CN108665060A - A kind of integrated neural network for calculating photoetching - Google Patents

Abstract

The invention discloses an integrated neural network for computational lithography, including a conjugate neural network and a forward neural network, and the output of the conjugate neural network is connected to the input of the forward neural network; the The conjugate neural network is used to extract the feature vector of computational lithography, and input the extracted feature vector into the forward neural network, wherein, the method for extracting the feature vector of computational lithography by the conjugate neural network is: Y _j ＝∑ _i W _ij X _i , Z _j ＝Y _j · Y _j ^* ; wherein, Z _j is the extracted feature vector, W _ij is the parameter of the conjugate neural network, and Xi is the _i -th points in the neighborhood, Y _j ^* is the conjugate of Y _j . An integrated neural network for computational lithography provided in the present invention combines the conjugate convolutional neural network structure and the forward neural network for extracting feature vectors to form an integrated neural network, which can be used for any type of Computational Lithography Learning.

Description

An integrated neural network for computational lithography

technical field

The invention relates to the field of integrated circuits, in particular to an integrated neural network for computational lithography.

Background technique

In order to continuously pursue performance enhancement, power consumption reduction, and chip area shrinkage of semiconductor chips, the minimum feature pitch and minimum feature size of the semiconductor chip need to be reduced accordingly. To support this endless trend, the semiconductor industry needs to develop lithography tools, such as scanners, with increasingly shorter exposure wavelengths and higher numerical apertures (NA) to achieve high optical resolution. The semiconductor industry has successfully followed this path prior to the 14nm technology node, however, the industry has found it very difficult to continue to advance hardware (scanner) technology along this path. Development is slow to see.

As a remedy, the development and application of computational lithography has allowed the semiconductor industry to move forward. This computational lithography technology includes source mask collaborative optimization, advanced OPC model, model-based auxiliary pattern generation, reverse lithography technology, etc. Most computational lithography techniques are computationally expensive when applied to full chips.

To alleviate this problem, deep convolutional neural network (DCNN) architectures have been proposed to learn reverse lithography, especially auxiliary pattern generation. The DCNN architecture is a powerful and general learning machine, however, it requires a large amount of data for training. This is because it requires DCNN to automatically extract feature detection kernels from input data training. Such a DCNN architecture should only be used in cases where no prior knowledge is available for the neural network architecture itself or the design of input feature vectors. Therefore, the deep convolutional neural network architecture has complex procedures and time-consuming procedures in the process of computational lithography learning, which cannot guarantee the current production efficiency.

For computational lithography, everything starts with optical imaging, and the structure of optical imaging equations is quite mature. Therefore, modern production urgently needs an integrated neural network for effective computational lithography learning.

Contents of the invention

The technical problem to be solved by the present invention is to provide an integrated neural network for computational lithography, combining the conjugate convolutional neural network structure and the forward neural network for extracting feature vectors to form an integrated neural network, which can be used for any type of computational lithography learning.

In order to achieve the above object, the present invention adopts the following technical solution: an integrated neural network for computational lithography, the integrated neural network includes a conjugate neural network and a forward neural network, and the output terminal of the conjugate neural network Connect the input end of the forward neural network; the conjugate neural network is used to extract the feature vector of computational lithography, and input the extracted feature vector into the forward neural network, wherein the conjugate neural network The method of network extraction and calculation of the feature vector of lithography is: Y _j = ∑ _i W _ij X _i , Wherein, Z _j is the extracted feature vector, W _ij is the parameter of the conjugate neural network, Xi is the adjacent environment of the _i -th point on the mask, is the conjugate of Y _j .

Further, the feature vector used for calculating lithography is a vector related to light intensity.

Further, the X _i uses a vector in real space or a vector based on spatial frequency as the surrounding environment of the i-th point on the mask.

Further, when X _i adopts vectors in real space, when extracting the feature vectors of computational lithography, it is necessary to input the number of vectors in real space Among them, a is the range of optical interaction in the lithography machine, which is divided into subunits of equal size; b is the size of the subunits within the range of optical interaction in the lithography machine.

Further, the size of the subunit within the optical interaction range Among them, NA is the numerical aperture of the lithography machine, σ _max is a parameter related to the maximum angle of the exposure light on the mask, and λ is the exposure wavelength of the lithography machine.

Further, the input value of the vector in the real space is Value _cell =t _bg ·Area _cell +(t _f -t _bg ) ·Area _{geo_in_cell} , wherein, t _bg is the complex transmission value of the background of the mask, t _f is the complex transmission value of the pattern on the mask, Area _cell is the area of the subunit, and Area _{geo_in_cell} is the area of the mask pattern in the subunit.

Further, when X _i adopts a vector based on spatial frequency, when extracting the feature vector of computational lithography, the number M of vectors in the input spatial frequency needs to be calculated by the following _formula :

Among them, M ₂ is the sum of all n _x and n _y numbers that conform to the above formula, n _x and n _y are the diffraction orders of the imaging system, NA is the numerical aperture of the photolithography machine, and σ _max is the Parameters related to the maximum angle on the reticle, λ is the exposure wavelength of the lithography machine, P=2*(the radius of the optical interaction range in optical imaging+the width of the safety belt), and the safety belt is set at the optical interaction The periphery of the range, used to ensure the calculation accuracy of the reticle pattern in the optical interaction range.

Further, the input value of the vector based on the spatial frequency is the value of the Fourier transform of the mask pattern on the grid point of λ/P within the optical interaction range plus the safety belt.

Further, the feedforward neural network structure has 3 or 4 hidden layers.

The beneficial effects of the present invention are as follows: first, the conjugate neural network is used to extract and calculate the feature vector of lithography, and then the extracted feature vector is input into the forward neural network, so that the output terminal of the conjugate neural network is consistent with the forward neural network. The input ends are combined to form an integrated neural network; the integrated neural network formed by the present invention can be used for any type of computational lithography learning, and the process of computational lithography learning is simple and fast, greatly improving the efficiency of computational lithography and reducing its complexity.

Description of drawings

Accompanying drawing 1 is the structural diagram of the conjugated neural network of the present invention.

Accompanying drawing 2 is the construction of the vector describing the surrounding environment of the point on the mask in the present invention.

Figure 3 is a schematic diagram of sampling of the input vector in the spatial frequency space when the input vector is a vector based on the spatial frequency in the present invention.

Accompanying drawing 4 is the structure diagram of the integrated neural network formed by the present invention.

Detailed ways

In order to make the purpose, technical solution and advantages of the present invention clearer, the specific implementation manners of the present invention will be further described in detail below in conjunction with the accompanying drawings.

An integrated neural network for computing lithography provided by the present invention, the integrated neural network includes a conjugated neural network and a forward neural network, and the output of the conjugated neural network is connected to the input of the forward neural network; the conjugated neural network The network is used to extract the feature vector of computational lithography, and input the extracted feature vector into the forward neural network, wherein the method of extracting the feature vector of computational lithography by the conjugate neural network is: Y _j = ∑ _i W _ij X _i , Wherein, Z _j is the extracted feature vector, W _ij is the parameter of the conjugate neural network, Xi is the adjacent environment of the _i -th point on the mask, is the conjugate of Y _j .

It is well known that any computational lithography starts from the light intensity distribution function. To this end, it can be assumed that the feature vectors for machine learning-based computational lithography should be light intensity-dependent vectors. To extract such intensity-dependent vectors from the input geometry, we conjugate neural networks for extraction. The structure of the conjugate neural network provided by the present invention is shown in FIG. 1 . Its input vector Xi is the surrounding environment of the _i -th point on the mask, and its output value is used to calculate the feature vector Z _m of photolithography.

In the present invention, X _i uses a vector in real space or a vector based on spatial frequency as the surrounding environment of the i-th point on the mask. The following two situations are introduced respectively:

If one decides to use real spatial quantities to describe the neighborhood of a point, one needs to first estimate the optical interaction range, and then divide the area within the optical interaction range into subunits, as shown in Figure 2. The extent to which light interacts depends on the imaging conditions, depending on the degree of spatial coherence in a given lighting condition. At this time, when extracting the feature vector of computational lithography, the number of vectors in the real space needs to be input Among them, a is the range of optical interaction in the lithography machine, which is divided into subunits of equal size; b is the size of the subunits within the range of optical interaction in the lithography machine. where the size of the subunit in the range of optical interaction Among them, NA is the numerical aperture of the lithography machine, σ _max is a parameter related to the maximum angle of the exposure light on the mask, and λ is the exposure wavelength of the lithography machine. The input value of the vector in the real space is Value _cell =t _bg Area _cell +(t _f -t _bg ) Area _{geo_in_cell} , where t _bg is the complex transmission value of the background of the mask, and t _f is the complex transmission value of the pattern Value, Area _cell is the area of the subunit, and Area _{geo_in_cell} is the area of the mask pattern in the subunit.

The following is illustrated by an immersion photolithography machine: the numerical aperture NA=1.35 of the photolithography machine, the parameter σ _max =0.95 related to the maximum angle of the exposure light on the mask in the photolithography machine, the photolithography machine The exposure wavelength λ=193nm, and the optical-optical interaction range a=1500nm of the lithography machine. Since the lithography scanner is an imaging system, the maximum spatial frequency of the light field that can pass through the scanner is NA(1+σ _max ), then the size of the subunit in the optical interaction range Further, the number of vectors in the real space needs to be input , the value of each subunit is determined by the following equation:

Value _cell =t _bg Area _cell +(t _f -t _bg ) Area _{geo_in_cell} , where t _bg is the complex transmission value of the background of the mask, t _f is the complex transmission value of the pattern, and Area _cell is the area of the subunit , Area _{geo_in_cell} is the area of the mask pattern in the subunit.

If you decide to use an input vector based on spatial frequency, you can estimate the number of elements of the input vector in the spatial frequency space shown in Figure 3, the maximum spatial frequency that can be passed through the imaging system in the present invention is NA(1+σ _max ) , as shown in Figure 3 circle radius. At this time, when extracting and calculating the feature vector of lithography, the number _M2 of the vectors in the real space needs to be input to be calculated by the following formula:

Among them, M ₂ is the sum of all n _x and n _y numbers that conform to the above formula, n _x and n _y are the diffraction orders of the imaging system, NA is the numerical aperture of the photolithography machine, and σ _max is the Parameters related to the maximum angle on the reticle, λ is the exposure wavelength of the lithography machine, P=2*(the radius of the optical interaction range in optical imaging+the width of the safety belt), and the safety belt is set at the optical interaction The periphery of the range, used to ensure the calculation accuracy of the reticle pattern in the optical interaction range.

The input value of the spatial frequency-based vector is the value of the Fourier transform of the pattern in the range of optical interaction plus a certain safety belt on the lattice point of λ/P, and it needs to be considered when performing the Fourier transform The retransmission information of the mask.

The following is illustrated by an immersion photolithography machine: the numerical aperture NA=1.35 of the photolithography machine, the parameter σ _max =0.95 related to the maximum angle of the exposure light on the mask in the photolithography machine, the photolithography machine The exposure wavelength λ=193nm, and the optical-optical interaction range a=1500nm of the lithography machine. In the present invention, the maximum spatial frequency that can pass through the imaging system is NA(1+σ _max ), as shown by the circle radius in FIG. 3 . The diffraction orders (n _x , _ny ) that can pass through the imaging system must satisfy the following equation

For NA=1.35, λ=193nm, _σmax =0.95, the required total number of elements of the input vector is about 5250. The input values for the spatial frequency-based vector are the values of the Fourier transform of the pattern on the λ/P grid over the range of optical interactions plus a certain safety band. It is worth noting that when using spatial frequency information to describe the adjacent environment, the complex transmission information of the mask needs to be considered when performing Fourier transform.

After the conjugate convolutional neural network is used for feature extraction, a feed-forward neural network structure with 3 or 4 hidden layers is used to approximate any nonlinear function that the user is interested in computational lithography, as shown in Figure 4 below An integrated neural network for any type of computational lithography learning. After forming the integrated neural network in the present invention, the above-mentioned integrated neural network can be used for computational lithography learning.

The above are only preferred embodiments of the present invention, and the embodiments are not intended to limit the scope of patent protection of the present invention, so all equivalent structural changes made by using the description and drawings of the present invention should be included in the same reason Within the protection scope of the appended claims of the present invention.

Claims (9)

1. An integrated neural network for computing lithography, characterized in that the integrated neural network comprises a conjugated neural network and a forward neural network, and the output of the conjugated neural network is connected to the forward neural network The input terminal of the network; the conjugate neural network is used to extract the feature vector of the calculation lithography, and input the extracted feature vector into the forward neural network, wherein, the conjugate neural network extracts the feature vector of the calculation lithography The method of feature vector is: Y _j =∑ _i W _ij X _i , Wherein, Z _j is the extracted feature vector, W _ij is the parameter of the conjugate neural network, Xi is the adjacent environment of the _i -th point on the mask, is the conjugate of Y _j . 2 . An integrated neural network for computational lithography according to claim 1 , wherein the feature vectors for computational lithography are vectors related to light intensity. 3 . 3. A kind of integrated neural network for computing lithography according to claim 1, characterized in that, said Xi adopts a vector in real space or a vector based on spatial frequency as the _i -th point on the reticle adjacent environment. 4. A kind of integrated neural network for computing lithography according to claim 3, is characterized in that, when Xi _adopts the vector in real space, when extracting the feature vector of computing lithography, input in real space number of vectors Among them, a is the range of optical interaction in the lithography machine, which is divided into subunits of equal size; b is the size of the subunits within the range of optical interaction in the lithography machine. 5. A kind of integrated neural network for computational lithography according to claim 4, characterized in that the size of the subunit in the optical interaction range Among them, NA is the numerical aperture of the lithography machine, σ _max is a parameter related to the maximum angle of the exposure light on the mask, and λ is the exposure wavelength of the lithography machine. 6. A kind of integrated neural network for computing lithography according to claim 4, characterized in that, the input value of the vector in the real space is Value _cell =t _bg ·Area _cell +(t _f -t _bg )Area _{geo_in_cell} , where, t _bg is the complex transmission value of the background of the mask, t _f is the complex transmission value of the pattern on the mask, Area _cell is the area of the subunit, and Area _{geo_in_cell} is the mask pattern in the subunit area. 7. A kind of integrated neural network for computing lithography according to claim 3, is characterized in that, when Xi _adopts the vector based on spatial frequency, when extracting the characteristic vector of computing lithography, in the input spatial frequency The number of vectors M ₂ is calculated by the following formula: Among them, M ₂ is the sum of all n _x and n _y numbers that conform to the above formula, n _x and n _y are the diffraction orders of the imaging system, NA is the numerical aperture of the photolithography machine, and σ _max is the Parameters related to the maximum angle on the reticle, λ is the exposure wavelength of the lithography machine, P=2*(the radius of the optical interaction range in optical imaging+the width of the safety belt), and the safety belt is set at the optical interaction The periphery of the range, used to ensure the calculation accuracy of the reticle pattern in the optical interaction range. 8. A kind of integrated neural network for computational lithography according to claim 7, characterized in that the input value of the vector based on spatial frequency is the The value of the Fourier transform on the grid point of λ/P. 9. An integrated neural network for computational lithography according to claim 1, wherein the forward neural network structure has 3 or 4 hidden layers.