material needs to be rotated by 1.1° or other magic angles, and the rotation accuracy can reach 0.01°. The rotation process uses acceleration and deceleration control to avoid mechanical vibration from affecting the positioning accuracy.

[0035] After the rotation, the original coordinate system changes. The system needs to perform coordinate transformation calculations to ensure that the material can be accurately moved to the design position. Specifically, if the material needs to move from the current position (x0, y0) to the design position (x1, y1) and rotates by an angle θ, the actual distance to be moved is Δx = (x1 - x0)cosθ + (y1 - y0)sinθ and Δy = -(x1 - x0)sinθ + (y1 - y0)cosθ. The system controls the sample displacement stage to move precisely in the X and Y directions based on the calculation results. During the movement, visual servo feedback is used to detect the material position in real time and make fine adjustments. Finally, the target material is positioned at the design position with the positioning error controlled within ±1 micrometer.

[0036] After the material is precisely positioned, the system begins the crucial transfer process, controlling the slide displacement stage to descend along the Z-axis from its current safe height at an initial speed of 0.1 mm / s. When the slide is approximately 0.5 mm from the silicon wafer surface, the speed is reduced to 0.01 mm / s. Simultaneously, the Newton's rings detection algorithm is activated, analyzing the images captured by the CCD camera in real time at a frequency of 30 fps. Newton's rings are colored concentric rings formed by the interference of light when the PC or PPC material at the bottom of the stamp comes into contact with the silicon wafer surface. Their appearance indicates that the stamp has contacted or is extremely close to the silicon wafer surface. The system precisely controls the stamp's downward pressing process by detecting the timing of the Newton's rings' appearance, their size changes, and their expansion speed.

[0037] When the system first detects the appearance of Newton's rings, it immediately reduces the Z-axis descent speed to an extremely slow 0.005 mm / s. At this point, Newton's rings typically appear as a small colored circular spot with a diameter of approximately 0.2-0.3 mm. As the stamp continues to descend slowly, Newton's rings gradually expand outwards. The system uses image processing algorithms to extract the outline of Newton's rings in real time and calculates the shortest distance between its outer edge and the outline of the target material. When this distance decreases to 0.05 mm, it indicates that Newton's rings are about to contact the target material, and the system immediately initiates the temperature control program.

[0038] After receiving the heating command, the temperature regulator uses a built-in PID control algorithm to raise the stamp temperature from room temperature (25℃) to the preset range of 45-55℃, with the heating rate controlled at 2-3℃ / s to avoid thermal stress caused by excessively rapid heating. As the temperature rises, the PC or PPC material undergoes thermal expansion. Its viscoelasticity above the glass transition temperature softens the material and increases its viscosity. Newton's rings continue to diffuse outward as the material expands. The system continuously monitors the diffusion process of Newton's rings. When it detects that the Newton's rings completely cover the target material, i.e., the outer edge of the Newton's rings extends at least 0.02mm beyond the outline of the target material, heating is stopped and the current temperature is maintained for 6-8 seconds. This heat preservation process ensures that a stable van der Waals interaction force is formed between the PC material and the target two-dimensional material.

[0039] After the heat preservation is completed, the system controls the temperature regulator to start cooling at a rate of 3-4℃ / s, with a target temperature of 10-15℃. The setting below room temperature is to utilize the thermal shrinkage characteristics of PC material. As the temperature decreases, the PC material shrinks, and Newton's rings gradually shrink from the outside to the inside. The system monitors the shrinkage process of Newton's rings in real time. When the inner edge of Newton's rings is completely separated from the outline of the target material, i.e., a gap of more than 0.01mm appears between them, it indicates that the shrinkage of the PC material has completely "grabbed" the target material. At this time, the Z-axis of the slide displacement stage is controlled to start rising at an extremely slow speed of 0.002mm / s. During this process, the adhesive force of the PC material overcomes the van der Waals forces between the target material and the silicon wafer, peeling the material from the silicon wafer surface and transferring it to the stamp. The rising process continues until Newton's rings are no longer detectable in the field of view, which usually requires a rise of 0.3-0.5mm. Then, the system performs an automatic focusing to confirm that the target material has been successfully transferred to the bottom of the stamp.

[0040] When the system detects that the user has imported two materials consecutively in the design interface, it automatically determines that it is in corner stacking mode and executes the corresponding control flow. After the first layer of material is successfully transferred to the stamp, the system needs to accurately determine its actual position on the stamp in order to accurately align with the second layer of material. To this end, the slide displacement stage is controlled to move horizontally 1mm to the side along the Y-axis. This movement moves the stamp to the side area of ​​the target material, avoiding the material already transferred at the bottom of the stamp from affecting the observation of the second layer of material on the silicon wafer. After the movement is completed, the focusing lift is controlled to move up about 0.1mm, adjusting the focal plane from the surface of the silicon wafer to the bottom surface of the stamp. At this time, the CCD camera captures the image of the bottom of the stamp.

[0041] After focusing, the machine vision system analyzes the image of the stamp surface, uses a trained model to identify the first layer of material transferred to the stamp, and extracts its contour information. Since the stamp is made of transparent PDMS material, the transferred two-dimensional material is clearly visible under transmitted light illumination. The system compares the actual detected material contour with the theoretical contour in the design drawing using a contour matching algorithm. The matching algorithm uses a shape matching method based on Hu moments to calculate the similarity between the two contours. At the same time, it calculates the optimal registration transformation parameters using the least squares method, including the X-direction offset Δx, the Y-direction offset Δy, and the angle deviation Δθ. These deviations reflect the positional errors generated during the actual transfer process.

[0042] Based on the calculated deviation value, the system generates a position compensation command, controlling the X-axis movement of the slide displacement stage by -Δx distance and the Y-axis movement by -Δy distance for position correction. If there is an angular deviation, the system will add compensation to the rotation angle of the subsequent second layer material to ensure the final angular accuracy. After the position calibration is completed, the slide displacement stage moves back to its original position along the Y-axis. Then, the system positions the second target material according to the design parameters and repeats the aforementioned rotation, positioning, pressing, heating, cooling, and peeling process to accurately transfer the second layer material to the stamp with the existing first layer material. The rotation angle and alignment accuracy between the two layers directly determine the performance of the heterojunction. Through this visual feedback and position compensation mechanism, the system can achieve sub-micron level alignment accuracy and 0.1° angle control accuracy.

[0043] In one possible embodiment of the invention, before the system begins material inspection and image analysis, it is essential to ensure that the microscope obtains a clear, focused image, which is fundamental to all subsequent machine vision operations. Since the CCD camera used in this system does not have an autofocus function, autofocus is achieved through precise control of the focusing platform in conjunction with image sharpness analysis.

[0044] At the start of the autofocus process, the system first sets the focus search range, determining the adaptive lowest and highest points based on the silicon wafer thickness and stamp height. Typically, the lowest point is set 0.5mm below the silicon wafer surface, and the highest point is set 2mm above the silicon wafer surface. This range ensures coverage of all possible focus positions. The system controls the focusing platform to move upwards from the adaptive lowest point in 0.05mm increments, pausing at each position for 50ms to allow mechanical vibrations to dissipate. Then, a 2048×2048 pixel image is acquired via a CCD camera. Throughout the search process, 60-80 images are continuously acquired and temporarily stored in a circular buffer in the system memory.

[0045] For each captured image frame, the system uses the Discrete Wavelet Transform (DWT) sharpness measurement algorithm to calculate its sharpness value. The core of this algorithm is to perform multi-scale wavelet decomposition on the image and extract high-frequency detail coefficients as a sharpness measure. Specifically, the RGB color image is first converted to a grayscale image. Then, a two-dimensional wavelet transform is performed on the image using the Daubechies wavelet basis, decomposing it into four sub-bands: approximation coefficient LL, horizontal detail coefficient LH, vertical detail coefficient HL, and diagonal detail coefficient HH. The sharpness value S_DWT is calculated using the formula S_DWT = Σ(|LH|). 2 + |HL| 2 + |HH| 2 The algorithm is calculated as (M×N), where M and N are the height and width of the image, respectively. This algorithm is sensitive to the high-frequency components of the image and can accurately reflect the image's clarity.

[0046] To accelerate the sharpness calculation process, the system implemented a GPU-parallel version of the DWT algorithm using CUDA programming. The wavelet transform operation, which originally required serial execution on the CPU, was decomposed into multiple parallel kernel functions. Each CUDA thread is responsible for processing a small region of the image. Through reasonable thread block partitioning and shared memory usage, the sharpness calculation time for a single frame image was reduced from 80ms on the CPU to 3ms on the GPU, significantly improving the speed of autofocus. Simultaneously, the system also implemented a morphological sharpness measurement algorithm as an alternative. This algorithm evaluates sharpness by calculating the sum of grayscale values ​​after morphological gradient operations on the image, with the formula S_morph = Σ|(I⊕B - I⊖B)| / (M×N), where ⊕ and ⊖ represent morphological dilation and erosion operations, respectively, and B is the structuring element. This method is more robust to noise and can provide more stable sharpness evaluation under certain less-than-ideal lighting conditions.

[0047] After completing image acquisition and sharpness calculation at all locations, the system constructs a position-sharpness curve in memory. This curve typically exhibits a single-peak characteristic, with the peak value corresponding to the optimal focusing position. The system determines the optimal focusing position Z_best by finding the maximum sharpness value. Considering the influence of mechanical backlash, the system does not move directly to the optimal position. Instead, it first moves the focusing platform back to the adaptive lowest point, and then moves unidirectionally from the lowest point to the Z_best position. This unidirectional movement strategy can eliminate the positioning error caused by the mechanical backlash of the platform's lead screw and guide rail, ensuring focusing accuracy.

[0048] After moving to the optimal focusing position, the system performs a secondary verification, acquiring the image again and calculating the sharpness value. If the sharpness value differs from the previously recorded maximum value by more than 5%, it indicates a possible positioning deviation or environmental change. The system then performs a fine search within ±0.1mm of the current position, re-finding the optimal focusing point in smaller steps of 0.01mm. This two-step focusing strategy, combining coarse and fine adjustments, ensures both focusing speed and accuracy. The entire automatic focusing process is typically completed within 5-8 seconds, with a focusing accuracy of ±0.5μm, providing a clear and stable image foundation for subsequent material identification and precise operations.

[0049] In practical applications, since the microscope may shift its focus due to temperature changes or mechanical drift during long-term operation, the system automatically performs a rapid focusing correction each time the sample stage is moved to a new position. This adaptive focusing mechanism ensures that high-quality microscopic images can be obtained throughout the entire automated preparation process.

[0050] This invention provides an automated van der Waals heterostructure fabrication instrument based on machine vision. It constructs a sophisticated hardware system integrating an optical microscope, sample displacement stage, rotary stage, slide displacement stage, focusing stage, CCD camera, and temperature regulator. A controller executes a computer program to overcome the bottleneck of manual operation in van der Waals heterostructure fabrication. The controller first acquires the microscope field-of-view image from the CCD camera, analyzes the image using the YOLOv11 instance segmentation model, and identifies the contours and positions of the two-dimensional material and Newton's rings, achieving automatic material detection and positioning. Based on this information, it coordinates the sample displacement stage to control the silicon wafer movement, the rotary stage to adjust its angle, and the slide displacement stage to perform pick-up actions. Combined with the temperature regulator to assist in material transfer, it achieves the picking up, precise positioning, and automated stacking of two-dimensional materials.

[0051] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions that fall within the scope of the present invention are within the scope of protection of the present invention.

Claims

1. A machine vision based automated preparation instrument for van der Waals heterojunctions, characterized by, The application relates to a two-dimensional material automatic stacking system based on optical microscopy, and belongs to the field of two-dimensional material automatic stacking. The system comprises the following: an optical microscope, a sample displacement table arranged below the optical microscope, a rotating table mounted on the sample displacement table, a slide displacement table arranged in an observation light path of the optical microscope, a focus lifting table connected with the optical microscope, a CCD camera mounted on an imaging end of the optical microscope, a temperature regulator in thermal connection with the slide displacement table, and a controller in electrical connection with a control end of the sample displacement table, a control end of the rotating table, a control end of the slide displacement table, a control end of the focus lifting table and a control end of the temperature regulator, and the CCD camera is in electrical connection with an input end of the controller. The controller is configured to execute a computer program stored therein to realize the following steps: acquiring a microscope field image collected by the CCD camera, and analyzing the image based on a YOLOv11 instance segmentation model to identify the contour and position information of a two-dimensional material and a Newton ring; and controlling the sample displacement table, the rotating table and the slide displacement table to move cooperatively according to the contour and position information, so as to realize picking, positioning and automatic stacking of the two-dimensional material. Before the microscope field image collected by the CCD camera is acquired, the following steps are further included: acquiring a plurality of images continuously collected by the CCD camera, and calculating a sharpness value of each image, wherein the sharpness value is obtained through a DWT sharpness measurement algorithm; and moving the focus lifting table from the adaptive lowest point to a position corresponding to the image with the maximum sharpness value, so as to complete automatic focusing. According to the contour and position information, the sample displacement table, the rotating table and the slide displacement table are controlled to move cooperatively, so as to realize picking, positioning and automatic stacking of the two-dimensional material, and the specific process is as follows: the sample displacement table is controlled to move along an S-shaped path to sequentially traverse the lower half and the upper half of a silicon wafer, and the position coordinates of all the detected two-dimensional materials are acquired and recorded; the sample displacement table is controlled to move the target material to the center of the field of view according to a target material position selected by a user; the slide displacement table is controlled to descend, and when the Newton ring is detected, the slide displacement table is controlled to remove the surrounding garbage materials from the left, right, top and bottom of the target material according to the relative position relationship between the Newton ring and the target material. After the slide displacement table removes the surrounding garbage materials of the target material from the four directions, the following steps are further included: the magnification of the optical microscope is reduced, the exposure value of the CCD camera is adjusted, the rotating table is controlled to rotate to a specified angle according to preset stacking design parameters; the sample displacement table is controlled to move to position the target material to a design position based on coordinate transformation after rotation; and the Z-axis of the slide displacement table is controlled to descend to detect the expansion state of the Newton ring in real time.

2. The machine vision based van der Waals heterojunction automated fabrication instrument according to claim 1, wherein, After the Z-axis of the slide displacement table is controlled to descend to detect the expansion state of the Newton ring in real time, the following steps are further included: when the edge of the Newton ring is detected to approach the contour of the target material, the temperature regulator is controlled to be heated to a preset temperature range, so that the stamp material is heat-expanded, and the Newton ring is diffused until the target material is completely covered; after the temperature is maintained for a preset time length, the temperature regulator is controlled to be cooled to below room temperature, so that the stamp material is cold-shrunk. ​ ​ 3. The machine vision based van der Waals heterojunction automated fabrication instrument of claim 1, wherein, ​ ​ ​ ​ 4. The machine vision based van der Waals heterojunction automated fabrication instrument according to claim 3, wherein, ​ ​ ​ ​ 5. The machine vision based van der Waals heterojunction automated fabrication instrument according to claim 4, wherein, ​ ​ ​ When the shrinkage of Newton's ring is detected to be separated from the target material profile, the Z-axis of the slide displacement stage is controlled to slowly rise, and the material transfer is completed.

6. The machine vision based van der Waals heterojunction automated fabrication apparatus of claim 4, wherein, Also included are: When two materials are continuously introduced, switch to the corner stacking mode, wherein the execution process of the corner stacking mode is: Control the Y-axis of the slide displacement stage to move to the target 1mm, and control the focusing stage to refocus to the surface of the stamp; Obtain the material image of the stamp surface, determine the actual position and angle of the transferred material through the contour matching algorithm, and calculate the position compensation amount according to the deviation of the matching result and the design parameters; Control the X and Y axes of the slide displacement stage to move the corresponding compensation amount, and realize the transfer of the next layer of material after alignment.

7. The machine vision based van der Waals heterojunction automated fabrication instrument of claim 1, wherein, The microscope field image collected by the CCD camera is obtained, and the image is analyzed based on the YOLOv11 instance segmentation model to identify the contour and position information of the two-dimensional material and Newton's ring, specifically: The YOLOv11 instance segmentation model is trained and optimized based on the self-built material and Newton's ring dataset to perform instance segmentation on the input image, generating an initial segmentation mask of the material and Newton's ring; When the segmentation contour is determined to be inaccurate, the initial segmentation mask is optimized through contour editing and repair, and the contour information is extracted from the optimized segmentation mask to calculate the center coordinates, area, perimeter, and circularity parameters of the contour; for the Newton's ring detection result, when the circularity, size, and color meet the preset threshold, it is determined as an effective Newton's ring.

Images