Photoetching dynamic reference sub-millisecond generation method and system

Abstract

The invention provides a photoetching dynamic reference sub-millisecond level generation method and system. The method comprises the following steps: firstly, constructing a grading curved surface grid database covering a full parameter domain based on a mirror ideal curved surface equation, carrying out fine granularity division on a high-curvature region and pre-storing a third-order derivative, and carrying out coarse granularity division on a non-high-curvature region; then receiving workpiece table pose data in real time to obtain a processing position parameter domain mapping point, and dynamically generating an interpolation block by taking the point as a center and combining the acceleration of the workpiece table; executing three-stage interpolation of bilinear interpolation positioning, Hermite spline interpolation and photoresist rheological equation correction through a GPU tensor core to obtain dynamic reference curved surface data; and finally, calculating an interpolation error, and feeding back and optimizing the grid database when the interpolation error exceeds a preset value, so as to realize dynamic reference sub-millisecond-level generation.

Description

Technical Field

[0001] This invention relates to the field of precision optical processing technology in semiconductor photolithography manufacturing, specifically to a method and system for generating sub-millisecond dynamic photolithography references. Background Technology

[0002] In the process of manufacturing lenses using lithography machines, the precision of the lithography process directly determines the surface accuracy and optical performance of the lens, and is the core link to ensure that the lens meets the imaging requirements of the lithography machine. As lithography machines evolve towards advanced technologies such as deep ultraviolet (DUV) and extreme ultraviolet (EUV), the precision requirements for the feature dimensions of the manufactured lenses have increased to the nanometer level. Nanometer-level deformation of the lens substrate during the exposure process can lead to surface deviations of the lens, thereby affecting the optical imaging quality of the subsequent lithography machine and causing serious optical aberrations.

[0003] To address these issues, the industry has widely integrated online inspection functions into lithography systems used in lens manufacturing (such as the ASMLTwinscan NXT:2050i lithography system). Through the system's built-in dynamic reference generation unit, the ideal surface equations of the lens design model are converted in real-time into reference data matching the current processing position, providing a basis for precision control during lens exposure processing. However, existing dynamic reference generation technologies for lithography lens manufacturing have the following key drawbacks, severely restricting the real-time performance and surface accuracy of lens processing, and failing to meet the manufacturing requirements of advanced lithography lenses: a. Computational bottleneck caused by high-order differential operations: In the process of generating dynamic references for lens manufacturing, it is necessary to perform inverse projection calculations on the ideal surface equations, which involves third-order and higher derivative operations. Taking the ideal surface equation of a typical aspherical lens as an example: The third derivative The calculation requires complex fractional operations, resulting in a computational complexity of up to [value missing]. In ASML's patent document US20200124231A1, the technical solution disclosed uses a CPU to serially calculate the third derivative at a single point, requiring 12.7 μs. This is for a 200 mm diameter lens (including 10mm). 6 The generation of a complete reference surface (with 1 calculation point) takes 2.3ms. However, the exposure cycle of current EUV lithography systems used for lens manufacturing has been compressed to less than 1ms (such as the Nikon NSR-S630D lithography system). The above calculation delay will cause the reference data to lag behind the workpiece stage displacement by more than 1mm (when the workpiece stage moving speed is 1.2m / s), and the reference inaccuracy rate is as high as 42%, which directly leads to deviations in the surface shape of the lens processing.

[0004] b. Mismatch between static computing architecture and dynamic operating conditions: Existing lithography systems used for lens manufacturing (such as the Canon FPA-1200NZ2C) employ a fixed-resolution mesh computing mode, which cannot adapt to the dynamic motion of the workpiece stage. During the uniform motion phase of the workpiece stage, a low-resolution mesh (100μm step size) is used to reduce computational load, but this leads to the loss of details on the lens surface, with the root mean square (RMS) error of the reference data exceeding 8nm. During the acceleration and deceleration phase of the workpiece stage, switching to a high-resolution mesh (10μm step size) results in a sharp increase in computational latency to 3.5ms, causing gaps in the reference data. Actual measurement data shows that during the start-up and shutdown phases when the workpiece stage acceleration is greater than 3m / s², the maximum deviation between the reference surface generated by the traditional method and the ideal surface of the lens reaches 23.6nm, far exceeding the manufacturing process tolerance (less than 5nm) of next-generation advanced lithography lenses.

[0005] c. Low Parallelization Efficiency: Some solutions attempt to use GPUs to accelerate benchmark generation (such as patent document US20210255543A1), but there are still two major drawbacks in lens manufacturing applications: First, data transmission bottlenecks, with the copying of lens surface data between the CPU and GPU taking more than 400μs (limited by PCIe 4.0 bandwidth); second, insufficient algorithm adaptation, directly porting the serial algorithm from the CPU to the GPU, failing to fully utilize the parallel computing characteristics of the GPU tensor core, with floating-point operation efficiency reaching only 18% of the theoretical peak. A test report released by Tokyo Electron in 2022 shows that the computational latency of such solutions at a grid size of 1024×1024 is still 1.2ms, which cannot meet the sub-millisecond benchmark generation requirements for lithography lens manufacturing.

[0006] d. Lack of physical process coupling: Existing technologies do not consider the influence of photoresist rheological properties on lens surface deformation during lens manufacturing. In processing scenarios with exposure energy greater than 30 mJ / cm², the photoresist viscosity changes with processing parameters. The viscosity change formula is as follows: This leads to micron-level elastic deformation on the lens surface. Traditional solutions ignore this physical effect, resulting in a deviation of more than 10nm between the generated dynamic reference surface and the actual physical state of the lens during processing (IMEC 2023 Technical Report), which seriously affects the lens processing accuracy.

[0007] In summary, existing dynamic benchmark generation technologies for lithography lens manufacturing fail to address the real-time bottleneck of high-order differential operations and the adaptive matching problem under dynamic conditions. The root causes are: the computational architecture does not decouple offline preprocessing from online real-time computation; it does not fully utilize the sparse computation characteristics of GPU tensor cores to improve parallel efficiency; and it lacks a closed-loop feedback mechanism between workpiece stage pose and the lithography physical process. These deficiencies make the dynamic benchmark generation unit a "performance black hole" in the online inspection system for lithography lens manufacturing, severely restricting the improvement of lithography process accuracy and failing to meet the high-precision, high-efficiency manufacturing requirements of advanced lithography lenses. Therefore, a dynamic benchmark generation solution that deeply integrates pre-computation optimization and hardware acceleration is urgently needed to adapt to the technical requirements of lithography lens manufacturing. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the purpose of this invention is to provide a sub-millisecond-level method and system for generating dynamic photolithography references.

[0009] A sub-millisecond-level generation method for photolithographic dynamic references provided by the present invention includes the following steps: Step S1: Based on the ideal surface equation of the mirror design model, construct a hierarchical parametric surface mesh database covering the entire parameter domain; wherein, for high curvature regions, perform fine-grained meshing and pre-store the third derivative data of the mesh nodes in this region; for non-high curvature regions, perform coarse-grained meshing. Step S2: Receive workpiece stage pose data in real time, and calculate the mapping point of the current processing position in the parameter domain based on the parameter domain mapping relationship of the ideal surface equation; Step S3: Using the mapping point as the center, dynamically generate an interpolation block based on the real-time acceleration data of the workpiece stage; Step S4: Call the GPU tensor core to perform three-stage interpolation calculation on the interpolation block: perform bilinear interpolation on the interpolation block region based on coarse-grained grid data to locate the fine-grained grid range around the mapping point; perform Hermite spline interpolation on the located region based on fine-grained grid data and pre-stored third derivative data to obtain the initial surface interpolation result; and correct the initial surface interpolation result by combining the photoresist rheological equation to obtain the final dynamic reference surface data. Step S5: Calculate the interpolation error of the dynamic reference surface data. When the error is greater than the preset value, feed the error information back to step S1 to optimize and adjust the hierarchical parameterized surface mesh database of the full parameter domain.

[0010] Preferably, in step S1, a fine-grained mesh with a step size of 5μm is used for high curvature regions, and a coarse-grained mesh with a step size of 50μm is used for non-high curvature regions.

[0011] Preferably, in step S3, when the workpiece stage acceleration is >0.5m / s², the interpolation block size is set to 64×64; otherwise, it is set to 32×32.

[0012] Preferably, in step S5, the preset value is 2nm.

[0013] Preferably, in step S1, the specific construction process of the hierarchical parametric surface mesh database includes: Calculate the curvature of the ideal surface of the mirror design model. ,in, Here, Z is the Laplacian operator, and Z is the surface height value. A curvature threshold is set. If the curvature K is greater than the threshold, the corresponding region is determined to be a high curvature region. If the curvature K is less than or equal to the threshold, the corresponding region is determined to be a non-high curvature region. Fine-grained meshing is performed on high-curvature regions, and coarse-grained meshing is performed on non-high-curvature regions, forming a hierarchical parametric surface mesh database covering the entire parameter domain.

[0014] Preferably, the threshold for curvature is set to 0.01 μm. -1 .

[0015] Preferably, the high curvature areas are subjected to mesh densification processing, and the densification level is dynamically adjusted based on historical processing data.

[0016] Preferably, if a region is repeatedly identified as a high curvature region in 10 consecutive processing cycles, the mesh step size of that region is increased from 5μm to 2.5μm.

[0017] Preferably, in step S4, Hermite spline interpolation is implemented using a GPU tensor core, and the specific process includes: Convert the Hermite spline interpolation formula to matrix multiplication form: Z interp =H×D, where H is the Hermite basis function matrix and D is the derivative matrix containing the coordinates of the fine-grained grid nodes and the pre-stored third derivative data; The above matrix multiplication operation is performed by calling the GPU tensor core, using FP16 precision and enabling Sparsity 2:4.

[0018] Preferably, in step S4, the initial surface interpolation result is corrected by combining the photoresist rheological equation, specifically including the following steps: Input the viscosity parameter η0, time constant λ, and power law exponent n of the photoresist; Calculate the shear rate The viscosity of the photoresist below The calculation expression is: ; Obtain the exposure energy distribution during lens processing According to the viscosity and exposure energy distribution The surface deformation compensation amount Δz is calculated using the following expression: Where k is an empirical coefficient; The initial surface interpolation result obtained by Hermite spline interpolation is corrected using the deformation compensation amount Δz to obtain the final dynamic reference surface data.

[0019] Preferably, in step S2, the parameter domain mapping relationship of the ideal surface equation is constructed by a coordinate transformation model between the pre-calibrated workpiece table plane coordinate system and the mirror parameter domain coordinate system; The equation of the ideal surface is expressed as: ,in, The coordinates of the ideal surface are given. Real-time acquisition of the workpiece stage's position coordinates (x, y) and attitude angles. Then, based on the coordinate transformation model, the current machining point is inversely transformed from the workpiece table coordinate system to the parameter domain coordinate system to obtain the parameter domain mapping point. ,in, , These are the horizontal and vertical coordinates of the current processing position within the parameter domain, respectively.

[0020] Preferably, in step S3, the real-time acceleration data of the workpiece stage is obtained by performing second-order difference calculation on the workpiece stage position data within a continuous sampling period, or is directly output by the workpiece stage servo controller. When dynamically generating interpolation blocks, the points are mapped using the parameter domain. Centered on the parameter domain, a local interpolation region is extracted, and the side length of this local interpolation region is adaptively determined based on the absolute value of the current workpiece stage acceleration. When the absolute value of the acceleration is greater than 0.5 m / s² 2 At that time, the local interpolation region is expanded to 64×64 grid cells; When the absolute value of the acceleration is less than or equal to 0.5 m / s² 2 In this case, the local interpolation region uses 32×32 grid cells; When the parameter domain mapping point When the parameter domain boundary is near, the portion exceeding the parameter domain is either truncated or mirrored for point filling.

[0021] Preferably, in step S5, the interpolation error is obtained by calculating the norm of the height difference between the dynamic reference surface data and the reference surface data directly calculated from the ideal surface equation on the same set of verification points. The interpolation error is characterized by the root mean square error, and its expression is: , Where N is the total number of verification points in the verification point set. Let be the height of the interpolated surface at the q-th verification point. Let be the ideal surface height at the q-th verification point.

[0022] Preferably, the dynamic adjustment of the encryption level is based on historical processing records established according to parameter domain sub-regions; The historical processing record includes at least the curvature determination results and corresponding interpolation error statistics of each parameter domain sub-region in the most recent processing. When a parameter domain sub-region is determined to be a high curvature region for 10 consecutive times, the grid step size of the parameter domain sub-region is adjusted from 5μm to 2.5μm; when the above continuous determination condition is not met, the fine-grained grid of 5μm is maintained.

[0023] The present invention also provides a sub-millisecond generation system for photolithography dynamic references, comprising the following modules: Module M1: Based on the ideal surface equation of the mirror design model, construct a hierarchical parametric surface mesh database covering the entire parameter domain; for high curvature regions, perform fine-grained meshing and pre-store the third derivative data of the mesh nodes in this region; for non-high curvature regions, perform coarse-grained meshing. Module M2: Receives workpiece stage pose data in real time, and calculates the mapping point of the current processing position in the parameter domain based on the parameter domain mapping relationship of the ideal surface equation; Module M3: Based on the mapping point, dynamically generate interpolation blocks by combining real-time acceleration data of the workpiece stage; Module M4: Calls the GPU tensor core to perform three-stage interpolation calculations on the interpolation block: Bilinear interpolation is performed on the interpolation block region based on coarse-grained grid data to locate the fine-grained grid range around the mapping point; Hermite spline interpolation is performed on the located region based on the fine-grained grid data and pre-stored third-order derivative data to obtain the initial surface interpolation result; The initial surface interpolation result is corrected by combining the photoresist rheological equation to obtain the final dynamic reference surface data. Module M5: Calculates the interpolation error of the dynamic reference surface data. When the error is greater than the preset value, it feeds the error information back to module M1 to optimize and adjust the hierarchical parameterized surface mesh database of the full parameter domain.

[0024] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs a hierarchical parameterized surface mesh database offline, pre-completing the fine-grained division of high-curvature regions of the lens and pre-storing the third derivative, effectively avoiding the computational bottleneck of online high-order differential operations and ensuring real-time response of reference data generation during lens processing. Combined with dynamic adaptation of interpolation block size to the stage acceleration, it accurately matches the dynamic motion of the stage during lens processing, avoiding the loss of lens surface details or reference data discontinuity caused by fixed meshes. By leveraging GPU tensor cores to perform three-stage interpolation and coupling photoresist rheological equation correction, it adapts to the lithography physics process of lens manufacturing and improves parallel computing efficiency, ensuring sub-millisecond real-time generation of dynamic reference surfaces, meeting the cycle requirements of lithography lens exposure processing. Through error feedback optimization of the closed-loop design of the mesh database, it continuously improves reference accuracy, significantly reduces lens processing surface deviation, and ensures that the lens surface accuracy meets the imaging requirements of advanced lithography machines, ultimately satisfying the high-precision and high-efficiency manufacturing needs of lithography lenses.

[0025] 2. This invention uses reverse projection calculation, which reduces the delay from 2.3ms in the traditional scheme to 487μs (measured value), meeting the 1ms exposure cycle requirement.

[0026] 3. In actual testing on the MT1217 lithography machine, the RMS error between the dynamic reference surface and the ideal surface decreased from 8.2nm to 1.3nm.

[0027] 4. With a workpiece stage speed of 1.5 m / s, the present invention reduces the reference misalignment rate from 42% to 4.7% and increases the lens processing yield by 13.8%. Attached Figure Description

[0028] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the dynamic benchmark generation system of the present invention, used to show the physical connection relationship and data flow of the parametric surface mesh pre-computation module, pre-computation database, dynamic block interpolation engine and GPU tensor core acceleration unit; Figure 2 This is a flowchart of the dynamic benchmark generation operation of the present invention, which is used to illustrate the two-stage operation logic of offline pre-calculation and online dynamic interpolation; Figure 3 This invention provides a sub-millisecond-level reference generation timing interaction diagram, used to present the dynamic interaction timing of key components during the lithography exposure cycle. Detailed Implementation

[0029] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0030] Example 1

[0031] This embodiment provides a sub-millisecond-level method for generating dynamic lithography references, mainly applied to the scenario of manufacturing lenses using lithography machines. It can generate dynamic reference surface data that precisely matches the lens processing position in real time, ensuring the processing accuracy and efficiency of the lens surface. The specific steps include the following: Step S1: Construct a hierarchical parametric surface mesh database covering the entire parameter domain.

[0032] Based on the ideal surface equations of the lens design model to be manufactured, a hierarchical parametric surface mesh database covering the entire parameter domain is constructed to provide basic data support for subsequent online generation of dynamic reference surfaces. Specifically, for the high curvature region of the lens surface, a fine-grained mesh with a step size of 5μm is used, and the third derivative data of the mesh nodes in this region are pre-stored; for the non-high curvature region of the lens surface, a coarse-grained mesh with a step size of 50μm is used, balancing computational efficiency while ensuring reference accuracy.

[0033] The specific construction process includes the following sub-steps: Step S11: Calculate the curvature of the ideal surface of the mirror design model. ,in, Z is the Laplacian operator, used to characterize the spatial distribution of surface curvature; Z is the surface height value, corresponding to the ideal height parameter of the lens to be manufactured at a certain plane coordinate (x, y).

[0034] Step S12: Set the curvature threshold to 0.01 μm. -1 The calculated curvature K is compared with this threshold: if curvature K > 0.01 μm -1 If the curvature K is less than or equal to 0.01μm, then the corresponding region is determined to be a high curvature region (such regions have a significant impact on the accuracy of the lens surface shape and require high-precision mesh support); if the curvature K ≤ 0.01μm -1 If so, the corresponding region is determined to be a non-high curvature region.

[0035] Step S13: Perform fine-grained mesh generation (initial step size 5μm) on the determined high curvature regions and coarse-grained mesh generation (step size 50μm) on the non-high curvature regions to initially form a hierarchical parametric surface mesh database covering the entire parameter domain.

[0036] Step S14: Perform mesh refinement processing on the identified high curvature areas, and dynamically adjust the refinement level based on the historical lens processing data: If a certain area is repeatedly identified as a high curvature area in 10 consecutive lens processing processes, it indicates that the area is a high-precision sensitive area for lens processing. The mesh step size of the area is further increased from 5μm to 2.5μm to enhance the accuracy of the baseline data of the area.

[0037] Specifically, the system maintains a historical processing count table for each parameter domain sub-region, recording the high curvature determination results and local interpolation errors of each sub-region in the most recent 10 processing operations. When the cumulative number of high curvature occurrences in the count table for a certain sub-region reaches 10, its encryption level is switched from Level 1 to Level 2, and the grid step size is adjusted from 5μm to 2.5μm. High curvature regions that do not meet the above conditions continue to maintain a step size of 5μm, thereby achieving a balance between reference accuracy and storage and computing resource consumption.

[0038] Step S2: Calculate the mapping point of the current processing position within the parameter domain.

[0039] During the lens processing using a lithography machine, the pose data (including position, attitude and other parameters) of the lithography machine workpiece stage is received in real time. Based on the parameter domain mapping relationship of the above-mentioned ideal surface equation, the actual processing position of the workpiece stage is converted into a mapping point in the parameter domain, realizing the accurate association between the physical processing position and the parameter domain data, and providing a central reference for the subsequent generation of interpolation blocks.

[0040] Specifically, the parameter domain mapping relationship can be established by a two-dimensional rigid transformation from the workpiece stage coordinate system to the parameter domain coordinate system: after reading the workpiece stage position coordinates x, y and attitude angle θ in real time, the current machining point is inversely transformed to the parameter domain according to the pre-calibrated coordinate transformation model to obtain the parameter domain mapping point (u0, v0); where x and y are the real-time translation coordinates of the workpiece stage, θ is the workpiece stage attitude angle, and u0 and v0 are the horizontal and vertical coordinates of the current machining position in the parameter domain, respectively.

[0041] Step S3: Dynamically generate interpolation blocks that adapt to the working conditions of the workpiece stage.

[0042] Centered on the mapping point obtained in step S2, interpolation blocks are dynamically generated by combining real-time acceleration data of the workpiece stage. The size of the interpolation blocks is adapted to the motion state of the workpiece stage, ensuring a balance between real-time performance and accuracy in the interpolation calculation. Specifically, when the workpiece stage acceleration is >0.5m / s² (i.e., acceleration and deceleration conditions in lens processing, where the workpiece stage motion changes rapidly and a larger range of data is needed to ensure the continuity of the reference), the size of the interpolation blocks is set to 64×64; when the workpiece stage acceleration is ≤0.5m / s² (i.e., uniform speed processing conditions), the size of the interpolation blocks is set to 32×32 to reduce the computational load.

[0043] Specifically, the real-time acceleration data of the workpiece stage can be obtained by performing second-order difference calculation on the workpiece stage position data at continuous sampling times, or by directly reading the acceleration estimate output by the workpiece stage servo controller. After determining the mapping point (u0, v0), a local mesh window is extracted in the parameter domain centered on the mapping point as an interpolation block, and a mesh cell range of 64×64 or 32×32 is selected according to the conditions |a|>0.5m / s² or |a|≤0.5m / s² respectively. When the interpolation block exceeds the boundary of the parameter domain, boundary truncation or mirror point filling is performed on the excess part.

[0044] Step S4: GPU tensor cores accelerate three-stage interpolation calculations.

[0045] The GPU tensor core is invoked to perform three-stage interpolation calculations on the above interpolation block. Through hardware acceleration and a step-by-step refinement strategy, high-precision initial surface interpolation results are quickly obtained. These results are then corrected in conjunction with the photolithography physical process, as detailed below: Step S41: First-stage localization: Based on the coarse-grained grid data constructed in step S1, bilinear interpolation is performed on the interpolation block region to quickly narrow the calculation range, accurately locate the fine-grained grid range around the mapping point, and improve the efficiency of subsequent interpolation.

[0046] Step S42: Second-stage high-precision interpolation: Based on fine-grained grid data and pre-stored third-order derivative data, Hermite spline interpolation is performed on the located region to obtain the initial surface interpolation result. Hermite spline interpolation is accelerated using GPU tensor cores. The specific process involves converting the Hermite spline interpolation formula into matrix multiplication form Z. interp =H×D (where H is the Hermite basis function matrix and D is the derivative matrix containing the coordinates of fine-grained grid nodes and pre-stored third-order derivative data); the above matrix multiplication operation is performed by calling the GPU tensor core, using FP16 precision and enabling Sparsity 2:4, which greatly improves the interpolation calculation speed and meets the sub-millisecond requirement.

[0047] Step S43: Third-stage physical correction: Correct the initial surface interpolation results by combining the photoresist rheology equation. During the lens photolithography process, the viscosity of the photoresist will undergo elastic deformation as the exposure parameters change. This equation can quantify the impact of this deformation on the lens surface shape, so that the final dynamic reference surface data can be accurately matched with the actual physical state of the lens processing.

[0048] The initial surface interpolation results are corrected by combining the photoresist rheological equation, specifically including the following steps: inputting the photoresist viscosity parameter η0, time constant λ, and power law exponent n; calculating the shear rate. The viscosity of the photoresist below The calculation expression is: ; Obtain the exposure energy distribution during lens processing According to the viscosity and exposure energy distribution The surface deformation compensation amount Δz is calculated using the following expression: Where k is an empirical coefficient; The initial surface interpolation result obtained by Hermite spline interpolation is corrected using the deformation compensation amount Δz to obtain the final dynamic reference surface data.

[0049] Step S5: Error feedback optimizes the grid database.

[0050] The interpolation error of the dynamic reference surface data obtained in step S4 is calculated and compared with a preset threshold (2nm). When the error > 2nm, it indicates that the accuracy of the current mesh database cannot meet the requirements of lens processing. This error information is fed back to step S1 to optimize and adjust the hierarchical parameterized surface mesh database of the entire parameter domain (such as further densifying the mesh in the corresponding region and correcting the curvature threshold). When the error ≤ 2nm, the dynamic reference surface data is directly output to the lithography machine control system as the accuracy benchmark for lens processing. Through the above closed-loop optimization mechanism, the accuracy of the dynamic benchmark is continuously improved, ensuring the stability of lens processing quality.

[0051] Specifically, the interpolation error is preferably calculated using the root mean square error, that is, N verification points are selected within the interpolation block, and the height of the dynamic reference surface obtained after Hermite spline interpolation and physical correction is compared point by point with the reference height directly calculated from the ideal surface equation, and the error is calculated according to the following formula; , Where e is the interpolation error, N is the total number of check points, and q is the check point index. Let be the height of the dynamic reference surface at the q-th verification point. Let be the ideal surface height at the q-th verification point.

[0052] This embodiment relates to precision optical processing technology in the field of semiconductor lithography manufacturing, specifically falling under the category of real-time generation of dynamic reference surfaces in lithography processes. In particular, addressing the processing delay issue caused by high-order differential operations of ideal surface equations in the dynamic reference generation unit during lens substrate exposure in ultraviolet (UV) / extreme ultraviolet (EUV) lithography machines, this invention provides a sub-millisecond dynamic reference generation scheme based on parametric mesh pre-computation and GPU tensor core acceleration. This technology can be widely applied to high-precision optical component processing, semiconductor lithography equipment, and nanoscale surface topography control systems, and has core value in improving the real-time response capability and processing accuracy of lithography processes. This invention is particularly suitable for high-speed lithography scenarios with workpiece stage movement speeds >1 m / s, solving the problem of processing error accumulation caused by reference data lag in existing technologies.

[0053] The present invention also provides a sub-millisecond level photolithography dynamic reference generation system. The sub-millisecond level photolithography dynamic reference generation system can be implemented by executing the process steps of the sub-millisecond level photolithography dynamic reference generation method. That is, those skilled in the art can understand the sub-millisecond level photolithography dynamic reference generation method as a preferred embodiment of the sub-millisecond level photolithography dynamic reference generation system.

[0054] Example 2

[0055] This embodiment provides a sub-millisecond-level dynamic reference generation system for photolithography, mainly applied in scenarios where lenses are manufactured using photolithography machines. It can generate dynamic reference surface data that precisely matches the lens processing position in real time, ensuring the accuracy and efficiency of lens surface processing. Specifically, it includes the following modules: Module M1: Constructs a hierarchical parametric surface mesh database covering the entire parameter domain.

[0056] Based on the ideal surface equations of the lens design model to be manufactured, a hierarchical parametric surface mesh database covering the entire parameter domain is constructed to provide basic data support for subsequent online generation of dynamic reference surfaces. Specifically, for the high curvature region of the lens surface, a fine-grained mesh with a step size of 5μm is used, and the third derivative data of the mesh nodes in this region are pre-stored; for the non-high curvature region of the lens surface, a coarse-grained mesh with a step size of 50μm is used, balancing computational efficiency while ensuring reference accuracy.

[0057] The specific construction process includes the following sub-modules: Module M11: Calculates the curvature of the ideal surface of the mirror design model. ,in, is the Laplacian operator, used to characterize the spatial distribution of surface curvature; z is the surface height value, corresponding to the ideal height parameter of the lens to be manufactured at a certain plane coordinate (x, y).

[0058] Module M12: Sets the curvature threshold to 0.01μm. -1 The calculated curvature K is compared with this threshold: if curvature K > 0.01 μm -1 If the curvature K is less than or equal to 0.01μm, then the corresponding region is determined to be a high curvature region (such regions have a significant impact on the accuracy of the lens surface shape and require high-precision mesh support); if the curvature K ≤ 0.01μm -1 If so, the corresponding region is determined to be a non-high curvature region.

[0059] Module M13: Performs fine-grained mesh generation (initial step size 5μm) on the determined high curvature regions and coarse-grained mesh generation (step size 50μm) on the non-high curvature regions, initially forming a hierarchical parametric surface mesh database covering the entire parameter domain.

[0060] Module M14: Performs mesh refinement processing on the identified high curvature areas, and the refinement level is dynamically adjusted based on the lens's historical processing data: If a certain area is repeatedly identified as a high curvature area in 10 consecutive lens processing processes, it indicates that the area is a high-precision sensitive area for lens processing. The mesh step size of the area is further increased from 5μm to 2.5μm to enhance the accuracy of the baseline data in the area.

[0061] The dynamic adjustment of the encryption level is based on historical processing records established by parameter domain sub-regions. The historical processing records include at least the curvature determination results and corresponding interpolation error statistics of each sub-region in the most recent processing. When a sub-region is determined to be a high curvature region for 10 consecutive times, the grid step size of the sub-region is adjusted from 5μm to 2.5μm. When the above continuous determination condition is not met, the fine-grained grid of 5μm is maintained to avoid storage and computational overhead caused by invalid encryption.

[0062] Module M2: Calculates the mapping point of the current processing position within the parameter domain.

[0063] During the lens processing using a lithography machine, the pose data (including position, attitude and other parameters) of the lithography machine workpiece stage is received in real time. Based on the parameter domain mapping relationship of the above-mentioned ideal surface equation, the actual processing position of the workpiece stage is converted into a mapping point in the parameter domain, realizing the accurate association between the physical processing position and the parameter domain data, and providing a central reference for the subsequent generation of interpolation blocks.

[0064] In module M2, the parameter domain mapping relationship of the ideal surface equation is constructed through a coordinate transformation model between the pre-calibrated workpiece table plane coordinate system and the mirror parameter domain coordinate system; the ideal surface is written as ,in For surface parameter coordinates; real-time acquisition of workpiece stage position coordinates. and attitude angle Then, based on the coordinate transformation model, the current machining point is inversely transformed from the workpiece table coordinate system to the parameter domain coordinate system to obtain the parameter domain mapping point. ,in These represent the horizontal and vertical coordinates of the current processing position within the parameter domain, respectively.

[0065] Module M3: Dynamically generates interpolation blocks that adapt to the working conditions of the workpiece stage.

[0066] Centered on the mapping point obtained from module M2, interpolation blocks are dynamically generated by combining real-time acceleration data of the workpiece stage. The size of the interpolation blocks is adapted to the motion state of the workpiece stage, ensuring a balance between real-time performance and accuracy in the interpolation calculation. Specifically, when the workpiece stage acceleration is >0.5m / s² (i.e., acceleration and deceleration conditions in lens processing, where the workpiece stage motion changes rapidly and a larger range of data is needed to ensure the continuity of the reference), the size of the interpolation block is set to 64×64; when the workpiece stage acceleration is ≤0.5m / s² (i.e., uniform speed processing conditions), the size of the interpolation block is set to 32×32 to reduce the computational load.

[0067] In module M3, the real-time acceleration data of the workpiece stage is obtained by performing second-order difference calculation on the workpiece stage position data within a continuous sampling period, or directly output by the workpiece stage servo controller; when dynamically generating interpolation blocks, the mapping points are used. Centered on the parameter domain, a local interpolation region is extracted, and the side length of the local interpolation region is adaptively determined based on the absolute value of the current workpiece stage acceleration: when the absolute value of acceleration is greater than 0.5 m / s², it is expanded to a 64×64 grid element; when the absolute value of acceleration is less than or equal to 0.5 m / s², a 32×32 grid element is used; when the mapping point is close to the parameter domain boundary, the part exceeding the parameter domain is truncated or mirrored for point filling.

[0068] Module M4: GPU Tensor Core accelerates three-stage interpolation calculations.

[0069] The GPU tensor core is invoked to perform three-stage interpolation calculations on the above interpolation block. Through hardware acceleration and a step-by-step refinement strategy, high-precision initial surface interpolation results are quickly obtained. These results are then corrected in conjunction with the photolithography physical process, as detailed below: Module M41: First-stage localization: Based on the coarse-grained grid data constructed by module M1, bilinear interpolation calculations are performed on the interpolation block region to quickly narrow the calculation range, accurately locate the fine-grained grid range around the mapping point, and improve the efficiency of subsequent interpolation.

[0070] Module M42: Second-stage high-precision interpolation: Based on fine-grained grid data and pre-stored third-order derivative data, Hermite spline interpolation is performed on the located region to obtain the initial surface interpolation result. The Hermite spline interpolation is accelerated using a GPU tensor core, specifically by converting the Hermite spline interpolation formula into a matrix multiplication form Z. interp =H×D (where H is the Hermite basis function matrix and D is the derivative matrix containing the coordinates of fine-grained grid nodes and pre-stored third-order derivative data); the above matrix multiplication operation is performed by calling the GPU tensor core, using FP16 precision and enabling Sparsity 2:4, which greatly improves the interpolation calculation speed and meets the sub-millisecond requirement.

[0071] Module M43: Third-stage physical correction: Correcting the initial surface interpolation results by combining the photoresist rheology equation. During the lens photolithography process, the photoresist viscosity will undergo elastic deformation as the exposure parameters change. This equation can quantify the impact of this deformation on the lens surface shape, so that the final dynamic reference surface data can be accurately matched with the actual physical state of the lens during processing.

[0072] The initial surface interpolation results are corrected by incorporating the photoresist rheological equation. Specifically, this involves the following steps: inputting the photoresist viscosity parameter η0, time constant λ, and power-law exponent n; calculating the shear rate. The viscosity of the photoresist below The calculation expression is: ; Obtain the exposure energy distribution during lens processing According to the viscosity and exposure energy distribution The surface deformation compensation amount Δz is calculated using the following expression: Where k is an empirical coefficient; The initial surface interpolation result obtained by Hermite spline interpolation is corrected using the deformation compensation amount Δz to obtain the final dynamic reference surface data.

[0073] Module M5: Error feedback optimizes the grid database.

[0074] The interpolation error of the dynamic reference surface data obtained by module M4 is compared with a preset threshold (2nm). When the error > 2nm, it indicates that the accuracy of the current mesh database cannot meet the requirements of lens processing. This error information is fed back to module M1 to optimize and adjust the hierarchical parameterized surface mesh database across the entire parameter domain (e.g., further densifying the mesh in the corresponding region, correcting the curvature threshold, etc.). When the error ≤ 2nm, the dynamic reference surface data is directly output to the lithography machine control system as the accuracy benchmark for lens processing. Through the above closed-loop optimization mechanism, the accuracy of the dynamic benchmark is continuously improved, ensuring the stability of lens processing quality.

[0075] The interpolation error is obtained by calculating the norm of the height difference between the dynamic reference surface data and the reference surface data directly calculated from the ideal surface equation on the same set of verification points; specifically, it can be characterized by the root mean square error, expressed as: , Where N is the total number of verification points. Let be the height of the interpolated surface at the q-th verification point. Let be the ideal surface height at the q-th verification point.

[0076] Example 3

[0077] Those skilled in the art can understand this embodiment as a more specific description of Embodiment 1.

[0078] This embodiment proposes a two-stage decoupled architecture of "pre-computation-dynamic interpolation," which compresses the inverse projection calculation to sub-millisecond levels (<500μs) through a parametric surface mesh pre-computation database and deep optimization of the GPU tensor core. Specific improvements are as follows: In this embodiment, x and y represent the position coordinates of the workpiece stage in the planar coordinate system, θ represents the workpiece stage attitude angle, u and v represent the mirror parameter domain coordinates, u0 and v0 represent the parameter domain mapping point coordinates corresponding to the current machining position, and z represents the ideal surface height. interp The z-axis represents the height of the interpolated surface. ideal κ represents the ideal surface reference height, a represents the curvature, H represents the Hermite basis function matrix, D represents the derivative matrix, η0, λ, and n represent the zero-shear viscosity, time constant, and power-law exponent of the photoresist, respectively, γ̇ represents the shear rate, Eexposure represents the exposure energy, Δz represents the deformation compensation amount, R represents the radius of curvature, A2 represents the aspherical coefficient, i and j represent the Hermite interpolation term index, q represents the error statistics sampling point index, and N represents the total number of calibration points.

[0079] I. Pre-calculation mechanism for hierarchical parametric surface mesh: The surface computation is decomposed into two stages: offline pre-computation and online interpolation. 1) Offline Stage: Based on the ISO 10110 optical standard, a hierarchical grid database covering the entire parameter domain (radius of curvature 0.5-500mm, aspheric coefficient 0-0.1) is constructed. Coarse-grained grid (50μm step size): stored in SSD cache for fast positioning; Fine-grained mesh (5μm step size): Only high curvature regions are pre-stored ( The third derivative data of ) is compressed at a ratio of 4:1.

[0080] 2) Adaptive optimization: Dynamically adjust the mesh density based on historical processing data. If a region repeatedly exhibits high curvature (such as the edge of a lens), the mesh density level is automatically increased, reducing the amount of online computation by more than 30%.

[0081] II. Dynamic block interpolation driven by the GPU tensor core: Designed tensor computation pipeline optimized specifically for Hermite interpolation: 1) Pose-driven module: Receives real-time pose of the workpiece stage. At the current processing position Dynamically generate a 32×32 interpolation block centered on the area (covering a ±1mm region); Where x and y are the real-time position coordinates of the workpiece stage in the planar coordinate system, θ is the current attitude angle of the workpiece stage, u0 and v0 are the coordinates of the parameter domain mapping point corresponding to the current processing position, and ±1mm represents the interpolation coverage range reserved on both sides of the corresponding physical processing area with the current processing position as the center.

[0082] 2) Three-stage accelerated interpolation: a. Coarse-grained rapid positioning: Bilinear interpolation initial screening (delay <50μs); b. Fine-grained Hermite interpolation: Calling Tensor Core to perform matrix operations: , Among them, z interp Let represent the surface height obtained by interpolation, and let hi(u) and hj(v) represent the Hermite basis functions in the u and v directions, respectively. u and v are the parameter domain coordinates, and i and j are the indices of the Hermite interpolation terms. The mixed partial derivative of the surface height z with respect to the i-th partial derivative with respect to the parameter u and the j-th partial derivative with respect to the parameter v represents the geometric information of the surface at the grid points, such as the slope and curvature. This indicates that the value is taken at a grid point. This represents the partial derivative data of the corresponding order at the grid node. is a bicubic Hermite basis function, a polynomial function of parameters u and v, used to weight the contribution of partial derivatives of different orders to the interpolation result.

[0083] By leveraging FP16 precision and Sparsity 2:4, the interpolation speed is increased by 17 times; c. Physical constraint correction: Integrating the photoresist rheological equation Correcting surface deformation (delay <80μs).

[0084] III. Closed-loop feedback adaptive operating condition mechanism: Constructing a dynamic closed loop of "pre-computation-interpolation-feedback": The FPGA interpolation engine monitors the workpiece stage acceleration in real time. Dynamically expand the interpolation block size (up to 64×64); The interpolation error is fed back to the pre-calculation module. If the error is greater than 2nm, mesh optimization is triggered to ensure long-term accuracy and stability.

[0085] Specifically, this embodiment provides a sub-millisecond-level generation method for photolithographic dynamic references, including the following steps: In the offline phase, based on the ideal surface equation of the mirror design model, a hierarchical parametric surface mesh database covering the entire parameter domain is constructed. The high curvature region adopts a 5μm step size fine-grained mesh and pre-stores the third derivative data, while the non-high curvature region adopts a 50μm step size coarse-grained mesh. During the online phase, workpiece stage pose data is received in real time, and the mapping point of the current machining position in the parameter domain is calculated based on the pose. An interpolation block is dynamically generated centered on the mapping point. When the workpiece stage acceleration is >0.5m / s², the size of the interpolation block is set to 64×64; otherwise, it is set to 32×32. Three-stage interpolation is performed using GPU tensor cores: first, bilinear interpolation with a coarse-grained grid is used for rapid localization; then, Hermite spline interpolation with a fine-grained grid is used; and finally, the surface deformation is corrected by combining the photoresist rheology equation. The interpolation results are output as dynamic reference surface data, and the interpolation error is calculated. When the error is greater than 2nm, it is fed back to the offline stage to optimize the mesh.

[0086] Furthermore, the construction of the hierarchical parametric surface mesh database includes: Calculate the curvature of a surface ,when It is then identified as a high curvature region; The mesh is refined in areas with high curvature, and the refinement level is dynamically adjusted based on historical processing data. If an area is identified as having high curvature in 10 consecutive processing runs, the mesh step size is increased to 2.5 μm.

[0087] Furthermore, the Hermite spline interpolation is accelerated using tensor kernels: the interpolation formula is converted into matrix multiplication. , where H is the Hermite basis function matrix and D is the derivative matrix; By leveraging the FP16 precision and Sparsity 2:4 characteristics of the GPU to perform matrix operations, the interpolation speed is increased by more than 17 times.

[0088] Furthermore, the correction of the photoresist rheological equation includes: Input the photoresist viscosity parameter η0, time constant λ, and power law exponent n; Calculate the shear rate viscosity below ; Based on exposure energy distribution Calculate deformation compensation , where k is an empirical coefficient.

[0089] More specifically, this embodiment is described in detail through the following hardware configuration, algorithm flow, and operation examples. The implementation environment is an MT1217 lithography machine, with a workpiece stage movement speed of 1.5m / s and an exposure cycle of 1ms.

[0090] I. Hardware deployment details.

[0091] 1. Parametric Surface Mesh Pre-calculation Module: Deployed on a dedicated server for the lithography control system (Dell PowerEdge R750), equipped with dual Intel Xeon Gold 6348 processors (28 cores / 56 threads). The module features a surface analysis engine developed based on the ISO 10110 standard, supporting 12 types of optical surface equations, including Zernike polynomials and Q-type aspherical surfaces. Key design features: 1) Hierarchical mesh generator: Employs an adaptive octree algorithm to automatically refine high-curvature regions. When When the 50μm step size coarse mesh is refined into a 5μm step size fine mesh, memory usage is reduced by 63% (actual measurement: pre-calculated data for a 200mm lens is compressed from 15.7GB to 5.8GB). 2) Third derivative pre-storage unit: Calculates and compresses the data for storage. Eight third-order derivative terms were compressed using the Zstandard algorithm (compression ratio 4:1), resulting in an SSD read latency of <15μs.

[0092] 2. Dynamic Block Interpolation Engine: Implemented on a Xilinx Versal ACAP XCVC1902 FPGA (occupying 18% of logic resources), physically mounted in the lithography control cabinet (0.18m from the workpiece stage displacement sensor). Core components include: 1) Pose Analysis Unit: Receives data from the Heidenhain encoder (update rate 10kHz, accuracy ±0.05μm), and transforms it using a coordinate transformation matrix. Calculate parameter domain mapping points ; 2) Dynamic segmentation controller: based on the workpiece stage acceleration Adjust the interpolation block size: ,

[0093] Ensure boundary continuity (actual measurement: interpolation error <0.5nm during abrupt acceleration changes); 3) Unified Memory Manager: Shares address space with GPU via PCIe 5.0 x16 (128GT / s bandwidth), RDMA direct connection latency <500ns.

[0094] 3. GPU Tensor Core Acceleration Unit: It uses an NVIDIA L20S GPU (40GB VRAM) and integrates the Tensor Core v3 architecture. Key optimizations: 1) Hermite interpolation tensor quantization: Converts the interpolation formula into matrix multiplication. , Where H is the Hermite basis function matrix (4×4) and D is the derivative matrix (4×4). The theoretical peak value reaches 312 TFLOPS when calculated using Tensor Core FP16 mixed precision. 2) Sparse Computing Activation: Enable the Sparsity 2:4 feature through the NVIDIA spTensor library to increase the utilization of effective computing units to 89% (compared to only 52% in traditional solutions).

[0095] II. Detailed Explanation of Operation Procedures

[0096] This solution is implemented in stages within the photolithography exposure cycle (1ms), and the specific steps are as follows: Phase 1: Offline pre-calculation (one-time configuration) 1) Model analysis (t=-∞): Receive the mirror design model (STEP format) and extract the ideal surface equation parameters (radius of curvature R=150mm, aspherical coefficient A2=0.002).

[0097] 2) Generation of hierarchical meshes (t=-∞): Parameter domain Divide into a 4000×4000 basic grid (50μm step size); Detecting high curvature regions: Calculating curvature ,when Time (such as the edge of the lens) At each location, a 5μm step size fine mesh (2000×2000 points) is generated. Pre-stored third derivative: Calculation Eight derivatives were compressed and stored in the SSD cache (actual measurement: 5.8GB of data for a 200mm lens).

[0098] 3) Historical data optimization (ongoing): If a certain area (such as High curvature was observed in 10 consecutive processing steps. The mesh refinement level of the area was automatically increased to a step size of 2.5μm, reducing the online computation by 32%.

[0099] 2. Stage 2: Online dynamic interpolation (within the exposure period, <500μs) 1) Pose-driven block segmentation (t=0-100μs): Receiving workpiece stage posture Calculate the parameter domain mapping point ; Detecting acceleration Dynamically set the interpolation block size to 64×64 (covering a ±2mm area); The active interpolation block data is requested via RDMA, and SSD returns the grid point coordinates and third derivatives (t=35μs).

[0100] 2) Tensor core acceleration interpolation (t=100-450μs): a. Coarse-grained positioning: Bilinear interpolation quickly determines the reference point (t=150μs, error <5nm); b. Fine-grained Hermite interpolation: Convert the derivative matrix D into an FP16 tensor and load it into a Tensor Core; Perform sparse matrix multiplication: (t=150-420μs); Actual measurement: 1024-point interpolation takes 270μs (traditional CPU solution takes 1.8ms); c. Physical constraint correction: Input photoresist parameters ( ); Calculate deformation compensation (k is an empirical coefficient); Corrected surface data output (t=430μs).

[0101] 3) Error feedback and output (t=450-500μs): Calculate the interpolation error;

[0102] like (Measured value 1.8nm < 2nm, no feedback required); The dynamic reference surface data packet is encapsulated and output to the control bus via the AXI4-Stream bus (t=487μs).

[0103] III. Verification of Implementation Examples

[0104] Resin lenses (radius of curvature 150mm, asphericity 0.002) were processed on an MT1217 lithography machine. The results are compared to traditional methods: 1) Computational performance comparison:

[0105] Key breakthrough: When the workpiece stage acceleration is >3m / s², this embodiment still maintains a delay of <500μs (the traditional solution has a delay of >3.5ms).

[0106] 2) Verification of process impact: 100 lenses were processed continuously, and their surface morphology was measured.

[0107] Data shows that this embodiment significantly suppresses reference inaccuracies under dynamic operating conditions, and the yield rate is improved by an average of 13.8%.

[0108] 3) Resource consumption analysis:

[0109] Hierarchical grids and sparse computing reduce resource consumption by more than 60%, which aligns with the trend of green manufacturing.

[0110] This invention provides a sub-millisecond-level generation method and system for dynamic lithography references. First, a hierarchical surface mesh database covering the entire parameter domain is constructed based on the ideal surface equation of a mirror. High-curvature regions are finely divided and their third derivatives are pre-stored, while non-high-curvature regions are coarsely divided. Then, workpiece stage pose data is received in real-time to obtain the machining position parameter domain mapping point. An interpolation block is dynamically generated using this point as the center and combined with the workpiece stage acceleration. Dynamic reference surface data is obtained through a three-stage interpolation process using the GPU tensor core, which includes bilinear interpolation positioning, Hermite spline interpolation, and correction of the photoresist rheological equation. Finally, the interpolation error is calculated, and if it exceeds a preset value, the mesh database is optimized to achieve sub-millisecond-level generation of the dynamic reference.

[0111] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0112] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for generating dynamic photolithographic references at the sub-millisecond level, characterized in that, Includes the following steps: Step S1: Based on the ideal surface equation of the mirror design model, construct a hierarchical parametric surface mesh database covering the entire parameter domain; wherein, for high curvature regions, perform fine-grained meshing and pre-store the third derivative data of the mesh nodes in this region; for non-high curvature regions, perform coarse-grained meshing. Step S2: Receive workpiece stage pose data in real time, and calculate the mapping point of the current processing position in the parameter domain based on the parameter domain mapping relationship of the ideal surface equation; Step S3: Using the mapping point as the center, dynamically generate an interpolation block based on the real-time acceleration data of the workpiece stage; Step S4: Call the GPU tensor core to perform three-stage interpolation calculation on the interpolation block: perform bilinear interpolation on the interpolation block region based on coarse-grained grid data to locate the fine-grained grid range around the mapping point; perform Hermite spline interpolation on the located region based on fine-grained grid data and pre-stored third derivative data to obtain the initial surface interpolation result; and correct the initial surface interpolation result by combining the photoresist rheological equation to obtain the final dynamic reference surface data. Step S5: Calculate the interpolation error of the dynamic reference surface data. When the error is greater than the preset value, feed the error information back to step S1 to optimize and adjust the hierarchical parameterized surface mesh database of the full parameter domain.

2. The method for generating a dynamic photolithographic reference at the sub-millisecond level according to claim 1, characterized in that, In step S1, a fine-grained mesh with a step size of 5μm is used for high curvature regions, and a coarse-grained mesh with a step size of 50μm is used for non-high curvature regions.

3. The method for generating a dynamic photolithographic reference at the sub-millisecond level according to claim 1, characterized in that, In step S3, when the workpiece stage acceleration is greater than 0.5 m / s², the interpolation block size is set to 64×64; otherwise, it is set to 32×32.

4. The method for generating a dynamic photolithographic reference at the sub-millisecond level according to claim 1, characterized in that, In step S5, the preset value is 2nm.

5. The method for generating a photolithographic dynamic reference at the sub-millisecond level according to claim 2, characterized in that, In step S1, the specific construction process of the hierarchical parametric surface mesh database includes: Calculate the curvature of the ideal surface of the mirror design model. ,in, Here, Z is the Laplacian operator, and Z is the surface height value. A curvature threshold is set. If the curvature K is greater than the threshold, the corresponding region is determined to be a high curvature region. If the curvature K is less than or equal to the threshold, the corresponding region is determined to be a non-high curvature region. Fine-grained meshing is performed on high-curvature regions, and coarse-grained meshing is performed on non-high-curvature regions, forming a hierarchical parametric surface mesh database covering the entire parameter domain.

6. The method for generating a photolithographic dynamic reference at the sub-millisecond level according to claim 5, characterized in that, The curvature threshold is set to 0.01 μm. -1 .

7. The method for generating a dynamic photolithographic reference at the sub-millisecond level according to claim 5, characterized in that, The high curvature regions are subjected to mesh densification, and the densification level is dynamically adjusted based on historical processing data.

8. The method for generating a dynamic photolithographic reference at the sub-millisecond level according to claim 7, characterized in that, If a region is repeatedly identified as a high curvature region in 10 consecutive processing cycles, the mesh step size for that region will be increased from 5 μm to 2.5 μm.

9. The method for generating a dynamic photolithographic reference at the sub-millisecond level according to claim 1, characterized in that, In step S4, Hermite spline interpolation is implemented using a GPU tensor core, and the specific process includes: Convert the Hermite spline interpolation formula to matrix multiplication form: Z interp =H×D, where H is the Hermite basis function matrix and D is the derivative matrix containing the coordinates of the fine-grained grid nodes and the pre-stored third derivative data; The above matrix multiplication operation is performed by calling the GPU tensor core, using FP16 precision and enabling Sparsity 2:

4.

10. The method for generating a photolithographic dynamic reference at the sub-millisecond level according to claim 1, characterized in that, In step S4, the initial surface interpolation result is corrected by combining the photoresist rheological equation, specifically including the following steps: Input the viscosity parameter η0, time constant λ, and power law exponent n of the photoresist; Calculate the shear rate The viscosity of the photoresist below The calculation expression is: ； Obtain the exposure energy distribution during lens processing According to the viscosity and exposure energy distribution The surface deformation compensation amount Δz is calculated using the following expression: Where k is an empirical coefficient; The initial surface interpolation result obtained by Hermite spline interpolation is corrected using the deformation compensation amount Δz to obtain the final dynamic reference surface data.

11. The method for generating photolithographic dynamic references at the sub-millisecond level according to claim 1, characterized in that, In step S2, the parameter domain mapping relationship of the ideal surface equation is constructed by the coordinate transformation model between the pre-calibrated workpiece table plane coordinate system and the mirror parameter domain coordinate system. The equation of the ideal surface is expressed as: ,in, , The coordinates of the ideal surface are given. Real-time acquisition of workpiece stage position coordinates , and attitude angle Then, based on the coordinate transformation model, the current machining point is inversely transformed from the workpiece table coordinate system to the parameter domain coordinate system to obtain the parameter domain mapping point. ,in, , These are the horizontal and vertical coordinates of the current processing position within the parameter domain, respectively.

12. The method for generating a photolithographic dynamic reference at the sub-millisecond level according to claim 11, characterized in that, In step S3, the real-time acceleration data of the workpiece stage is obtained by performing second-order difference calculation on the workpiece stage position data within a continuous sampling period, or is directly output by the workpiece stage servo controller. When dynamically generating interpolation blocks, the points are mapped using the parameter domain. Centered on the parameter domain, a local interpolation region is extracted, and the side length of this local interpolation region is adaptively determined based on the absolute value of the current workpiece stage acceleration. When the absolute value of the acceleration is greater than 0.5 m / s² 2 At that time, the local interpolation region is expanded to 64×64 grid cells; When the absolute value of the acceleration is less than or equal to 0.5 m / s² 2 In this case, the local interpolation region uses 32×32 grid cells; When the parameter domain mapping point When the parameter domain boundary is near, the portion exceeding the parameter domain is either truncated or mirrored for point filling.

13. The method for generating a dynamic photolithographic reference at the sub-millisecond level according to claim 1, characterized in that, In step S5, the interpolation error is obtained by calculating the norm of the height difference between the dynamic reference surface data and the reference surface data directly calculated from the ideal surface equation on the same set of verification points. The interpolation error is characterized by the root mean square error, and its expression is: ， Where N is the total number of verification points in the verification point set. Let be the height of the interpolated surface at the q-th verification point. Let be the ideal surface height at the q-th verification point.

14. The method for generating a photolithographic dynamic reference at the sub-millisecond level according to claim 7, characterized in that, The dynamic adjustment of the encryption level is based on historical processing records established according to parameter domain sub-regions; The historical processing record includes at least the curvature determination results and corresponding interpolation error statistics of each parameter domain sub-region in the most recent processing. When a parameter domain sub-region is determined to be a high curvature region for 10 consecutive times, the grid step size of the parameter domain sub-region is adjusted from 5μm to 2.5μm; when the above continuous determination condition is not met, the fine-grained grid of 5μm is maintained.

15. A sub-millisecond-level photolithography dynamic reference generation system, characterized in that, Includes the following modules: Module M1: Based on the ideal surface equation of the mirror design model, construct a hierarchical parametric surface mesh database covering the entire parameter domain; for high curvature regions, perform fine-grained meshing and pre-store the third derivative data of the mesh nodes in this region; for non-high curvature regions, perform coarse-grained meshing. Module M2: Receives workpiece stage pose data in real time, and calculates the mapping point of the current processing position in the parameter domain based on the parameter domain mapping relationship of the ideal surface equation; Module M3: Based on the mapping point, dynamically generate interpolation blocks by combining real-time acceleration data of the workpiece stage; Module M4: Calls the GPU tensor core to perform three-stage interpolation calculations on the interpolation block: Bilinear interpolation is performed on the interpolation block region based on coarse-grained grid data to locate the fine-grained grid range around the mapping point; Hermite spline interpolation is performed on the located region based on the fine-grained grid data and pre-stored third-order derivative data to obtain the initial surface interpolation result; The initial surface interpolation result is corrected by combining the photoresist rheological equation to obtain the final dynamic reference surface data. Module M5: Calculates the interpolation error of the dynamic reference surface data. When the error is greater than a preset value, it feeds the error information back to module M1 to optimize and adjust the hierarchical parameterized surface mesh database of the full parameter domain.

Images