Rapid inversion photoetching method and system based on bandwidth awareness

Through the optimization path of sparse sampling cost gradient and momentum direction, the problems of high computational complexity and slow convergence rate in inversion lithography technology are solved, and the efficiency and speed of inversion lithography technology are improved.

CN119002166APending Publication Date: 2024-11-22BEIJING INST OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411077891.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-07
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

The existing inversion lithography technology has a high time complexity in calculating cost gradients during optimization, resulting in too long running time and difficulty in applying it in large-scale integrated circuit layout corrections, and the convergence rate and calculation overhead of the optimization algorithm are poor.

Method used

The rapid inversion lithography method based on bandwidth perception is adopted to reduce the computational complexity by sparse sampling cost gradients, and the optimal iteration path is determined using the cost gradients of momentum direction and prospective position during the iteration process, optimizing the efficiency of the ILT algorithm.

Benefits of technology

It significantly reduces the computational redundancy of ILT optimization, improves the convergence rate and algorithm efficiency, reduces the computational complexity, and improves the ILT mask solution speed of large-scale integrated circuit layouts.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119002166A_ABST
    Figure CN119002166A_ABST
Patent Text Reader

Abstract

The invention provides a rapid inversion photoetching method based on bandwidth perception, which is characterized in that sparse sampling is carried out on the gradient of a cost function by analyzing the bandwidth limited characteristic of a photoetching imaging system, so that the calculation redundancy of ILT optimization is reduced, the calculation complexity of each iteration of an ILT algorithm is reduced, and the algorithm efficiency of the ILT algorithm is optimized; an ILT mask efficient iteration path method is adopted, each iteration continues to advance on the basis of the previous iteration direction, and then the cost gradient is calculated at the look-ahead position, so that the optimization direction of the iteration is obtained. On the premise that the calculation complexity is not increased, the method is beneficial for increasing the convergence rate of ILT optimization, and the efficiency of the ILT algorithm is further improved; according to the method, a path which can be rapidly and stably iterated to the optimal point of an ILT cost function is designed for an ILT algorithm; and the calculation complexity is reduced by eliminating redundant calculation during running of the ILT algorithm.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical fields of inversion lithography, optical proximity effect correction and computational lithography, and in particular relates to a fast inversion lithography method and system based on bandwidth perception. Background Art

[0002] The photolithography process is an important part of current integrated circuit manufacturing. Photolithography uses light waves as a carrier to transfer the electronic circuit pattern on the mask to the image-side wafer, presenting it as a light intensity distribution reduced by a predetermined ratio, and exposing the photoresist. After the photoresist is developed, a three-dimensional photoresist outline of the layout is formed, and then the design pattern is transferred to the silicon wafer by etching. As the critical dimension (CD) of the mask in advanced nodes gradually becomes smaller, the diffraction effect of the light wave is very obvious, resulting in the optical proximity effect of the mask imaging on the wafer becoming more and more serious, which causes serious distortion of the lithography pattern, resulting in the inability to achieve the expected integrated circuit design function.

[0003] Inverse Lithography Technology (ILT) aims to improve the imaging quality produced during lithography. It regards the mask as a pixel map and pre-distorts the mask pattern by optimizing the transmittance (for deep ultraviolet lithography masks) or reflectivity (for extreme ultraviolet lithography masks) of all pixels of the mask, thereby compensating for the lithography errors produced during imaging. At the same time, a larger lithography process window can be obtained than the traditional optical proximity correction (OPC) method.

[0004] Although ILT technology can provide a mask solution with high imaging fidelity, it needs to continuously calculate the gradient of the cost function relative to a large number of mask pixels (hereinafter referred to as "cost gradient") during the optimization process. The time complexity of this calculation step grows superlinearly at the order of O(NlogN) (where N is the number of mask pixels calculated). Therefore, in large-scale integrated circuit layout correction, ILT solutions are difficult to put into practical application due to their long running time.

[0005] In addition, the ILT optimization algorithm using gradient descent has a linear convergence rate, which requires a large number of iterations when optimizing the ILT mask, further limiting the efficiency of the ILT solution. On the other hand, although the Newton-type algorithm has a quadratic convergence rate, which helps to reduce the number of ILT iterations, this type of algorithm needs to calculate the second-order derivative of the cost function (i.e., the Hessian matrix) in each iteration step, which introduces a lot of additional computational overhead and is not conducive to shortening the ILT running time.

[0006] In summary, it is necessary to invent a fast, accurate, and low-complexity cost gradient calculation method and system, and at the same time provide an optimization method and system with fast convergence rate and low computational overhead, so as to synergistically improve the operating efficiency of the ILT solution and solve the application difficulties of ILT technology in large-scale integrated circuit layout correction. Summary of the invention

[0007] In view of this, an object of the present invention is to provide a fast inversion lithography method and system based on bandwidth perception.

[0008] A fast inversion lithography method comprises the following steps:

[0009] Step S1: determining a cost function;

[0010] Step S2: determining the cost gradient of the cost function for the ILT mask, where the cost gradient represents the direction in which the cost function changes fastest in the mask variable space;

[0011] Step S3: determining the optimal iteration path of the ILT mask in the mask space according to the cost gradient;

[0012] Step S4: Iterate by repeatedly executing step S3 to approximate the optimal solution of the cost function, thereby obtaining the final ILT mask pattern.

[0013] Preferably, the method for determining the cost function is:

[0014] One of the pattern error of lithography imaging, edge placement error, normalized image logarithmic slope, mask error enhancement factor, and process window is set as the cost function or a part of the cost function.

[0015] Preferably, the method for determining the cost gradient of the cost function for the ILT mask is:

[0016] The cost function is expressed as a function of the lithography pattern, which is a function of the aerial image, so the cost function is calculated as a function of the aerial image; when determining the gradient function of the cost function relative to the mask, first determine the derivative of the cost function relative to the aerial image, and then determine the derivative of the aerial image relative to the ILT mask; then multiply the two derivatives to determine the cost gradient of the cost function with respect to the mask.

[0017] Preferably, the photolithography pattern in the cost function is calculated according to a forward imaging model of the photolithography system; the forward imaging model of the photolithography system can be expressed as a cascade model of a spatial image model and a photoresist model.

[0018] Preferably, when determining the cost gradient of the cost function for the ILT mask, the effective bandwidth of the continuous cost gradient function is determined based on the lithography light source and projection objective parameters used, and then the Nyquist sampling interval of the continuous cost gradient function is determined, that is, the maximum sampling interval allowed for the cost gradient; during the ILT optimization process, the cost gradient is calculated on a sparse Nyquist grid.

[0019] Preferably, when calculating the cost gradient, the lithography spatial image formed on the wafer and the wafer complex amplitude distribution corresponding to different coherent light sources are predetermined, and according to the bandwidth-limited characteristics of the lithography imaging system, the cost gradient is calculated on a sparse Nyquist grid.

[0020] Preferably, the sparsely sampled cost gradient and lithography spatial image are upsampled by an interpolation algorithm to obtain a cost gradient or lithography spatial image with a high sampling frequency when required by the ILT algorithm.

[0021] Preferably, the method for determining the optimal iteration path of the ILT mask in the mask space according to the cost gradient in step S3 includes:

[0022] The optimization direction of each iteration is synthesized from the following two vectors: the first part is to move forward a certain step along the iterative optimization direction of the previous step, which is the momentum direction; the second part is to move forward a set step along the cost gradient determined by the forward position in the momentum direction.

[0023] Preferably, the cost gradient determined by the forward position in the momentum direction is calculated on a sparse Nyquist grid.

[0024] Preferably, during the iteration process, the termination condition of the iteration is that the number of iterations reaches a set upper limit or the cost function value is less than a given error limit.

[0025] The present invention has the following beneficial effects:

[0026] The present invention provides a fast inversion lithography method based on bandwidth perception, which performs sparse sampling on the gradient of the cost function by analyzing the bandwidth-limited characteristics of the lithography imaging system, thereby reducing the computational redundancy of ILT optimization, reducing the computational complexity of each iteration of the ILT algorithm, and optimizing the algorithm efficiency of the ILT;

[0027] Among them, the ILT mask efficient iteration path method is adopted. Each iteration continues to move forward based on the direction of the previous iteration, and then calculates the cost gradient at the forward position to obtain the optimization direction of this iteration. Without increasing the computational complexity, this method helps to accelerate the convergence rate of ILT optimization and further improve the efficiency of the ILT algorithm.

[0028] In summary, the present invention designs a path for the ILT algorithm that can quickly and smoothly iterate to the global optimum (or satisfactory local optimum) of the ILT cost function. At the same time, by eliminating redundant calculations during the operation of the ILT algorithm, the computational complexity is reduced. Combining the above methods, the operating efficiency of the ILT algorithm can be improved, and the speed of solving the ILT mask of the large-scale integrated circuit layout can be accelerated. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A flowchart of a bandwidth-aware fast ILT method provided by the present invention;

[0030] Figure 2 is a layout used as an ILT optimization target in an embodiment of the present invention;

[0031] FIG3(a) is a densely sampled gradient map during the first iteration in an embodiment of the present invention; FIG3(b) is a sparsely sampled gradient map during the first iteration in an embodiment of the present invention;

[0032] Figure 4 This is a convergence curve diagram of ILT optimization using multiple optimization algorithms in an example of the present invention. DETAILED DESCRIPTION

[0033] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0034] Figure 1 A flowchart of a fast ILT method and system based on bandwidth perception is shown, which specifically includes the following steps:

[0035] S1: Determine a cost function, which evaluates the lithography performance of the optimized ILT mask under a given lithography system. The lithography performance represents the quality of lithography imaging formed on a wafer using the ILT mask, or the difference between the printed lithography pattern and the target layout or the desired pattern.

[0036] In step 1, the specific method for determining the cost function is:

[0037] In some embodiments, the lithographic pattern Z printed by the ILT mask and the pattern to be printed can be The pattern error (PE) between them is determined as the cost function:

[0038]

[0039] Wherein, represents the cost function of using the ILT mask M; ||·||2 represents the L2 norm. It should be noted that the embodiments of the present invention are not limited to specifying the cost function as the function described in formula (1), but may have multiple variations. For example, the edge placement error (EPE), the normalized image log slope (NILS), the mask error enhancement factor (MEEF), the process window (PW), etc. of the lithography imaging may be set as the cost function of the ILT or a part of the cost function.

[0040] The calculation of the photolithography pattern Z in the cost function (1) is further described below. The photolithography pattern is a pattern printed by the ILT mask through the photolithography system, which can be described by a forward imaging model. In some embodiments, the forward imaging model of the photolithography system can be expressed as a cascade model of an aerial image model and a photoresist model. The aerial image model can use, but is not limited to, the Abbe imaging model and the Hopkins imaging model. Specifically, the aerial image model can be calculated by the following formula:

[0041]

[0042] Where I(x,y) represents the light intensity sampling value of the spatial image at position (x,y); Represents the kernel function of coherent imaging. Since the lithography imaging system is a diffraction-limited system, It has the characteristic of limited bandwidth; n Represents the kernel function The weight is related to the intensity of the exposure source; N k represents the number of kernel functions; m(x,y) represents the complex transmittance (for transmissive masks) or reflectance (for reflective masks) of the mask at position (x,y).

[0043] For the photoresist model, in some embodiments, it can be expressed as a sigmoid function, so as to derive the aerial image during ILT optimization. Specifically, the relationship between the photolithography pattern on the wafer and the aerial image is:

[0044]

[0045] Where Z(x,y) represents the sampling value of the lithography pattern at position (x,y); a r is the steepness factor of the sigmoid function; t r Represents the threshold of photoresist.

[0046] S2: Determine the cost gradient of the cost function relative to the mask, where the cost gradient represents the direction in which the cost function changes fastest in the mask variable space. Determining the cost gradient includes: first determining the effective bandwidth of the continuous cost gradient function, and then determining the Nyquist sampling interval of the continuous cost gradient function, that is, the maximum sampling interval allowed for the cost gradient. When calculating the cost gradient in each subsequent iteration, only the sampled value of the cost gradient is calculated on the above sparse Nyquist grid.

[0047] In step S2, determining the gradient of the cost function relative to the mask, in some embodiments, includes the following specific steps:

[0048] As shown in the above formula (1), the cost can be represented by PE, which can be expressed as a function of the lithography pattern, and the lithography pattern can be expressed as a function of the aerial image, so the cost function can be calculated as a function of the aerial image. Therefore, the gradient function of the cost function relative to the mask can be calculated according to the chain rule as follows: the first part is to calculate the derivative of the cost function relative to the aerial image, and the second part is to calculate the derivative of the aerial image relative to the mask. Then the cost gradient can be obtained by multiplying the two derivatives.

[0049] In some embodiments, the derivative of the cost function with respect to the mask may be calculated as follows:

[0050]

[0051] Among them, g(I(x,y)) represents a function with the spatial image formed on the wafer as a variable. After calculating dcost / dm(x,y) for all sampling points on the mask, their values ​​are used as coefficients to construct the gradient vector of the mask variable space, and its direction represents the direction in which the cost function in the mask space changes fastest at the current ILT mask position. It should be noted that the embodiments of the present invention are not limited to specifying the derivative of the cost function as the form described in formula (4). Depending on the different types of cost functions selected, the derivative of the cost function with respect to the mask may have other variations. For example, with the kernel function of coherent imaging The convolution function g may also be a function with the photolithography pattern Z obtained on the wafer as a variable, that is:

[0052]

[0053] According to formula (4) or (5), the sampling value of the cost gradient at each sampling point can be calculated. Next, we will further consider the sampling method of the cost gradient in the ILT process, that is, the sampling interval. As mentioned above, the kernel function of the coherent imaging of the lithography system is is bandwidth-limited, so the cost function relative to the gradient function of the mask, that is, the cost gradient, is also necessarily bandwidth-limited, as can be seen from formula (5). Therefore, due to the diffraction-limited characteristics of the lithography imaging system, the high-frequency components of the cost gradient function will be filtered out during the ILT process. According to the Nyquist-Shannon Sampling Theorem, the continuous cost gradient function does not need to be oversampled during the sampling process to avoid unnecessary computational redundancy. Calculating the partial derivatives of the cost function with respect to the mask sampling points at as few sampling positions as possible can reduce the computational complexity of a single iteration of the ILT algorithm, thereby accelerating the optimization efficiency of the ILT.

[0054] Specifically, the bandwidth limitation of the cost gradient depends on the kernel function of the coherent imaging The bandwidth of the cost gradient function is limited to a range that further depends on the illumination light source used and the projection objective lens parameters of the lithography system. In the case of a commonly used partially coherent illumination light source, the bandwidth of the cost gradient function is proportional to the partial coherence factor of the illumination light source and the numerical aperture NA of the projection objective lens, and is inversely proportional to the wavelength λ of the illumination light source. In some embodiments, the bandwidth of the cost gradient function can be described by the following formula:

[0055]

[0056] Among them, B x and B y Respectively represent the bandwidth of the gradient function in the x-direction and y-direction of the mask plane; (σ x ,σ y ) represents the normalized coordinates of the partially coherent light source. Further, 2B x ,2B y Represents the Nyquist Rate of the continuous gradient function.

[0057] The following further describes the method and principle of sampling the cost gradient according to the bandwidth of the cost gradient function. Assuming a continuous gradient function in the spatial domain The Fourier transform corresponding to the frequency domain is G(f x ,f y ). In the x and y directions of the mask plane, T x and T y The sampling interval is Sampling is performed, and the discrete gradient function obtained is The dual function in the frequency domain will be G(f x ,f y ), and the period is the inverse of the spatial domain sampling interval.

[0058] In some embodiments of the present invention, the gradient function Downsampling is implemented to save the computational cost of the ILT algorithm. The sampling intervals of the downsampled discrete gradient function in the x-direction and y-direction of the mask plane are αT and x and βT y , that is, with the initial discrete gradient function In comparison, the downsampling rates in the x and y directions are α and β respectively, where α and β are positive real numbers greater than 1. After downsampling, the periodic spectrum of the discrete gradient function can be expressed as follows:

[0059]

[0060] In order to make the downsampled spatial domain gradient function retain all the information of the original continuous gradient function, in the dual frequency domain, it should be ensured that the period of the periodic spectrum is greater than the Nyquist rate of the gradient function, that is:

[0061]

[0062] According to formula (8), with the largest possible sampling interval αT x ,βT y Sampling the continuous gradient function to obtain the cost gradient of small dimensions can meet the calculation requirements of the ILT algorithm for the cost gradient. Figure 2 Figure 3(a) shows the cost gradient graph of dense sampling at the first iteration of the ILT optimization process, compared with the cost gradient graph of sparse sampling shown in Figure 3(b). The cost gradient calculated on a sparse Nyquist grid is obviously cheaper than that calculated on a grid with a high sampling frequency.

[0063] When calculating the sampling value of the cost gradient at each sampling point (as shown in formula (4)), it is necessary to first determine the light spatial image formed on the wafer. Similarly, due to the bandwidth-limited characteristics of the lithography imaging system, in some embodiments, the wafer complex amplitude distribution corresponding to the coherent light source and the continuous spatial image formed by the incoherent light source can also be sampled at a sparse Nyquist interval, thereby eliminating computational redundancy.

[0064] In some embodiments, when the ILT algorithm requires a higher sampling frequency cost gradient or lithography aerial image, the sparsely sampled cost gradient and aerial image can be upsampled to a high-dimensional space by interpolation. During the upsampling process, based on the sparse sampling points, although many kinds of continuous functions can be fitted, only one of them has a bandwidth limited by the diffraction limit range of the lithography system, that is, the original continuous gradient function. Therefore, the cost gradient and aerial image upsampled using the interpolation algorithm have no error with the direct sampling results of the continuous cost gradient function and the continuous aerial image function.

[0065] S3: Determine the optimal iteration path of the ILT mask in the mask space. The optimal iteration direction of each iteration consists of two parts: the first part is to advance a certain step length along the iteration optimization direction of the previous step (called the momentum direction) to reach the forward position; the second part is to advance a certain step length on the cost gradient determined by the forward position in the momentum direction, and the cost gradient is calculated on the sparse Nyquist grid. The synthesis of these two parts constitutes the optimal iteration path of the ILT mask in the mask space of the present invention.

[0066] In step S3, determining the optimal iteration direction of the ILT mask in the mask space specifically includes the following steps:

[0067] The following briefly describes the shortcomings of the traditional gradient-based ILT algorithm in determining the direction of mask optimization. In ILT optimization, the gradient of the cost function with respect to the mask is obtained. In theory, the direction in which the cost function changes fastest at the current position in the mask space is obtained. However, the gradient method approximates the morphology of its neighborhood through the first-order Taylor expansion of the objective function, lacks a description of the high-order characteristics of the objective function, and therefore it is difficult to perceive the global changes of the objective function. In addition, the ILT problem is a non-convex optimization problem, and the cost function usually has multiple local minimum points. Therefore, in actual ILT optimization, the gradient descent method usually finds it difficult to guide the ILT mask to quickly move to the global optimal point of the cost function from a global perspective.

[0068] In comparison, some embodiments disclosed in the present invention do not calculate the cost gradient at the current iteration position in the ILT optimization, but continue to move in the direction of the previous iteration (i.e., the momentum direction), and then calculate the cost gradient at the forward-looking position. Therefore, in some embodiments disclosed in the present invention, the optimal moving direction of each iteration step of the ILT optimization consists of two parts: the first part is a vector composed of a momentum direction with a certain modulus, and the second part is a cost gradient vector at the forward-looking position with a certain modulus, and the combined vector of these two parts constitutes the optimal iteration direction of the ILT optimization. Through forward-looking, some embodiments of the present invention will obtain a convergence rate that exceeds that of traditional ILT algorithms, and help eliminate the oscillation effect of the ILT optimization convergence process.

[0069] Further, in the first part of the iteration direction, a certain step length is continued to be advanced in the momentum direction, and the step length may be a constant step length or an optimal step length obtained by the line search algorithm. In addition, in the second part of the iteration direction, a certain step length is advanced, which may be a constant step length or a constant step length that is proportional to the step length of the first part, or an optimal step length obtained by the line search algorithm. Specifically, in some embodiments disclosed in the present invention, the optimal iteration direction of the ILT mask in the mask space may be calculated as follows:

[0070]

[0071] Among them, p k-1 represents the momentum direction, p k It represents the optimal moving direction of the current iteration step in ILT optimization. The optimal moving direction of the current iteration step consists of two parts: the first part is the vector of the momentum direction with a modulus of γ, γ is called the momentum coefficient; the second part is the forward constant step length s based on the cost gradient determined by the forward position in the momentum direction k . The resultant vector of these two parts constitutes the optimal iteration direction of the ILT mask in the ILT optimization. The update method of ILT optimization described in the formula is also called the Nesterov Accelerated Gradient (NAG) method. It should be noted that the embodiments of the present invention are not limited to specifying the iterative update strategy of ILT optimization as the method described in formula (9), but there can be multiple variants. For example, the step size of the first part and the step size of the second part can also be solved by a line search algorithm. These variants should all fall within the scope of protection of the present invention.

[0072] S4: Further, the scheme is continuously iterated to approach the global optimal solution (or satisfactory local optimal solution) of the ILT cost function, and a satisfactory ILT mask pattern can be obtained. The termination condition of the iteration is that the number of iterations reaches the set upper limit or the cost function value is less than the given error limit. Figure 4 1 shows the convergence curve of ILT optimization in some embodiments of the present invention, the horizontal axis of the curve represents the number of iterations, and the vertical axis of the curve represents the cost function value. Figure 4 At the same time, the steepest descent (SD) method, the quasi-Newton (QN) method and the method of the present invention are shown. Figure 2 From the convergence curve of the optimized layout shown in FIG. 1 , it can be seen that the method of the present invention can solve a better ILT mask than the SD method in a shorter time; at the same time, the method of the present invention can obtain an optimization effect equivalent to that of the QN method, but the solution time of the method of the present invention is much shorter than that of the QN method.

[0073] Based on the above-mentioned inversion lithography method, the present invention also provides a bandwidth-aware fast inversion lithography system, comprising a first module, a second module, a third module and a fourth module; wherein the first module executes the above-mentioned step S1; the second module executes the above-mentioned step S2; the third module executes the above-mentioned step S3; and the fourth module executes the above-mentioned step S4.

[0074] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A fast inversion lithography method, It is characterized in that The following steps are involved: Step S1: determining a cost function; Step S2: determining the cost gradient of the cost function for the ILT mask, where the cost gradient represents the direction in which the cost function changes fastest in the mask variable space; Step S3: determining the optimal iteration path of the ILT mask in the mask space according to the cost gradient; Step S4: Iterate by repeatedly executing step S3 to approximate the optimal solution of the cost function, thereby obtaining the final ILT mask pattern.

2. A fast inversion lithography method as claimed in claim 1, It is characterized in that The method for determining the cost function is: One of the pattern error of lithography imaging, edge placement error, normalized image logarithmic slope, mask error enhancement factor, and process window is set as the cost function or a part of the cost function.

3. A fast inversion lithography method as claimed in claim 1, It is characterized in that The method to determine the cost gradient of the cost function for the ILT mask is: The cost function is expressed as a function of the lithography pattern, which is a function of the aerial image, so the cost function is calculated as a function of the aerial image; when determining the gradient function of the cost function relative to the mask, first determine the derivative of the cost function relative to the aerial image, and then determine the derivative of the aerial image relative to the ILT mask; then multiply the two derivatives to determine the cost gradient of the cost function with respect to the mask.

4. A fast inversion lithography method as claimed in claim 3, It is characterized in that The photolithography pattern in the cost function is calculated according to the forward imaging model of the photolithography system; the forward imaging model of the photolithography system can be expressed as a cascade model of a spatial image model and a photoresist model.

5. A fast inversion lithography method as claimed in claim 3, It is characterized in that When determining the cost gradient of the cost function for the ILT mask, the effective bandwidth of the continuous cost gradient function is determined based on the lithography light source and projection objective parameters used, and then the Nyquist sampling interval of the continuous cost gradient function is determined, that is, the maximum sampling interval allowed for the cost gradient; during the ILT optimization process, the cost gradient is calculated on a sparse Nyquist grid.

6. A fast inversion lithography method as claimed in claim 3, It is characterized in that When calculating the cost gradient, the lithography spatial image formed on the wafer and the wafer complex amplitude distribution corresponding to different coherent light sources are predetermined. According to the bandwidth-limited characteristics of the lithography imaging system, the cost gradient is calculated on a sparse Nyquist grid.

7. A fast inversion lithography method as claimed in claim 5 or 6, It is characterized in that The sparsely sampled cost gradient and lithography spatial image are upsampled by an interpolation algorithm to obtain a cost gradient or lithography spatial image with a high sampling frequency when required by the ILT algorithm.

8. A fast inversion lithography method as claimed in claim 1, It is characterized in that The method for determining the optimal iterative path of the ILT mask in the mask space according to the cost gradient in step S3 includes: The optimization direction of each iteration is synthesized from the following two vectors: the first part is to move forward a certain step along the iterative optimization direction of the previous step, which is the momentum direction; the second part is to move forward a set step along the cost gradient determined by the forward position in the momentum direction.

9. A fast inversion lithography method as claimed in claim 8, It is characterized in that The cost gradient determined by the look-ahead position in the momentum direction is computed on a sparse Nyquist grid.

10. A fast inversion lithography method as claimed in claims 1 to 9, It is characterized in that During the iteration process, the termination condition of the iteration is that the number of iterations reaches the set upper limit or the cost function value is less than the given error limit.