Residual network-based quantitative phase imaging aberration correction method

Abstract

The application discloses a kind of quantitative phase imaging aberration correction methods based on residual network, comprising: step S10, the dynamic multispectral quantitative phase video stream data under the continuous time sequence of post-exposure un-developed photoresist film is obtained;Step S20, the dynamic multispectral quantitative phase video stream data is input into the double-channel residual network model pre-trained;Step S30, the time-domain fluctuation characteristics of the multi-wavelength initial two-dimensional phase gradient distribution data on the continuous time sequence is extracted.The application extracts the time-domain fluctuation characteristics of the multi-wavelength initial two-dimensional phase gradient distribution data on the continuous time sequence, uses principal component analysis fitting and solves first-order spatial partial derivative, generates dimensionless dynamic phase gradient noise data.The application improves the stability and signal-to-noise ratio of the actual measurement data in industrial production site.

Description

Technical Field

[0001] This invention relates to the fields of semiconductor manufacturing inspection and computational optical imaging technology, and in particular to a quantitative phase imaging aberration correction method based on residual networks. Background Technology

[0002] In semiconductor manufacturing processes, achieving non-destructive, high-precision measurement of the three-dimensional refractive index distribution data of latent images of undeveloped photoresist films after exposure is crucial for evaluating photolithography quality and intercepting defective wafers before development. Quantitative phase imaging (QPI), as a non-destructive characterization technique, is commonly used for morphological measurement of micro- and nanostructures. However, in practical industrial inspection environments, traditional QPI methods face the following technical bottlenecks closely related to existing optical and physical mechanisms when applied to the detection of latent images of undeveloped photoresist films after exposure:

[0003] First, due to the physical limitation that optical detectors can only record light intensity, the raw phase data directly obtained through interferometric holographic demodulation will be strictly truncated. Within a finite interval, a two-dimensional encapsulated phase temporal field information with discontinuous jump characteristics is formed. Traditional mathematical unwrapping algorithms such as path tracing or least squares are computationally inefficient and prone to global error propagation when dealing with the high-frequency steep edge structure of photoresist latent images, making it difficult to accurately and quickly extract multi-wavelength initial two-dimensional phase gradient distribution data without phase jumps.

[0004] Secondly, high-frequency micro-vibration interference from fans, pipeline fluids, and robotic arms is unavoidable in semiconductor manufacturing workshops. This interference introduces dynamically changing global common-mode noise during continuous time-series data acquisition. Simultaneously, lens defects or transient focus drift in the optical imaging system itself introduce systematic spatial aberrations. Furthermore, the Ewald spherical shell coverage of a single illumination wavelength in three-dimensional frequency space is too narrow, resulting in severe degradation of the axial resolution of the reconstructed three-dimensional refractive index distribution data. Traditional hardware vibration damping solutions and conventional imaging algorithms are not only costly but also unable to adaptively remove dynamic phase gradient noise data introduced by environmental vibrations, thus making it difficult to achieve high-fidelity three-dimensional tomographic reconstruction without spatial aberrations. Summary of the Invention

[0005] The present invention aims to at least partially solve one of the technical problems in the above-mentioned technologies.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A quantitative phase imaging aberration correction method based on residual networks, the method being executed by a computing device, the method comprising: step S10, acquiring dynamic multispectral quantitative phase video stream data of an undeveloped photoresist film after exposure in a continuous time series, wherein the dynamic multispectral quantitative phase video stream data includes two-dimensional encapsulated phase temporal field information of the undeveloped photoresist film after exposure under multiple different illumination wavelengths.

[0008] Step S20: Input the dynamic multispectral quantitative phase video stream data into a pre-trained dual-channel residual network model. Process the dynamic multispectral quantitative phase video stream data through the feature extraction layer of the dual-channel residual network model to obtain the initial two-dimensional phase gradient distribution data of multiple wavelengths corresponding to multiple different illumination wavelengths.

[0009] Step S30: Extract the temporal fluctuation characteristics of the multi-wavelength initial two-dimensional phase gradient distribution data on the continuous time series, calculate the dynamic phase gradient noise data introduced by environmental vibration, and use the dynamic phase gradient noise data to perform differential compensation operation on the multi-wavelength initial two-dimensional phase gradient distribution data to generate denoised two-dimensional phase gradient distribution data.

[0010] Step S40: Combine the denoised two-dimensional phase gradient distribution data with the Fourier diffraction theorem and the Ewald sphere frequency coverage model to perform a three-dimensional spectrum stitching operation, and reconstruct the three-dimensional spatial refractive index distribution data of the latent image of the undeveloped photoresist film after exposure with spatial aberrations eliminated.

[0011] Furthermore, in step S10, acquiring dynamic multispectral quantitative phase video stream data of the undeveloped photoresist film after exposure in a continuous time series includes: receiving the original interferometric holographic video stream data of the undeveloped photoresist film after exposure, which is continuously acquired at high frequency within a set time window by a multi-wavelength interferometric measurement system.

[0012] Each frame of the original interferometric holographic video stream is sequentially processed by a two-dimensional fast Fourier transform to obtain spatial spectral feature data corresponding to each discrete sampling time point.

[0013] In the spatial spectral feature data, high-pass filtering is used to separate the positive first-order diffraction order sub-spectral data containing optical field phase information, and the positive first-order diffraction order sub-spectral data is subjected to two-dimensional inverse Fourier transform processing to extract complex amplitude distribution information.

[0014] Calculate the arctangent function value of the complex amplitude distribution information to generate a function with... The dynamic multispectral quantitative phase video stream data with phase wrapping characteristics within the interval.

[0015] Furthermore, obtaining multi-wavelength initial two-dimensional phase gradient distribution data in step S20 includes: splitting the dynamic multispectral quantitative phase video stream data into long-wavelength group phase channel data and short-wavelength group phase channel data according to wavelength frequency bands.

[0016] The long-wavelength group phase channel data and the short-wavelength group phase channel data are respectively input into the first residual encoder path and the second residual encoder path of the dual-channel residual network model to perform multi-scale downsampling feature extraction operations, thereby extracting the long-wavelength deep semantic feature matrix and the short-wavelength deep semantic feature matrix.

[0017] The feature matrix is ​​reconstructed by cross-channel cascaded splicing and transposed convolution upsampling through the densely connected decoder path of the dual-channel residual network model, and the multi-wavelength initial two-dimensional phase gradient distribution data at each spatial pixel is output.

[0018] Furthermore, in step S30, the dynamic phase gradient noise data introduced by environmental vibration is calculated, including: performing principal component analysis projection processing on the time series matrix of the multi-wavelength initial two-dimensional phase gradient distribution data to isolate the low-frequency global common-mode phase fluctuation function representing the background mechanical structure micro-vibration interference. , where variables This represents the dynamic sampling time point in the continuous time series. Represents a two-dimensional spatial coordinate point.

[0019] The low-frequency global common-mode phase ripple function in the spatial domain By performing first-order spatial partial derivative calculations, the corresponding dynamic global phase gradient bias function can be obtained. .

[0020] According to the dynamic global phase gradient bias function And the corresponding acquisition timestamp, mapped to generate the dynamic phase gradient noise data corresponding to each spatial pixel coordinate, realizing the unified conversion of the fluctuation error dimension.

[0021] Furthermore, in step S30, generating denoised two-dimensional phase gradient distribution data includes: selecting a spatial coordinate point as a first pixel node from the multi-wavelength initial two-dimensional phase gradient distribution data, and extracting the initial phase gradient matrix value of the first pixel node at the current time.

[0022] Read the noise gradient bias compensation value corresponding to the first pixel node from the dynamic phase gradient noise data.

[0023] The initial phase gradient matrix value and the noise gradient bias compensation value are input into the arithmetic logic unit for pixel-by-pixel matrix subtraction iterative operation to calculate the residual gradient tensor data that eliminates the interference of the environmental vibration error.

[0024] The subtraction iteration operation is repeated by traversing all spatial coordinate points and using each point as a new first pixel node to reconstruct the denoised two-dimensional phase gradient distribution data.

[0025] Furthermore, in step S40, the denoised two-dimensional phase gradient distribution data is combined with the Fourier diffraction theorem and the Ewald sphere frequency coverage model to perform a three-dimensional spectrum stitching operation, including: performing a two-dimensional spatial Fourier transform on the denoised two-dimensional phase gradient distribution data to obtain two-dimensional frequency domain scattering potential function data corresponding to multiple illumination wavelengths.

[0026] Based on the Ewald sphere frequency coverage model, the two-dimensional frequency domain scattering potential function data corresponding to different wavelengths are mapped and projected onto the surfaces of different hemispherical shells in the three-dimensional spatial frequency coordinate system.

[0027] The frequency domain iterative convex set projection algorithm is used to perform phase consistency constraints and interpolation filling fusion operations on the overlapping and missing spectral components in the three-dimensional spatial frequency coordinate system, so as to synthesize latent image broadband three-dimensional integrated frequency domain scattering potential tensor data with isotropic high resolution characteristics, and complete the dimensional upgrade from planar image to three-dimensional tomographic spectrum.

[0028] Furthermore, in step S40, the three-dimensional spatial refractive index distribution data of the latent image of the undeveloped photoresist film after exposure, which eliminates spatial aberrations, is reconstructed, including: performing a three-dimensional inverse Fourier transform calculation operation on the broadband three-dimensional integrated frequency domain scattering potential tensor data of the latent image to restore the three-dimensional dielectric constant perturbation spatial distribution matrix of the latent image.

[0029] A refractive index polarization conversion model based on the Lorentz-Lorentz equation is established, and the parameters of the initial matrix macroscopic polarizability constant of the undeveloped photoresist film after exposure and the microscopic characteristic polarizability attenuation coefficient of the photogenerated acid molecules are introduced.

[0030] The spatial distribution matrix of the three-dimensional dielectric constant perturbation of the latent image is input into the refractive index polarization conversion model based on the Lorentz-Lorentz equation to perform a nonlinear microscopic response algebraic solution mapping operation, and finally outputs the high-fidelity reconstructed three-dimensional spatial refractive index distribution data of the latent image of the undeveloped photoresist film after exposure.

[0031] Furthermore, after reconstructing the three-dimensional spatial refractive index distribution data of the undeveloped photoresist film latent image after exposure with spatial aberrations eliminated in step S40, the method further includes: retrieving the standard photomask layout design geometry file and performing forward optical exposure and development calculation simulation using a rigorous coupled-wave analysis algorithm to generate a three-dimensional refractive index distribution reference model of the target ideal aberration-free latent image.

[0032] The reconstructed three-dimensional spatial refractive index distribution data and the three-dimensional refractive index distribution reference model of the target ideal aberration-free latent image are synchronously input into the spatial three-dimensional registration analyzer to perform rigid coordinate alignment and difference operations, and the three-dimensional refractive index absolute error deviation matrix is ​​calculated.

[0033] When the local peak variance in the three-dimensional refractive index absolute error deviation matrix exceeds the linewidth critical tolerance extreme value, feedforward interception warning signal data characterizing the risk of aberration distortion failure in the current exposure process is generated in real time.

[0034] Furthermore, the pre-trained dual-channel residual network model is trained and constructed through the following steps: collecting multiple sets of multispectral interferometric images of historical wafers under preset aberration perturbation conditions, and strictly performing two-dimensional fast Fourier transform and single-level extraction operations to generate sample two-dimensional wrapped phase time-series data consistent with the physical properties of the inference stage.

[0035] Based on Maxwell's law of forward propagation of electromagnetic waves, the true three-dimensional refractive index topography label of the wafer obtained through destructive measurement is reverse-calculated into theoretically ideal aberration-free two-dimensional phase gradient reference data.

[0036] The sample two-dimensional wrapped phase time-series data is input into the model to be initialized to output the predicted phase gradient distribution data. The network weight parameters are updated using the theoretical ideal aberration-free two-dimensional phase gradient benchmark data as the target until the total loss function converges, thus obtaining the pre-trained model.

[0037] Furthermore, after generating feedforward interception warning signal data in real time to characterize the risk of aberration distortion failure in the current exposure process, the method further includes: calling the inter-process asynchronous communication interface protocol library of the device operating system kernel layer to encapsulate the feedforward interception warning signal data and the three-dimensional refractive index absolute error deviation matrix of high-dimensional features into a standard manufacturing execution system instruction control data packet.

[0038] The standard manufacturing execution system instruction control data packets are sent to the high-priority control bus buffer queue of the lithography machine master computer node through a high-stability industrial real-time Ethernet physical link channel.

[0039] The unmasked hardware interrupt exception request sequence operation of the lithography machine master control computer node is triggered, which forcibly blocks and stops the extreme ultraviolet scanning exposure action of the current wafer batch on the moving mask stage, and outputs the error wafer feedback information to the graphical interface for reminder.

[0040] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0041] (1): This invention extracts the temporal fluctuation characteristics of multi-wavelength initial two-dimensional phase gradient distribution data on a continuous time series, uses principal component analysis to fit and solve the first-order spatial partial derivatives, and generates dimensionlessly uniform dynamic phase gradient noise data. Subsequently, a hardware arithmetic logic unit is used to perform pixel-by-pixel matrix subtraction iteration operations. This mechanism can effectively decouple and remove low-frequency mechanical common-mode vibration errors generated by the stepping motion of the lithography machine mask stage or scanning stage. Thus, this invention gets rid of the strong dependence of traditional interferometry on vibration isolation in extreme environments and improves the stability and signal-to-noise ratio of measured data in industrial production sites.

[0042] (2): This invention combines the Fourier diffraction theorem and the Ewald sphere frequency coverage model to perform hemispherical shell projection and frequency domain interpolation on two-dimensional frequency domain scattering potential function data of different wavelengths, realizing the dimensional upgrade from a two-dimensional plane to a three-dimensional frequency coordinate system. Furthermore, by introducing the Lorentz-Lorentz equation, the reconstructed three-dimensional dielectric constant perturbation spatial distribution matrix of the latent image is converted into microscopic hydrogen ion (photogenerated acid) concentration data. This technical solution does not require destructive physical slicing of the wafer, nor does it require waiting for development and imaging. It can directly output latent image three-dimensional spatial refractive index distribution data containing depth resolution information, meeting the online non-destructive testing requirements of advanced photolithography processes for sub-nanometer-level latent structure features.

[0043] (3): This invention employs an offline pre-trained dual-channel residual network model to process long and short wavelength group phase data. Compared to traditional interferometric measurement techniques that rely on multi-step mechanical phase shifting or complex optical path compensation, this scheme transfers the complex nonlinear optical aberration calculation process to the weight parameters of the neural network. In the online execution phase, only forward inference reconstruction through a densely connected decoder is required to simultaneously complete phase roll-up recovery and high-frequency aberration elimination. This significantly reduces the processing cycle of a single frame of data while preserving the high-frequency edge details of the photoresist latent image, meeting the timeliness requirements of large-volume online wafer measurement.

[0044] (4): This invention not only completes the quantitative evaluation of latent image quality, but also constructs a complete manufacturing execution feedback chain. By spatially registering and calculating the variance of the three-dimensional spatial refractive index data reconstructed from the measured data with the ideal benchmark model generated by forward simulation based on the standard lithography mask layout, an early warning can be triggered when the error deviates from the critical tolerance. The system directly sends a hardware interrupt request data packet to the lithography machine master control node by calling the underlying communication protocol, forcibly blocking the exposure action. This feedforward interception mechanism can detect and isolate wafers with focus drift or abnormal dosage in time before photoresist development, blocking defective wafers from flowing to downstream high-cost process links, thereby effectively controlling wafer scrap costs and improving the overall production line yield. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is the main flowchart of the quantitative phase imaging aberration correction method based on residual networks proposed in this invention;

[0047] Figure 2 This is a diagram of the dual-channel residual network feature extraction and reconstruction architecture proposed in this invention. Detailed Implementation

[0048] 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 some embodiments of the present invention, and not all embodiments. 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.

[0049] The following describes, with reference to the accompanying drawings, a quantitative phase imaging aberration correction method based on residual networks according to an embodiment of the present invention.

[0050] like Figure 1 As shown, Figure 1 This is the main flowchart of the quantitative phase imaging aberration correction method based on residual networks proposed in this invention. This invention provides a quantitative phase imaging aberration correction method based on residual networks, comprising the following steps:

[0051] Step S10: Obtain dynamic multispectral quantitative phase video stream data of the undeveloped photoresist film after exposure in a continuous time series, wherein the dynamic multispectral quantitative phase video stream data includes two-dimensional encapsulated phase temporal field information of the undeveloped photoresist film after exposure under multiple different illumination wavelengths.

[0052] In practical applications, this step aims to acquire initial raw data on the microscopic refractive index changes within the photoresist caused by the exposure chemical reaction using non-destructive optical interferometry. "Undeveloped photoresist film after exposure" refers to the wafer surface coating that has been exposed to ultraviolet light in a photolithography machine but has not yet been washed away by chemical developer. At this stage, only a slight difference in refractive index exists between the exposed and unexposed areas (i.e., a latent image). "Dynamic multispectral quantitative phase video stream data" refers to a sequence of images containing phase information acquired continuously over time using a multi-wavelength interferometric microscopy system. Because optical detectors can only record light intensity and cannot directly record phase, the phase value is truncated when calculating the phase using the arctangent function in interferometry. Within a given interval, discontinuous abrupt changes occur in the phase distribution. This original phase field information containing these phase abrupt changes is termed "two-dimensional wrapped phase temporal field information." By acquiring this two-dimensional wrapped phase temporal field information across multiple wavelengths in a continuous time series, a rich high-frequency spatiotemporal data foundation can be provided for subsequent elimination of single-wavelength phase wrapping ambiguity and separation of environmental vibration noise.

[0053] For example, in the photolithography manufacturing process of 14nm silicon-based wafers, a wafer coated with a 100nm thick photoresist layer is sent to the inspection station after being exposed by a 193nm argon fluoride (ArF) excimer laser. The interferometric microscopy system is activated, emitting wavelengths of [wavelength values ​​missing]. , as well as The system continuously acquires multi-wavelength interferometric images for 2 seconds at a sampling rate of 100 frames per second. After preliminary processing of the images, the computing device generates dynamic multispectral quantitative phase video stream data containing latent image features of a 14-nanometer mask pattern. At this time, the phase distribution map at different wavelengths is uniformly filled with the detection beam. Enveloping stripes caused by truncation.

[0054] In this invention, the "set time window" is preset to a specific value range of 10ms to 50ms to ensure sufficient temporal freedom is captured without introducing excessive redundant data. The "raw interferometric holographic video stream data" recorded by the off-axis interferometry system is actually the intensity fringe pattern formed after the object light wave and the reference light wave interfere. A "two-dimensional fast Fourier transform" (2D-FFT) is performed on it to convert the image from the spatial domain to the spatial frequency domain. In the frequency domain, the central spectrum containing the zero-order background intensity and the "positive first-order diffraction order sub-spectral data" containing the true object light phase are physically separated. A digital high-pass filter is used to separately cut out the positive first-order diffraction order sub-spectral data and shift it to the center of the frequency domain, then an inverse transform (2D-IFFT) is performed to obtain the "complex amplitude distribution information" (containing real and imaginary parts) of the object light wave. Finally, the phase value is calculated by taking the arctangent of the ratio of the imaginary part to the real part. The advantage of this method is that it can acquire amplitude and phase simultaneously in a single exposure, eliminating the need for multi-step phase shifting mechanisms and significantly improving the throughput of industrial inspection.

[0055] For example, in the aforementioned 14nm lithography measurement system, the interferometer captured the striped original interferometric holographic video stream data within the set time window of 20ms. The computing device performed a two-dimensional fast Fourier transform on the first frame image, observing a bright zero-order spot in the center and symmetrically distributed first-order diffraction spots on both sides in the frequency domain image. The system automatically located and used a circular high-pass filter window to cut out the positive first-order diffraction order sub-spectral data on the right side, and then performed an inverse transform to obtain a complex amplitude tensor. Subsequently, for each matrix element in the tensor, the arctangent values ​​of the imaginary and real parts were calculated to generate a dynamic multispectral quantitative phase video stream data frame with a value range strictly limited to [-3.1415, 3.1415].

[0056] Step S20: Input the dynamic multispectral quantitative phase video stream data into a pre-trained dual-channel residual network model. Process the dynamic multispectral quantitative phase video stream data through the feature extraction layer of the dual-channel residual network model to obtain the initial two-dimensional phase gradient distribution data of multiple wavelengths corresponding to multiple different illumination wavelengths.

[0057] In practical applications, the core of this step lies in using an artificial intelligence model to replace the traditional complex physical phase unwrapping algorithm, directly extracting continuous phase gradient information from the truncated wrapped phase. The "pre-trained dual-channel residual network model" is a deep learning-based convolutional neural network architecture, consisting of an input layer, a hidden feature extraction layer containing multi-level residual blocks, and an output layer.

[0058] Before inputting the dynamic multispectral quantitative phase video stream data into the dual-channel residual network model, a data preprocessing operation is performed: the values ​​in the wrapped phase tensor are within the range of... The values ​​are normalized to the [-1, 1] interval using a linear function to prevent gradient divergence in the early stages of network training. In terms of the specific network architecture configuration, each residual block in the first and second residual encoder paths adopts a standard structure including a batch normalization layer, specifically: a 3×3 convolutional layer, a batch normalization layer, and a LeakyReLU activation function (with a negative half-axis slope). The input features are then processed by a second 3×3 convolutional layer and a second batch normalization layer. The input features and the output of the residual block are added element-wise through a short-circuit connection.

[0059] When a residual block spans a downsampling layer, the short-circuit connection branch uses a 1×1 convolution with a stride of 2 for spatial downsampling and channel alignment. Regarding the channel number configuration rules, the number of feature channels in the first and second residual encoder paths are set to 64, 128, 256, and 512 respectively during the four downsampling processes with a stride of 2. In the densely connected decoder path, the number of channels is halved at each level during upsampling via transposed convolution; during cross-channel concatenation, a 1×1 convolution is used to compress the number of channels in the encoder-side feature map to match the number of channels in the corresponding layer on the decoder side before concatenation is performed along the channel dimension.

[0060] Specifically, this invention aims to solve the dimensionality mapping problem between dynamic video stream processing and two-dimensional physical network architecture. The dynamic multispectral quantitative phase video stream data physically exhibits a dimension of... The four-dimensional tensor, in which and Spatial pixel resolution, The number of channels for illumination wavelengths. This is a continuously sampled time series. To preserve temporal continuity for subsequent common-mode noise stripping while fully utilizing the advantages of two-dimensional spatial networks in extracting high-frequency latent image features, the feature extraction layer adopts a processing logic of "frame-by-frame slicing of the time series and parallel input of spatial channels." That is, at any discrete timestamp... Below, the network reception dimension is The two-dimensional spatial feature slices are used as input. The output layer of the network model uses a linear activation function, selecting each pixel in the output space as an input. direction and The initial two-dimensional phase gradient distribution data of multiple wavelengths in the direction is used instead of directly outputting the absolute unwrapped phase. The physical consideration is that the phase gradient is the spatial first-order partial derivative of the absolute phase. This derivative operation not only naturally filters out the global constant phase offset constant brought about by the lithography system, but also greatly amplifies the high-frequency steep characteristics of the photoresist edge. Thus, in terms of physical principle, it connects to the subsequent differential stripping step based on principal component analysis for low-frequency time-domain common-mode vibration noise.

[0061] For example, the computing device inputs the dynamic multispectral quantitative phase video stream data, containing the three wavelength channels and a resolution of 1024×1024 pixels, into the pre-trained dual-channel residual network model. The hidden layer of this model performs convolution operations on 64 initial feature maps, autonomously identifies the edge features of the wrapping stripes, and smoothly connects across transition boundaries. After millisecond-level forward propagation calculations, the output layer outputs three sets of initial two-dimensional phase gradient distribution data matrices corresponding to wavelengths of 450nm, 532nm, and 633nm. This matrix clearly presents the contour gradient changes of the 14nm lithographic latent image without any artificial phase transition truncation.

[0062] like Figure 2 As shown, Figure 2 This is a diagram of the dual-channel residual network feature extraction and reconstruction architecture proposed in this invention. Long-wavelength light waves have greater penetrating power and fewer phase wrapping abrupt changes, while short-wavelength light waves have extremely high lateral spatial resolution. Dividing the data into "long-wavelength groups" and "short-wavelength groups" and feeding them into independent residual encoder paths allows the network to autonomously learn the most suitable features for wavelengths with different physical characteristics during training. After splitting the data by wavelength band, for single-channel long wavelengths (e.g.,...)... Construct the first residual encoder path for dual-channel short wavelengths (e.g.) and A second residual encoder path is constructed. Each of the two residual encoder paths contains a combination of four levels of residual blocks, and all convolutional kernels are set to a size of 3×3. To preserve the high-frequency spatial location information of the photoresist micro / nano structure, the downsampling process abandons the max pooling operation, which easily leads to loss of detail, and instead uses a stride-2 convolution with stride to compress the feature dimension, generating a deep semantic feature matrix. In the decoding stage, the densely connected decoder path contains four transposed convolutional layers that are completely symmetrical to the encoder, with a transposed convolutional kernel size of 4×4 and an upsampling factor of 2. When performing cross-channel cascade stitching, the network extracts the feature maps of the corresponding levels in the encoder, first precisely aligns the channel dimensions of the feature maps through a 1×1 convolution operation, and then performs depth stitching with the feature maps of the decoder in the channel dimensions. The end of the dual-channel residual network model has a 1×1 linear output layer, which maps the fused deep features to a strict set of six output channels, where the first two channels correspond to longer wavelengths. and The gradient tensor, with the last four channels corresponding to two short wavelengths respectively. and The gradient tensor enables an unambiguous physical mapping from multi-wavelength phase to multi-channel spatial gradient.

[0063] For example, the computing device inputs the 633nm data from the previous example as the long-wavelength group phase channel data into the first residual encoder path, and inputs the 450nm and 532nm data as the short-wavelength group phase channel data into the second residual encoder path. After four layers of pooling downsampling, the long-wavelength deep semantic feature matrix and the short-wavelength deep semantic feature matrix with dimensions of 64×64 are extracted respectively. Subsequently, the densely connected decoder path stacks and merges these two matrices along the channel dimension, and undergoes multiple transpose convolutions to enlarge the size, finally outputting the high-resolution, dual-band 14nm photoresist latent image multi-wavelength initial two-dimensional phase gradient distribution data at the end of the network.

[0064] During the sample generation phase, to ensure model generalization ability, the true three-dimensional refractive index topography labels of the wafer, which are relied upon for the inverse calculation of the Rigorous Coupled Wave Analysis (RCWA) algorithm, are uniformly sampled within the following boundaries: the latent image linewidth critical dimension is 10nm to 20nm, and the sidewall angle is 85 degrees to 90 degrees. The value range of the preset aberration perturbation conditions is randomly generated based on the first 9 Zernike polynomials, where the defocus aberration is set to... to Spherical aberration is set to to .

[0065] During model training, the network weights were initialized using the Kaiming normal distribution, and the batch size was set to 16. The optimizer used the Adam algorithm, with an initial learning rate of 1×10⁻⁶. -3 It is coupled with a cosine annealing learning rate decay strategy, with an annealing period set to 50 epochs and a minimum learning rate threshold set to 1×10. -6 Total loss function Depend on Mean square error of directional gradient The formula is a combination of the mean square error of the directional gradient and the regularization term, and its specific calculation formula is as follows:

[0066]

[0067] The weighting coefficients are configured as follows: , , Regularization penalty coefficient .

[0068] For example, to prepare an AI model suitable for 14nm lithography lines, the research team deliberately introduced defocus aberrations within the aforementioned preset range into the lithography machine for batch residual exposures, collecting interferometric images containing encapsulation mutations to create 100,000 sample two-dimensional encapsulated phase time-series data. Simultaneously, identical wafers were selected and cut using FIB (Fiber Optic Injection) and placed under a TEM (Transmission Electron Microscope) to obtain nanoscale real physical cross-sections, which were then used to derive correct theoretical ideal aberration-free two-dimensional phase gradient reference data labels. These 100,000 samples were fed into an untrained dual-channel residual network model and subjected to 24 hours of continuous gradient descent backpropagation calculations for up to 300 rounds on a high-performance GPU server equipped with an Adam optimizer, until the mean square error converged to the order of 0.001. Saving the connection weights in this state constitutes the core pre-trained model for successfully extracting gradients in the aforementioned steps.

[0069] Step S30: Extract the temporal fluctuation characteristics of the multi-wavelength initial two-dimensional phase gradient distribution data on the continuous time series, calculate the dynamic phase gradient noise data introduced by environmental vibration, and use the dynamic phase gradient noise data to perform differential compensation operation on the multi-wavelength initial two-dimensional phase gradient distribution data to generate denoised two-dimensional phase gradient distribution data.

[0070] In practical applications, this step aims to eliminate unavoidable mechanical micro-vibration errors in precision optical measurements purely based on data post-processing algorithms. Since the structure of the latent image on the measured wafer is statically constant, environmental vibrations (such as platform vibrations caused by cleanroom fans and cooling water pipes) can lead to dynamic shifts in the imaging field of view. Extracting "temporal fluctuation features" involves statistically analyzing the gradient value sequence of each pixel along the time axis. By separating the signal components that exhibit low-frequency periodic global changes over time, "dynamic phase gradient noise data" characterizing the system's mechanical jitter is calculated. Subsequently, a differential compensation operation is performed, subtracting the dynamic phase gradient noise data tensor from the original gradient tensor at the corresponding spatial position in each frame. Through these operations, without the need to install expensive physical vibration isolation stages or hardware displacement sensors on the lithography inspection equipment, adaptive suppression of environmental common-mode noise can be achieved at the algorithmic level, significantly improving the data signal-to-noise ratio of extremely small latent image features.

[0071] For example, within a continuous 2-second acquisition cycle, the aforementioned computing device extracts the time series of the initial two-dimensional phase gradient distribution data of the multi-wavelength array. Data analysis reveals that all pixel gradient values ​​within the entire field of view are accompanied by a periodic fluctuation component with a frequency of approximately 12Hz. The system defines this 12Hz global fluctuation component as the dynamic phase gradient noise data. Subsequently, in the computation of each frame, the original 14-nanometer line gradient matrix is ​​subtracted pixel-by-pixel from the 12Hz noise bias matrix, thereby eliminating the edge blurring caused by the mechanical micro-vibration of the wafer stage and generating the denoised two-dimensional phase gradient distribution data with extremely high static stability.

[0072] Since the neural network in the preceding steps outputs a static phase gradient distribution for inference on a single frame image, a temporal matrix needs to be reconstructed first to extract temporal fluctuation features. Specifically, for a spatial resolution of... The number of consecutive sampling frames is Phase gradient video stream, at each independent moment ,Will The two-dimensional spatial gradient matrix is ​​flattened row by row into a length of A one-dimensional row vector. Then, The row vectors at each time step are stacked layer by layer along the time dimension to construct a structure with dimension . The two-dimensional time-series covariance evaluation matrix. After inputting the matrix into the PCA algorithm, singular value decomposition is performed, and the left singular vector corresponding to the largest eigenvalue is extracted as the low-frequency global common-mode phase ripple function. Discrete representation of .

[0073] For example, for the aforementioned 200-frame continuous matrix containing 14nm latent image data, after the algorithm performs PCA projection processing, it is found that the first principal component occupies [a certain position]. The signal energy variance exhibits sinusoidal fluctuation characteristics at a specific frequency, and this feature vector is extracted as the low-frequency global common-mode phase fluctuation function. The computing device then applied the function to... and The derivative is calculated in the direction of the problem to obtain the dynamic global phase gradient bias function corresponding to the jitter of the wafer stage. Since the derivative operation unifies the physical dimensions, this function is time-stamped. The specific dynamic phase gradient noise data matrix (e.g., global lateral offset dimension is generated at the location) is generated. This provides a benchmark for the next step of precise noise reduction.

[0074] The Arithmetic Logic Unit (ALU) is the core computing component within a central processing unit (CPU) or graphics processing unit (GPU) that performs basic arithmetic operations. Anchored in a spatial coordinate system, at a determined first pixel node position, the initial gradient value of the static image inferred by the network is subtracted from the noise bias value calculated in the previous PCA step. This pixel-by-pixel matrix subtraction iterative operation not only achieves precise removal of physically significant vibration errors, but also, because its traversal loop structure is highly compatible with the parallel computing thread model of the GPU, it enables real-time processing of extremely high-resolution industrial inspection images.

[0075] For example, in the detection image of a 14nm latent image, the computing device designates the pixel at coordinates (500, 500) as the first pixel node. The GPU thread reads this node... The initial phase gradient matrix value at time t is 0.38. Subsequently, the GPU immediately retrieves the noise gradient bias compensation value (i.e., 0.015 in the previous example) corresponding to that location at that time from memory. A fast subtraction operation is performed in the GPU's arithmetic logic unit, yielding a value of 0.365, which is the residual gradient tensor data after jitter removal. The GPU activates thousands of parallel threads, traversing all millions of pixels in less than 1 ms, completely generating a clear, denoised two-dimensional phase gradient distribution data map.

[0076] Step S40: Combine the denoised two-dimensional phase gradient distribution data with the Fourier diffraction theorem and the Ewald sphere frequency coverage model to perform a three-dimensional spectrum stitching operation, and reconstruct the three-dimensional spatial refractive index distribution data of the latent image of the undeveloped photoresist film after exposure with spatial aberrations eliminated.

[0077] In practical applications, this step elucidates the core physical conversion principle for upgrading from two-dimensional planar information to three-dimensional spatial morphology. The Fourier diffraction theorem states that the two-dimensional forward scattered field generated by a weakly scattering object under plane wave illumination in a specific direction is equivalent to the cross-sectional distribution of the object's three-dimensional scattering potential function on a hemisphere (i.e., the "Ewald sphere") in frequency space. Since a single wavelength can only cover a very small thin layer in frequency space, the axial (Z-axis) resolution is extremely poor. Therefore, this step utilizes the previously acquired data of multiple different illumination wavelengths and, based on the Ewald sphere frequency coverage model, aligns and stitches hemispheres of different sizes corresponding to different wavelengths in three-dimensional frequency space, filling in the missing parts of the frequency space. After stitching, an inverse transform operation is performed to convert it from the frequency domain back to the spatial domain, thus reconstructing the true three-dimensional refractive index distribution data. By comprehensively expanding the spatial frequency bandwidth of the optical detection system through algorithmic synthesis, three-dimensional tomographic imaging of the latent image inside the photoresist was achieved without moving the objective lens during mechanical axial scanning, and systematic spatial aberrations such as defocus and coma caused by optical lens defects were eliminated.

[0078] For example, the computing device performs a spatial two-dimensional Fourier transform based on the denoised two-dimensional phase gradient distribution data. The high-frequency scattering field corresponding to 450nm is mapped to an Ewald sphere with a larger radius, and the low-frequency scattering field corresponding to 633nm is mapped to a sphere with a smaller radius. By seamlessly stitching and fusing the spectral data of these three hemispherical surfaces with different radii in a three-dimensional coordinate system, a broadband three-dimensional frequency sphere data is formed. Subsequently, a three-dimensional inverse Fourier transform is performed, and a three-dimensional morphological image of the exposed area of ​​the 14nm node wafer is visually output on the display of the computing device. The image accurately shows the three-dimensional spatial refractive index distribution data, where the refractive index inside the latent image region is 1.65, while the refractive index of the surrounding unexposed region is 1.61, achieving non-destructive penetration observation of the latent image sidewall morphology.

[0079] Before performing a two-dimensional spatial Fourier transform on the denoised two-dimensional phase gradient distribution data, the gradient distribution is mapped to a phase distribution using the frequency domain integration method, and then transformed into two-dimensional frequency domain scattering potential function data by combining the Born approximation relationship.

[0080] When performing spectrum stitching using the Iterative Convex Set Projection (POCS) algorithm in the Ewald sphere three-dimensional frequency coordinate system, the phase consistency constraint operation specifically performed is as follows: In the... In this iteration, for known frequency coordinate regions in the frequency space that are already covered by the effective illumination wavelength, the amplitude of the original effective spectrum is forced to be used. Replace the complex spectral magnitude generated in the current iteration, while retaining the phase information of the current iteration. This constitutes a new projection operator feature. .

[0081] Subsequently, a nonlinear refractive index polarization conversion model based on the Lorentz-Lorenz equation was established. Let the absolute refractive index of the target medium at a certain point in space be... The optical refractive index constant of the initial matrix before exposure is The micro-polarization attenuation coefficient of the photoresist material is The concentration of photogenerated acid at the corresponding location was... The algebraic mapping formula is specifically defined as follows:

[0082]

[0083] Taking ArF photoresist, commonly used at the 14nm node, as an example, substituting constants... Experimental calibration coefficients for corresponding materials This allows for the precise mapping of the three-dimensional refractive index distribution data within the microscopic volume after exposure by solving for the nonlinear algebraic roots in three-dimensional space.

[0084] For example, for a 14nm photoresist system, the two-dimensional frequency domain scattering potential function data corresponding to the aforementioned three wavelengths are mapped to a three-dimensional spatial frequency coordinate system. The 450nm data constructs a large-radius outer shell, while the 633nm data constructs a small-radius inner shell, leaving large areas of unknown blank frequency domain data between the shells. The computing device initiates the frequency domain iterative convex set projection algorithm, setting the spatial constraint that the scattering potential outside the photoresist latent image region is strictly equal to zero. After 200 rounds of interpolation and constraint correction iterations, the algorithm fills in the blank bands between the shells, successfully synthesizing complete and dense 14nm latent image broadband three-dimensional integrated frequency domain scattering potential tensor data, thus establishing a three-dimensional spectrum.

[0085] The three-dimensional inverse Fourier transform can only reconstruct the weak changes in the relative electromagnetic field caused by exposure, i.e., the dielectric constant perturbation. To allow semiconductor process engineers to understand this intuitively, it must be converted to the absolute refractive index. The Lorentz-Lorentz equation reveals the quantitative relationship between the macroscopic dielectric constant (the square of the refractive index) and the polarizability of the microscopic molecules within the medium. When photoresist is exposed to extreme ultraviolet light, the photoacid-generating agent decomposes to produce photoacid molecules, which leads to the breaking and recombination of chemical bonds in local polymer molecules, thereby altering the local microscopic characteristic polarizability of the material. Introducing these specific chemical prior constants and solving them using nonlinear algebraic mappings allows for the precise conversion of the purely abstract optical scattering potential into a specific material refractive index spectrum. This operation directly breaks down the barrier between optical measurement and photochemical material analysis, providing a direct physical criterion for evaluating the rationality of exposure dose allocation.

[0086] For example, regarding the broadband three-dimensional integrated frequency domain scattering potential tensor data of the latent image of the aforementioned 14nm chip, after the computing device performs a three-dimensional inverse transformation, the perturbation of the latent image dielectric constant deviates from the reference by 0.024. The system then retrieves the material database, substituting the initial matrix macroscopic polarizability constant of this type of ArF photoresist and the corresponding microscopic characteristic polarizability attenuation coefficient of the photogenerated acid molecule. After nonlinear algebraic operations in the refractive index polarization conversion model based on the Lorentz-Lorentz equation, the final output three-dimensional matrix result shows that the true refractive index of the exposed core region reaches 1.652, while the edge gradient region smoothly transitions to 1.611. This result is the three-dimensional spatial refractive index distribution data of the latent image of the undeveloped photoresist film after exposure, intuitively demonstrating the three-dimensional spatial chemical diffusion morphology of the latent image.

[0087] Standard photolithography mask layout design geometry files typically refer to chip layout files in GDSII format. Using rigorous coupled-wave analysis (RCWA) algorithms from the field of computational lithography, the theoretically aberration-free light intensity distribution within the photoresist can be simulated based on pure Maxwell's electromagnetic field theory, thereby generating a three-dimensional refractive index distribution reference model. The actual detected and reconstructed distribution data is then registered and aligned point-to-point with the theoretical reference data in virtual three-dimensional space. The resulting three-dimensional refractive index absolute error deviation matrix represents the deviation caused by the actual exposure process. The "linewidth critical tolerance extreme value" is specifically preset to a range of 0.03 to 0.08. An early warning is triggered when the error exceeds the set tolerance threshold extreme value. Its superior technical effect lies in its ability to detect optical projection system deviation or focal plane drift before the wafer undergoes expensive chemical development and etching processes, allowing for early scrapping or reprocessing, thus saving significant wafer material and time costs.

[0088] For example, the server retrieves the GDSII file of the 14nm SRAM memory cell mask and uses the RCWA algorithm to quickly calculate the corresponding three-dimensional refractive index distribution reference model of the target ideal aberration-free latent image (assuming the refractive index step at the line edge should be a vertical 90° sidewall). The system then aligns and subtracts the actual measured and reconstructed three-dimensional refractive index data with tilt angles. The result shows that in a certain region at the edge of the chip array, the local peak variance calculated by the three-dimensional refractive index absolute error deviation matrix reaches 0.06, exceeding the system's preset linewidth critical tolerance extreme value of 0.05. The system immediately judges that there is a risk of latent image collapse in this region due to edge aberrations of the lithography machine projection lens, and instantly generates the feedforward interception warning signal data to prevent the 14nm process from continuing to advance to the development process.

[0089] Asynchronous inter-process communication allows monitoring algorithms to operate independently without blocking or freezing the main process. Standard Manufacturing Execution System (MES) instructions are the standardized workshop management language for semiconductor factories. After merging and encapsulating the aforementioned generated early warning signals and detailed error matrices, data packets can bypass the conventional Internet layer and reach the nerve center of the lithography machine—the main control computer node—at extremely high speed via an industrial real-time Ethernet (e.g., EtherCAT) physical link. Most importantly, it employs an unmasked hardware interrupt exception request sequence instruction, meaning that the interception alarm has the highest physical execution priority. It directly suspends any other tasks being executed by the lithography machine's CPU, immediately cutting off the power supply to the servo motors of the mechanical table through a physical emergency stop. This achieves a true leap from pure software detection results to hard real-time physical protection, ensuring the avoidance of the tragedy of hundreds of consecutive wafers being scrapped due to homogeneous distortion.

[0090] For example, when the monitoring algorithm generates an early warning signal for the 14nm SRAM chip edge latent image collapse as described in the previous example, the server running the detection program immediately uses inter-process asynchronous communication to convert the alarm signal into a standard manufacturing execution system instruction control data packet using the SECS / GEM protocol. This packet reaches the main control computer node of the ASML lithography machine in the wafer fab within 2ms via an independent industrial real-time Ethernet physical link channel of EtherCAT within the factory. This packet forcibly sends a high-priority unmasked hardware interrupt exception request to the lithography machine motherboard. At the instant the instruction is executed, the next 14nm wafer step-scan exposure operation being performed by the lithography machine immediately stops and locks. At the same time, a red "3D latent image aberration exceeds limit, batch processing blocked" error feedback information window pops up on the graphical interface in the operator's room, and the process engineer immediately stops the line to intervene and investigate.

[0091] The present invention has the following advantages:

[0092] (1): This invention extracts the temporal fluctuation characteristics of multi-wavelength initial two-dimensional phase gradient distribution data on a continuous time series, uses principal component analysis to fit and solve the first-order spatial partial derivatives, and generates dimensionlessly uniform dynamic phase gradient noise data. Subsequently, a hardware arithmetic logic unit is used to perform pixel-by-pixel matrix subtraction iteration operations. This mechanism can effectively decouple and remove low-frequency mechanical common-mode vibration errors generated by the stepping motion of the lithography machine mask stage or scanning stage. Thus, this invention gets rid of the strong dependence of traditional interferometry on vibration isolation in extreme environments and improves the stability and signal-to-noise ratio of measured data in industrial production sites.

[0093] (2): This invention combines the Fourier diffraction theorem and the Ewald sphere frequency coverage model to perform hemispherical shell projection and frequency domain interpolation on two-dimensional frequency domain scattering potential function data of different wavelengths, realizing the dimensional upgrade from a two-dimensional plane to a three-dimensional frequency coordinate system. Furthermore, by introducing the Lorentz-Lorentz equation, the reconstructed three-dimensional dielectric constant perturbation spatial distribution matrix of the latent image is converted into microscopic hydrogen ion (photogenerated acid) concentration data. This technical solution does not require destructive physical slicing of the wafer, nor does it require waiting for development and imaging. It can directly output latent image three-dimensional spatial refractive index distribution data containing depth resolution information, meeting the online non-destructive testing requirements of advanced photolithography processes for sub-nanometer-level latent structure features.

[0094] (3): This invention employs an offline pre-trained dual-channel residual network model to process long and short wavelength group phase data. Compared to traditional interferometric measurement techniques that rely on multi-step mechanical phase shifting or complex optical path compensation, this scheme transfers the complex nonlinear optical aberration calculation process to the weight parameters of the neural network. In the online execution phase, only forward inference reconstruction through a densely connected decoder is required to simultaneously complete phase roll-up recovery and high-frequency aberration elimination. This significantly reduces the processing cycle of a single frame of data while preserving the high-frequency edge details of the photoresist latent image, meeting the timeliness requirements of large-volume online wafer measurement.

[0095] (4): This invention not only completes the quantitative evaluation of latent image quality, but also constructs a complete manufacturing execution feedback chain. By spatially registering and calculating the variance of the three-dimensional spatial refractive index data reconstructed from the measured data with the ideal benchmark model generated by forward simulation based on the standard lithography mask layout, an early warning can be triggered when the error deviates from the critical tolerance. The system directly sends a hardware interrupt request data packet to the lithography machine master control node by calling the underlying communication protocol, forcibly blocking the exposure action. This feedforward interception mechanism can detect and isolate wafers with focus drift or abnormal dosage in time before photoresist development, blocking defective wafers from flowing to downstream high-cost process links, thereby effectively controlling wafer scrap costs and improving the overall production line yield.

[0096] In the description of this specification, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0097] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0098] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A quantitative phase imaging aberration correction method based on residual networks, characterized in that, The method is performed by a computing device, and the method includes: Step S10: Obtain dynamic multispectral quantitative phase video stream data of the undeveloped photoresist film after exposure in a continuous time series, wherein the dynamic multispectral quantitative phase video stream data includes two-dimensional encapsulated phase temporal field information of the undeveloped photoresist film after exposure under multiple different illumination wavelengths. Step S20: Input the dynamic multispectral quantitative phase video stream data into a pre-trained dual-channel residual network model, and process the dynamic multispectral quantitative phase video stream data through the feature extraction layer of the dual-channel residual network model to obtain the initial two-dimensional phase gradient distribution data of multiple wavelengths corresponding to multiple different illumination wavelengths. Step S30: Extract the temporal fluctuation characteristics of the multi-wavelength initial two-dimensional phase gradient distribution data on the continuous time series, calculate the dynamic phase gradient noise data introduced by environmental vibration, and use the dynamic phase gradient noise data to perform differential compensation operation on the multi-wavelength initial two-dimensional phase gradient distribution data to generate denoised two-dimensional phase gradient distribution data. Step S40: Combine the denoised two-dimensional phase gradient distribution data with the Fourier diffraction theorem and the Ewald sphere frequency coverage model to perform a three-dimensional spectrum stitching operation, and reconstruct the three-dimensional spatial refractive index distribution data of the latent image of the undeveloped photoresist film after exposure with spatial aberrations eliminated.

2. The method according to claim 1, characterized in that, In step S10, dynamic multispectral quantitative phase video stream data of the undeveloped photoresist film after exposure is obtained in a continuous time series, including: Receive the raw interferometric holographic video stream data of the undeveloped photoresist film after exposure, which is continuously acquired at high frequency within a set time window by a multi-wavelength interferometric measurement system; Each frame of the original interferometric holographic video stream data is sequentially processed by a two-dimensional fast Fourier transform to obtain spatial spectral feature data corresponding to each discrete sampling time point; In the spatial spectrum feature data, high-pass filtering is used to separate the positive first-order diffraction order sub-spectral data containing optical field phase information, and the positive first-order diffraction order sub-spectral data is subjected to two-dimensional inverse Fourier transform processing to extract complex amplitude distribution information. Calculate the arctangent function value of the complex amplitude distribution information to generate a function with... The dynamic multispectral quantitative phase video stream data with phase wrapping characteristics within the interval.

3. The method according to claim 1, characterized in that, In step S20, the initial two-dimensional phase gradient distribution data of multiple wavelengths is obtained, including: The dynamic multispectral quantitative phase video stream data is divided into long-wavelength group phase channel data and short-wavelength group phase channel data according to wavelength band. The long-wavelength group phase channel data and the short-wavelength group phase channel data are respectively input into the first residual encoder path and the second residual encoder path of the dual-channel residual network model to perform multi-scale downsampling feature extraction operations, and extract the long-wavelength deep semantic feature matrix and the short-wavelength deep semantic feature matrix. The feature matrix is ​​reconstructed by cross-channel cascaded splicing and transposed convolution upsampling through the densely connected decoder path of the dual-channel residual network model, and the multi-wavelength initial two-dimensional phase gradient distribution data at each spatial pixel is output.

4. The method according to claim 1, characterized in that, In step S30, the dynamic phase gradient noise data introduced by environmental vibration is calculated, including: Principal component analysis projection processing is performed on the time series matrix of the initial two-dimensional phase gradient distribution data of the multi-wavelength system to isolate the low-frequency global common-mode phase ripple function representing the background mechanical structure micro-vibration interference. , where variables This represents the dynamic sampling time point in the continuous time series. Represents a two-dimensional spatial coordinate point; The low-frequency global common-mode phase ripple function in the spatial domain By performing first-order spatial partial derivative calculations, the corresponding dynamic global phase gradient bias function can be obtained. ; According to the dynamic global phase gradient bias function And the corresponding acquisition timestamp, mapped to generate the dynamic phase gradient noise data corresponding to each spatial pixel coordinate, realizing the unified conversion of the fluctuation error dimension.

5. The method according to claim 1, characterized in that, In step S30, generating denoised two-dimensional phase gradient distribution data includes: Select a spatial coordinate point from the multi-wavelength initial two-dimensional phase gradient distribution data as the first pixel node, and extract the initial phase gradient matrix value of the first pixel node at the current time. Read the noise gradient bias compensation value corresponding to the first pixel node from the dynamic phase gradient noise data; The initial phase gradient matrix value and the noise gradient bias compensation value are input into the arithmetic logic unit for pixel-by-pixel matrix subtraction iterative operation to calculate the residual gradient tensor data that has eliminated the environmental vibration error interference. The subtraction iteration operation is repeated by traversing all spatial coordinate points and using each point as a new first pixel node to reconstruct the denoised two-dimensional phase gradient distribution data.

6. The method according to claim 1, characterized in that, In step S40, the denoised two-dimensional phase gradient distribution data is combined with the Fourier diffraction theorem and the Ewald sphere frequency coverage model to perform a three-dimensional spectrum stitching operation, including: Perform a two-dimensional spatial Fourier transform on the denoised two-dimensional phase gradient distribution data to obtain two-dimensional frequency domain scattering potential function data corresponding to multiple illumination wavelengths; According to the Ewald sphere frequency coverage model, the two-dimensional frequency domain scattering potential function data corresponding to different wavelengths are mapped and projected onto the surface of different hemispherical shells in the three-dimensional spatial frequency coordinate system. The frequency domain iterative convex set projection algorithm is used to perform phase consistency constraints and interpolation filling fusion operations on the overlapping and missing spectral components in the three-dimensional spatial frequency coordinate system, so as to synthesize latent image broadband three-dimensional integrated frequency domain scattering potential tensor data with isotropic high resolution characteristics, and complete the dimensional upgrade from planar image to three-dimensional tomographic spectrum.

7. The method according to claim 6, characterized in that, In step S40, the three-dimensional spatial refractive index distribution data of the latent image of the undeveloped photoresist film after exposure, which has been freed from spatial aberrations, is reconstructed, including: Perform a three-dimensional inverse Fourier transform calculation on the broadband three-dimensional composite frequency domain scattering potential tensor data of the latent image to reconstruct the three-dimensional dielectric constant perturbation spatial distribution matrix of the latent image; A refractive index polarization conversion model based on the Lorentz-Lorentz equation is established, and the initial matrix macroscopic polarizability constant of the undeveloped photoresist film after exposure and the microscopic characteristic polarizability attenuation coefficient of the photogenerated acid molecules are introduced as parameters. The spatial distribution matrix of the three-dimensional dielectric constant perturbation of the latent image is input into the refractive index polarization conversion model based on the Lorentz-Lorentz equation to perform a nonlinear microscopic response algebraic solution mapping operation, and finally outputs the high-fidelity reconstructed three-dimensional spatial refractive index distribution data of the latent image of the undeveloped photoresist film after exposure.

8. The method according to claim 7, characterized in that, After reconstructing the three-dimensional spatial refractive index distribution data of the latent image of the undeveloped photoresist film after exposure, which has been freed from spatial aberrations, in step S40, the method further includes: The standard photomask layout design geometry file is retrieved and the forward optical exposure and development calculation simulation is performed using a rigorous coupled-wave analysis algorithm to generate a three-dimensional refractive index distribution reference model of the target ideal aberration-free latent image. The reconstructed three-dimensional spatial refractive index distribution data and the three-dimensional refractive index distribution reference model of the target ideal aberration-free latent image are synchronously input into the spatial three-dimensional registration analyzer to perform rigid coordinate alignment and difference operations, and the three-dimensional refractive index absolute error deviation matrix is ​​calculated. When the local peak variance in the three-dimensional refractive index absolute error deviation matrix exceeds the linewidth critical tolerance extreme value, feedforward interception warning signal data characterizing the risk of aberration distortion failure in the current exposure process is generated in real time.

9. The method according to claim 1, characterized in that, The pre-trained dual-channel residual network model is constructed through the following steps: Collect multiple sets of multispectral interferometric images of historical wafers under preset aberration perturbation conditions, and strictly perform two-dimensional fast Fourier transform and single-level extraction operations to generate sample two-dimensional wrapped phase time-series data consistent with the physical properties of the inference stage; Based on Maxwell's law of forward propagation of electromagnetic waves, the real three-dimensional refractive index topography label of the wafer obtained through destructive measurement is reverse-calculated into theoretical ideal aberration-free two-dimensional phase gradient reference data. The sample two-dimensional wrapped phase time-series data is input into the model to be initialized to output the predicted phase gradient distribution data. The network weight parameters are updated using the theoretical ideal aberration-free two-dimensional phase gradient benchmark data as the target until the total loss function converges, thus obtaining the pre-trained model.

10. The method according to claim 8, characterized in that, After generating feedforward interception warning signal data in real time, which characterizes the risk of aberration and distortion failure in the current exposure process, the method further includes: The inter-process asynchronous communication interface protocol library of the device operating system kernel layer is invoked to encapsulate the feedforward interception and early warning signal data and the three-dimensional refractive index absolute error deviation matrix of high-dimensional features into a standard manufacturing execution system instruction control data packet; The standard manufacturing execution system instruction control data packets are sent to the high-priority control bus buffer queue of the lithography machine main control computer node through a high-stability industrial real-time Ethernet physical link channel; The unmasked hardware interrupt exception request sequence operation of the lithography machine master control computer node is triggered, which forcibly blocks and stops the extreme ultraviolet scanning exposure action of the current wafer batch on the moving mask stage, and outputs the error wafer feedback information to the graphical interface for reminder.

Images