A neural network-based laser direct writing gray exposure parameter prediction method
By using a forward model based on neural networks to predict grayscale exposure parameters, the problems of time-consuming, labor-intensive, and low-accuracy traditional methods are solved, achieving efficient and accurate grayscale exposure parameter prediction and etching depth control.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SUZHOU LABORATORY
- Filing Date
- 2026-04-27
- Publication Date
- 2026-06-23
AI Technical Summary
Traditional grayscale exposure parameter testing methods are time-consuming, labor-intensive, and have low accuracy. Furthermore, the prediction models have weak generalization ability and cannot efficiently construct high-precision process models.
A forward model based on neural networks is adopted to predict the etching depth through the training dataset, directly determine the exposure time and grayscale value, eliminate the need for repeated experiments, and establish a continuous mapping relationship using neural networks.
It significantly saves time and cost, improves grayscale control accuracy and etching depth prediction accuracy, and ensures that the optimal solution with the minimum grayscale step resolution is achieved.
Smart Images
Figure CN122260728A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of grayscale exposure technology, and particularly relates to a method for predicting laser direct-write grayscale exposure parameters based on neural networks. Background Technology
[0002] With the rapid development of integrated circuits and micro-nano optics, higher demands are being placed on the fabrication technology of complex three-dimensional microstructures. Maskless laser direct writing technology based on digital micromirrors has attracted widespread attention in this field due to its advantages of low cost, high efficiency, and flexible control. The core of this technology lies in controlling the flipping of micromirror units using digital signals to determine the exposure position, and controlling the exposure dose through pixel grayscale values, thereby precisely controlling the etching depth of the photoresist and ultimately achieving the direct fabrication of complex three-dimensional structures.
[0003] In this technology, the key to achieving precise depth control lies in establishing an accurate correspondence between grayscale values and etching depth, and determining the optimal exposure parameters. Currently, the traditional method commonly used in the industry is the dose testing method. This method first requires extensive experiments using grayscale step maps, exposing the material under different combinations of exposure times and light intensities, and then measuring the actual etching depth corresponding to each grayscale level under each parameter combination after development. Subsequently, the parameter range in which the etching depth and grayscale value have a linear relationship is manually selected from all experimental results, and the exposure parameter with the smallest grayscale step resolution within this range is chosen as the processing benchmark. Finally, the grayscale-etching depth relationship curve is fitted using the experimental data under this set of parameters, thereby inferring the required grayscale value based on the target depth.
[0004] However, this traditional method has several significant drawbacks in practical applications. First, the entire dose testing process requires exposure and testing experiments on over a hundred patterns with different exposure doses, consuming a significant amount of time, manpower, and material costs, severely restricting the efficiency of process development and adjustment. Second, due to the massive experimental workload, the test step size of the exposure parameters cannot be set too small; for example, the exposure time step is usually set to 20 milliseconds instead of the 1 millisecond supported by the equipment. This coarse-grained search means that the "optimal" parameters found may not be the truly optimal solution under the equipment's precision limits, thus affecting the final accuracy of grayscale control. Furthermore, the prediction model ultimately used to guide production relies only on the curve fitted by a single set of experimental data. The data sample is limited, the model's generalization ability is weak, resulting in limited accuracy in predicting different target depths. The root cause of these drawbacks lies in the fact that the traditional method relies on exhaustive experiments and manual analysis, failing to efficiently utilize limited experimental data to construct a high-precision global process model, thus becoming a major obstacle to improving the efficiency and accuracy of laser direct-write grayscale exposure processes. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention proposes a neural network-based method for predicting grayscale exposure parameters in laser direct writing, thereby resolving the issues present in the prior art.
[0006] In a first aspect, to achieve the above objective, the present invention provides a method for predicting laser direct-write grayscale exposure parameters based on neural networks, comprising the following steps: Construct a positive neural network model for predicting etching depth; The target etching depth and the preset exposure light intensity are input into the forward neural network model to determine the corresponding exposure time; Based on the determined exposure intensity and exposure time, the grayscale value matching the target etching depth is determined using the forward neural network model.
[0007] Optionally, the process of constructing the forward neural network model includes: Obtain the training dataset, where each set of training data is associated with exposure time, exposure intensity, grayscale value, and actual etching depth obtained through experimental measurement; The neural network is trained using the exposure time, exposure intensity, and grayscale value as model inputs and the actual etching depth as the model output target.
[0008] Optionally, the process of obtaining the training dataset includes: Using a test pattern containing multiple gray levels, laser direct writing exposure experiments were conducted under different combinations of exposure time and exposure light intensity. The photoresist etching depth corresponding to each gray level in the test pattern is measured under each combination of exposure parameters to form the training dataset.
[0009] Optionally, the process of determining the exposure time includes: The preset exposure intensity and maximum gray value are combined with multiple candidate exposure times to form an input parameter group; Each of the input parameter groups is input into the forward neural network model to obtain the predicted etching depth corresponding to each candidate exposure time; From the candidate exposure times, select the exposure time that makes the predicted etching depth greater than the maximum value of the target etching depth.
[0010] Optionally, multiple candidate exposure times are generated by iterating through a preset time range with the smallest time adjustment step that the device can set.
[0011] Optionally, the process of determining the grayscale value includes: The determined exposure intensity and exposure time are combined with multiple candidate grayscale values to form an input parameter group. Each of the input parameter groups is input into the forward neural network model to obtain the predicted etching depth corresponding to each candidate gray value; From the candidate grayscale values, select a grayscale value that makes the difference between the predicted etching depth and the target etching depth less than a preset error threshold.
[0012] Optionally, multiple candidate grayscale values are generated by traversing the entire grayscale range with the smallest integer step size of the digital grayscale values.
[0013] Optionally, the forward neural network model is a multilayer perceptron model containing multiple fully connected layers.
[0014] In a second aspect, the present invention also provides a computer terminal device, comprising: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the neural network-based laser direct-write grayscale exposure parameter prediction method in the first aspect described above.
[0015] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of the laser direct-write grayscale exposure parameter prediction method based on neural networks in the first aspect described above.
[0016] Compared with the prior art, the present invention has the following advantages and technical effects: This invention provides a neural network-based method for predicting grayscale exposure parameters in laser direct writing. By utilizing a trained forward neural network model combined with a backpropagation algorithm, it directly predicts exposure parameters and grayscale values, effectively overcoming the inherent defects of traditional dose testing methods. First, this invention completely eliminates the need for repetitive, large-scale exposure and measurement experiments to find suitable parameters; results are quickly output simply by running the prediction program, significantly saving time and manpower costs required for process development and adjustment. Second, based on the continuous mapping relationship established by the neural network, this invention optimizes virtual parameters with the highest precision supported by the equipment, ensuring that the obtained exposure parameters are the truly optimal solution that minimizes grayscale step resolution, thus improving the precision of depth control. Finally, because the prediction model is based on broad-spectrum experimental data covering multiple sets of different exposure parameters, the learned process model is more comprehensive and accurate than traditional methods based solely on fitting a single experimental curve, thereby achieving higher precision in target etching depth prediction and reproduction. Attached Figure Description
[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a grayscale step image used for preliminary experimental dose testing in an embodiment of the present invention; Figure 2 This is a schematic diagram of the forward neural network model structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the MSE loss of the forward neural network model in an embodiment of the present invention; Figure 4 This is a flowchart illustrating a method for predicting laser direct-write grayscale exposure parameters based on a neural network, according to an embodiment of the present invention. Figure 5 This is a schematic diagram comparing the linear ranges of different exposure parameters in an embodiment of the present invention. (a) is the grayscale-etching depth curve when the exposure time is 300ms and the light intensity is 29%, with a linear range of grayscale of 68-255, an etching depth of 64nm-1881nm, and a grayscale step resolution of 9.7nm; (b) is the grayscale-etching depth curve when the exposure time is 360ms and the light intensity is 37%, with a grayscale linear range of 34-221, an etching depth of 34nm-2158nm, and a grayscale step resolution of 11.3nm. Figure 6 The diagram shows the test results of the grayscale exposure parameter prediction model in an embodiment of the present invention. In (a), the exposure time is 322ms and the exposure intensity is 28% as predicted by the model from 50 randomly generated target depth data with a maximum etching depth of 1700nm. The grayscale image is shown on the left. In (b), the exposure parameters and grayscale image are predicted by the model from randomly generated target depth data with a maximum etching depth of 1800nm. The error between the measured depth and the designed depth does not exceed 100nm. Detailed Implementation
[0018] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0019] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0020] This embodiment provides a method for predicting laser direct-write grayscale exposure parameters based on neural networks, including: constructing a forward neural network model for predicting etching depth; The target etching depth and the preset exposure light intensity are input into the forward neural network model to determine the corresponding exposure time; Based on the determined exposure intensity and exposure time, the grayscale value matching the target etching depth is determined using the forward neural network model.
[0021] Furthermore, the process of constructing the forward neural network model includes: Obtain the training dataset, where each set of training data is associated with exposure time, exposure intensity, grayscale value, and actual etching depth obtained through experimental measurement; The neural network is trained using the exposure time, exposure intensity, and grayscale value as model inputs and the actual etching depth as the model output target.
[0022] Furthermore, the process of obtaining the training dataset includes: Using a test pattern containing multiple gray levels, laser direct writing exposure experiments were conducted under different combinations of exposure time and exposure light intensity. The photoresist etching depth corresponding to each gray level in the test pattern is measured under each combination of exposure parameters to form the training dataset.
[0023] Specifically, the implementation process of this embodiment includes: S1. Preliminary dose testing experiments to obtain data for training the neural network: use Figure 1 The step pattern with 16 increasing gray levels shown was used for dose testing of exposure parameter variations. The exposure time range was 300ms-800ms with a step size of 20ms, and the light intensity range was 25%-41% with a step size of 2%. AR300-80new was used as the tackifier, with a spin-coating speed of 2000rpm and a pre-baking temperature of 180℃ for 2 minutes. AZ5124 was used as the photoresist, with a spin-coating speed of 5000rpm and a pre-baking temperature of 105℃ for 1 minute. After exposure, MIF-300 was used for development for 35 seconds, followed by water fixing for 35 seconds.
[0024] The depth of the step etching was measured using a profilometer.
[0025] S2, Data Preprocessing: The etching depth was combined with the corresponding exposure time, light intensity, and grayscale to form a set of training data. Exposure time, light intensity, and grayscale were used as inputs, and etching depth was used as the true output, resulting in a total of 1856 sets of data. Normalization was applied to the data to eliminate the influence of dimensions and ensure that each set of data conformed to the same distribution, accelerating model convergence. 80% of all data was used as the training set, and 20% as the validation set to monitor model convergence during training.
[0026] Furthermore, the forward neural network model is a multilayer perceptron model containing multiple fully connected layers.
[0027] Specifically, the implementation process of this embodiment includes: S3. Construct a positive neural network model and train the network: Forward neural network structure such as Figure 2 As shown, a 5-layer neural network is used. The input is a one-dimensional vector composed of grayscale, time, and light intensity, and the output is the etching depth. The middle layers are all fully connected layers, with the number of neurons marked in the figure. The activation function for both the input and hidden layers is the ReLU function. The loss function uses mean squared error (MSE) to measure the difference between the predicted etching depth and the true depth. The iterative optimizer uses stochastic gradient descent (SGD), with a learning rate (lr) of 0.01 and momentum of 0.9. The batch size is set to 64 samples per training iteration. The total number of training epochs is set to 2000. The data order is shuffled before each training iteration to enhance data robustness. The forward neural network is then trained.
[0028] Training set loss and validation set loss are as follows Figure 3 As shown, the training set loss is less than 0.00075 and the validation set loss is less than 0.001, which proves the accuracy of the positive neural network model in predicting the etching depth by exposure parameters and grayscale.
[0029] Furthermore, the process of determining the exposure time includes: The preset exposure intensity and maximum gray value are combined with multiple candidate exposure times to form an input parameter group; Each of the input parameter groups is input into the forward neural network model to obtain the predicted etching depth corresponding to each candidate exposure time; From the candidate exposure times, select the exposure time that makes the predicted etching depth greater than the maximum value of the target etching depth.
[0030] Furthermore, multiple candidate exposure times are generated by iterating through a preset time range using the minimum time adjustment step that the device can set.
[0031] Furthermore, the process of determining the grayscale value includes: The determined exposure intensity and exposure time are combined with multiple candidate grayscale values to form an input parameter group. Each of the input parameter groups is input into the forward neural network model to obtain the predicted etching depth corresponding to each candidate gray value; From the candidate grayscale values, select a grayscale value that makes the difference between the predicted etching depth and the target etching depth less than a preset error threshold.
[0032] Furthermore, multiple candidate grayscale values are generated by traversing the entire grayscale range with the smallest integer step size of the digital grayscale values.
[0033] Specifically, the implementation process of this embodiment includes: S4. Use a forward neural network model to predict exposure parameters and grayscale in reverse: The process of reverse prediction of exposure parameters and grayscale is as follows: Figure 4 As shown.
[0034] S4.1 Obtain the exposure parameters that minimize the grayscale step resolution within the linear range: The explanation of linear intervals and grayscale step resolution is as follows: Figure 5 As shown. Figure 5 The grayscale-etch depth curves are shown under two sets of exposure parameters, where Figure 5 (a) Gray-etch depth curves with exposure time = 300ms and light intensity = 29%, with a linear range of grayscale of 68-255, an etching depth of 64nm-1881nm, and a grayscale step resolution of 9.7nm. Figure 5 (b) The grayscale-etching depth curves at an exposure time of 360ms and a light intensity of 37% show a grayscale linear range of 34-221nm, an etching depth of 34nm-2158nm, and a grayscale step resolution of 11.3nm. It can be seen that at lower grayscale values, the etching depth approaches zero, and this range is not practically meaningful for determining the grayscale value corresponding to the target depth. However, when the exposure parameters are larger (…), the etching depth… Figure 5 (b) When the grayscale value exceeds a certain threshold, the etching depth no longer changes with increasing grayscale, indicating that the photoresist has been completely exposed, and this section of the curve is also not of reference value. Therefore, in this embodiment, the region where the etching depth changes approximately linearly with grayscale between two invalid intervals is defined as the linear interval under this exposure parameter. The ratio of the etching depth difference to the grayscale value difference within this interval is the grayscale step resolution, which represents the amount of etching depth change caused by a unit change in grayscale. The smaller the grayscale step resolution, the higher the accuracy of controlling the etching depth through grayscale.
[0035] Combination Figure 5 (a) and Figure 5 As can be seen from (b), the grayscale range widths corresponding to the two sets of parameters are similar within the linear interval, but due to Figure 5 In (a), the exposure parameters are relatively small, the photoresist is not fully exposed, the etching depth variation range is small, and the calculated grayscale step resolution is smaller. Therefore, to obtain a smaller grayscale step resolution, it is necessary to select exposure parameters that do not fully expose the photoresist, while ensuring that the etching depth at the maximum grayscale value is greater than the maximum value of the target depth.
[0036] Because different exposure parameters can yield the same etching depth—for example, an exposure time of 300ms with a light intensity of 33% and an exposure time of 360ms with a light intensity of 27%, both with a grayscale value of 255, result in etching depths of 1938nm and 1940nm respectively, which are very close—this embodiment first fixes the exposure light intensity at 1 and the grayscale value at 255, iterating through the exposure times from 300ms to 800ms with a step size of 1ms. The light intensity, time, and grayscale value are input into a forward neural network to obtain the predicted etching depth. Since the predicted etching depth increases with exposure time when the photoresist is not fully exposed, the exposure parameters that minimize the grayscale step resolution when the predicted depth is just greater than the maximum target depth are recorded as t.
[0037] S4.2. Under the given exposure parameters (light intensity l, time t), find the grayscale value such that the etching depth predicted by the forward neural network is close to the target depth: With fixed light intensity l and time t, iterate through grayscale values, with a grayscale range of 0-255 and a step size of 1.
[0038] Light intensity, time, and grayscale are input into a forward neural network to obtain the predicted etching depth. The absolute value of the difference between the predicted depth and the target depth is taken, and it is checked whether it is less than a set error threshold. If it is less than the threshold, the grayscale value at this time is recorded as g. In this way, the exposure parameters (exposure time, light intensity) and the corresponding grayscale value are predicted in reverse from the target depth.
[0039] S5. Accuracy test of predicted exposure parameters and grayscale value model: Fifty target depth data points were randomly generated, with two sets of maximum values of 1700nm and 1800nm respectively. These target depth data were input into the model, with a fixed light intensity of 28% and an allowable error threshold of 10nm. The exposure parameters and grayscale values predicted by the model are as follows: Figure 6 As shown, where Figure 6 (a) Using 50 randomly generated target depth data points with a maximum etching depth of 1700 nm, the model predicts an exposure time of 322 ms and an exposure intensity of 28%, as shown in the left grayscale image. The right image compares the measured depth of the pattern after exposure and development using these exposure parameters and the grayscale image with the designed depth; the error is no more than 100 nm. Figure 6 (b) The maximum etching depth of the randomly generated target depth data is 1800 nm, and the error between the measured depth and the designed depth does not exceed 100 nm.
[0040] In this embodiment, a computer terminal device is provided, including: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the above-described neural network-based laser direct-write grayscale exposure parameter prediction method.
[0041] In this embodiment, a computer-readable storage medium is also provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the above-described neural network-based laser direct-write grayscale exposure parameter prediction method.
[0042] This invention provides a neural network-based method for predicting grayscale exposure parameters in laser direct writing. By utilizing a trained forward neural network model to directly predict exposure parameters and grayscale values, it effectively overcomes the inherent defects of traditional dose testing methods. First, this invention completely eliminates the need for repetitive, large-scale exposure and measurement experiments to find suitable parameters; results are quickly output simply by running the prediction program, significantly saving time and manpower costs required for process development and adjustment. Second, based on the continuous mapping relationship established by the neural network, this invention optimizes virtual parameters with the highest precision supported by the equipment, ensuring that the obtained exposure parameters are the truly optimal solution that minimizes grayscale step resolution, thus improving the precision of depth control. Finally, since the prediction model is based on broad-spectrum experimental data covering multiple sets of different exposure parameters, the learned process model is more comprehensive and accurate than traditional methods based solely on fitting a single experimental curve, thereby achieving higher precision in target etching depth prediction and reproduction.
[0043] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for predicting laser direct-write grayscale exposure parameters based on neural networks, characterized in that, Includes the following steps: Construct a positive neural network model for predicting etching depth; The target etching depth and the preset exposure light intensity are input into the forward neural network model to determine the corresponding exposure time; Based on the determined exposure intensity and exposure time, the grayscale value matching the target etching depth is determined using the forward neural network model.
2. The method according to claim 1, characterized in that, The process of constructing the aforementioned forward neural network model includes: Obtain the training dataset, where each set of training data is associated with exposure time, exposure intensity, grayscale value, and actual etching depth obtained through experimental measurement; The neural network is trained using the exposure time, exposure intensity, and grayscale value as model inputs and the actual etching depth as the model output target.
3. The method according to claim 2, characterized in that, The process of obtaining the training dataset includes: Using a test pattern containing multiple gray levels, laser direct writing exposure experiments were conducted under different combinations of exposure time and exposure light intensity. The photoresist etching depth corresponding to each gray level in the test pattern is measured under each combination of exposure parameters to form the training dataset.
4. The method according to claim 1, characterized in that, The process of determining the exposure time includes: The preset exposure intensity and maximum gray value are combined with multiple candidate exposure times to form an input parameter group; Each of the input parameter groups is input into the forward neural network model to obtain the predicted etching depth corresponding to each candidate exposure time; From the candidate exposure times, select the exposure time that makes the predicted etching depth greater than the maximum value of the target etching depth.
5. The method according to claim 4, characterized in that, Multiple candidate exposure times are generated by iterating through a preset time range using the minimum time adjustment step that the device can set.
6. The method according to claim 1, characterized in that, The process of determining the grayscale value includes: The determined exposure intensity and exposure time are combined with multiple candidate grayscale values to form an input parameter group. Each of the input parameter groups is input into the forward neural network model to obtain the predicted etching depth corresponding to each candidate gray value; From the candidate grayscale values, select a grayscale value that makes the difference between the predicted etching depth and the target etching depth less than a preset error threshold.
7. The method according to claim 6, characterized in that, Multiple candidate grayscale values are generated by traversing the entire grayscale range using the smallest integer step size of the digital grayscale values.
8. The method according to claim 1, characterized in that, The forward neural network model is a multilayer perceptron model containing multiple fully connected layers.
9. A computer terminal device, characterized in that, include: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the steps of the method as described in any one of claims 1-8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-8.