Laser direct writing overlay precision improving method based on deep learning

Abstract

The invention discloses a laser direct writing overlay precision improving method based on deep learning, which belongs to the technical field of micromachining optics, and comprises the following steps: designing a main vernier, an auxiliary vernier, a standard cross mark and a defect cross mark; carrying out layout generation and laser direct writing data processing; completing defect cross marking and main vernier processing through a laser direct writing system based on a data processing result; defect cross mark recognition and auxiliary vernier machining are completed through a laser direct writing system; the overlay error measurement of the main vernier and the auxiliary vernier is completed; training a neural network model based on the measured main and auxiliary vernier overlay error experiment data set; predicting an overlay error of the to-be-processed graph based on the neural network model; and compensating and adjusting the position coordinates of the to-be-processed graph according to the predicted overlay error of the to-be-processed graph. According to the invention, the overlay alignment precision can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of micro-machining optical technology, and specifically relates to a method for improving the accuracy of laser direct writing overlay based on deep learning. Background Art

[0002] In the field of traditional microfabrication optical technology, laser direct writing lithography technology has the advantages of high precision, high efficiency, and low cost for small-batch lithography of large-aperture micron and submicron feature sizes. However, there is still a certain gap in overlay accuracy compared to projection lithography, making it difficult to meet the needs of higher-precision applications. Existing laser direct writing overlay technology mainly relies on the precise displacement measurement of dual-frequency laser interferometers, combined with the graphic recognition of positioning markers to determine the absolute position of the markers. Among them, the main alignment error caused by the manufacturing error of the markers comes from the graphic recognition error of the markers. Therefore, how to reduce the alignment error caused by the graphic recognition error of the markers is an urgent problem to be solved. Summary of the Invention

[0003] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0004] A method for improving laser direct writing overlay accuracy based on deep learning, comprising:

[0005] Step 1: Design the main and auxiliary cursors, standard cross marks and defect cross marks;

[0006] Step 2: Generate layout and process laser direct writing data based on step 1;

[0007] Step 3: Based on the data processing results of step 2, the defect cross mark and main cursor processing are completed by the laser direct writing system;

[0008] Step 4: Based on the data processing results of step 2, the defect cross mark recognition and auxiliary cursor processing are completed through the laser direct writing system;

[0009] Step 5, complete the main and auxiliary cursor engraving error measurement;

[0010] Step 6: training a neural network model based on the experimental data set of the primary and secondary cursor overlay errors measured in step 5;

[0011] Step 7, predicting the overlay error of the graphics to be processed based on the neural network model in step 6;

[0012] Step 8: Compensate for the overlay error of the to-be-processed pattern predicted in step 7 and adjust the position coordinates of the to-be-processed pattern.

[0013] The present invention has the following beneficial effects:

[0014] (1) The present invention introduces a neural network model to predict the alignment error in advance and compensate for it, thereby overcoming the problem of reduced accuracy caused by the alignment error caused by the graphic recognition error of the marking point in the laser direct writing overlay process, and can effectively improve the overlay alignment accuracy.

[0015] (2) The present invention has the advantages of better compatibility with laser direct writing and no need for equipment modification, compared with traditional physical methods, through process improvements at the software algorithm level. It is convenient and practical. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of the method for improving laser direct writing overlay accuracy based on deep learning of the present invention;

[0017] Figure 2 A schematic diagram of a typical alignment mark. DETAILED DESCRIPTION

[0018] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only intended to illustrate the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0019] This paper proposes a deep learning-based method for improving the accuracy of laser direct writing overlay. By building a neural network model, a large number of defective marking patterns and their alignment errors are trained. The optimized network model is then used to predict the actual marking patterns. The position coordinates of the pattern to be exposed are adjusted based on the prediction results to compensate for alignment errors caused by pattern recognition errors at the marking points. Compared to traditional pattern recognition technology, this deep learning-based method can predict recognition accuracy in advance, reduce alignment errors, and meet the needs of higher-precision applications.

[0020] like Figure 1 As shown, the method for improving the laser direct writing overlay accuracy based on deep learning of the present invention includes the following eight steps:

[0021] Step 1: Design the main and auxiliary cursors, standard cross marks and defect cross marks;

[0022] Step 2: Generate layout and process laser direct writing data based on step 1;

[0023] Step 3: Based on the data processing results of step 2, the defect cross mark and main cursor processing are completed by the laser direct writing system;

[0024] Step 4: Based on the data processing results of step 2, the defect cross mark recognition and auxiliary cursor processing are completed through the laser direct writing system;

[0025] Step 5, complete the main and auxiliary cursor engraving error measurement;

[0026] Step 6: training a neural network model based on the experimental data set of the primary and secondary cursor overlay errors measured in step 5;

[0027] Step 7, predicting the overlay error of the graphics to be processed based on the neural network model in step 6;

[0028] Step 8: Compensate for the overlay error of the to-be-processed pattern predicted in step 7 and adjust the position coordinates of the to-be-processed pattern.

[0029] like Figure 2 As shown in Figure 1, in step 1.1, the design of the main and auxiliary verniers involves two sets of main and auxiliary verniers, one in the X direction and one in the Y direction. Both the main and auxiliary verniers are arranged with multiple micrometer-scale scale intervals and maintain center alignment. The vernier scale lines are 2μm wide, with the main vernier scale interval being 2.1μm and the auxiliary vernier scale interval being 2μm. The minimum resolvable alignment error between the main and auxiliary verniers is 100nm.

[0030] In step 1.2, the standard cross mark is designed to be in the shape of a crosshair, with a cross width of 10 μm and a length of 100 μm. The material used to make the cross mark is opaque metal chromium Cr.

[0031] In step 1.3, the defective cross mark design introduces irregular defects based on the standard cross mark, simulating the production of defective marks in real process scenarios. Using computer image processing techniques, edge roughness and line width errors are added to the image of the standard cross mark, resulting in a cross mark that approximates the actual cross mark observed under a microscope (computer simulation generates a realistic cross mark image). To ensure optimal training results for deep learning, the number of defective cross mark samples is 10,000.

[0032] Step 2 includes: combining the aforementioned defect cross mark and the main cursor to form the exposure pattern required for the first laser direct writing, forming the secondary cursor to form the exposure pattern required for the second laser direct writing, and converting the above-mentioned graphic file into a data format recognizable by laser direct writing. The defect cross mark and the main and secondary cursors are strictly aligned according to the designed position coordinates. The typical position coordinate distribution is: the defect cross mark array is distributed over the entire area of ​​the substrate, and the main and secondary cursors in the X direction are distributed directly above each defect cross mark, and the main and secondary cursors in the Y direction are distributed to the left. The central axis of the main and secondary cursors just passes through the center position of the cross mark. The goal is to achieve perfect alignment of the main and secondary cursors under ideal conditions.

[0033] Step 3 involves transferring the defect cross mark and main cursor onto the substrate surface through photolithography and etching using a laser direct writing system and supporting lithography equipment. The main process steps include coating (such as chromium), coating, baking, exposure, development, etching, and stripping, among other mature processes.

[0034] Step 4 involves identifying the defective cross mark using a laser direct writing system and transferring the secondary vernier to the substrate surface through photolithography. The main process steps include coating, baking, exposure, development, etching, coating (such as chromium), and stripping, all in a mature process. (The combination of steps 3 and 4 involves processing the primary vernier and defective cross mark before processing the secondary vernier.)

[0035] Step 5 includes: measuring the overlay error of the main and auxiliary vernier is achieved by observing the alignment error of the main and auxiliary vernier lines through a microscope. Specifically, observe the deviation between the main and secondary cursor axes (the centermost scale line, usually 21 in total) in the microscope's field of view, and observe which scale line of the secondary cursor aligns with the corresponding scale line of the main cursor (ideally, the centermost scale lines of the main and secondary cursors are aligned, and the remaining scale lines are misaligned. When deviated from the center, they will align with another scale line). When the secondary cursor is aligned with the centermost scale line of the main cursor, the deviation is 0. When the secondary cursor is aligned with the Nth scale line in the positive direction (rightward) of the main cursor, the deviation is the difference between the single scale lengths of the main and secondary cursors (for example, the single scale length of the main cursor is 2.1μm, and the single scale length of the secondary cursor is 2μm) (100nm) × N. When the secondary cursor is aligned with the Nth scale line in the negative direction of the main cursor, the deviation is the difference between the single scale lengths of the main and secondary cursors (100nm) × -N. Use this method to measure and record the overlay error of the two sets of main and secondary cursors near each cross mark. Wherein, N is a positive integer.

[0036] Step 6 includes:

[0037] Step 6.1: Photograph each defect cross mark under the microscope field of view and save it as training data. Save the corresponding primary and secondary cursor overlay errors (including the X and Y directions) as labels. The number of samples in the training data set is 10,000.

[0038] In step 6.2, based on the deep learning model framework being used, divide the dataset of primary and secondary cursor overlay errors into a training set, a validation set, and a test set in a specific ratio (e.g., 8:1:1); the training set contains 8,000 data sets, the validation set contains 1,000 data sets, and the test set contains 1,000 data sets. The deep learning model framework can be a verified open source deep learning framework from existing technologies.

[0039] Step 6.3: Train the neural network model; specifically:

[0040] In step 6.3.1, preprocess the training data, converting the defect cross-mark image into a grayscale image and then removing noise using Gaussian blur. Binarize the training data using an appropriate threshold to obtain a black-and-white image of the defect cross mark. Convert the image data into a four-dimensional array, while maintaining the label data as a two-dimensional array.

[0041] In step 6.3.2, build an appropriate neural network model. The neural network model usually contains multiple convolutional layers, pooling layers, and fully connected layers. The loss function is set to mean square error (MSE). The computer automatically updates the parameters of the neural network through the backpropagation algorithm. After multiple iterations, the hyperparameter settings of the neural network model are optimized.

[0042] Step 6.3.3: Use the optimized neural network model to perform prediction verification on the test set data.

[0043] In step 6.3.4, the loss value of the prediction result and the predicted overlay error of the main and auxiliary cursors in the X and Y directions are calculated and compared with the overlay error of the main and auxiliary cursors in the X and Y directions in the label to analyze the prediction effect of the neural network method for predicting overlay error.

[0044] Step 7 includes: in the actual graphic overlay process, when the laser direct writing has completed one exposure and etching, and left a cross mark and the first layer of graphic structure on the surface of the substrate, and before the second exposure, the cross mark is photographed with a microscope, and the cross mark image is passed into the neural network model trained in step 6, and the trained neural network model is used to predict the overlay error of the graphic to be processed.

[0045] The neural network model of the present invention can be MobileNet (mobile network), EfficientNet (rethinking model extended convolutional neural network), VGG (Visual Geometry Graph Group), etc.

[0046] Step 8 includes compensating and adjusting the position coordinates of the pattern to be processed based on the overlay error predicted in step 7 to obtain a new pattern to be processed. This is then processed by laser direct writing data before starting secondary exposure and subsequent etching.

[0047] The present invention optimizes the neural network model by training a large number of defective marking patterns and their alignment errors. The optimized neural network model is then used to predict the actual marking pattern. The position coordinates of the pattern to be exposed are adjusted based on the prediction results to compensate for alignment errors caused by pattern recognition errors at the marking points. This method overcomes the existing problem of reduced precision in the overlay process due to marking defects, effectively improving overlay alignment accuracy and meeting the needs of higher-precision applications. Furthermore, compared to traditional physical methods, this method offers advantages such as better compatibility with laser direct writing and the elimination of equipment modification, making it convenient and practical.

[0048] The above descriptions are merely embodiments of the present invention and are not intended to limit the scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied to other related system fields, are also included in the scope of protection of the present invention.

[0049] The contents not described in detail in the specification of the present invention belong to the prior art known to those skilled in the art.

Claims

1. A method for improving laser direct writing overlay accuracy based on deep learning, characterized in that: include: Step 1: Design the main and auxiliary cursors, standard cross marks and defect cross marks; Step 2: Generate layout and process laser direct writing data based on step 1; Step 3: Based on the data processing results of step 2, the defect cross mark and main cursor processing are completed by the laser direct writing system; Step 4: Based on the data processing results of step 2, the defect cross mark recognition and auxiliary cursor processing are completed through the laser direct writing system; Step 5, complete the main and auxiliary cursor engraving error measurement; Step 6: training a neural network model based on the experimental data set of the primary and secondary cursor overlay errors measured in step 5; Step 7, predicting the overlay error of the graphics to be processed based on the neural network model in step 6; Step 8: Compensate for the overlay error of the to-be-processed pattern predicted in step 7 and adjust the position coordinates of the to-be-processed pattern.

2. The method for improving laser direct writing overlay accuracy based on deep learning according to claim 1, characterized in that: Step 1 includes: Step 1.1, design of the main and auxiliary cursors; this involves two sets of main and auxiliary cursors, one in the X direction and one in the Y direction. Both the main and auxiliary cursors are arranged according to multiple micron-scale intervals and keep their centers aligned. Step 1.2, design of a standard cross mark; the standard cross mark is designed in the shape of a crosshair, and the material for making the cross mark is opaque metal chromium Cr; Step 1.3, design of defect cross marks: introduce irregular defects based on the standard cross mark to simulate the production of defect marks in real process scenarios. Specifically, through computer image processing technology, add edge roughness and line width error to the image of the standard cross mark to obtain a real cross mark close to that observed under a microscope.

3. The method for improving laser direct writing overlay accuracy based on deep learning according to claim 1, characterized in that: Step 2 includes: combining the defect cross mark and the main cursor designed in step 1 to form the exposure pattern required for the first laser direct writing, forming the exposure pattern required for the second laser direct writing with the auxiliary cursor, and converting the exposure pattern file into a data format recognized by laser direct writing; strictly aligning the defect cross mark with the main and auxiliary cursors according to the designed position coordinates.

4. The method for improving laser direct writing overlay accuracy based on deep learning according to claim 3, characterized in that: In step 2, the position coordinates are distributed as follows: the defect cross mark array is distributed over the entire area of ​​the substrate, the main and auxiliary cursors in the X direction are distributed directly above each defect cross mark, and the main and auxiliary cursors in the Y direction are distributed to the left, and the central axes of the main and auxiliary cursors just pass through the center position of the cross mark.

5. The method for improving laser direct writing overlay accuracy based on deep learning according to claim 4, characterized in that: Step 3 includes: transferring the above-mentioned defect cross mark and main cursor to the substrate surface through the laser direct writing system and supporting photolithography equipment through photolithography and etching processes; the process steps include coating, coating, baking, exposure, development, corrosion, and degumming.

6. The method for improving laser direct writing overlay accuracy based on deep learning according to claim 5, characterized in that: Step 4 includes: identifying the defective cross mark through the laser direct writing system and transferring the secondary cursor to the substrate surface through the photolithography process; the process steps include coating, baking, exposure, development, etching, coating, and degumming.

7. The method for improving laser direct writing overlay accuracy based on deep learning according to claim 1, characterized in that: In step 5, the overlay error of the main and auxiliary verniers is measured by observing the alignment error of the main and auxiliary verniers through a microscope.

8. The method for improving laser direct writing overlay accuracy based on deep learning according to claim 1, characterized in that: Step 6 includes: Step 6.1: Photograph each defect cross mark under the microscope field of view and save it as training data, and save the corresponding main and auxiliary cursor overlay errors in both the X and Y directions as labels; Step 6.2: Divide the dataset of primary and secondary cursor overlay errors into training, validation, and test sets according to a specific ratio based on the deep learning model framework used. Step 6.3, train the neural network model.

9. The method for improving laser direct writing overlay accuracy based on deep learning according to claim 1, characterized in that: Step 7 includes: in the actual graphic overlay process, when the laser direct writing has completed one exposure and etching, and left a cross mark and the first layer of graphic structure on the surface of the substrate, and before the second exposure, the cross mark is photographed with a microscope, and the cross mark image is passed into the neural network model trained in step 6, and the trained neural network model is used to predict the overlay error of the graphic to be processed.

10. The method for improving laser direct writing overlay accuracy based on deep learning according to claim 1, characterized in that: Step 8 includes: based on the overlay error of the pattern to be processed predicted in step 7, compensating and adjusting the position coordinates of the pattern to be processed to obtain a new pattern to be processed; then, after laser direct writing data processing, starting secondary exposure and subsequent etching processing.

Images