OPC(Optical Proximity Correction) method, and methods for manufacturing mask comprising the OPC method

Abstract

The technical concept of the present invention provides an OPC method capable of minimizing the Edge Placement Error (EPE) of a curvilinear full-chip mask, and a mask manufacturing method including the OPC method. The OPC method comprises the steps of: obtaining a curvy Neural Jacobian Model through Artificial Neural Network (ANN) learning using training data; converting a target mask into a multigon mask through a process of obtaining a first baseline curvilinear mask and a vertex reduction; extracting first geometric features from the multigon mask; and calculating the Jacobian (dEPEi / dMaskj) between the multigon mask (Maskj) and the Edge Placement Error (EPEi) using the curvy Neural Jacobian Model and the first geometric features. and the step of optimizing the multigon mask using the Jacobian;

Description

Technology Field

[0001] The technical concept of the present invention relates to a method for manufacturing a mask, and in particular to an OPC method and a method for manufacturing a mask including the OPC method. Background Technology

[0002] In semiconductor manufacturing processes, a photolithography process using a mask can be performed to form a pattern on a semiconductor substrate, such as a wafer. A mask can be described as a pattern transfer medium in which a pattern shape made of an opaque material is formed on a transparent substrate. To fabricate such a mask, a layout for the required pattern is first designed, and then the OPC-corrected layout data obtained through OPC (Optical Proximity Correction) is transmitted as MTO (Mask Tape-Out) design data. Subsequently, based on the MTO design data, Mask Data Preparation (MDP) is performed, and an exposure process can be carried out on the mask substrate. The problem to be solved

[0003] The problem that the technical concept of the present invention aims to solve is to provide an OPC method capable of minimizing the Edge Placement Error (EPE) of a curved full-chip mask, and a method for manufacturing a mask including said OPC method.

[0004] Furthermore, the problems that the technical concept of the present invention aims to solve are not limited to those mentioned above, and other problems can be clearly understood by a person skilled in the art from the description below. means of solving the problem

[0005] The technical concept of the present invention, in order to solve the above problem, comprises the steps of: obtaining a curvy Neural Jacobian Model through Artificial Neural Network (ANN) learning using training data; converting a target mask into a multigon mask through a process of obtaining a first baseline curvilinear mask and a vertex reduction; extracting first geometric features from the multigon mask; and using the curvy Neural Jacobian Model and the first geometric features to obtain a multigon mask. j ) and Edge Placement Error (EPE) i ) between Jacobian (dEPE i / dMask j The present invention provides an OPC method comprising: a step of calculating ); and a step of optimizing the multigon mask using the Jacobian.

[0006] Furthermore, the technical concept of the present invention, in order to solve the above problem, comprises the steps of: obtaining a Brute Jacobian using a learning mask; obtaining training data in the step of obtaining the Brute Jacobian; performing ANN training using the training data to obtain a neural Jacobian model of a curve; receiving data for the target mask; performing curve OPC on the target mask to obtain a first basic curve mask; performing vertex reduction on the first basic curve mask and converting the first basic curve mask into a multigon mask using the vertices; extracting first geometric features from the multigon mask; and using the neural Jacobian model of the curve and the first geometric features to obtain a multigon mask (Mask j ) and EPE(EPEi ) between Jacobian (dEPE i / dMask j The present invention provides an OPC method comprising: a step of calculating ); and a step of optimizing the multigon mask using the Jacobian.

[0007] Furthermore, the technical concept of the present invention provides a mask manufacturing method comprising the steps of: obtaining an OPC-enhanced layout by performing an OPC method using a neural Jacobian model of a curve on a target mask to solve the above problem; performing an Optical Rule Check (ORC) on the OPC-enhanced layout; transmitting the final OPC-enhanced layout that has passed the ORC as Mask Tape-Out (MTO) design data; preparing mask data based on the MTO design data; and exposing a substrate for a mask based on the mask data; wherein, in the step of obtaining the OPC-enhanced layout, a neural Jacobian matrix model of the curve is obtained through ANN learning, and a predicted value based on the neural Jacobian model of the curve is applied to the optimization of the mask. Effects of the invention

[0008] The OPC method based on the technical concept of the present invention can minimize EPE while sufficiently implementing vertex reduction using a multigon mask. For example, by obtaining a base curve mask through ILT, performing vertex reduction on the base curve mask and conversion to a multigon mask using vertices, and subsequently optimizing the mask using a neural Jacobian model of the curve, the minimization of EPE can be carried out while minimizing the computational load. Furthermore, by introducing the concepts of centroid and area of ​​perforation to extract geometric features as training data and utilizing such training data for ANN training, a neural Jacobian model of the curve with excellent performance can be obtained despite the non-uniform vertex spacing of the multigon mask. Therefore, the Jacobian can be accurately predicted through such a neural Jacobian model of the curve. Moreover, by applying accelerated mask optimization, EPE can be minimized through a small number of iterations, thereby minimizing the computational load. Brief explanation of the drawing

[0009] Figure 1 is a flowchart schematically showing the process of an OPC method using a neural Jacobian model of a curve according to one embodiment of the present invention. Figure 2 is a block structure diagram showing the OPC method of Figure 1 in more detail. Figure 3 is a conceptual diagram showing the change in surrounding EPE according to the perforation of the control point in a multigon mask. Figures 4a and 4b are conceptual diagrams showing the vertices of a curve mask and a multigon mask. Fig. 5 is a conceptual diagram showing the ANN used to generate the neural Jacobian model of the curve in the OPC method of Fig. 1. Figures 6a and 6b are conceptual diagrams showing the approximation of the area and center of gravity of a perforated multigon mask. Figures 7a and 7b are conceptual diagrams to explain the concept of geometric features used for AAN learning in the OPC method of Figure 1. Figure 8 is a conceptual diagram showing the accelerated mask optimization process that takes into account even the minute EPE change caused by the minute mask change. Figure 9 shows graphs illustrating the performance of the neural Jacobian model of the curve. Figures 10a and 10b are graphs showing the effects of the OPC method of Figure 1. FIG. 11 is a flowchart schematically showing the process of a mask manufacturing method including an OPC method according to one embodiment of the present invention. Specific details for implementing the invention

[0010] Embodiments of the present invention will be described in detail below with reference to the attached drawings. Identical components in the drawings are denoted by the same reference numerals, and redundant descriptions thereof are omitted.

[0011] FIG. 1 is a flowchart schematically showing the process of an OPC method using a neural Jacobian model of a curve according to an embodiment of the present invention, and FIG. 2 is a block structure diagram showing the OPC method of FIG. 1 in more detail. FIG. 3 is a conceptual diagram showing the change in surrounding EPE according to the perforation of control points in a multigon mask, and FIG. 4a and FIG. 4b are conceptual diagrams showing the vertices of a curve mask and a multigon mask. FIG. 5 is a conceptual diagram showing the ANN used to generate the neural Jacobian model of a curve in the OPC method of FIG. 1.

[0012] Referring to FIGS. 1 to 5, the OPC method using the Curvy Neural Jacobian Model of the present embodiment (hereinafter simply referred to as the 'OPC method') first obtains a Curvy Neural Jacobian Model through the learning of an Artificial Neural Network (ANN) using training data (S110). Here, the Jacobian may refer to the change in Edge Placement Error (EPE) (dEPE) at surrounding evaluation points (EP) according to mask variation (dMask), i.e., dEPE / dMask. For example, the Jacobian may refer to a value obtained by applying perturbation to the control point (CP) of the mask and measuring or calculating the response at surrounding evaluation points (EP) accordingly, i.e., the change in EPE at surrounding evaluation points (EP). Meanwhile, since multiple control points (CPs) and multiple variations of EPE corresponding to each of them are calculated, the Jacobian is dEPE j / dMask i It can be represented in the form of a two-dimensional matrix as shown. Accordingly, the Jacobian may also be referred to as the Jacobian matrix.

[0013] Meanwhile, the control point (CP) refers to a point where perforation is applied on the mask (M), and EPE may refer to the value obtained by subtracting the target mask (TM) from the mask contour (Con) at the evaluation point (EP). Referring to FIG. 3 for a more detailed explanation, the target mask (TM) refers to a specific shape of pattern to be actually formed on a substrate such as a wafer, and in FIG. 3, it may correspond to a rectangle. The mask (M) may refer to a pattern on the mask corresponding to the target mask (TM), or a layout for that pattern. In FIG. 3, the mask (M) may correspond to the large outer circle. The mask contour (Con) is the result obtained through the OPC method and may correspond to the small circle in FIG. 3. The purpose of the OPC method may be to make the mask contour (Con) as similar as possible to the shape of the target mask (TM).

[0014] Meanwhile, control points (CP) are points on the mask (M) where perforation is applied, and are indicated by black dots on the mask (M) in FIG. 3. Additionally, in the large circle in FIG. 3, the inner part may correspond to the mask before perforation (M-be) and the outer part may correspond to the mask after perforation (M-af). Also, in the small circle in FIG. 3, the inner part may correspond to the mask contour before perforation (Con-be) and the outer part may correspond to the mask contour after perforation (Con-af). Evaluation points (EP) may refer to points on the target mask (TM). For example, in FIG. 3, evaluation points (EP) correspond to the center points of each side of the rectangular target mask (TM), and EPE can be calculated at those evaluation points (EP).

[0015] Additionally, as shown in FIG. 3, when perforation is applied to the control point (CP0) of the central right mask (M), the corresponding mask contour (Con) changes relatively significantly, and therefore, the EPEs caused by the mask contour (Con) can also change relatively significantly. Furthermore, the perforation of the control point (CP0) affects the mask contour (Con) of the surrounding masks (M), and accordingly, the EPEs in the surrounding masks (M) can also change slightly. In FIG. 3, the EPE change values ​​at the evaluation points (EP) according to the perforation of the control point (CP0) are indicated in small squares. Meanwhile, in the case of the surrounding masks (M), the change in the mask contour (Con) is slight, so the mask contour (Con) appears as a single small circle, but in reality, it may correspond to two small circles.

[0016] Meanwhile, in the neural Jacobian model of a curve, "neural" means that the Jacobian is obtained using machine learning such as Artificial Neural Network (ANN) learning. Additionally, the preceding "curvilinear" may mean that the Jacobian is obtained for a curvilinear mask rather than a rectangular mask. For reference, the mask (M) can be represented in the form of a curve as shown in Fig. 3, but it can also be represented in the form of a combination of straight line segments. A mask represented by such a combination of segments is called a rectangular mask, and to distinguish it from the Jacobian for a rectangular mask, the Jacobian for a curvilinear mask is called the curvilinear Jacobian.

[0017] The training data for generating the neural Jacobian model of the curve may include input training data used as input values ​​in subsequent ANN training and output training data used as output values. In the OPC method of the present embodiment, the input training data may refer to geometric features. Here, the geometric feature may refer to the relative relationship between a control point (CP) and surrounding evaluation points (EP) affected by the perforation of the control point (CP). The geometric feature may include, for example, the relative position, relative angle, optical parameters, etc., between the perforation and the evaluation point (EP). Geometric features are described in more detail in the description section of FIGS. 6a to 7b.

[0018] Meanwhile, the output training data is utilized as target data or output data in ANN training, and can refer to the actual measured or calculated values ​​for the Jacobian mentioned earlier.

[0019] Referring to Figure 2, the process of obtaining the neural Jacobian model of the curve is explained in more detail as follows.

[0020] First, regarding the process of acquiring training data, a Train GDS is received as input (S10). 'Train GDS' may refer to Graphic Data System (GDS) data for a clip that is part of a full chip. Here, 'Train' means for training purposes, and the GDS data may include information such as the size and shape of the training mask. Subsequently, based on the Train GDS, a baseline curve mask can be derived through OPC. For example, a baseline curve mask can be introduced through curvilinear OPC such as Inverse Lithography Technology (ILT). Here, the baseline curve mask may appear to be a curve but may not be an actual curve. For example, as shown in Fig. 4, the baseline curve mask may be composed of a combination of fine straight line segments.

[0021] Subsequently, regarding the base curve mask, vertex reduction is performed, and the base curve mask is converted into a multigon mask using the reduced vertices (S20). Vertex reduction may refer to the process of reducing the number of vertices by selecting some of the infinite vertices of the base curve mask. Vertex reduction may be performed, for example, through a distance tolerance method. Here, the distance tolerance method may refer to selecting vertices such that the line segments between the selected vertices have a distance difference within the allowable limit with the line segments, i.e., segments, of the original base curve mask.

[0022] Subsequently, a multigon mask is formed by constructing a Bezier curve based on the reduced vertices. That is, the basic curve mask is converted into a multigon mask. Here, the Bezier curve may correspond to, for example, an implicit Bezier curve formed by only two vertices. In Fig. 4b, the thin curve between the vertices may correspond to the Bezier curve.

[0023] Subsequently, a first OPC simulation is performed on the multigon mask (S30). Through the first OPC simulation, the mask contour corresponding to the multigon mask can be calculated, and also the EPE can be calculated. Here, the first OPC simulation may refer to an OPC simulation prior to perforation of the multigon mask.

[0024] Next, geometric features are extracted from the multigon mask (S40). As previously described, the geometric features may represent the relative relationship between the control points of the multigon mask and the surrounding evaluation points (EP) that are affected by the perforation of the control points.

[0025] Geometric features are described in more detail in the description section of Figures 6a to 7b. Meanwhile, geometric features can constitute input training data (Train X) in ANN training to derive a neural Jacobian model of a curve.

[0026] Next, curve point perturbation is performed at the control points of the multigon mask (S50). Curve point perturbation may mean deforming the curve shape of the multigon mask through fine movement of the control points.

[0027] Next, a brute Jacobian (dEPE / dMask) between the multigon mask and EPE is calculated (S60). Since the brute Jacobian (dEPE / dMask) is calculated based on the actual mask variation (dMask) and the corresponding actual EPE variation (dEPE), it can be calculated as an accurate value. This brute Jacobian can constitute the output training data (Train Y) in ANN training to derive the neural Jacobian model of the curve.

[0028] However, regarding the perforation of the mask in a multigon mask, the following problem may arise. In the case of the basic curve mask shown in Fig. 4a, it consists of vertices with uniform spacing and corresponding uniform segments, making it easy to derive the optimal movement of segments to minimize surrounding EPE. However, in the case of the multigon mask shown in Fig. 4b, the distribution of vertices is irregular and does not have uniform spacing, making it difficult to derive the optimal movement of vertices to minimize surrounding EPE. Here, the moving vertices may correspond to control points. Accordingly, in the OPC method of this embodiment, when an irregular control point moves in the normal direction (dMask j ), change in EPE at surrounding evaluation points (dEPE i ), that is, the value of the Jacobian (dEPE i / dMask j To accurately deduce ), the concepts of the centroid and area of ​​the perforation are introduced. That is, geometric features are extracted using the centroid of the perforation rather than the control points, and the Jacobian is calculated using the concept of a normalized area. The concepts of the centroid and area of ​​the perforation are explained in more detail in the explanatory section of Figures 6a and 6b.

[0029] Subsequently, a second OPC simulation is performed on the multigon mask (S70). Through the second OPC simulation, the mask contour corresponding to the multigon mask can be calculated, and also the EPE can be calculated. Here, the second OPC simulation may refer to an OPC simulation after perforation of the multigon mask.

[0030] Next, it is determined whether the number of iterations (Nit) is greater than or equal to the reference number of iterations (Nref) (S80). Here, the reference number of iterations (Nref) can be set to a number of iterations sufficient to extract training data for ANN training. If the number of iterations (Nit) is greater than or equal to the reference number of iterations (Nref) (YES), the process of extracting training data (Train X, Train Y) is terminated. If the number of iterations (Nit) is less than the reference number of iterations (Nref) (NO), the process proceeds to the step of extracting geometric features (S40) and the following steps are repeated.

[0031] In some embodiments, instead of determining the number of iterations at this step, it may be determined whether the calculated EPE is less than or equal to a reference EPE. Perturbation after the calculated EPE has become less than or equal to the required reference EPE may not be meaningful in ANN learning.

[0032] Through the aforementioned process, training data (Train X, Train Y) used for ANN training can be extracted. Subsequently, the training data (Train X, Train Y) is input into the ANN (S112), and ANN training is performed based on the training data (S114). Through training, a neural Jacobian model of the curve is derived (S116).

[0033] To explain ANN in more detail with reference to Fig. 5, ANN is a network that mimics biological neural networks. In Fig. 5, for the convenience of explanation, ANN is depicted as having one hidden layer. However, ANN is not limited to this and can include various numbers of hidden layers.

[0034] In FIG. 5, the hidden layer is depicted as including four nodes, but the number of nodes included in the hidden layer is not limited thereto. For example, the hidden layer may include tens to hundreds of nodes. In the OPC method of the present embodiment, for example, the hidden layer may include 100 nodes. Also, in FIG. 5, the ANN is depicted as including a separate input layer of two nodes for receiving input data, but according to the embodiment, input data may be input directly to the hidden layer without an input layer. Additionally, according to the embodiment, the input layer may include three or more nodes.

[0035] In an ANN, nodes in layers excluding the output layer can be connected to nodes in the next layer through links that transmit output signals. Through these links, a node can be input with values ​​obtained by multiplying the node values ​​of the nodes included in the previous layer by the weights assigned to each link. The node values ​​of the previous layer correspond to axon values, and the weights correspond to synaptic weights. Weights can be referred to as parameters of the ANN. Activation functions include the sigmoid function, hyperbolic tangent (tanh) function, and ReLU (Rectified Linear Unit) function, and non-linearity can be implemented in the neural network through these activation functions.

[0036] The output of any node included in the ANN can be represented as shown in equation (1) below.

[0037] ...............................Equation (1)

[0038] Equation (1) is the output value y of the i-th node for m input values ​​in an arbitrary layer. i It can represent x j can represent the output value of the j-th node of the previous layer, and w j,i can represent the weight applied to the connection between the j-th node of the previous layer and the i-th node of the current layer. f() can represent the activation function. As shown in Equation (1), the activation function includes the input value x j and weight w j,i The cumulative result of the multiplication of can be used. In other words, the input value x at each node j and weight w j,i An operation to multiply and sum the results, i.e., a Multiply Accumulate (MAC) operation, can be performed.

[0039] Meanwhile, a neural model can be generated through such ANN training. In other words, a neural model is generated through ANN training, and when a specific value is input to the neural model, a corresponding predicted value or result value can be output. For example, in the OPC method of the present embodiment, the input value for ANN training can be the input training data (Train X), i.e., geometric features, and the output value can be the output training data (Train Y), i.e., a Brute Jacobian. Additionally, a neural Jacobian model of a curve can be obtained through ANN training. When data corresponding to the input training data is input into the obtained neural Jacobian model of the curve, a predicted value corresponding to the output training data can be calculated and output.

[0040] After obtaining the neural Jacobian model of the curve, the target mask (TM) is converted into a multigon mask through the process of obtaining a base curve mask and reducing vertices (S130). The process of converting the target mask (TM) into a multigon mask may be substantially the same as the process of converting the previously learned mask into a multigon mask. Here, the target mask (TM) may refer to the pattern to be actually formed on the substrate.

[0041] With reference to FIG. 2, the step of converting a target mask (TM) into a multigon mask (S130) is explained in more detail. First, GDS data for the target mask is received (S132). The GDS data for the target mask may refer to GDS data for a full chip. Subsequently, curve OPC is performed based on the GDS data (S134). Curve OPC can be performed, for example, through ILT. A basic curve mask is generated through curve OPC (S136). Subsequently, for the basic curve mask, vertex reduction is performed, and the basic curve mask is converted into a multigon mask using the reduced vertices (S138). The process of converting to a multigon mask may be substantially the same as the process of converting from a basic curve mask to a multigon mask for the training mask mentioned earlier. The steps from the input of GDS data (S132) to the conversion to a multigon mask (S138) may correspond to the step of converting to a multigon mask (S130) for the target mask (TM).

[0042] Subsequently, geometric features are extracted from the multigon mask (S150). As previously described, geometric features may represent the relative relationship between the control points of the multigon mask and the surrounding evaluation points (EP) that are affected by the perforation of the control points. Geometric features are described in more detail in the description section of FIGS. 6a to 7b. Meanwhile, the geometric features here may correspond to geometric features used to derive a predicted Jacobian by inputting the neural Jacobian model of the curve obtained through AAN learning.

[0043] Therefore, the extracted geometric features are input into the neural Jacobian model of the curve, and the Jacobian i,j Calculates ) (S170). Here, the Jacobian i,j ) is the output of the neural Jacobian model of the curve, and can correspond to the predicted value corresponding to the training data (Brute Jacobian) of the ANN output, i.e., the prediction Jacobian. In FIG. 2, the hat (⌒) on the dEPE within the block of S170 is the Jacobian i,j This can mean that it is a predicted value rather than an exactly calculated value. Such a Jacobian i,j ) can be calculated as in the following formula (2).

[0045] Jacobian i,j = dEPE i / dMask j ANN(Feature) * ECL ....Equation(2)

[0047] Here, ANN(Feature) refers to the output value of the ANN and corresponds to the predicted Jacobian output when geometric features are input into the neural Jacobian model. Meanwhile, in ANN training, an area-normalized brute Jacobian can be used as the output training data. Accordingly, the original brute Jacobian can be calculated by multiplying by the area. Furthermore, the area can be approximated as the Effective Curve Length (ECL) by approximating the minute variations of the mask to zero; accordingly, an accurate predicted Jacobian can be calculated by multiplying the ANN(Feature) by the ECL. The calculation of the Jacobian through specific formulas is explained in more detail in the explanatory section of Figures 6a and 6b.

[0048] After calculating the Jacobian, mask optimization is performed using the Jacobian (S190). For example, gradient descent can be used for the optimization of the multigon mask. Accordingly, mask optimization using the Jacobian can be expressed by the following equation (3) through gradient descent.

[0050] Mask new = Mask old - α (dEPE i / dMask j ) * EPE i ......Equation (3)

[0052] In Equation (3), α can be a constant related to the learning rate. Meanwhile, since the Jacobian can be approximated as the output value of the ANN by Equation (2), Equation (3) can be expressed as Equation (4).

[0054] Mask new = Mask old - α (ANN(Feature) * ECL) * EPEi ......Equation (4)

[0056] The mask optimization step (S190) may include a step of obtaining EPE through OPC simulation (192), a step of optimizing the mask using gradient descent (S194), and a step of deriving an optimized multigon mask (S196). The step of deriving an optimized multigon mask (S196) can be implemented by minimizing EPE by repeating the step of obtaining EPE (192) and the step of optimizing the mask using gradient descent (S194) tens to hundreds of times.

[0057] Meanwhile, assuming that the Jacobian predicted using ANN (i.e., the predicted Jacobian used by the neural Jacobian model of the curve obtained through ANN learning) is very accurate, accelerated mask optimization can be applied, and the change in EPE due to mask variation is expressed as in the following equation (5), thereby allowing EPE to be reduced using less simulation time.

[0059] EPE predicted = EPE true + (dEPE i / dMask j ) * dMask .......Equation(5)

[0061] Here, EPE predicted represents the predicted value for the change in EPE due to mask fluctuations, and EPE true may refer to the current EPE. Acceleration mask optimization is explained in more detail in the description section of Fig. 8.

[0062] The OPC method of the present embodiment can minimize EPE while sufficiently implementing vertex reduction using a multigon mask. For example, by obtaining a base curve mask through ILT, performing vertex reduction on the base curve mask and conversion to a multigon mask using vertices, and then optimizing the mask using a neural Jacobian model of the curve, the EPE can be minimized while minimizing computational load. Furthermore, by introducing the concepts of centroid and area of ​​perforation to extract geometric features as training data and utilizing such training data for ANN training, a neural Jacobian model of the curve with excellent performance can be obtained despite the non-uniform vertex spacing of the multigon mask. Therefore, the Jacobian can be accurately predicted through such a neural Jacobian model of the curve. Moreover, by applying accelerated mask optimization, EPE can be minimized through a small number of iterations, thereby minimizing computational load. For example, when performing a general curved OPC method, 60GB of capacity is required, but in the case of the OPC method of this embodiment, the capacity can be significantly reduced to about 23GB.

[0063] For reference, the OPC method of this embodiment may be related to an optical OPC model in a general OPC method. To briefly explain general OPC methods, they are broadly divided into two types: one is a rule-based OPC method, and the other is a simulation-based or model-based OPC method. The model-based OPC method can be advantageous in terms of time and cost because it utilizes only the measurement results of representative patterns without the need to measure all of the large number of test patterns.

[0064] Meanwhile, general OPC methods may include not only variations in the layout of the pattern, but also methods for adding sub-lithographic features called serifs on the corners of the pattern, or methods for adding sub-resolution assist features (SRAFs) such as scattering bars.

[0065] The execution of a general OPC method involves first preparing basic data for OPC. Here, the basic data may include data regarding the shape of the pattern on the sample, the location of the pattern, the type of measurement such as measurements of the pattern's space or line, and basic measurement values. Additionally, the basic data may include information regarding the thickness, refractive index, and dielectric constant of the photoresist (PR), and may include a source map regarding the shape of the illumination system. Of course, the basic data is not limited to the data exemplified above.

[0066] After preparing the basic data, an optical OPC model is generated. The generation of the optical OPC model may include the optimization of the Defocus Stand (DS) position and Best Focus (BF) position in the exposure process. Additionally, the generation of the optical OPC model may include the creation of optical images that consider light diffraction phenomena or the optical conditions of the exposure equipment itself. Of course, the generation of the optical OPC model is not limited to the above contents. For example, the generation of the optical OPC model may include various contents related to optical phenomena in the exposure process.

[0067] After generating an optical OPC model, an OPC model for the PR is generated. The generation of the OPC model for the PR may include the optimization of the PR threshold value. Here, the PR threshold value refers to the threshold value at which a chemical change occurs during the exposure process; for example, the threshold value can be given as the intensity of the exposure light. The generation of the OPC model for the PR may also include selecting an appropriate model form from among several PR model forms.

[0068] The optical OPC model and the OPC model for PR are combined and generally referred to as the OPC model. After generating the OPC model, a simulation using the OPC model is performed to generate an OPC-enabled layout. As mentioned earlier, the OPC simulation can be considered a simulation using the optical OPC model.

[0069] FIGS. 6a and 6b are conceptual diagrams showing the approximation of the area and center of gravity of a perforated multigon mask, and FIG. 6b shows the perforated portion of FIG. 6a approximated in a rectangular shape. The center control point (CP0) before perforation in FIG. 6a corresponds to b in the rectangle of FIG. 6b, the control points (CP1, CP2) adjacent to both sides of the center control point (CP0) correspond to vertices a and c in the rectangle of FIG. 6b, and the control point (CP0) after perforation corresponds to vertex b' in the rectangle of FIG. 6b.

[0070] Referring to FIG. 6a and FIG. 6b, in FIG. 6b, when the length of the side between vertices a and b is l and the length of the side between vertices b and c is m, the centroid of the first triangle formed by vertices a, b, and b' according to perforation is located inside the first triangle, and the centroid of the second triangle formed by vertices b, b', and c according to perforation can be located inside the second triangle. If the mask variation (dMask) is nearly zero, that is, when the perforation of the center control point (CP0) is very fine, the length of the side between vertices b and b' can be negligibly small compared to l and m. Additionally, the centroid can be approximated as a point on the side between vertices a and b for the first triangle, and as a point on the side between vertices b and c for the second triangle. In FIG. 6b, the centroid of the first triangle (COM of abb') is marked as a point on the side between a and b, and the centroid of the second triangle (COM of cbb') is marked as a point on the side between b and c. By applying a similar approximation concept to the centroid of the first triangle (COM of abb') and the centroid of the second triangle (COM of cbb'), the centroid of the area due to perforation (COM of Aptb) is approximated as a point on the dotted line connecting the centroid of the first triangle (COM of abb') and the centroid of the second triangle (COM of cbb'), and is marked as a point in FIG. 6b.

[0071] For reference, the center of gravity of the area (COM of Aptb) due to perforation can be calculated using the following equation (6).

[0073] COM of Aptb = [l*{(1-1 / 2 1 / 2 )*a + (1 / 2 1 / 2 )*b} + m*{(1-1 / 2 1 / 2 )*c + (1 / 2 1 / 2)*b}] / (l + m) ...........Equation(6)

[0075] Meanwhile, when the mask variation (dMask) is nearly zero, the area due to perforation can be approximated by the effective curve length (ECL). The effective curve length (ECL) can be calculated using the following equation (7).

[0077] ECL = 0.5*{l*sin(θ / 2) + m*sin(θ / 2)} ......Equation (7)

[0079] Here, θ may correspond to the angle formed by line segments when connecting the central control point (CP0) to adjacent control points (CP1, CP2) with line segments, as shown in FIG. 6b. This θ may correspond to one of the geometric features. In the OPC method of the present embodiment, the geometric features may be defined by the relative relationship between the centroid of the mask variation (dMask), rather than the control point, and the adjacent evaluation point (EP) adjacent thereto. Consequently, a neural Jacobian model of a curve with excellent performance can be generated, and furthermore, through this neural Jacobian model of the curve, the accuracy of the prediction value can be improved.

[0080] Figures 7a and 7b are conceptual diagrams for explaining the concept of geometric features used for AAN learning in the OPC method of Figure 1, and Figure 7b shows an enlarged view of a part of Figure 7a.

[0081] Referring to FIGS. 7a and 7b, FIG. 7a schematically shows the EPE variation (△EPE) at an evaluation point on the left adjacent target mask according to the mask variation (△Mask), and FIG. 7b shows the concepts of geometric features by zooming in on the corresponding portion. In the OPC method of the present embodiment, as geometric features, TDIR, CDIR, e (-(DIST)) , e -(DX) , e -(DY)The 8-dimensional elements of ), DDIR, ILS (Image Log Slope), and CUIR can be defined. To explain each element of the geometric features in detail,

[0082] TDIR is the normal vector of the evaluation point (EP) TDIR The absolute value of the y-component of )(abs( TDIR · y It means )). Here, bold text represents a vector, and abs represents an absolute value. The same applies below.

[0083] CDIR is the normal vector of the centroid (COM) CDIR) The absolute value of the y-component of (abs( CDIR · y This means )). For reference, the center of gravity (COM) is shown as the same point as the control point (CP0) in FIG. 7b for convenience, but as can be seen from FIG. 6b, the positions of the control point (CP0) and the center of gravity (COM) may differ slightly.

[0084] e (-(DIST)) is the distance vector from the center of gravity (COM) to the evaluation point (EP) DISTRICT The absolute value of ) (∥ DISTRICT ∥ = DIST) can mean an exponential value having a negative exponent value. For reference, the center of gravity (COM) is shown as the same point as the control point (CP0) in FIG. 7b for convenience, but as can be seen from FIG. 6b, the location of the control point (CP0) and the center of gravity (COM) may be different.

[0085] e -(DX) is the distance vector( DISTRICT The absolute value of the x component of )(abs( DISTRICT · x It means an exponential value having ) = DX) as a (-) exponent value, and e -(DY) is the distance vector( DISTRICT y component of ) DISTRICT · y )'s absolute value(abs( DISTRICT · y It means an exponential value that has ) = DY) as a (-) exponent value.

[0086] DDIR is the normal vector of the evaluation point (EP) TDIR ) and the normal vector of the centroid (COM) CDIR) It refers to the angle between them and can be expressed as follows. DDIR = ( TDIR CDIR ) / ∥ TDIR ∥*∥ CDIR ∥, where “∥ ∥” is the absolute value symbol for a vector.

[0087] ILS is the logarithmic slope of the image at the evaluation point, and CURV is the curvature of the mask contour, which can be derived from θ in Fig. 6b above.

[0088] Meanwhile, in the OPC method of the present embodiment, 8-dimensional elements were described as geometric features, but geometric features are not limited thereto. For example, geometric features can be defined in various ways to derive a neural Jacobian model of a curve with excellent performance through an ANN. Accordingly, geometric features may be defined to be less than 8 dimensions or to be defined to be more than 8 dimensions.

[0089] Figure 8 is a conceptual diagram showing the accelerated mask optimization process that takes into account even the minute EPE change caused by the minute mask change.

[0090] Referring to FIG. 8, the optimization of the mask can be performed through the aforementioned equation (4), where △Mask is indicated instead of dMask in the middle box of the large box in FIG. 8, and EPEi is the predicted value by the neural Jacobian model of the curve ( It is recorded as ). In addition, these predicted values ​​( ) can be calculated through Equation (5) by considering the change in micro EPE according to the change in micro mask. Therefore, by repeating the process inside the large box, the predicted value ( ) and mask fluctuations(△Maskj ) can be calculated precisely. When the process of Equation (4) is called the mask optimization process, the process of optimizing the mask using Equation (5) together is called the accelerated mask optimization process. X300 inside the large box may mean 300 iterations. The number of such iterations is not limited to 300.

[0091] Subsequently, mask fluctuations (△Mask j ) can be used to calculate the mask (Maskj) in the last box. Subsequently, the mask contour and EPE can be calculated for the mask (Maskj) calculated through OPC simulation in the first box. In this way, through the accelerated mask optimization process of the intermediate box, the predicted value ( ) and mask fluctuations(△Mask j By precisely calculating ) first, the number of total iterations (X Itrn) can be minimized.

[0092] For reference, to minimize EPE, the entire simulation process of calculating EPE in Fig. 8 can be repeated (X Itrn). However, sufficient iterations are performed during the accelerated mask optimization process to obtain the optimal mask △Mask j After obtaining ), proceeding with the entire simulation process may be advantageous in terms of speed and accuracy in minimizing EPE. For example, in the OPC method of the present embodiment, to minimize EPE, the acceleration mask optimization process may be repeated about 300 times, and the entire simulation process may be repeated about 10 times. Of course, the number of repetitions of the acceleration mask optimization process M and the entire simulation process is not limited to the aforementioned figures.

[0093] Figure 9 is a graph showing the performance of the neural Jacobian model of the curve, where the x-axis represents the actual Jacobian value obtained through actual measurement and the y-axis represents the Jacobian value predicted by the neural Jacobian model, both of which have no units.

[0094] In the graph of Figure 9, the R² coefficient of determination is 0.98, which is close to 1. Therefore, it can be seen that the neural Jacobian model and the predicted values ​​are very consistent.

[0095] Figures 10a and 10b are graphs showing the effects of the OPC method of Figure 1.

[0096] Referring to Fig. 10a, in the graph of Fig. 10a, the x-axis represents the total number of simulation iterations of Fig. 8, and the y-axis represents the EPE value. Additionally, the solid line represents the maximum value (Max) of the EPE value, and the dotted line represents the minimum value (Min) of the EPE value. As can be seen from the graph of Fig. 10a, the EPE value decreases as the number of iterations increases. Specifically, after conversion to a multigon mask, the EPE value is approximately -1 to 3 at 0 iterations, and can decrease to approximately -0.2075 to 0.2226 through 8 iterations. Therefore, it can be seen that the final EPE value decreases to about 1 / 10 of the initial EPE value.

[0097] Referring to FIG. 10b, in the graph of FIG. 10b, the x-axis represents the EPE value and the y-axis represents the number of evaluation points. Additionally, light hatching represents the distribution of EPE according to the OPC method of the comparative example, and dark hatching represents the distribution of EPE according to the OPC method of the present embodiment. As can be seen from the graph of FIG. 10b, in the distribution of EPE according to the OPC method of the present embodiment, it can be confirmed that the EPE for most evaluation points is minimized and approaches almost zero.

[0098] FIG. 11 is a flowchart schematically illustrating the process of a mask manufacturing method including an OPC method according to an embodiment of the present invention. Content already explained in the description of FIG. 1 to FIG. 10b is briefly explained or omitted.

[0099] Referring to FIG. 11, a mask manufacturing method including an OPC method (hereinafter simply referred to as the "mask manufacturing method") first obtains an OPC-enhanced layout by performing an OPC method using a neural Jacobian model of a curve on a target mask (S210). The OPC method using a neural Jacobian model of a curve is as described in the description section of FIG. 1 to 5. Meanwhile, the OPC-enhanced layout may correspond to a mask layout with minimized EPE obtained through the OPC method using a neural Jacobian model of a curve.

[0100] After generating an OPC layout, an Optical Rule Check (ORC) is performed on the OPC layout (S230). The ORC may include, for example, calculation of Root Mean Square (RMS) for CD errors, calculation of Edge Placement Error (EPE), pinch error checking, bridge error checking, etc. However, the items checked in the ORC are not limited to the aforementioned items.

[0101] During the execution of ORC, it is determined whether a defect exists. Here, a defect may include cases where the RMS for CD error is greater than a set threshold value, where EPE is greater than a set threshold value, where pinch error exists, or where bridge error exists. Additionally, if there are other items in ORC, cases where those items deviate from the criteria may also be considered defects.

[0102] In the execution of ORC, if a defect exists, the process proceeds to the step of generating an OPC layout (S210) depending on the cause. Therefore, before proceeding to the step of generating an OPC layout (S210), a step of analyzing the cause of the defect and reflecting the cause in the corresponding OPC model may be performed first.

[0103] In the ORC execution, if no defects exist, the OPC layout is determined as the final OPC layout, and the final OPC layout image is transmitted to the mask production team as MTO (Mask Tape-Out) design data (S250). Generally, MTO may mean handing over the final mask data obtained through the OPC method to the mask production team to request mask production. Therefore, the MTO design data can be substantially identical to the data for the final OPC layout image obtained through the OPC method. Such MTO design data may have a graphic data format used in Electronic Design Automation (EDA) software, etc. For example, the MTO design data may have a data format such as GDS2 (Graphic Data System II) or OASIS (Open Artwork System Interchange Standard).

[0104] Subsequently, Mask Data Preparation (MDP) is performed (S270). Mask data preparation may include, for example, i) format conversion, ii) augmentation of a barcode for mechanical reading, a standard mask pattern for inspection, a job deck, etc., called fracturing, and iii) verification in automatic and manual ways. Here, a job deck may refer to creating a text file containing a series of instructions such as batch information of multiple mask files, a reference dose, and exposure speed or method.

[0105] Meanwhile, format conversion, or fracturing, may refer to a process of dividing MTO design data into regions and converting it into a format suitable for an electron beam lithography machine. Fracturing may include data manipulations such as scaling, sizing, rotation, pattern reflection, and color inversion. During the conversion process through fracturing, data regarding numerous systematic errors that may occur at any point during the transfer from design data to the image on the wafer can be corrected. The data correction process for these systematic errors is called Mask Process Correction (MPC), and may include operations such as linewidth adjustment (CD adjustment) and increasing pattern placement precision. Therefore, fracturing can contribute to improving the quality of the final mask and may also be a process performed prior to Mask Process Correction. Here, systematic errors may be caused by distortions occurring during the exposure process, mask development and etching processes, and wafer imaging processes.

[0106] Meanwhile, mask data preparation may include MPC. As previously mentioned, MPC refers to a process that corrects errors, specifically systematic errors, that occur during the photolithography process. Here, the photolithography process may be a concept that generally includes electron beam writing, development, etching, baking, etc. Additionally, data processing may be performed prior to the photolithography process. Data processing is a type of preprocessing step for mask data and may include syntax checking of the mask data and prediction of exposure time.

[0107] After preparing the mask data, the substrate for the mask is exposed based on the mask data (S290). Here, exposure may refer to, for example, electron beam writing. Here, electron beam writing can be performed using a gray exposure method, for example, with a Multi-Beam Mask Writer (MBMW). Additionally, electron beam writing may be performed using a Variable Shape Beam (VSB) exposure machine.

[0108] Meanwhile, after the mask data preparation stage and before the exposure process, a process of converting the mask data into pixel data may be performed. Pixel data is data directly used for actual exposure and may include data regarding the shape to be exposed and data regarding the dots assigned to each. Here, the data regarding the shape may be bit-map data converted from shape data, which is vector data, through rasterization or the like.

[0109] After the photolithography process, a series of processes are performed to complete the mask. These processes may include, for example, development, etching, and cleaning. Additionally, the series of processes for mask manufacturing may include metrology, defect inspection, or defect repair processes. Furthermore, a pellicle application process may be included. Here, the pellicle application process refers to the process of attaching a pellicle to the mask surface to protect it from subsequent contamination during shipping and throughout its service life, once it is confirmed through final cleaning and inspection that there are no contaminants or chemical stains.

[0110] The mask manufacturing method of the present embodiment may employ an OPC method utilizing a neural Jacobian curve model. Accordingly, by acquiring a neural Jacobian curve model through ANN training and utilizing the neural Jacobian curve model for mask optimization, the execution time of the entire OPC can be accelerated, and the EPE can also be minimized. Specifically, a multigon mask is acquired, and perforation is performed at control points to acquire a large amount of training data, consisting of input training data (geometric features such as relative coordinates, relative angles, optical parameters, etc.) and output training data (a brute Jacobian (dEPE / dMask)). This training data can then be utilized in ANN training to generate a neural Jacobian curve model. Furthermore, by introducing the concept of the centroid to the perforation of control points, a neural Jacobian curve model with excellent performance for multigon masks can be acquired. By utilizing the predicted values ​​from the neural Jacobian model of these curves and performing mask optimization through gradient descent, EPE can be minimized with a small number of iterations, and the amount of computation can also be minimized.

[0111] Up to this point, the present invention has been described with reference to the embodiments illustrated in the drawings, but this is merely illustrative, and those skilled in the art will understand that various modifications and equivalent alternative embodiments are possible therefrom. Accordingly, the true technical scope of protection of the present invention should be determined by the technical spirit of the appended claims.

Claims

Claim 1 A step of obtaining a curvy Neural Jacobian Model through Artificial Neural Network (ANN) training using training data; a step of converting a target mask into a multigon mask through obtaining a first baseline curvilinear mask and a vertex reduction process; a step of extracting first geometric features from the multigon mask; and a step of using the curvy Neural Jacobian Model and the first geometric features to obtain a multigon mask. j ) and Edge Placement Error (EPE) i ) between Jacobian (dEPE i / dMask j An OPC method comprising: a step of calculating ); and a step of optimizing the multigon mask using the Jacobian. Claim 2 An OPC method according to claim 1, wherein the step of obtaining a neural Jacobian model of the curve comprises: obtaining a brute Jacobian for a learning mask; extracting training data in the step of obtaining the brute Jacobian; performing ANN training using the training data; and deriving a neural Jacobian model of the curve. Claim 3 In claim 2, the step of obtaining the brute Jacobian comprises: receiving data for the training mask; obtaining a second basic curve mask for the training mask; performing vertex reduction on the second basic curve mask and converting the second basic curve mask into a training multigon mask using the vertices; performing a first OPC simulation on the training multigon mask; extracting second geometric features from the training multigon mask; performing curve point perturbation at the control points of the training multigon mask; performing a second OPC simulation on the training multigon mask; and deriving the brute Jacobian between the training multigon mask and EPE, wherein after the step of deriving the brute Jacobian, if a set criterion is satisfied, the process terminates, and if it is not satisfied, proceeds to the step of extracting the second geometric features. Claim 4 An OPC method according to claim 3, wherein the second geometric features are obtained based on the relative relationship between the control point of the learning multigon mask and the surrounding evaluation point influenced by the perforation (dMask) of the control point. Claim 5 An OPC method according to claim 1, wherein the step of converting to a multigon mask comprises: receiving data for the target mask; performing curve OPC on the target mask to obtain the first basic curve mask; and performing vertex reduction on the first basic curve mask; and converting the first basic curve mask into the multigon mask using the vertices. Claim 6 A step of obtaining a Brute Jacobian using a training mask; a step of obtaining training data in the step of obtaining the Brute Jacobian; a step of performing ANN training using the training data to obtain a Neural Jacobian model of the curve; a step of receiving data for the target mask; a step of obtaining a first basic curve mask by performing curve OPC on the target mask; a step of vertex reduction on the first basic curve mask and converting the first basic curve mask into a multigon mask using the vertices; a step of extracting first geometric features from the multigon mask; and a multigon mask (Mask) using the Neural Jacobian model of the curve and the first geometric features. j ) and EPE(EPE i ) between Jacobian (dEPE i / dMask j An OPC method comprising: a step of calculating ); and a step of optimizing the multigon mask using the Jacobian. Claim 7 In claim 6, the step of obtaining the brute Jacobian comprises: receiving data for the learning mask; obtaining a second basic curve mask for the learning mask; performing vertex reduction on the second basic curve mask and converting the second basic curve mask into a learning multigon mask using the vertices; performing a first OPC simulation on the learning multigon mask; extracting second geometric features from the learning multigon mask; performing curve point perforation at the control points of the learning multigon mask; performing a second OPC simulation on the learning multigon mask; and deriving the brute Jacobian between the learning multigon mask and EPE, wherein after the step of deriving the brute Jacobian, if a set criterion is satisfied, the process terminates, and if it is not satisfied, proceeds to the step of extracting the second geometric features. Claim 8 A mask manufacturing method comprising: a step of obtaining an OPC-enhanced layout by performing an OPC method using a neural Jacobian model of a curve on a target mask; a step of performing an Optical Rule Check (ORC) on the OPC-enhanced layout; a step of transmitting the final OPC-enhanced layout that has passed the ORC as Mask Tape-Out (MTO) design data; a step of preparing mask data based on the MTO design data; and a step of exposing a substrate for a mask based on the mask data; wherein, in the step of obtaining the OPC-enhanced layout, a neural Jacobian matrix model of the curve is obtained through ANN learning, and a predicted value based on the neural Jacobian model of the curve is applied to the optimization of the mask. Claim 9 In claim 8, the step of obtaining an OPC-enhanced layout comprises: obtaining a neural Jacobian model of the curve through the ANN learning using the training data; obtaining a first basic curve mask for the target mask and converting it into a multigon mask through a vertex reduction process; extracting first geometric features from the multigon mask; and using the neural Jacobian model of the curve and the first geometric features to obtain a multigon mask (Mask j ) and EPE(EPE i ) between Jacobian (dEPE i / dMask j A mask manufacturing method characterized by including the step of calculating ) and the step of optimizing the multigon mask through gradient descent using the Jacobian. Claim 10 A method for manufacturing a mask according to claim 9, wherein the step of obtaining a neural Jacobian model of the curve comprises: a step of obtaining a brute Jacobian for a learning mask; a step of obtaining the training data in the step of obtaining the brute Jacobian; a step of performing the ANN training using the training data; and a step of deriving a neural Jacobian model of the curve.

Images