Multi-lattice calculation photoetching system self-correction method and device based on AI driving optimization

Abstract

The invention discloses a multi-dot-matrix calculation photoetching system self-correction method and device based on AI driving optimization, the method utilizes a DMD and a camera of the system to directly put an optimization result of an optimizer into an actual optical system, and compared with a prediction result obtained through a mathematical model, a more accurate result can be efficiently obtained, and the accuracy of the system is improved. Influences of aberration, disturbance, environment and other factors which are difficult to predict in a complex system can be overcome. In addition, the deviation between the optimized result and the target value is quantified through the reward function, the advantages and disadvantages of the current DMD hologram group can be accurately evaluated, the DMD hologram with the high reward value is evolved in the updating process, certain updating exists while the advantages of the last round of DMD hologram are inherited, and the updating efficiency is improved. Therefore, the DMD hologram with a higher reward value can be obtained with a larger probability in the next iteration process.

Description

Technical Field

[0001] This invention belongs to the field of laser direct writing technology, specifically involving a self-calibration method for a multi-array computational lithography system based on AI-driven optimization. Background Technology

[0002] Currently, the fabrication of large-area complex micro / nano structures has broad practical application prospects in terms of industrial demand, such as biomimetic drag reduction for aircraft, large-size flexible touch screens, and space thin-film lenses. However, its fabrication faces many challenges, such as how to achieve efficient photolithography processing of micro / nano-scale feature structures at millimeter-scale or larger dimensions, and how to flexibly fabricate and arrange various microstructures without frequently changing masks and other devices.

[0003] High-throughput laser direct-write systems, especially multi-point parallel processing based on digital micromirror devices (DMDs), have attracted much attention due to their writing flexibility, processing accuracy, and processing throughput. However, the final processing quality of parallel systems is severely constrained by optical imaging problems such as uneven light intensity distribution, geometric distortion, and astigmatism. These problems stem from the contradiction between a large field of view and low aberrations in the optical system. Simultaneously, the processing errors of the optical components themselves, subtle deviations during assembly and adjustment, and the non-ideal nature of the illumination source collectively constitute a complex and difficult-to-model system forward transfer function. Traditional optimization methods heavily rely on accurate physical modeling of the system, followed by optical system design and construction based on the model. This process is not only cumbersome and requires high expertise, but also often fails to balance lithographic accuracy and processing throughput due to the mismatch between the model and reality, becoming a key pain point restricting the widespread application of high-throughput laser direct-write systems.

[0004] Patent application CN110221421A discloses a structured light illumination super-resolution microscopy imaging system and method based on machine learning. The system comprises, sequentially along the optical path, a digital micromirror device (DMD), a mirror (M), a first lens (L1), an active phase modulation system (APM), a second lens (L2), and a charge-coupled device (CCD). The method utilizes distorted fringe patterns and their corresponding scattering phase reconstruction coefficients to form a training dataset. A deep convolutional neural network algorithm is then used to construct a corresponding learning model, thereby rapidly correcting distorted fringe patterns with unknown scattering phase reconstruction coefficients to obtain high-precision corrected fringe patterns.

[0005] Patent application CN120510301A discloses an ultrafast three-dimensional refractive index reconstruction method based on compressed sensing illumination. This invention employs a dual-DMD-based multiplexed ODT imaging system to acquire holographic images of multiplexed illumination; modulates and spatially filters the illumination beam using a dual digital micromirror array; deholographically reconstructs the holographic image to obtain composite object light fields at multiple illumination angles, which are then used to train a deep learning neural network for de-multiplexing the light fields; the composite object light fields at each illumination angle are obtained through the deep learning neural network, and the three-dimensional refractive index of the sample is reconstructed using an inverse reconstruction algorithm based on scalar diffraction theory.

[0006] All of the above patent applications require pre-training models in a computer, which utilizes traditional reinforcement learning training. This involves complex mathematical modeling of the system, and the accuracy of the results needs to be improved. Summary of the Invention

[0007] This invention provides a self-calibration method for a multi-array computational lithography system based on AI-driven optimization. This self-calibration method learns and optimizes directly on the actual physical response, enabling efficient and accurate self-calibration of holographic images.

[0008] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention utilizes the system's DMD and camera to directly apply the optimizer's results to the actual optical system. Compared with the prediction results obtained through mathematical models, it can efficiently obtain more accurate results and overcome the influence of unpredictable aberrations, disturbances, environmental factors, etc. in complex systems.

[0009] This invention also uses a reward function to quantify the deviation between the optimization result and the target value, which can accurately evaluate the quality of the current DMD hologram group. This allows DMD holograms with higher reward values ​​to evolve during the update process, achieving a certain degree of update while inheriting the advantages of the previous round of DMD holograms. As a result, DMD holograms with higher reward values ​​are obtained with a greater probability in the next iteration. Attached Figure Description

[0010] Figure 1 A flowchart illustrating a self-calibration method for a multi-array computational lithography system based on AI-driven optimization, provided as a specific embodiment of the present invention; Figure 2 This is a block diagram illustrating a self-calibration method for a multi-array computational lithography system based on AI-driven optimization, provided as a specific embodiment of the present invention. Detailed Implementation

[0011] To better understand the purpose, technical solution, and advantages of this application, the application is described and illustrated below in conjunction with the accompanying drawings and embodiments.

[0012] This invention proposes a data-driven, goal-oriented intelligent optimization method. In this method, the entire optical system is treated as an unknown black box environment. The DMD (Discrete Modulation Device) is endowed with the ability to generate arbitrary modulation patterns, and each pattern setting can be considered as an action performed by an intelligent agent. The camera at the end of the system acts as an "observer," and the final light field image it acquires is the state observation. This invention carefully designs a reward function based on the optimization objective. This function quantitatively analyzes the image acquired by the camera and calculates a scalarized reward value to accurately evaluate the quality of the current action (DMD pattern).

[0013] This invention employs a highly efficient evolutionary policy optimizer. The optimizer's task is to explore a vast action space and find the optimal policy that maximizes the reward (i.e., the final image quality). Unlike traditional reinforcement learning, which requires pre-training a parameterized "value network" or "policy network" in a computer, this invention replaces the simulated environment model with the optical hardware system itself, providing unique and realistic positive feedback. This completely avoids the need for complex and imprecise mathematical modeling of the system, allowing learning and optimization to be performed directly on the actual physical response.

[0014] The method involved in this invention can be applied to a DMD optical system as disclosed in patent CN118778380B, but is not limited to that specific system. This method not only provides a novel and efficient solution for the calibration of DMD computational lithography systems, but also offers important insights for the intelligent self-calibration and adaptive optimization of a wider range of precision instruments.

[0015] This invention provides a self-calibration method for a multi-slot computational lithography system based on AI-driven optimization. The multi-slot computational lithography system includes a DMD and a camera. The method is characterized by... Figure 1 As shown, it includes: Step S1: First, create an initial DMD hologram group and load the holograms onto the digital micromirror array (DMD).

[0016] Subsequently, the system's camera is used to capture the light spot array corresponding to the DMD hologram. The feature information of the light spot array, such as light intensity, concentration, and distortion, is extracted through image recognition processing algorithms. In a specific embodiment, the feature information of the light spot array provided in this embodiment is one or more of the light intensity value, light intensity concentration, and light spot position.

[0017] Step S2: Set a reward function according to the target requirements. Calculate the reward value of each spot based on the feature information of the spot array using the reward function, ensuring that each spot has a corresponding reward. Spots closer to the target will receive a larger reward value. Summing the reward values ​​of the entire spot array yields the reward value of the corresponding DMD hologram. If the reward value of the DMD hologram is lower than the target value, update the DMD hologram group and continue to the next round of iteration optimization. If the reward value of any DMD hologram in the initial or updated DMD hologram group is not lower than the target value, stop the iteration. The DMD hologram with a reward value not lower than the target value is taken as the optimal hologram, completing the self-calibration of the multi-spot computational lithography system.

[0018] The reward function provided in this specific embodiment of the invention is determined by the target requirement and serves to characterize the quality of the current light spot. The calculated reward value is negatively correlated with the difference between the light spot's feature information and the target. The independent variable of the reward function can be single or multiple, determined by the light spot feature information corresponding to the optimization target, such as the light intensity value of the light spot, the light intensity concentration of the light spot, and the offset of the light spot from the ideal grid position. The importance of different light spot feature information is characterized by the weight coefficients of the corresponding independent variables, thereby ensuring that the optimization results fit various optimization targets.

[0019] In one specific embodiment, the reward function is negatively correlated with the deviation between the characteristic information of the extracted spot array and the characteristic information of the target spot array. The deviation between the characteristic information of the extracted spot array and the characteristic information of the target spot array is obtained by weighted summation of the light intensity value, light intensity concentration and spot position offset of the corresponding spot in the extracted spot array and the spot based on the target spot array.

[0020] The evolutionary strategy optimizer provided in this specific embodiment aims to generate the next hologram group based on the previous hologram group. First, a new base hologram is selected according to the reward function corresponding to the previous hologram. Holograms with higher reward values ​​receive a greater chance of selection, allowing more excellent holograms to participate in the evolution. Then, the arrangement of pixels in a portion of the randomly selected holograms is swapped, and a random region of the resulting hologram is flipped to obtain a new hologram group, thus completing the update of the DMD hologram group. The new hologram obtained using this method will largely inherit the advantages of the previous hologram (higher reward value) while incorporating some random updates, ensuring a greater probability of generating better holograms (higher reward values) in the next iteration.

[0021] The method for updating a DMD hologram group provided in a specific embodiment of the present invention includes: calculating the reward value of the spot array to obtain the reward value of its corresponding DMD hologram, and performing evolutionary iteration on the DMD hologram through an optimizer based on the reward value of the DMD hologram to obtain an updated DMD hologram group.

[0022] In one specific embodiment, this embodiment obtains an updated DMD hologram group by iteratively evolving the DMD hologram group through an optimizer, including: The optimizer re-selects DMD holograms based on the reward value of the initial DMD hologram or the previous round of DMD holograms. The selection probability of DMD holograms is determined based on the magnitude of the reward value, with holograms with higher reward values ​​having a greater probability of being selected. Subsequently, some hologram pixels in the selected DMD hologram group are randomly arranged and swapped, and some regions of the swapped DMD holograms are randomly switched on and off to obtain a new DMD hologram group.

[0023] It should be noted that the specific embodiments of the present invention use an optimizer based on optimization algorithms to update the hologram group. These optimization algorithms include gradient descent, particle swarm optimization, ant colony optimization, etc. The evolutionary strategy is only one embodiment, and other algorithms are also within the scope of protection of this patent.

[0024] On the other hand, such as Figure 2 As shown in the figure, a self-calibration device for a multi-slot computational lithography system based on AI-driven optimization is provided in a specific embodiment of the present invention, including an intelligent optimization framework and a hardware framework, wherein the hardware framework is a multi-slot computational lithography system.

[0025] The multi-spot computational lithography system provided in the specific embodiments of the present invention is used to convert an initial DMD hologram or an updated DMD hologram into a corresponding spot array pattern.

[0026] The intelligent optimization framework provided in this specific embodiment of the invention is used to obtain an initial DMD hologram group, load the DMD holograms onto the DMD, obtain a spot array image corresponding to the DMD hologram captured by a camera, realize state observation, extract the characteristic information of the spot array through image recognition processing algorithm based on the spot array image, calculate the reward value of each spot through a reward function based on the feature information of the spot array, and sum the reward values ​​of the entire spot array to obtain the reward value of the DMD hologram. When the reward value of the DMD hologram is lower than the target value, the DMD hologram group is updated and the next iteration is performed to realize the evolution strategy. If the reward value of the DMD hologram is not lower than the target value, the optimal hologram is obtained, and the self-calibration of the multi-spot computational lithography system is completed. The method of updating the DMD hologram group includes: calculating the reward value of the spot array to obtain the reward value of its corresponding DMD hologram, and performing evolutionary iteration on the DMD hologram through an optimizer based on the reward value to obtain the updated DMD hologram group.

[0027] The multi-spot computational lithography system provided in this specific embodiment of the invention comprises, in sequence according to the light propagation direction, a femtosecond laser 1, a digital micromirror array (DMD2), a first lens 3, a second lens 4, a microlens array 5, a third lens 6, an image rotator 7, a dichroic mirror 8, a microscope objective 9, a displacement stage 10, a fourth lens 11, and a camera 12. The femtosecond laser generated by the femtosecond laser 1 is irradiated onto the DMD2 and its amplitude is modulated. Then, it is imaged onto the MLA5 microlens array by a 4F system composed of the first lens 3 and the second lens 4, generating a focal point array on the focal plane of the MLA5 microlens array. This focal point array then passes through the third lens 6, the image rotator 7, the dichroic mirror 8, and the objective lens 9, finally being imaged onto the focal plane of the objective lens on the large-area, high-precision displacement stage 10 for high-throughput, super-resolution two-photon direct-write processing. A small portion of the leaked light passing through the dichroic mirror 8 is imaged onto the target surface of the camera 12 through the fourth lens 11, enabling real-time monitoring of the light spot array.

[0028] The digital micromirror array (DMD2) provided in this specific embodiment modulates the amplitude of the incident light spot. The DMD2 divides the effective pixel area into M×N micromirror array units, with one micromirror array unit corresponding to one unit light spot. Each micromirror array unit includes k×k micromirrors. The DMD2 independently switches the k×k micromirrors in each micromirror array unit to "on" and "off" states, thereby independently controlling the intensity and uniformity of each unit light spot.

[0029] The first lens 3 and the second lens 4 provided in the specific embodiment of the present invention form a 4F system, which is used to image the light field on the digital micromirror array (DMD2) onto the front focal plane of the microlens array (MLA5). This allows each micromirror array unit of the DMD2, which contains k×k micromirrors, to independently and correspondingly control a microlens beam on the MLA5. By controlling the number and distribution of micromirror switches on the DMD2, precise frequency domain control of the microlens beam can be achieved, making the intensity of the focal point obtained through each microlens uniform, while suppressing aberrations, thus achieving higher lithography accuracy and effect.

[0030] The microlens array MLA5 provided in the specific embodiment of the present invention includes M×N microlenses for focusing M×N incident unit light spots. One unit light spot corresponds to one microlens, and a maximum of M×N focal arrays can be generated on the focal plane of the microlens array MLA5.

Claims

1. A self-calibration method for a multi-slot computational lithography system based on AI-driven optimization, wherein the multi-slot computational lithography system includes a DMD and a camera, characterized in that, include: Step S1: Obtain the initial DMD hologram group and load the initial DMD hologram onto the DMD. Obtain the spot array image corresponding to the initial DMD hologram captured by the camera. Extract the characteristic information of the spot array based on the spot array image using an image recognition processing algorithm. Step S2: Based on the feature information of the spot array, calculate the reward value of each spot using the reward function. Sum the reward values ​​of the entire spot array to obtain the reward value of the corresponding initial DMD hologram. If the reward value of the initial DMD hologram is lower than the target value, update the DMD hologram group. If the reward value of the initial DMD hologram is not lower than the target value, obtain the optimal hologram and complete the self-calibration of the multi-spot computational lithography system. The method for updating the DMD hologram group includes: using an optimizer to iterate and evolve the DMD hologram based on the reward value of the initial DMD hologram to obtain the updated DMD hologram group; Step S3: Repeat steps S1-S2 for the updated DMD hologram group.

2. The self-calibration method for a multi-slot computational lithography system based on AI-driven optimization according to claim 1, characterized in that, The reward function is negatively correlated with the deviation between the extracted characteristic information of the spot array and the characteristic information of the target spot array.

3. The self-calibration method for a multi-slot computational lithography system based on AI-driven optimization according to claim 2, characterized in that, The characteristic information of the light spot array includes one or more of the following: light intensity value, light intensity concentration, and light spot position.

4. The self-calibration method for a multi-slot computational lithography system based on AI-driven optimization according to claim 1 or 3, characterized in that, The deviation between the extracted characteristic information of the light spot array and the characteristic information of the target light spot array is obtained by weighted summation of the light intensity value, light intensity concentration and light spot position offset of the corresponding light spot in the extracted light spot array and the light spot based on the target light spot array.

5. The self-calibration method for a multi-slot computational lithography system based on AI-driven optimization according to claim 1, characterized in that, Based on the reward value of the initial DMD hologram, the optimizer iterates through the DMD hologram, including: The optimizer re-selects DMD holograms based on the reward value of the initial DMD holograms, with holograms having a higher probability of being selected. Then, some hologram pixels in the selected DMD hologram group are randomly arranged and swapped, and some areas of the swapped DMD holograms are randomly switched on and off to obtain an updated DMD hologram group.

6. The self-calibration method for a multi-slot computational lithography system based on AI-driven optimization according to claim 1 or 5, characterized in that, An optimizer based on an optimization algorithm completes the update of the hologram group. The optimization algorithm includes gradient descent, particle swarm optimization, or ant colony optimization.

7. A self-calibration device for a multi-slot computational lithography system based on AI-driven optimization, characterized in that, This includes the multi-slot computational lithography system and the intelligent optimization framework; The multi-spot computational lithography system is used to convert the initial or updated DMD hologram into a corresponding spot array pattern. The intelligent optimization framework is used to obtain an initial DMD hologram group, load the initial DMD hologram onto the DMD, obtain a spot array image corresponding to the initial DMD hologram captured by a camera, and extract the feature information of the spot array based on the spot array image using an image recognition processing algorithm. Based on the feature information of the spot array, the reward value of each spot is calculated using a reward function. The reward values ​​of the entire spot array are summed to obtain the reward value of the corresponding initial DMD hologram. When the reward value of the initial DMD hologram is lower than the target value, the DMD hologram group is updated and the next iteration is performed. If the reward value of the initial DMD hologram is not lower than the target value, the optimal hologram is obtained, and the self-calibration of the multi-spot computational lithography system is completed. The method of updating the DMD hologram group includes: based on the reward value of the initial DMD hologram, the DMD hologram is iterated through evolution by an optimizer to obtain the updated DMD hologram group.

8. The self-calibration device for a multi-slot computational lithography system based on AI-driven optimization according to claim 7, characterized in that, The multi-point computational lithography system is arranged in sequence according to the direction of light propagation: femtosecond laser, DMD, 4F system, microlens array, third lens, image rotator, dichroic mirror, objective lens, displacement stage, fourth lens and camera; The femtosecond laser is used to generate femtosecond lasers; The DMD is used to modulate the amplitude of the femtosecond laser based on the received initial DMD hologram or the updated DMD hologram, and the amplitude-modulated femtosecond laser is irradiated onto the microlens array through the 4F system. The microlens array is used to focus the amplitude-modulated femtosecond laser to obtain a focal array. The focal array passes sequentially through the third lens, image rotator, dichroic mirror, and objective lens, and is imaged on the focal plane of the objective lens on the displacement stage. Part of the focal array passes through the dichroic mirror and then through the fourth lens to be imaged onto the target surface of the camera. The camera captures a light spot array pattern corresponding to the initial DMD hologram or the updated DMD hologram.

9. The self-calibration device for a multi-slot computational lithography system based on AI-driven optimization according to claim 8, characterized in that, The DMD comprises m×n micromirror array units, and the sub-beams pass through the m×n micromirror array units to obtain m×n unit light spots; Each micromirror array unit includes k×k micromirrors. By independently switching the k×k micromirrors on and off, the intensity and uniformity of the light spot in each unit can be independently controlled.

10. The self-calibration device for a multi-array computational lithography system based on AI-driven optimization according to claim 8, characterized in that, The microlens array includes m×n microlenses, which are used to focus the incident m×n unit light spots to generate an m×n focal array.

Images