Interference-resistant wafer surface curvature measurement method and apparatus
By employing three phase mask wavefront encoding and deep learning processing, the problem of curvature calculation failure caused by laser interference destructive phases was solved, achieving high-precision and stable wafer surface curvature measurement, which is suitable for semiconductor industry field applications.
Patent Information
- Application Number
- CN202611079588.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-21
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-07-21
AI Technical Summary
Existing wafer surface curvature measurement methods cannot effectively capture the light spot during laser interference destructive phases, leading to curvature calculation failures and affecting thin film quality monitoring.
By employing a cubic phase mask for wavefront coding and deep learning processing, combined with a U-shaped network and heatmap regression algorithm, and through polarization beam splitting and optical module design, high-precision extraction of the center coordinates of the light spot and curvature calculation are achieved.
It effectively suppresses the influence of interference cancellation on measurement, ensures the continuity and high accuracy of wafer surface curvature data, adapts to complex semiconductor industrial environments, and reduces data acquisition costs.
Smart Images

Figure CN122590758B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of semiconductor testing technology, and in particular to an interference-resistant method and apparatus for measuring the curvature of a wafer surface. Background Technology
[0002] Metal-organic chemical vapor deposition (MOCVD) is the core equipment for fabricating epitaxial wafers of compound semiconductors such as gallium nitride and gallium arsenide. During epitaxial growth, warpage occurs due to lattice and thermal mismatches between the substrate and the epitaxial layer. Real-time monitoring of warpage curvature is crucial for controlling growth parameters and ensuring film quality.
[0003] In related technologies, the two-point laser reflection method is commonly used to measure warpage: two parallel laser beams are incident at a small angle onto the surface of a growing epitaxial wafer, and after reflection, they are imaged onto a camera. The radius of curvature is calculated by the change in the distance between the two laser spots. However, due to the high coherence of the laser, the two reflected beams and stray light from each beam interfere on the camera's target surface. When the growth of the thin film on the epitaxial wafer surface causes a change in the optical path difference, the interference pattern shows alternating bright and dark areas. At the moment of decoherence, the local brightness of the laser spot drops sharply or even disappears completely. At this time, if the camera uses a fixed exposure time and analog gain, it cannot effectively capture the laser spot, leading to the failure of the curvature calculation and causing a data breakpoint.
[0004] Therefore, there is an urgent need for an interference-resistant method and apparatus for measuring wafer surface curvature to solve the above problems. Summary of the Invention
[0005] The technical problem solved by this invention is to provide an interference-resistant wafer surface curvature measurement method and device, which can effectively suppress the influence of external interference factors such as ambient light and vibration on the measurement results, and realize high-precision and high-stability real-time wafer curvature detection.
[0006] According to a first aspect of the present invention, an interference-resistant method for measuring the curvature of a wafer surface is provided. The method includes: controlling a laser source to output linearly polarized light; inputting the linearly polarized light into a polarization beam splitting and displacement module to output two parallel orthogonally polarized beams with a fixed spacing; inputting the two orthogonally polarized beams into a wavefront encoding module for wavefront encoding to output two encoded first light waves; the wavefront encoding module includes a cubic phase mask; inputting the two first light waves into an optical module to output two second light waves with enlarged spot diameters and spacing; inputting the two second light waves into a reaction chamber to illuminate the wafer surface; acquiring reflected light returning from the wafer surface along the original path; acquiring an encoded image based on the reflected light; sending the encoded image into a deep learning processing module to obtain the center coordinates of the reconstructed two light spots; and calculating the curvature of the wafer surface based on the center coordinates and a formula for calculating curvature based on the spot positions.
[0007] Its beneficial effects lie in the fact that by using a cubic phase mask for wavefront encoding, the modulation transfer function (MTF) of the cubic phase mask remains approximately constant over a large defocus range, and its point spread function (PSF) exhibits a "trident"-shaped dispersion, preventing the appearance of completely zero dark spots in the energy distribution. Even if interference cancellation occurs, the energy is merely redistributed, rather than causing the signal to completely disappear, ensuring effective spot capture, avoiding curvature calculation failures, and ensuring continuous wafer surface curvature data during the process.
[0008] In one implementation, the surface profile height function z(x, y) of the cubic phase mask satisfies: (1) Where α is the modulation intensity coefficient; r0 is the half-aperture of the cubic phase mask; x is the abscissa with the center of the cubic phase mask as the origin; and y is the ordinate with the center of the cubic phase mask as the origin.
[0009] Its beneficial effects lie in the fact that the surface profile height function of the cubic phase mask adopts a cubic polynomial form, enabling precise and controllable design of the phase modulation depth through the modulation intensity coefficient and the half-aperture normalization factor. The modulation intensity coefficient determines the amplitude of phase modulation, and by adjusting this coefficient, a balance can be achieved between expanding the depth of field and the difficulty of image decoding. This cubic function form introduces a continuous and smooth phase change on the mask surface, causing the system's point spread function to exhibit an asymmetric diffusion morphology with multiple side lobes. The energy is continuously distributed in space without any completely zero dark spots. Even when coherent light interference is destructive, the reflected light signal can still maintain sufficient contrast and detectability. At the same time, this phase modulation method keeps the system's optical transfer function approximately unchanged over a large defocus range, significantly expanding the depth of field of the measurement system and effectively overcoming the problem of spot blurring or even signal loss caused by changes in wafer surface height or positional offset in the traditional two-point laser reflection method.
[0010] In one implementation, the deep learning processing module is equipped with a U-shaped network and a heatmap regression algorithm; the U-shaped network is used to generate two heatmaps based on the encoded image; the heatmap regression algorithm is used to obtain the position of the maximum pixel based on the two heatmaps, take a neighborhood of a preset size with the position of the maximum pixel as the center, and use the centroid method to obtain the center coordinates of the two reconstructed light spots.
[0011] Its beneficial effect lies in the fact that the two-stage localization strategy combining a U-shaped network with a heatmap regression algorithm significantly improves the accuracy and robustness of spot center coordinate extraction. The U-shaped network, with its encoder-decoder structure, can effectively learn multi-scale feature representations of spots in the encoded image. The two generated heatmaps not only accurately reflect the spatial distribution probability of the two spots but also retain rich contextual information and edge details, providing a reliable confidence distribution for subsequent localization.
[0012] In one implementation, the U-shaped network is obtained through the following training method: establishing an optical simulation model including a polarization beam splitter, a cubic phase mask, a reaction chamber window, and an epitaxial wafer; generating training pairs of coded images and real spot center coordinates through parameter scanning; performing convolution blurring, noise superposition, and background radiation addition processing on the coded images; generating real Gaussian heatmap labels using the real spot centers, and training the U-shaped network through the mean square error loss function.
[0013] Its beneficial effects lie in the fact that by establishing a complete optical simulation model including a polarization beam splitter, a cubic phase mask, a reaction chamber window, and an epitaxial wafer, the optical propagation process of a real measurement system can be accurately reproduced in a virtual environment, generating massive training samples without relying on a large amount of expensive experimental data. The parameter scanning strategy systematically traverses various possible combinations of operating parameters, including different polarization angles, phase modulation intensities, window thicknesses, and epitaxial wafer positions, ensuring that the training data covers various boundary conditions that may be encountered in practical applications, significantly enhancing the model's generalization ability. Data augmentation processing, such as convolutional blurring, noise superposition, and background radiation addition, is applied to the encoded image to effectively simulate complex factors such as optical aberrations, electronic noise, and ambient light interference present in the real measurement environment. This gives the trained network strong anti-interference capabilities, maintaining stable performance even under harsh industrial conditions. Using Gaussian heatmaps as training labels, combined with a mean squared error loss function, not only provides a smooth and continuous supervision signal, which is beneficial for the network to learn the probability distribution characteristics of the spot position, but also avoids the overfitting problem that may be caused by hard labels, giving the model's output heatmap better spatial continuity and physical interpretability. This simulation-based data-driven training method ensures the quality and diversity of training data while significantly reducing data acquisition and time costs, providing an efficient and feasible technical path for the application of deep learning models in the field of precision optical measurement.
[0014] In one embodiment, parameter scanning includes scanning with at least one of the following parameters: wafer surface curvature radius, interference phase difference between two light spots, wafer surface reflectivity, and thermal noise standard deviation.
[0015] Its beneficial effects lie in the fact that this multi-parameter, multi-dimensional scanning strategy not only significantly expands the coverage of training data but also discovers the coupling effects between parameters through cross-validation of parameter combinations, providing important references for optimizing optical system design. More importantly, this method can quickly generate massive amounts of labeled data in a virtual simulation environment, avoiding the tedious process of extensive experimental measurements and manual annotation required in traditional methods. This significantly reduces data acquisition costs and time costs, providing a solid data foundation for the efficient training and rapid iteration of deep learning models.
[0016] In one implementation, noise simulation during training includes superimposing shot noise and readout noise, and adding thermal radiation background, wherein the intensity of the thermal radiation background is 1% to 10% of the signal peak value.
[0017] Its beneficial effect is that by superimposing shot noise, readout noise and adding thermal radiation background with an intensity of 1% to 10% of the signal peak, this noise simulation strategy accurately reproduces the main interference factors in the real measurement environment, and significantly improves the robustness and anti-interference ability of the model.
[0018] In one implementation, pixel regions containing the first and second light spots in the encoded image are extracted as samples, and the coordinates of the center of the real light spot in each sample are recorded. The center of the real light spot is taken as the center of the heat map, the standard deviation σ = 3 pixels, the Gaussian kernel size is 15×15, and the peak value is normalized to 1, so that the heat map output by the U-shaped network presents a Gaussian distribution peak at the center of the light spot.
[0019] Its beneficial effect lies in the fact that this heatmap generation method provides accurate and smooth supervision signals for the U-shaped network by constructing a normalized heatmap with a standard deviation σ=3 pixels and a 15×15 Gaussian kernel based on the actual spot center. The Gaussian distribution peak not only avoids the overfitting risk caused by hard labels, but also enables the network to learn the probability distribution characteristics of the spot position and output prediction results with spatial continuity.
[0020] In one implementation, before training the U-shaped network, the method further includes online data augmentation of the encoded image, including at least one of: random rotation by a preset angle, random translation by a preset pixel, random brightness adjustment, random contrast adjustment, and random mirror flipping.
[0021] Its beneficial effects lie in significantly expanding the diversity of training data through online data augmentation strategies (including random rotation, translation, brightness adjustment, contrast adjustment, and mirror flipping), effectively improving the generalization ability and robustness of the U-shaped network. This method enables the model to learn invariant features to geometric transformations and lighting changes, avoiding the risk of overfitting and ensuring that the network maintains stable performance when faced with wafer images under various poses, positions, and lighting conditions in real-world applications.
[0022] In one implementation, a real Gaussian heatmap label is generated using the center of the real light spot, and a U-shaped network is obtained through training with a mean squared error loss function. This includes: comparing the pixel-level differences between the predicted heatmap output by the U-shaped network and the real Gaussian heatmap label; the two light spots include a first light spot and a second light spot; the mean squared error loss function L satisfies: (2) Where N is the number of samples in the current training batch, for each sample i, The label for the true Gaussian heatmap corresponding to the first light spot. The label for the actual Gaussian heatmap corresponding to the second light spot. and These are two predicted heatmaps output by a U-shaped network for the same input encoded image. It is the squared norm of the matrix elements.
[0023] Its beneficial effect is that the mean square error loss function is used to directly measure the pixel-level difference between the predicted heatmap and the real Gaussian heatmap, enabling the U-shaped network to optimize the accurate reconstruction of the spot position end-to-end. The loss is modeled and summed for the two spots respectively, which not only ensures the independence of multi-target localization, but also improves the overall localization accuracy and stability through global supervision.
[0024] In one embodiment, the curvature of the wafer surface is calculated based on the center coordinates and a formula for calculating curvature based on the spot position, including: calculating the real-time distance d between the center coordinates during the process; and calculating the curvature change E based on the initial center distance d0 between the two spots and the real-time distance d, wherein the curvature change E satisfies: (3) Where K is a system constant, satisfying: K=2LDcosθ, L is the equivalent optical path length from the wafer surface to the camera, D is the distance between the incident points of the two laser beams on the wafer surface, and θ is the laser incident angle.
[0025] Its beneficial effect lies in the fact that the formula for calculating curvature based on the position of the light spot can be used to shift the center displacement (d) of the light spot. The method explicitly correlates d0 with the wafer curvature change E, enabling interpretable and high-precision inversion from image measurement to physical deformation. Combined with the calibration of the system constant K=2LDcosθ, this method can directly calculate curvature from visual measurement results without additional sensors, thus balancing measurement accuracy and engineering practicality.
[0026] In one implementation, the U-shaped network includes an encoder, a bottleneck layer, and a decoder; the encoder includes at least 5 max pooling layers, and the decoder includes at least 4 upsampling layers; the encoder is used to perform downsampling operations on the encoded image to obtain a downsampling result; the bottleneck layer is used to generate a feature map based on the downsampling result; the decoder is used to perform upsampling operations on the feature map to obtain an upsampling result; the upsampling result is used to generate two heatmaps after Gaussian peak detection.
[0027] Its beneficial effects lie in the fact that this U-shaped network architecture achieves multi-level feature extraction through five layers of max pooling in the encoder, gradually capturing deep features from local details to global semantics; the bottleneck layer efficiently compresses and represents the downsampled features; and the decoder gradually restores spatial resolution through four layers of upsampling. This symmetrical structure not only ensures high-precision regression capability for spot positions but also possesses powerful feature representation capabilities, making it suitable for spot center detection tasks requiring pixel-level positioning accuracy.
[0028] In one implementation, based on two heatmaps, the location of the maximum pixel is obtained. Then, a neighborhood of a preset size is taken centered on the maximum pixel location. The centroid method is used to obtain the center coordinates of the two reconstructed light spots. This includes: for the output heatmap H...k Let k represent the k-th heatmap; first, find the maximum pixel position (i0, j0), then take a 5×5 neighborhood centered on the maximum pixel position, and use the centroid method to obtain the sub-pixel coordinates, which are used as the center coordinates of the two reconstructed spots; where the sub-pixel level x-coordinate is... c and sub-pixel level y-coordinate c satisfy: (4) Where i is the x-axis variable, j is the y-axis variable, and x... c The x-coordinate is at the sub-pixel level, y c The vertical coordinate is at the sub-pixel level.
[0029] Its beneficial effect is that by using the centroid method to calculate the sub-pixel level spot center coordinates in the neighborhood (5×5) of the maximum pixel, the integer pixel positioning error caused by image discretization is effectively overcome. This method uses the heat map response intensity as weight to make the center estimation closer to the peak position of the real Gaussian distribution, which significantly improves the positioning accuracy to the sub-pixel level and provides highly reliable input data for subsequent curvature calculation.
[0030] According to a second aspect of the present invention, an interference-resistant wafer surface curvature measurement device is provided, used in any of the methods of the first aspect. The device includes: a laser source, a polarization beam splitting and displacement module, a wavefront encoding module, an optical module, an image acquisition module, and a deep learning processing module, arranged sequentially along an optical path; the laser source is used to output linearly polarized light; the polarization beam splitting and displacement module is used to obtain two parallel orthogonally polarized beams with a fixed spacing based on the linearly polarized light; the wavefront encoding module includes a cubic phase mask and is used to perform wavefront encoding on the two orthogonally polarized beams to obtain two encoded first light waves; the optical module is used to obtain two second light waves with enlarged spot diameters and spacing based on the two first light waves, and input the two second light waves into a reaction chamber to irradiate the wafer surface; the image acquisition module is used to obtain an acquired encoded image based on the reflected light returned from the wafer surface along the original path; the deep learning processing module is used to obtain the center coordinates of the reconstructed two light spots based on the encoded image, and calculate the curvature of the wafer surface based on the center coordinates and a formula for calculating curvature based on the spot positions.
[0031] In one embodiment, the laser source is a single-mode fiber-coupled semiconductor laser; the polarization beam splitting and shifting module includes a passive element made of birefringent crystal; the optical module includes a beam expander group and a beam splitter; the beam expander group is located in the optical path between the laser source and the beam splitter, and is used to expand the beam from the laser source; the beam splitter is located in the optical path between the beam expander group and the window of the reaction chamber, and is used to reflect the expanded beam to the window; the beam expander group includes a first convex lens and a second convex lens, the first convex lens and the second convex lens are arranged confocally, and the distance between them is adjustable.
[0032] Its beneficial effects are as follows: the use of a single-mode fiber-coupled semiconductor laser ensures high spatial coherence and stability of the light source; the polarization beam splitting and displacement module uses a birefringent crystal to achieve passive, high-precision beam separation and displacement, improving system reliability; the beam expander group consists of a first and a second lens arranged confocally with adjustable spacing, which not only effectively controls the spot size to match imaging requirements, but also enhances the flexibility of optical path alignment and system adaptability, thereby improving the overall accuracy and robustness of wafer curvature measurement.
[0033] In one embodiment, the optical module further includes an adjustable stop located between the beam expander group and the beam splitter. The adjustable stop is used to limit the beam diameter to avoid introducing stray light.
[0034] Its beneficial effects are that the adjustable aperture is set between the beam expander group and the beam splitter, which effectively suppresses the generation of stray light by precisely limiting the beam diameter, and significantly improves the imaging contrast and signal-to-noise ratio. At the same time, the adjustable design allows the beam size to be flexibly adjusted according to the actual measurement needs, which avoids optical crosstalk caused by an excessively large beam and prevents signal attenuation caused by an excessively small beam, thereby optimizing the performance stability and measurement accuracy of the entire optical system.
[0035] This invention provides an interference-resistant wafer surface curvature measurement method and apparatus. Its overall beneficial effect is that by combining cubic phase mask wavefront coding technology with deep learning image processing algorithms, a high-precision and highly robust wafer surface curvature measurement method is constructed.
[0036] At the optical design level, the introduction of the cubic phase mask enables the system modulation transfer function to remain stable over a large defocus range, and the point spread function exhibits a diffuse distribution. This effectively avoids the signal loss problem caused by interference cancellation in the traditional two-point laser reflection method, expands the measurement depth of field, and ensures the continuity and reliability of wafer surface curvature data during the process.
[0037] At the image processing level, a simulation-based data-driven training method was adopted. A two-stage localization strategy combining a U-shaped network and a heatmap regression algorithm was used to achieve sub-pixel-level accurate extraction of the spot center coordinates. This method not only significantly reduced the cost of experimental data acquisition but also improved the model's generalization ability and anti-interference performance through multi-parameter scanning and noise simulation, enabling it to adapt to the complex semiconductor industrial environment.
[0038] At the system integration level, through optimized optical module design and precise curvature calculation formula, a high-precision curvature inversion process with clear physical mapping from image measurement to physical deformation is achieved. Curvature can be calculated directly from visual measurement results without additional sensors, balancing measurement accuracy and engineering practicality. Attached Figure Description
[0039] Figure 1 This is a flowchart illustrating an interference-resistant wafer surface curvature measurement method according to an exemplary embodiment.
[0040] Figure 2 It is a three-dimensional surface pattern of a cubic phase mask according to an exemplary embodiment.
[0041] Figure 3 This is a curvature versus time series curve of a comparative experimental result, as shown in another exemplary embodiment.
[0042] Figure 4 This is an interference-resistant wafer surface curvature measurement device according to an exemplary embodiment.
[0043] Figure 5 This is a schematic diagram of the optical path structure of an in-situ warp measurement system according to another exemplary embodiment.
[0044] Explanation of the reference numerals in the figure: 1. Laser source; 2. Polarization beam splitter and displacement module; 3. Wavefront coding module; 4. Optical module; 5. Wafer; 6. Image acquisition module; 7. Deep learning processing module; 11. Fiber collimator; 12. Polarization beam splitter; 13. Beam expander assembly; 14. Focusing lens; 15. Wafer; 16. Camera; 17. Cubic phase mask; 18. Beam splitter; 19. Half-wave plate. Detailed Implementation
[0045] Unless otherwise defined, the technical or scientific terms used in this specification should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. Specific embodiments of the invention will be described below with reference to the accompanying drawings. It should be noted that, in order to provide a concise description, this specification cannot provide a detailed description of all features of the actual embodiments. Without departing from the spirit and scope of the invention, those skilled in the art can make modifications and substitutions to the embodiments of the invention, and the resulting embodiments are also within the protection scope of the invention.
[0046] like Figure 1 As shown, the first embodiment of the present invention provides an interference-resistant wafer surface curvature measurement method, which includes the following steps S1-S6: S1 controls the output of linearly polarized light from the laser source.
[0047] S2, the linearly polarized light is input into the polarization beam splitting and displacement module, and two parallel orthogonally polarized beams with a fixed spacing are output.
[0048] S3, input the two beams of orthogonally polarized light into the wavefront encoding module for wavefront encoding, and output the encoded two beams of first light waves. The wavefront encoding module includes a cubic phase mask.
[0049] S4, the two first light waves are input into the optical module, and two second light waves with enlarged spot diameter and spacing are output. The two second light waves are input into the reaction chamber and irradiate the wafer surface.
[0050] S5, acquire the reflected light returning from the wafer surface along the original path, and collect an encoded image based on the reflected light. The encoded image is then fed into a deep learning processing module to obtain the center coordinates of the two reconstructed light spots.
[0051] S6, calculate the curvature of the wafer surface based on the center coordinates and the formula for calculating curvature based on the spot position.
[0052] In one embodiment, the surface profile height function z(x, y) of the cubic phase mask satisfies: (1) Where α is the modulation intensity coefficient, r0 is the half-aperture of the cubic phase mask, x is the abscissa with the center of the cubic phase mask as the origin, and y is the ordinate with the center of the cubic phase mask as the origin.
[0053] In some examples, for the wavelength λ = 635 nm of the first light wave, α = 8λ and α = 5.08 μm are taken. The half-aperture r0 of the cubic phase mask is taken as 5 mm.
[0054] It is worth noting that the diameter of the expanded beam spot is 4mm, and the half-aperture of the cubic phase mask is larger than the diameter of the expanded beam spot to ensure that the beam spot completely passes through the cubic phase mask.
[0055] In other examples, the substrate for the triple-phase mask is fused silica, a material with high transmittance in the ultraviolet to near-infrared bands, low coefficient of thermal expansion, suitable for the external optical path environment of MOCVD, and capable of withstanding an operating temperature of 60°C. The fabrication method employs photolithography combined with reactive ion etching (RIE) or single-point diamond turning, with surface shape error controlled within λ / 10 (i.e., 63.5nm). A double-sided antireflection coating is deposited on the surface to ensure a reflectivity of less than 0.5% at 635nm wavelength. The effective aperture is designed as a 20mm × 15mm elongated structure to cover the entire area of the two measurement beams.
[0056] like Figure 2 The image shows a three-dimensional surface diagram of a cubic phase mask, illustrating the aspherical shape of the mask. The cubic phase mask has a peak value of +2.5 μm and a valley value of -2.5 μm in the height direction, i.e., the z-direction.
[0057] In one embodiment, the deep learning processing module is equipped with a U-shaped network and a heatmap regression algorithm. The U-shaped network is used to generate two heatmaps based on the encoded image. The heatmap regression algorithm is used to obtain the position of the maximum pixel based on the two heatmaps, and then, using the maximum pixel position as the center, take a neighborhood of a preset size and use the centroid method to obtain the center coordinates of the two reconstructed light spots.
[0058] In some specific embodiments, the U-shaped network is obtained through the following training method: An optical simulation model is established, including a polarization beam splitter, a cubic phase mask, a reaction chamber window, and an epitaxial wafer. Training pairs of coded images and real spot center coordinates are generated through parameter scanning. The coded images are processed with convolution blurring, noise superposition, and background radiation addition. Real Gaussian heatmap labels are generated using the real spot centers, and the U-shaped network is obtained through training with a mean squared error loss function.
[0059] In one embodiment, the parameter scanning includes scanning with at least one of the following parameters: wafer surface curvature radius, interference phase difference between two light spots, wafer surface reflectivity, and thermal noise standard deviation.
[0060] In some specific embodiments, the radius of curvature R of the epitaxial wafer is set to range from -10 meters to +10 meters, with a step size of 0.2 meters, containing a total of 101 sampled values, where negative values indicate concave surfaces and positive values indicate convex surfaces. The interference phase difference φ ranges from 0 to 2π radians, with a step size of 0.05 radians, containing a total of 126 sampled values. The reflectivity of the epitaxial wafer ranges from 0.1 to 0.8, with a step size of 0.1, containing a total of 8 sampled values. The thermal noise adopts a Gaussian white noise model, with a standard deviation range from 0 to 10 digital numbers (DN), a step size of 2DN, containing a total of 6 sampled values. Theoretically, the complete combination of these four parameters will generate approximately 610,848 samples, but considering computational efficiency, this embodiment uses the Latin hypercube sampling method for random sampling, ultimately generating 200,000 pairs of training data. Each pair of data contains an coded image and the corresponding real spot center coordinate label.
[0061] In one embodiment, the noise simulation during the training process includes: superimposing shot noise and readout noise, and adding a thermal radiation background, wherein the intensity of the thermal radiation background is 1% to 10% of the signal peak value.
[0062] In one implementation, pixel regions containing the first and second light spots in the encoded image are extracted as samples, and the coordinates of the actual light spot center in each sample are recorded. Using the actual light spot center as the heatmap center, with a standard deviation σ = 3 pixels, a Gaussian kernel size of 15×15, and peak value normalized to 1, the heatmap output by the U-shaped network exhibits a Gaussian peak distribution at the light spot center.
[0063] In one embodiment, before training the U-shaped network, the method further includes performing online data augmentation on the encoded image, including at least one of: random rotation by a preset angle, random translation by a preset pixel, random brightness adjustment, random contrast adjustment, and random mirror flipping.
[0064] In some specific embodiments, regarding the training strategy, this embodiment uses the PyTorch 2.0 deep learning framework, and the training hardware configuration is an NVIDIA A100 40GB graphics processor. The loss function chosen is Mean Squared Error Loss (MSE Loss), which optimizes the model parameters by comparing the difference between the predicted heatmap and the actual heatmap. The optimizer used is the AdamW optimizer, with the weight decay coefficient set to 1x10. -4 The learning rate strategy uses cosine annealing decay, with the initial learning rate set to 1x10. -3After 50 training epochs, the performance gradually decays to 1x10. -6 The batch size is set to 32, the total training epochs are 100, and early stopping is enabled with a patience value of 10. Training will automatically terminate if the validation set performance does not improve for 10 consecutive epochs. The validation set contains 20,000 samples; this data is not used in the model training process and is only used to monitor the model's generalization ability. Expected performance metrics are: an average center localization error of less than 0.2 pixels on the test set (corresponding to an actual physical displacement of less than 1.2 micrometers); and, even under extreme conditions where the signal-to-noise ratio (SNR) is 0 dB (i.e., signal strength equals noise strength), the localization error should still be controlled within 0.5 pixels.
[0065] For inference deployment, this embodiment uses the NVIDIA Jetson AGX Orin embedded computing platform as the hardware carrier, equipped with 32GB of memory and a 2048-core graphics processor. Inference performance test results show that for a single-frame 512×512 pixel input image, the average inference time using the heatmap method is approximately 12 milliseconds, achieving a real-time output rate of 80 frames per second. On the software side, the TensorRT inference acceleration engine is used for model optimization, and half-precision floating-point numbers (16-bit floating-point numbers, or FP16) quantization are employed. This significantly improves computational efficiency while ensuring inference accuracy, meeting the dual requirements of real-time performance and accuracy in industrial settings.
[0066] In one implementation, a real Gaussian heatmap label is generated using the center of a real light spot, and the U-shaped network is trained using a mean squared error loss function. This includes comparing the pixel-level differences between the predicted heatmap output by the U-shaped network and the real Gaussian heatmap label. The two light spots include a first light spot and a second light spot. The mean squared error loss function L satisfies: (2) Where N is the number of samples in the current training batch, for each sample i, The label for the true Gaussian heatmap corresponding to the first light spot. The label for the actual Gaussian heatmap corresponding to the second light spot. and These are two predicted heatmaps output by a U-shaped network for the same input encoded image. It is the squared norm of the matrix elements.
[0067] In one embodiment, calculating the curvature of the wafer surface based on the center coordinates and a formula for calculating curvature based on the spot position includes: calculating the real-time distance d between the center coordinates during the process, based on the center coordinates. The curvature change E is calculated based on the initial center distance d0 between the two spots and the real-time distance d, wherein the curvature change E satisfies: (3) Where K is a system constant, satisfying: K=2LDcosθ, L is the equivalent optical path length from the wafer surface to the camera, D is the distance between the incident points of the two laser beams on the wafer surface, and θ is the laser incident angle.
[0068] In some examples, the calibration of the system constant K is satisfied.
[0069] ; Among them, two pre-selected standard plates with known curvature have curvature values of E1 and E2 (unit: m). -1 Under the same optical path conditions, measure the center-to-center distances d1 and d2 (in pixels) of the two corresponding light spots. Store the calibrated K value in the system parameters for direct use in subsequent measurements.
[0070] In one embodiment, the U-shaped network includes an encoder, a bottleneck layer, and a decoder. The encoder contains at least five max-pooling layers, and the decoder contains at least four upsampling layers. The encoder performs a downsampling operation on the encoded image to obtain a downsampling result. The bottleneck layer generates a feature map based on the downsampling result. The decoder performs an upsampling operation on the feature map to obtain an upsampling result. The upsampling result is used to generate the two heatmaps after Gaussian peak detection.
[0071] In some examples, the encoder structure is shown in Table 1: where Con is downsampling convolution, BN is normalization, ReLU is linear rectification, and MaxPool is max pooling.
[0072] Table 1 Encoder Structure Table
[0073] In some examples, the structure table of the bottleneck layer is shown in Table 2: where Dropout is the drop rate.
[0074] Table 2 Bottleneck Layer Structure
[0075] In some examples, the decoder structure is shown in Table 3: UpConv is the upsampling convolution operation, and Concatenate is the feature fusion process.
[0076] Table 3 Decoder Structure Table
[0077] In one implementation, based on the two heatmaps, the position of the maximum pixel is obtained. Then, a neighborhood of a preset size is taken centered on the maximum pixel position, and the center coordinates of the two reconstructed light spots are obtained using the centroid method. This includes: for the output heatmap H... k Let k represent the k-th heatmap. First, find the maximum pixel position (i0, j0). Then, take a 5×5 neighborhood centered on the maximum pixel position and use the centroid method to obtain the sub-pixel coordinates, which are used as the center coordinates of the two reconstructed spots. The sub-pixel level x-coordinate... c and sub-pixel level y-coordinate c satisfy: (4) Where i is the x-axis variable, j is the y-axis variable, and x... c The x-coordinate is at the sub-pixel level, y c The vertical coordinate is at the sub-pixel level.
[0078] This embodiment provides a specific image processing flow for wafer surface curvature measurement. The specific steps are as follows: First, the original coded image is acquired with a resolution of 512×512 pixels and a grayscale depth of 8 bits (i.e., the value range of each pixel is 0–255). Then, preprocessing is performed by subtracting the dark field background image (i.e., the reference image acquired under conditions without laser illumination) to eliminate the inherent bias of the optical system and the influence of sensor dark current. Furthermore, 3×3 neighborhood median filtering is used to suppress impulse noise and improve the image signal-to-noise ratio. The processed image is input into a trained deep learning model, such as the U-Net architecture. The model outputs two normalized heatmaps, corresponding to the response distributions of the two interference spots. For each heatmap, the centroid method is used to calculate its peak position coordinates, obtaining the center coordinates of the first spot (x1, y1) and the center coordinates of the second spot (x2, y2). Then, according to the Euclidean distance formula between the two points d=(x1, y1)... x2) 2 +(y1 y2) 2Calculate the pixel spacing *d* between the centers of the two light spots. Substitute this distance into a calibrated curvature calculation formula (e.g., a curvature radius inversion model derived from the geometry of two-beam interference) to calculate the local curvature radius *E* of the wafer surface. To further improve measurement stability, a first-order low-pass filter is applied to the obtained curvature sequence with a time constant of 0.1 seconds to effectively suppress high-frequency random jitter. Finally, the smoothed curvature change curve is output, and the relative reflectance of the corresponding region is recorded simultaneously. This relative reflectance, obtained by normalizing the average pixel grayscale value within the light spot coverage area, is used to assist in judging the surface condition (e.g., whether there is contamination, scratches, or film abnormalities). The entire process can be completed within a single frame image processing time, meeting the requirements of real-time online detection.
[0079] During the system calibration phase, a standard planar silicon wafer was first used as a reference sample for measurement, obtaining a reference spacing d0 = 512.3 pixels. Subsequently, calibration tests were performed using standard curvature wafers with radii of curvature of +5 meters and -5 meters, respectively, obtaining corresponding spacings d0 and d0. +5 =525.1 pixels and d -5 =499.5 pixels. Based on the above calibration data, according to the formula K=(d +5 -d -5 The system calibration coefficient K = (525.1 - 499.5) / 0.4 = 64.0 pixels per meter is obtained by calculating K = (525.1 - 499.5) / 0.4. This coefficient is used to convert the pixel spacing into the actual curvature value in the subsequent curvature inversion calculation.
[0080] To verify the performance advantages of the method of this invention in a real-world process environment, a metal-organic chemical vapor deposition (MOCVD) epitaxial growth experiment was conducted. The experimental conditions were as follows: an AIXTRON CCS 3×2-inch reaction chamber system was used; the epitaxial structure was gallium nitride / sapphire (GaN on Sapphire); the growth temperature was set at 1050°C; the reaction chamber pressure was maintained at 100 mbar; and the total growth time was 120 minutes. The experimental results showed that the traditional measurement method (without wavefront encoding, a fixed exposure time of 2 ms, and a gain of 0 dB) experienced 23 data loss events during the growth process, with a cumulative loss time of 1080 seconds, severely affecting the continuity of process monitoring. In contrast, the method of this invention (triple phase plate wavefront encoding combined with a thermogram regression algorithm) achieved zero data loss throughout the entire growth process, with a curvature measurement noise standard deviation of only 0.019 meters. -1The reflectance monitoring curve is smooth and continuous, without obvious jumps or abnormal fluctuations. Furthermore, the measured system inference frame rate reaches 82 frames per second, fully meeting the technical requirements for real-time online monitoring of the MOCVD process.
[0081] In some examples, comparative experimental data tables are shown in Table 4.
[0082] Table 4 Comparative Experiment Data Table
[0083] like Figure 3 As shown, this figure is a curvature versus time series curve comparing experimental results, used to evaluate the performance difference between the proposed method (wavefront encoding + deep learning) and the traditional method (fixed exposure) in curvature monitoring during MOCVD epitaxial growth. The horizontal axis represents time (in minutes), ranging from 0 to 120 minutes, covering the complete 120-minute growth cycle. The vertical axis represents curvature (in meters). -1 ), range from 0.5 to +0.5m -1 This reflects the curvature of the wafer surface (positive values indicate convex curvature, and negative values indicate concave curvature). This invention improves the dynamic range through wavefront encoding and enhances anti-interference capabilities by combining deep learning thermal mapping regression, significantly outperforming traditional fixed exposure schemes. It maintains high precision and high continuity in real-time curvature monitoring even under complex process scenarios such as strong light changes and abrupt surface reflection changes.
[0084] It is worth noting that this embodiment uses a single laser in conjunction with a yttrium vanadate (YVO4) polarization beam splitter to directly generate two parallel, fixed-spacing, orthogonally polarized outgoing beams, resulting in a very simple structure that requires no complex alignment. A 4x beam expander group enlarges the spot diameter from 1mm to 4mm and the beam spacing from 2.5mm to 10mm, maintaining parallelism and spacing ratios. A cubic phase mask (surface shape parameter α=8λ, fused silica material, photolithography + reactive ion etching process) is installed in the expanded parallel optical path to transform the originally interference-sensitive point spread function into an interference-insensitive diffuse point spread function. The beam is incident on the external window of the reactor chamber via a beam splitter (90% transmitted to the camera, 10% reflected to the power detector), and after reflection, is imaged onto an industrial camera by a focusing lens. The power detector monitors the laser power in real time and provides feedback to control the drive current, ensuring power fluctuations are <1%.
[0085] In terms of data processing, the acquired coded images, after dark field correction and median filtering preprocessing, are input into a dedicated U-Net variant network for end-to-end reconstruction. The network outputs a Gaussian heatmap, and then the centroid method is used to calculate the sub-pixel-level spot center coordinates (positioning error < 0.2 pixels). The wafer curvature is inverted based on the distance between the two spots and the calibration coefficient (K = 64.0 pixels·meter), and the average gray value of the spot region is extracted and normalized to the relative reflectance. The training data is generated through Zemax optical simulation, covering random sampling of multiple parameters such as radius of curvature, interference phase difference, reflectance, and thermal noise, and combined with geometric enhancement to improve the model's generalization ability.
[0086] Experimental verification was conducted using an AIXTRON CCS 3×2-inch MOCVD machine, with GaN on Sapphire epitaxial structures grown at 1050°C, 100 mbar pressure, and for 120 minutes. Results showed that the traditional method (fixed exposure time 2 ms, gain 0 dB) resulted in 23 data loss events, totaling 1080 seconds of data loss. The method of this invention exhibited no data loss throughout the entire process, with a curvature noise standard deviation of 0.019 m. -1 The reflectivity curve is smooth without jumps, and the inference frame rate reaches 82fps. Even when the signal-to-noise ratio drops to 0dB (signal and noise are of equal strength), the positioning error is still <0.5 pixels, which is significantly better than traditional image processing methods.
[0087] The core advantages of this invention are as follows: First, by leveraging the energy dispersion effect of wavefront encoding and the dual guarantee of orthogonal polarization interference elimination, the data interruption problem caused by laser interference cancellation is completely solved, achieving 100% measurement continuity. Second, stable operation is achieved by fixing camera parameters (exposure time, gain), avoiding delays and measurement artifacts caused by dynamic adjustments and simplifying system control. Third, the optical structure is extremely simple (single laser + polarization beam shifter), reducing setup time from several hours to 10 minutes and lowering costs by approximately 40%. Finally, the non-invasive modification scheme requires no alteration to the reaction chamber itself, is compatible with mainstream models such as AIXTRON and Veeco, and can be quickly deployed to existing equipment. The system can also be extended to the infrared band to achieve synchronous temperature measurement, forming an in-situ monitoring capability for three parameters: warp, reflectivity, and temperature.
[0088] like Figure 4As shown, according to a second embodiment of the present invention, an interference-resistant wafer surface curvature measurement device is provided for use with the method described in any one of the above embodiments. The device includes: a laser source 1, a polarization beam splitting and displacement module 2, a wavefront encoding module 3, an optical module 4, an image acquisition module 6, and a deep learning processing module 7, sequentially distributed along the optical path. The laser source 1 is used to output linearly polarized light. The polarization beam splitting and displacement module 2 is used to obtain two parallel orthogonally polarized beams with a fixed spacing based on the linearly polarized light. The wavefront encoding module 3 includes a cubic phase mask for wavefront encoding the two orthogonally polarized beams to obtain two encoded first light waves. The optical module 4 is used to obtain two second light waves with enlarged spot diameters and spacing based on the two first light waves, and input the two second light waves into a reaction chamber to irradiate the surface of the wafer 5. The image acquisition module 6 is used to obtain an acquired encoded image based on the reflected light returning from the surface of the wafer 5 along the original path. The deep learning processing module 7 is used to obtain the center coordinates of the two reconstructed light spots based on the encoded image, and to calculate the curvature of the surface of the wafer 5 according to the center coordinates and the formula for calculating curvature based on the position of the light spots.
[0089] In one embodiment, the laser source 1 is a single-mode fiber-coupled semiconductor laser. The polarization beam splitting and shifting module 2 includes passive elements made of birefringent crystal. The optical module 4 includes a beam expander group and a beam splitter. The beam expander group is located in the optical path between the laser source and the beam splitter, and is used to expand the beam from the laser source. The beam splitter is located in the optical path between the beam expander group and the window of the reaction chamber, and is used to reflect the expanded beam to the window. The beam expander group includes a first convex lens and a second convex lens, which are confocally arranged, and the distance between them is adjustable.
[0090] In some specific embodiments, the phase mask is mounted and fixed using a five-dimensional precision adjustment frame, which provides three translational degrees of freedom (X, Y, and Z) and two angular degrees of freedom (pitch and yaw). Specifically, the adjustment range in the X and Y directions is ±5mm, the adjustment range in the Z direction is ±10mm, and the angular adjustment range is ±3°. During adjustment, precise adjustments are made to ensure that the two incident light spots are accurately located in the central region of the mask, and to ensure good symmetry in wavefront modulation. The phase mask is mounted in the parallel optical path after the beam expander and before the beam splitter. At this position, the light spot diameter is maximized, thus providing the highest tolerance for mask processing errors and improving the stability and reliability of the system.
[0091] Furthermore, the curvature measurement device also includes a non-polarizing beam splitter with a transmission / reflection ratio of 90:10, an operating wavelength of 635 nm, and a diameter of 1 inch (25.4 mm), mounted at a 45° angle in the optical path. 90% of the transmitted light enters the MOCVD reaction chamber for the process, while 10% of the reflected light is deflected upwards and enters a power monitoring detector. The power monitoring detector uses a silicon photodiode with integrated amplification circuitry to monitor the laser power in real time and control the laser drive current through feedback, ensuring that laser power fluctuations are kept within 1%, thereby guaranteeing the stability and repeatability of the process.
[0092] In addition, the curvature measurement device also includes a focusing lens group and an imaging lens group. The focusing lens is a plano-convex lens with a focal length f=150mm and a diameter of 25.4mm. Its surface is coated with an anti-reflection film, and its function is to precisely converge two parallel beams of light onto the surface of the epitaxial wafer, forming two small light spots with a diameter of approximately 100μm. The imaging lens is a cemented doublet achromatic lens with a focal length f=100mm and a diameter of 25.4mm. Its function is to image the reflected light, after being reflected from the epitaxial wafer and then passed through the focusing lens and beam splitter, onto the camera target surface, enabling precise observation and recording of the position and shape of the light spots.
[0093] In other specific embodiments, such as this one, the imaging acquisition module includes an industrial-grade complementary metal-oxide-semiconductor (CMOS) global shutter camera, such as the Basler acA4024-29μm model. This camera has a resolution of 2048×2048 pixels, a single pixel size of 5.86μm, a dynamic range of 70dB, and a frame rate of up to 50fps at full resolution, meeting the requirements of high-speed imaging. To effectively suppress the strong thermal radiation interference generated when the MOCVD reaction chamber operates at temperatures above 1000℃ (covering the visible to infrared bands), a bandpass filter is configured in front of the camera. This filter has a center wavelength of 635nm and a full width at half maximum (FWHM) of 10nm, effectively filtering out background noise and improving the imaging signal-to-noise ratio.
[0094] Regarding imaging parameter settings, the camera exposure time is fixed at 2ms to avoid motion blur caused by epitaxial wafer movement or vibration. The analog gain is fixed at 0dB to avoid introducing additional electronic noise and ensure the stability of image quality. It is worth noting that although the spot morphology becomes diffuse after wavefront coding, the total luminous flux of the system remains unchanged according to the principle of energy conservation. Therefore, there is no need to compensate for signal strength by increasing the exposure time or increasing the gain; the imaging signal-to-noise ratio is mainly determined by the total luminous flux. The advantage of wavefront coding technology is that it only redistributes the spatial distribution of spot energy without reducing the total energy, thus achieving robust improvement in aberrations such as defocus while ensuring image quality, providing a stable and reliable image input for subsequent spot center localization algorithms.
[0095] This embodiment provides an in-situ warpage measurement system, the optical path structure of which is as follows: Figure 5 As shown, the laser beam output by the fiber collimator 11 is a parallel beam with a diameter of 1 mm and a divergence angle of <0.5 mrad. Subsequently, an optional half-wave plate 19 adjusts the incident polarization direction to 45° with the optical axis of the polarization beam displacementr 12, achieving equal power splitting of the ordinary ray (o-ray) and the extraordinary ray (e-ray). The half-wave plate 19 can be mounted on an electrically driven rotating frame and fixed after initial calibration. The polarization beam displacementr (PBD) utilizes the birefringence effect to directly split the incident light into two orthogonally polarized, spatially parallel, and fixed-spacing output beams (typically 2.5 mm apart), eliminating the need for a reflector or translation stage, resulting in a compact structure and long-term stability.
[0096] Two beams of light enter the beam expander group 13 simultaneously, magnifying the spot diameter to 4 mm and the beam spacing to 10 mm while maintaining parallelism and proportionality. A cubic phase mask 17 is placed in the expanded optical path to simultaneously perform wavefront encoding on both beams, transforming the point spread function into an interference-insensitive diffuse distribution. Subsequently, the transmitted light passes through a beam splitter 18: it is focused by a focusing lens 14, illuminating the surface of the wafer 15 in the reaction chamber and then reflected. The reflected light is converged by the same lens and imaged by a camera 16. The 10% reflected light path of the beam splitter is connected to a power detector to monitor the laser power in real time and provide feedback to adjust the drive current, ensuring intensity stability <1%.
[0097] This solution does not include a post-half-wave plate by default, maintaining the orthogonal polarization of the two beams and completely eliminating the risk of destructive interference at the physical level. Since warpage measurement relies solely on the geometric position of the light spot, no interference information is required. If the user needs to retain interference capability, such as for other sensing functions, a local half-wave plate can be added to one of the beam paths and rotated 90° to make its polarization parallel to the other beam, achieving flexible configuration.
[0098] The entire system is installed outside the MOCVD reactor, using a non-invasive modification and is compatible with mainstream equipment. Image data is input into a customized U-Net network, outputting a heat map and using the centroid method to accurately locate the center of the light spot, achieving sub-pixel-level (<0.2 pixels) positioning and 100% continuous measurement. The measured inference frame rate reaches 82fps, meeting the requirements for real-time closed-loop control.
[0099] In one embodiment, the optical module further includes an adjustable aperture located between the beam expander group and the beam splitter, the adjustable aperture being used to limit the beam diameter to avoid introducing stray light.
[0100] In this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. The term "multiple" refers to two or more unless otherwise expressly defined.
[0101] The above description of the embodiments is intended to enable those skilled in the art to understand and apply the present invention. It will be apparent to those skilled in the art that various modifications can be made to these embodiments, and the general principles described herein can be applied to other embodiments without creative effort. Therefore, the present invention is not limited to the embodiments described herein, and any improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope and spirit of the invention are within the scope of the present invention.
Claims
1. A method of anti-interference wafer surface curvature measurement, characterized in that, The method includes: Control the linearly polarized light output from the laser source; The linearly polarized light is input into the polarization beam splitting and displacement module, which outputs two parallel orthogonally polarized beams with a fixed spacing. Two beams of orthogonally polarized light are input into a wavefront encoding module for wavefront encoding, and the encoded two beams of first light waves are output; the wavefront encoding module includes a cubic phase mask. The two first light waves are input into the optical module, and two second light waves with enlarged spot diameter and spacing are output. The two second light waves are input into the reaction chamber and irradiate the wafer surface. The reflected light returning from the wafer surface along the original path is acquired, and an coded image is collected based on the reflected light. The coded image is then fed into a deep learning processing module to obtain the center coordinates of the two reconstructed light spots. The deep learning processing module is equipped with a U-shaped network and a heatmap regression algorithm. The U-shaped network is used to generate two heatmaps based on the coded image. The heatmap regression algorithm is used to obtain the position of the maximum pixel based on the two heatmaps, and then, using the maximum pixel position as the center, takes a neighborhood of a preset size and uses the centroid method to obtain the center coordinates of the two reconstructed light spots. The U-shaped network is obtained through the following training method: An optical simulation model is established, including a polarization beam splitter, a cubic phase mask, a reaction chamber window, and an epitaxial wafer. Training pairs of coded images and real spot center coordinates are generated through parameter scanning. The coded images are processed by convolution blurring, noise superposition, and background radiation addition. Real Gaussian heatmap labels are generated using the real spot centers, and the U-shaped network is obtained by training with the mean square error loss function. The curvature of the wafer surface is calculated based on the center coordinates and the formula for calculating curvature based on the spot position.
2. The method according to claim 1, characterized in that, The surface profile height function z(x, y) of the cubic phase mask satisfies: ; Where α is the modulation intensity coefficient; r0 is the half-aperture of the cubic phase mask; x is the abscissa with the center of the cubic phase mask as the origin; and y is the ordinate with the center of the cubic phase mask as the origin.
3. The method according to claim 1, characterized in that, The parameter scanning includes scanning with at least one of the following parameters: wafer surface curvature radius, interference phase difference between two light spots, wafer surface reflectivity, and thermal noise standard deviation.
4. The method according to claim 1, characterized in that, The noise simulation during training includes superimposing shot noise and readout noise, and adding thermal radiation background, where the intensity of thermal radiation background is 1% to 10% of the signal peak value.
5. The method according to claim 1, characterized in that, Two heatmaps are generated based on the encoded image, including: The pixel regions containing the first and second light spots in the encoded image are extracted as samples, and the center coordinates of the real light spot in each sample are recorded. The center of the real light spot is taken as the center of the heat map, the standard deviation σ = 3 pixels, the Gaussian kernel size is 15×15, and the peak value is normalized to 1, so that the heat map output by the U-shaped network presents a Gaussian distribution peak at the center of the light spot.
6. The method according to claim 1, characterized in that, Before training the U-shaped network, the method further includes online data augmentation of the encoded image, including at least one of: random rotation by a preset angle, random translation by a preset pixel, random brightness adjustment, random contrast adjustment, and random mirror flipping.
7. The method according to claim 1, characterized in that, Generating realistic Gaussian heatmap labels using real light spot centers, and training the U-shaped network using a mean squared error loss function, includes: comparing the pixel-level differences between the predicted heatmap output by the U-shaped network and the real Gaussian heatmap labels; the two light spots include a first light spot and a second light spot; the mean squared error loss function L satisfies: Where N is the number of samples in the current training batch, for each sample i, The label for the true Gaussian heatmap corresponding to the first light spot. The label for the actual Gaussian heatmap corresponding to the second light spot. and These are two predicted heatmaps output by a U-shaped network for the same input encoded image. It is the squared norm of the matrix elements.
8. The method according to claim 1, characterized in that, The curvature of the wafer surface is calculated based on the center coordinates and the formula for calculating curvature based on the spot position, including: Based on the center coordinates, calculate the real-time distance d between the center coordinates during the process; calculate the curvature change E based on the initial center distance d0 of the two light spots and the real-time distance d, wherein the curvature change E satisfies: Where K is a system constant, satisfying: K=2MDcosθ, M is the equivalent optical path length from the wafer surface to the camera, D is the distance between the incident points of the two laser beams on the wafer surface, and θ is the laser incident angle.
9. The method according to claim 1, characterized in that, The U-shaped network includes an encoder, a bottleneck layer, and a decoder; the encoder contains at least 5 max pooling layers, and the decoder contains at least 4 upsampling layers; the encoder is used to perform downsampling operations on the encoded image to obtain downsampling results; the bottleneck layer is used to generate feature maps based on the downsampling results. The decoder is used to perform an upsampling operation based on the feature map to obtain an upsampling result; the upsampling result is used to generate the two heatmaps after Gaussian peak detection.
10. The method according to claim 1, characterized in that, According to the two heat maps, a maximum pixel position is obtained, a neighborhood of a preset size is taken with the maximum pixel position as a center, and a center coordinate of the reconstructed two light spots is obtained by using a centroid method, including: for the output heat map H k , k represents the kth heat map, a maximum pixel position (i0, j0) is first found, then a neighborhood with a pixel size of 5*5 is taken with the maximum pixel position as a center, and a sub-pixel coordinate is obtained by using a centroid method as a center coordinate of the reconstructed two light spots; wherein a sub-pixel level horizontal coordinate x c and a sub-pixel level vertical coordinate y c satisfy: Where i is the x-axis variable and j is the y-axis variable.
11. An interference-resistant wafer surface curvature measurement device, used in the method according to any one of claims 1 to 10, characterized in that, The device includes: a laser source, a polarization beam splitting and shifting module, a wavefront encoding module, an optical module, an image acquisition module, and a deep learning processing module, all arranged sequentially along the optical path. The laser source is used to output linearly polarized light; The polarization beam splitting and displacement module is used to obtain two parallel orthogonally polarized beams with a fixed spacing based on the linearly polarized light. The wavefront encoding module includes a cubic phase mask, which is used to perform wavefront encoding on the two orthogonally polarized beams to obtain two encoded first light waves. The optical module is used to obtain two second light waves with magnified spot diameter and spacing based on the two first light waves, and input the two second light waves into the reaction chamber to irradiate the wafer surface; The image acquisition module is used to obtain a captured coded image based on the reflected light returned from the original path on the wafer surface; The deep learning processing module is used to obtain the center coordinates of the two reconstructed light spots based on the encoded image. The deep learning processing module is equipped with a U-shaped network. The U-shaped network is obtained through the following training method: establishing an optical simulation model including a polarization beam splitter, a cubic phase mask, a reaction chamber window, and an epitaxial sheet; generating training pairs between the encoded image and the real light spot center coordinates through parameter scanning; performing convolution blurring, noise superposition, and background radiation addition processing on the encoded image; generating real Gaussian heatmap labels using the real light spot centers, and training the U-shaped network through the mean square error loss function. The curvature of the wafer surface is calculated based on the center coordinates and the formula for calculating curvature based on the spot position.
12. The apparatus according to claim 11, characterized in that, The laser source is a single-mode fiber-coupled semiconductor laser; the polarization beam splitting and shifting module includes passive elements made of birefringent crystals; the optical module includes a beam expander group and a beam splitter; the beam expander group is located on the optical path between the laser source and the beam splitter, and is used to expand the beam from the laser source; the beam splitter is located on the optical path between the beam expander group and the window of the reaction chamber, and is used to reflect the expanded beam to the window; the beam expander group includes a first convex lens and a second convex lens, the first convex lens and the second convex lens are arranged confocally, and the distance between them is adjustable.
13. The apparatus according to claim 12, characterized in that, The optical module also includes an adjustable aperture, which is located between the beam expander group and the beam splitter. The adjustable aperture is used to limit the beam diameter to avoid introducing stray light.
Citation Information
Patent Citations
Surface shape detection device and detection method
CN106052585A
Wafer warping degree measuring method and system
CN118328894A
Object key point positioning method, device and equipment
CN119850736A
