Frequency domain regression network for lithographic alignment measurements and methods of use thereof

Abstract

The invention belongs to the technical field of semiconductor processing, and discloses a frequency domain regression network for photoetching alignment measurement and a use method thereof, and the frequency domain regression network comprises a space-frequency domain transform used for cutting an image and then carrying out Fourier transform preprocessing; the frequency domain attention mechanism comprises four stacked self-attention modules, and each self-attention module comprises a trunk branch and a mask branch. According to the method, the data is processed through the frequency domain regression network formed by combining space-frequency domain transformation and a frequency domain attention mechanism, and sub-nano-scale photoetching alignment can be realized.

Description

Technical Field

[0001] The invention belongs to the technical field of semiconductor processing, and in particular relates to a frequency domain regression network for photolithography alignment measurement and a use method thereof. Background Art

[0002] Photolithography alignment is a key step in the photolithography process, and its purpose is to ensure that the pattern on the mask is accurately aligned with the pattern on the silicon wafer to ensure accurate overlay between the patterns after exposure. In the process of realizing the present invention, the inventors found that the existing photolithography alignment measurement algorithm has at least the following problems: it is unable to automatically fit and eliminate the error model of various error sources, resulting in large measurement errors; it is unable to effectively demodulate complex alignment signals, resulting in additional measurement errors brought by the algorithm; it is unable to effectively deal with nonlinear quantities in the measurement system, resulting in a limit to the accuracy of the algorithm strategy. Summary of the invention

[0003] The present invention aims to solve the above technical problems at least to a certain extent. To this end, the present invention aims to provide a frequency domain regression network for lithography alignment measurement and a method for using the network.

[0004] The technical solution adopted by the present invention is:

[0005] A frequency domain regression network for lithography alignment measurement, comprising:

[0006] The spatial frequency domain transform is used to crop the image, and then Fourier transform preprocessing is performed. The amplitude and phase information of the frequency domain after processing is:

[0007] Amplitude I amp =Norm(log(|FFTshift(FFT(I M ))|));

[0008] Phase I pha =Norm(∠(FFTshift(FFT(I M )))),

[0009] Among them, FFT is fast Fourier transform, FFTshift is used to move the zero-frequency component to the center of the spectrum, and Norm is normalization processing;

[0010] Frequency-domain attention mechanism, consisting of four stacked self-attention modules.

[0011] Preferably, the self-attention module includes a trunk branch and a mask branch, the trunk branch includes a residual unit, and the mask branch includes an encoder-decoder structure.

[0012] Preferably, the output H of the self-attention module i,c (I) = T i,c(I)*(1+M i,c (I)), where I is the input of the trunk branch, i and c represent the pixels and channels of the feature map respectively, M is the mask, and T is the feature map.

[0013] A method for using a frequency domain regression network, comprising:

[0014] S1, overlap the silicon wafer and the mask, and let the translation stage start to perform step scanning. At each scanning position, the camera is used to collect and image the formed moiré fringes to form a data set;

[0015] S2, taking the image as the input image and the position value of the displacement stage as the true value, inputting the frequency domain regression network for training, and completing the training by minimizing the loss through repeated training;

[0016] S3. Use the trained frequency domain regression network to perform regression prediction on the actually collected alignment signal.

[0017] Preferably, the mask surface is etched with linear or circular alignment marks. The step scan is a one-dimensional scan or a two-dimensional scan.

[0018] The present invention processes data by combining space-frequency domain transformation and a frequency domain regression network composed of a frequency domain attention mechanism to achieve sub-nanometer lithography alignment. The beneficial effects of the present invention are:

[0019] The frequency domain regression network can effectively extract effective information, filter out error data, and improve measurement accuracy by automatically fitting frequency domain filters; it can process various complex alignment signals and effectively deal with nonlinear problems in the strategy system; it uses global signals for alignment, has good robustness, and has no stringent requirements on beam quality and measurement environment; it has fast alignment speed and high alignment efficiency; BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the network structure of the frequency domain regression network used for lithography alignment measurement in the present invention. DETAILED DESCRIPTION

[0021] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] It should be understood that it should also be noted that in the embodiment, the functions / actions that appear may be different from the order in which the drawings appear. For example, depending on the functions / actions involved, they can actually be executed substantially concurrently, or two figures shown in succession can sometimes be executed in the reverse order.

[0023] like Figure 1 As shown, a frequency domain regression network for lithography alignment measurement in this embodiment includes space-frequency domain transformation and frequency domain attention mechanism.

[0024] The spatial frequency domain transform is used to crop the image. The original moiré fringe image is first cropped to a size of 512×512 pixels, and then preprocessed by Fourier transform. The amplitude and phase information in the frequency domain after processing are:

[0025] Amplitude I amp =Norm(log(|FFTshift(FFT(I M ))|));

[0026] Phase I pha =Norm(∠(FFTshift(FFT(I M )))),

[0027] Among them, Im represents the moiré fringe image, FFT(.) represents fast Fourier transform, FFTshift(.) is used to move the zero-frequency component to the center of the spectrum, |.| represents amplitude extraction, log(.) is the logarithmic function, ∠(.) is used to extract the phase, and Norm(.) performs normalization. In order to ensure the integrity and consistency of frequency domain information, the amplitude I amp and Phase I pha They are concatenated into a two-channel magnitude-phase input image of size 512×512 pixels for network processing.

[0028] The frequency domain attention mechanism is used to extract phase information. Since the linearly distributed moiré fringes contain multiple separated peaks in the spectrum domain, the self-attention (SA) module is introduced into the network to focus on the areas around these peaks for feature extraction. Specifically, the frequency domain attention mechanism includes four stacked self-attention modules. Each self-attention module includes a trunk branch and a mask branch. The feature extraction of the trunk branch can use a variety of excellent network structures, such as VGG, residual unit, and Inception. After performance comparison, the pre-activated residual unit is selected as the basic unit of the attention module. When the input I and output T(I) of the trunk branch are the same size, the mask branch uses an encoder-decoder structure to simulate the feedforward and feedback attention mechanism to generate a mask M(x) for slightly weighted processing of the feature map T(x).

[0029] Since simply stacking self-attention modules may reduce the expressiveness of feature maps and the performance of trunk branches, residual learning is introduced to alleviate this problem. Therefore, the output H(x) of the attention module can be expressed as:

[0030] H i，c (I) = T i，c (I)*(1+M i，c (I)),

[0031] Among them, i and c represent the pixels and channels of the feature map respectively.

[0032] The method for using the above frequency domain regression network includes the steps of:

[0033] S1. Overlap the silicon wafer and the mask, etch linear or circular alignment marks on the mask surface, and let the translation stage start one-dimensional or two-dimensional step scanning. Use the camera to collect and image the formed moiré fringes at each scanning position to form a data set.

[0034] S2, taking the image as the input image and the position value of the displacement stage as the true value, inputting the above-mentioned frequency domain regression network for training, and completing the training by minimizing the loss (the standard is L1 or L2 loss) through repeated training;

[0035] S3. Use the trained frequency domain regression network to perform regression prediction on the alignment signal actually collected to realize the alignment measurement function.

[0036] The method of using this embodiment uses a frequency domain regression network to improve the accuracy of the signal analysis algorithm of the lithography alignment system.

[0037] The present invention is not limited to the above-mentioned optional implementation modes. Anyone can derive other various forms of products under the inspiration of the present invention. However, no matter what changes are made in the shape or structure, all technical solutions that fall within the scope defined by the claims of the present invention fall within the protection scope of the present invention.

Claims

1. A frequency domain regression network for lithography alignment measurement, characterized in that: include: The spatial frequency domain transform is used to crop the image, and then Fourier transform preprocessing is performed. The amplitude and phase information of the frequency domain after processing is: Amplitude I amp = Norm(log(|FFTshift(FFT(I M ))|)); Phase I pha =Norm(∠(FFTshift(FFT(I M )))), Among them, FFT is fast Fourier transform, FFTshift is used to move the zero-frequency component to the center of the spectrum, and Norm is normalization processing; Frequency-domain attention mechanism, consisting of four stacked self-attention modules.

2. The frequency domain regression network according to claim 1, characterized in that: The self-attention module includes a trunk branch and a mask branch, the trunk branch includes a residual unit, and the mask branch includes an encoder-decoder structure.

3. The frequency domain regression network according to claim 2, characterized in that: The output H of the self-attention module i，c (I) = T i，c (I)*(1+M i，c (I)), where I is the input of the trunk branch, i and c represent the pixels and channels of the feature map respectively, M is the mask, and T is the feature map.

4. A method for using the frequency domain regression network according to any one of claims 1 to 3, characterized in that: include: S1, overlap the silicon wafer and the mask, and let the translation stage start to perform step scanning. At each scanning position, the camera is used to collect and image the formed moiré fringes to form a data set; S2, taking the image as the input image and the position value of the displacement stage as the true value, inputting the frequency domain regression network for training, and completing the training by minimizing the loss through repeated training; S3. Use the trained frequency domain regression network to perform regression prediction on the actually collected alignment signal.

5. The method of use according to claim 4, characterized in that: A linear or circular alignment mark is etched on the mask surface.

6. The method of use according to claim 4, characterized in that: The step scan is a one-dimensional scan or a two-dimensional scan.

Images