KR20210017431A - Apparatus and method for generating control data of polishing process and control device comprising the same - Google Patents

Abstract

The present invention provides an apparatus and a method for generating process control information and a process control apparatus including the same, which improve the yield of a semiconductor production process and improve the reliability of a process. According to an embodiment of the present invention, the apparatus for generating process control information required for a substrate manufacturing process comprises: an input unit to receive target scale information associated with a scale of areas belonging to a substrate to be manufactured, and process parameter information for at least one piece of equipment for manufacturing the substrate; a generation network to generate first process control information required for a process for manufacturing the substrate by using a plurality of layers based on the target scale information and the process parameter information; and a determination network to derive evaluation results for the first process control information by learning based on the first process control information and second process control information which is training data.

Description

Process control information generating apparatus, method, and process control device including the same {Apparatus and method for generating control data of polishing process and control device comprising the same}

The present invention relates to an apparatus and a method for generating process control information, and a process control apparatus including the same.

In the semiconductor manufacturing process, it is difficult to consistently obtain uniform process results due to very complex physical and chemical reactions and non-linear process variations according to equipment conditions. For precise process control, the run-to-run (R2R) control proposed in the 1990s estimates the process variation for each run (processing unit such as a single wafer or wafer lot) and reflects it in the input so that the process result does not deviate from the target. do.

Process drift, which is a type of process change, refers to a slow deterioration in process performance as equipment ages. Another type, process shift, is a rapid change in process performance, for example. It means a rapid performance improvement of the process by replacing the consumables of the equipment.

Traditionally, R2R control has used the exponential weighted moving average (EWMA) method and the double EWMA method, which are widely used for linear system control. The EWMA method estimates process drift by weighted average of the control results of past runs, and has been considered preferentially in the R2R controller due to its low computational complexity. The Double EWMA method copes with process drift by using a combination of an EWMA filer that estimates process results and an EWMA filter that estimates process drift.

Among the many unit processes required for semiconductor manufacturing, the chemical mechanical planarization (CMP) process has a distinct process drift, so R2R control is essential. In addition, there is a study that proved through experiment that the R2R control model using three EWMA filters shows superior control performance compared to the double EWMA model. However, these EWMA models are difficult to apply to systems with nonlinear process drift, so it is recommended to use only when the drift does not change rapidly.

Neural network-based R2R (NNR2R) control method is very useful for controlling nonlinear systems where process factors have complex correlations. Neural network is effective in modeling and controlling nonlinear systems because it can learn various types of functions because it adjusts network weights with only the input/output data of the system. In the CMP process, it was demonstrated that it is possible to control the input so that the thickness of the wafer has a uniform profile by using the NNR2R controller.

The NNR2R controller consists of two neural networks, a process model and a control model, where the process model means virtual metrology (VM). The VM model can reduce the cost and time required for actual measurement, detect process drift quickly and accurately, and improve final production by detecting defective wafers early. Due to this advantage, recent studies have been conducted to maximize the predictive ability of VMs using machine learning methods. In the VM study, a study that proposed a VM model structure combining neural networks to approximate process drift, a VM modeling study using trace data of the etching process in CNN, and a study that improved R2R control performance through a Gaussian process regression-based VM model. There is this. Until now, research on the VM model has been actively conducted, but care must be taken when integrating the R2R control model and the VM model. The performance of the R2R controller is determined by the prediction error of the VM model and the error of the control model, so the error is higher than that of a single R2R control model. In particular, when various input factors are intercorrelated like a semiconductor manufacturing process and a nonlinear process drift phenomenon is present, it is difficult to guarantee excellent control performance unless the prediction accuracy of the VM model is high.

Korean Patent Publication No. 10-2019-0041908 (Public)

The present invention is a system encompassing wafer measurement, R2R control, abnormality detection, etc. in the semiconductor manufacturing industry in order to solve the above-described conventional problems, and an apparatus for generating process control information, which improves the yield and reliability of the semiconductor production process, It provides a method and a process control device including the same.

The apparatus for generating process control information according to an embodiment of the present invention for achieving the above object includes target scale information related to the scale of regions belonging to a substrate to be manufactured and process parameter information for at least one equipment for manufacturing the substrate. A generation network for generating first process control information necessary for a process for manufacturing the substrate by using a plurality of layers based on the input unit receiving the input, the target scale information and the process parameter information, and the first process control information, and It may include a discrimination network that derives an evaluation result for the first process control information by learning based on the second process control information that is training data.

A method of generating process control information according to another embodiment of the present invention for achieving the above object includes target scale information related to the scale of regions belonging to a substrate to be manufactured, and a process for at least one equipment for manufacturing the substrate. Receiving parameter information, generating first process control information necessary for a process for manufacturing the substrate using a plurality of layers based on the target scale information and the process parameter information, and the first process control information And deriving an evaluation result for the first process control information by learning based on the second process control information that is training data.

In order to achieve the above object, a process control apparatus according to another embodiment of the present invention receives target scale information related to scales of regions belonging to a substrate to be polished, and includes at least one device for polishing the substrate. Including; a process control information generating unit for generating process control information including pressure setting information and time setting information related to the polishing time of the equipment, and a control unit for controlling the process conditions of the equipment based on the process control information; The control unit may transmit the process conditions of the controlled equipment and the accumulated use time of the equipment to the process control information generation unit.

In addition, the present invention proposes a computer program stored in a computer-readable recording medium for executing the method for generating process control information according to the above method.

The apparatus and method for generating process control information and a process control apparatus including the same according to an embodiment of the present invention provide control values for effectively polishing wafers by accurately estimating process changes caused by process performance deterioration caused by equipment aging. There is an effect that can be done

1 is a diagram illustrating components and equipment used in a substrate (wafer) manufacturing process according to an exemplary embodiment.
2 is a block diagram schematically showing the configuration of a process control apparatus according to an embodiment of the present invention.
3 is a conceptual diagram illustrating a learning process for outputting process control information according to an embodiment of the present invention.
4 is a block diagram schematically showing the configuration of a generation network according to an embodiment of the present invention.
5 is an enlarged view of FIG. 3 illustrating a drift characteristic information generation unit according to an embodiment of the present invention.
6 is a block diagram schematically showing the configuration of a process control apparatus according to an embodiment of the present invention.
7 is a diagram showing an index for performance evaluation.
FIG. 8 is a result of an experiment to improve prediction performance by optimizing a hyper-parameter to BO for each of the two comparison models in FIG. 7.
9 is a diagram showing an experiment result for hyper-parameter optimization.
10 is a diagram showing an experiment result for hyper-parameter optimization.
11 is a diagram showing a result of an experiment of an R2R control algorithm.

The present invention will be described in detail by describing a preferred embodiment of the present invention with reference to the accompanying drawings. However, the present invention may be implemented in various different forms, and is not limited to the described embodiments. Further, in order to clearly describe the present invention, parts not related to the description are omitted, and the same reference numerals in the drawings indicate the same members.

Throughout the specification, when a certain part "includes" a certain component, it means that other components may be further included, rather than excluding other components unless otherwise stated. In addition, terms such as "... unit", "... group", "module", and "block" described in the specification mean a unit that processes at least one function or operation, which is hardware, software, or hardware. And software.

Hereinafter, an embodiment of the present invention will be described in detail with reference to the accompanying drawings. In addition, in describing the present invention, when it is determined that a detailed description of a related known configuration or function may obscure the subject matter of the present invention, a detailed description thereof may be omitted.

In the semiconductor business, a structure in which microcircuit layers are stacked on a wafer is required to make more semiconductor devices per unit area. However, as the lithography process of forming the microcircuit pattern and the etching process of removing the remaining layer portions except for the necessary circuit pattern are repeated, a surface level difference between the high portion where the circuit is generated and the remaining low portion occurs. The surface level difference acts as a critical factor for the resolution in the lithograpy process of forming the circuit pattern of the subsequent layer, causing a defect in the semiconductor device. Therefore, in order to implement a stacked structure, a CMP process that enables local flattening of the wearer and global flattening required for fine circuit patterns is required. The chemical-mechanical polishing (CMP) process uses rotary equipment as shown in FIG. 1. This equipment is a device that supplies a carrier for wafers, a platen for a polishing pad, a dresser for cleaning the pad, and a polishing slurry. Can be configured. In the CMP process, a film of a wafer is flattened through a chemical action occurring between a slurry containing abrasive particles and a material on the wafer surface, and a mechanical action caused by direct contact between the polishing pad and the wafer. It is a typical process using CMP's consumable parts (polishing pads, retainer rings, membranes, and dressers). Consumable parts generate a drift that causes the material removal rate to vary depending on their lifespan, and a process shift that improves the removal rate occurs when the consumable parts of the equipment are replaced. Because of the various variations in these processes, the CMP process is a typical nonlinear system that is difficult to control.

The quality of the CMP process is the thickness of the material remaining on the wafer. Rotary CMP equipment is designed so that the membrane of the carrier in contact with the wafer can apply a different pressure to each area to control the thickness of each area of the wafer, so it is possible to manage the thickness of the material by area. Z1 to Z7 shown in FIG. 1 denote zones of the wafer to which pressure is applied.

Control variables and equipment parameters are required for CMP process control. Control variables include the relative speed of the platen and carrier, polishing time, slurry flow rate, dresser speed, and pressure applied to the pad by region of the wafer (substrate), but the actual production process Except for the pressure and polishing time in the area, the control variables are strictly controlled so that the amount of change is not large from the existing value. Part parameters record the life of the consumable and are used as information for process variation modeling.

Hereinafter, the apparatus and process for generating process control information according to an embodiment of the present invention, which optimizes the CMP process, which is a nonlinear system as described above, so that the desired thickness for each region of the wafer (substrate) can be realized despite various environmental changes. The configuration of the control device and method will be described in detail with reference to the related drawings.

The apparatus for generating process control information according to an embodiment of the present invention establishes a control model that determines the pressure and polishing time of the region in consideration of component parameters in order to obtain a target thickness for each region of the wafer (substrate) for each run. It aims to do.

A process control apparatus according to an embodiment of the present invention may include an input unit, a process control information generation unit, a control unit, and an update unit. The process control device is a device that is used when the filter coefficients of each internal layer are set and applied to an actual process as machine learning has already been completed. Such a process control device will be described in detail after the process control information generating device applied to the learning process before being applied to the process is described.

2 is a block diagram schematically showing a configuration of an apparatus for generating process control information according to an embodiment of the present invention. Referring to FIG. 2, an apparatus for generating process control information according to an embodiment of the present invention may include an input unit 100, a learning unit 200, a control unit 300, and an update unit 400.

The input unit 100 according to an embodiment of the present invention may receive target scale information related to the scale of regions belonging to the substrate to be manufactured and process parameter information for at least one component for manufacturing the substrate (hereinafter, referred to as wafer). have.

Here, the target scale information means a target thickness value for each area of the wafer to be achieved in polishing the wafer, and can be input from the user, and the process parameter information is used to polish the wafer as shown in FIG. This can mean the cumulative time the parts have been used to polish the wafer.

The input unit 100 according to an embodiment is provided separately, as shown in FIG. 2, to receive information on the target thickness value and the cumulative use time of the part, and transmit it to the generation network 210 and the discrimination network 220. It may be implemented, and each of the input information (target thickness value, accumulated part usage time, etc.) may be directly input to the generation network 210 or the discrimination network 220 for learning without the input unit 100.

In addition, the learning unit 200 according to an embodiment of the present invention performs machine learning in order to set the filter coefficient of each layer of the process control information generation unit of the process control device to apply to the actual process after the above-described machine learning is completed. As such, the learning unit 200 of the present invention may be composed of a generation network 210 and a discrimination network 220.

The generation network 210 may generate first process control information necessary for a process for manufacturing a wafer by using a plurality of layers based on scale information input to the input unit 100 and process parameter information.

The discrimination network 220 derives an evaluation result for the first process control information generated by the generation network 210 based on the second process control information applied to the actual substrate manufacturing process as first process control adjustment and training data. can do.

The process control information generating apparatus having the structure as shown in FIG. 2 can learn by transmitting the recently processed process data to the input unit and the learning unit using the update unit 400 whenever each run process is completed. I can.

Here, the learning unit 200 of the present invention is a Least Squares Generative Adversarial Network (LSGAN) composed of a generation network 210 and a discrimination network 220 that helps the generation network learning through competition. While the conventional neural network-based run to run control (NNR2R) described above is an indirect control method that finds a control value using a virtual metrology (VM) model, the generation network 210 of the present invention has a control-output relationship. Is learned through competition with the discriminant network 220 and used directly for R2R control. Accordingly, when the learned generation network 210 receives the target output value (target scale information), it may output process control information (control value) to be achieved.

The generation network 200 of the present invention may generate process control information using a control function as shown in Equation 1 below.

Here, u _k is an output control value output from the generation network 210, y is an output vector, u is a control vector, and n is a processed process referenced by the generation network 210 ( It means the number of run). The generation network 210 receives the most recent process control history information (control value) from the update unit 400, output data, and the cumulative use time of parts used in the polishing process, and the output thickness for each area of the final polished wafer It is possible to derive process control information u _k that makes the value y _k and the target thickness y ^target input to the input unit 100 from the user become the same.

In this case, the process control history information may include process sequence information and control parameter information for controlling respective equipment. In addition, the process control history information may further include target scale information in the sequence. Here, the control parameter information is, for example, the control value of each of the past multiple processes, that is, the pressure value and polishing time of the part set in the past process, the filter coefficient (weight) of each label of the generation network, and the final polishing. This may mean all parameter information in the wafer polishing process including thickness values for each region.

A control function for generating the first process control information by the generation network 210 using the process control information as described above may be defined as shown in Equation 2 below.

는 생성망(210)이 생성하는 공정 제어정보이며, Φ는 공정 파라미터 정보인 부품의 누적 사용시간이고, W_drift 및 W_R2R은 생성망에서의 학습 가중치(레이어들의 필터계수)이다. Here, G(·) is a function used in the generation network 210,

Is the process control information generated by the generating network 210, Φ is the cumulative use time of the part, which is the process parameter information, and W _drift and W _R2R are the learning weights (filter coefficients of layers) in the generating network.

In order to cope with the process drift of the CMP process, the weight (W) of the generation network 210 is based on the process parameter information of the latest process whenever each run in the process control information generation apparatus of the present invention is completed. _drift ) can be updated. Subsequently, the weight (W _R2R ) for achieving target scale information (target thickness value) is updated through competitive learning of the generation network 210 and the discrimination network 220, and when the learning is completed, the target thickness value (y ^target ) is Process control information (u _k ) to achieve can be output.

The GAN, which is a conventional neural network, is trained so that an output value corresponding to an arbitrary latent vector (z) defined within [0, 1] is the same as an actual output value. On the other hand, the control variables of the current process in the process control information generating apparatus of the present invention have a temporally dependent relationship with the control variables of previous processes. Accordingly, the apparatus for generating process control information of the present invention does not learn control variable values (output values) for the latent vector z like a conventional GAN, but based on feature information on process control variables in previous processes. This makes it possible to predict the value of the process control variable in the current process.

3 is a conceptual diagram illustrating a learning process for outputting process control information according to an embodiment of the present invention, and FIG. 4 is a schematic diagram of a configuration of a generating network 210 according to an embodiment of the present invention. It is a block diagram shown.

3 and 4, the generation network 210 of the present invention may further include a drift characteristic information generation unit 211, an FC characteristic information generation unit 212, and an output unit 213. I can. When the information 11 on the cumulative use time of parts in the current process and the target thickness value 12 for each area of the wafer are input to the FC feature information generator 212, the learned FC feature information generator 212 Outputs the control information, polishing time and polishing pressure value for each area. In order to reflect process changes to process control, the generation network 210 configures a drift characteristic information generation unit 211 in front of the FC characteristic information generation unit 212. Here, the drift characteristic information generation unit 211 of the present invention may be a convolution neural network (CNN). As each run proceeds, the drift characteristic information generation unit 211 receives data of the past processes that have been recently processed using a sliding window from the update unit 400 to provide a process control history. Information can be updated. After that, the FC characteristic information generation unit 212 receives the drift characteristic information output from the drift characteristic information generation unit 211 based on the updated process control history information, and performs competitive learning with the discrimination network 220. , When such learning is completed, process control information (control value) for controlling a part for polishing may be output.

The above-described contents will be described in more detail with reference to FIGS. 3 to 5 in order to explain the operation of each component. 5 is an enlarged view of FIG. 3 illustrating a drift characteristic information generation unit according to an embodiment of the present invention. Referring to FIG. 5, the drift feature information generation unit 211 according to an embodiment of the present invention further includes an encoding unit 510, a second feature map generator 520, and a third feature map generator 530. It can be configured.

The input unit 100 of the present invention inputs information 11 on the cumulative use time of parts, which are process parameter information, a target thickness value 12 for each area of the wafer, and process control history information and scale history information 13 Receive.

In addition, the encoding unit 510 of the drift characteristic information generation unit 211 according to an embodiment of the present invention includes sequence information of each of previously manufactured wafers, and process parameters for parts used when manufacturing the wafers. The first feature map 511 may be generated by using the information, the polishing time required to manufacture each of the previously manufactured wafers, and the process control history information 13 including the polishing pressure value.

The first feature map 511 according to an embodiment includes the process parameter information, polishing time, polishing pressure, and output thickness values (horizontal axis) for each region of the polished wafer according to each sequence information (vertical axis) for each wafer. As expressed, it may be a feature map expressed as a feature value between 0 and 1.

In addition, the second feature map generation unit 520 according to an embodiment of the present invention filters the first feature map 511 (convolution layer) using the first filter 521, A second feature map may be generated in consideration of correlation between process parameter information.

As an example, the second feature map generation unit 520 convolutionally filters the first feature map 511 using the first filter 521 having a size of 1 X 2, so that a second feature related to a process change between wafers You can create a map.

In addition, the third feature map generator 530 according to an embodiment of the present invention convolutionally filters the first feature map 511 using the second filter 531 (convolution layer), so that the same process parameters Based on the information, a third feature map may be generated in consideration of correlation between successive process processes.

As an example, the third feature map generator 530 convolutionally filters the first feature map 511 using a second filter 531 having a size of 2×1, Third feature maps may be generated.

The drift characteristic information generation unit 211, which is a CNN, has a structure in which a plurality of convolution layers and pooling layers are stacked. The drift feature information generation unit 211 generates a second feature map and a third feature map, which are convolutional layers, using the first filter and the second filter, and the pooling layers are in the second feature map and the third feature map. Select a representative feature value to reduce the size of each feature map. In this way, the drift characteristic information generation unit 211 generates drift characteristic information in consideration of process drift, and transmits it to the FC characteristic information generation unit 212. The drift feature information generated in this way is transmitted to the FC layer (fully connected layer) of the FC feature information generator 212 and is used for image recognition or classification.

That is, the process control information polished in the past n pieces as shown in FIG. 3 can be considered as a two-dimensional process image. The first filter 521 may generate a feature map in consideration of correlation between variables in the same process, while the second filter 531 may generate a feature map in consideration of correlation between successive processes.

Each of the second feature map and the third feature map may be generated by Equation 3 below.

는 1X2 크기의 제1 필터를 사용한 제2 특징 맵의 노드(i, j)이며,

는 2X1 크기의 제2 필터를 사용한 제3 특징 맵의 노드(i, j)이다. α는 활성 함수이고, H와 W는 입력 데이터의 열과 행의 크기이며, p와 q는 각 필터의 행과 열이고, s는 각 필터의 이동 간격인 stride length를 의미한다. 또한,

는 입력 데이터의 (si+p, sj+q) 좌표를 의미하고, W_p,q와 b는 각 필터의 p와 q에 위치하는 가중치(weight)와 바이어스(bias)를 의미한다. 또한, 위 <수학식2>에서의 W_drift는 컨볼루션 레이어들의 가중치들을 의미한다.here,

Is the node (i, j) of the second feature map using the first filter of 1X2 size,

Is the node (i, j) of the third feature map using the second filter of 2X1 size. α is the activation function, H and W are the sizes of the columns and rows of the input data, p and q are the rows and columns of each filter, and s is the stride length, which is the moving interval of each filter. Also,

Denotes the (si+p, sj+q) coordinates of the input data, and W _p,q and b denote the weights and biases located at p and q of each filter. In addition, W _drift in Equation 2 above means weights of convolutional layers.

As shown in FIG. 3, the FC characteristic information generation unit 212 according to an embodiment of the present invention accumulates the drift characteristic information output from the drift characteristic information generation unit 211 and the process parameter information from the input unit 100. Receive information on usage time and target thickness value for each area. Each of the input information may be encoded and converted into feature information (a feature vector or a feature matrix), and the converted feature information may be convolutionally filtered using at least convolutional layers to generate FC feature information. For example, the FC feature information may be generated as a feature vector or a feature matrix consisting of feature vectors, and the FC feature information includes a polishing time to be applied to each part by decoding the generated FC feature information and a pressure value for each region. The first process control information 15 may be output.

Then, the first process control information 15 output from the FC characteristic information generation unit 212 is transmitted to the armbyeol network 220.

As shown in FIG. 3, the discrimination network 220 includes the first process control information 15 and the second process control information applied to the actual wafer manufacturing process including the actual polishing time and pressure values for each region as training information. (14) can be entered.

The second process control information 14 is not separately shown in the drawings, but may be stored in a DB included in the process control information generating apparatus of the present invention, and the DB is the first process control information 15 in the generation network 210. When is outputted, the stored second process control information 14 can be transmitted to the discrimination network 220.

The second process control information 14 is actual information including the use time of the part when polishing each of the past wafers, the pressure value of the part, etc., and the input unit 100 received by the update unit 400 is It may be information that has been stored as.

Whenever the wafer process is completed, the control information (control value) of the process, the thickness value (output value) for each region of the polished wafer accordingly, and the usage time of the component are stored in the update unit 400.

Accordingly, the discrimination network 220 converts the first process control information 15 and the second process control information 14 into feature information, and converts the converted feature information into the first process using convolution layers. As the control information becomes 0 and the second process control information becomes 1, it is possible to determine each process control information.

That is, the discrimination network 220 may derive a value between 0 and 1 by receiving each process control information and discriminating for each process control information using internal convolutional layers. The generation network 210 may change and set filter coefficients (weights) of the inner layers in a later learning step so that the output value of the first process control information 15 from the discrimination network 220 converges to 1.

As described above, the learning unit 200 of the present invention composed of the generation network 210 and the discrimination network 220 uses the least square loss function based on the least squares method of LSGANs to backpropagation ( back propagation), and the loss function required for learning is as shown in Equation 4 below.

과

은 판별망(220)과 생성망(210)의 손실(loss)이며, p_d와 p_g는 제2 공정 제어정보(u_k)와 생성망(210)에 의해 생성된 제1 공정 제어정보(

)의 분포이다. 판별망(220)은 제2 공정 제어정보(u_k)에 대해 D(u_k)=1, 제1 공정 제어정보 D(

)=0이 되도록 학습한다. 반면, 생성망(210)은 제2 공정 제어정보(u_k)와 아주 유사한 정보를 생성하여 D(

)=1이 되도록 학습한다. 궁극적으로 학습이 성공적으로 완료되면 분포 p_g는 p_d와 같아지게 된다.here,

and

Is the loss of the discrimination network 220 and the generation network 210, and p _d and p _g are the second process control information (u _k ) and the first process control information generated by the generation network 210 (

) Is the distribution. The discrimination network 220 is D(u _k )=1 for the second process control information (u _k ), and the first process control information D(

Learn to become )=0. On the other hand, the generation network 210 generates information very similar to the second process control information (u _k ),

Learn to be )=1. Ultimately, upon successful completion of training, the distribution p _g becomes equal to p _d .

The reason why the process control information generation apparatus of the present invention uses the learning method of LSGANs is that the control performance is improved by making the distribution p _g in the multidimensional control space learned by the generation network 210 equal to the actual distribution p _d. to be.

Theoretically, the loss functions of the generation network 210 and the discriminant network 220 can converge through learning, but in reality, convergence is not guaranteed due to the structural complexity of the neural network. To solve this problem, the learning unit 200 of the present invention adjusts the filter coefficients (hyper-parameters) based on the Bayesian optimization model to change the structure of the model, so that learning may converge. Since the learning unit 200 of the present invention is based on the Bayesian optimization model (BO), it intelligently samples the solution for each iteration through iterative search, so filter coefficients (weighted values) that guarantee excellent model performance in a short time. ) Can be found.

<Equation 5> below is for an algorithm for adjusting the filter coefficients (hyper-parameter) based on the Bayesian optimization model (BO) in the learning unit 200 of the present invention.

Here, t denotes a hyper-parameter solution, T denotes a solution space, and c(t) denotes a cost function after completion of learning. T ^* in Equation 5 above can be defined as a hyper-parameter solution constituting the model with the lowest error, if the performance indicator of the model is an error.

The learning unit 200 of the present invention defines a network function c(t) _{LRC as} shown in Equation 6 below in order to apply the Bayesian optimization model BO.

와

는 에폭(i)에서 판별망(220)과 생성망(210)의 손실을 뜻하며, l은 LSGANs의 학습이 안정되는 에폭 시점을 의미한다. LSGANs을 학습시킬 때 초기 에폭에서는 d_loss와 g_loss의 진동이 심하기 때문에 c(t)_URC 값의 오류를 줄이기 위해 학습이 안정화되는 l 시점 이후의 에폭을 선택한다. 본 발명의 학습부의 전체 손실을 줄이고 d_loss와 g_loss의 적절한 균형을 맞추기 위해 d_loss와 g_loss의 합과 d_loss와 g_loss 의 절대 차이를 곱하고 에폭 수인 m-l을 나눈 평균값을 c(t)_URC로 정의한다. c(t)_URC의 값이 작을수록 생성망(210)이나 판별망(220)의 손실차이가 줄어든다는 의미로 생성망과 판별망의 어느 한쪽으로 치우치지 않는 상태에 도달한다고 볼 수 있다.Here, m denotes the number of epochs required for learning LSGANs, and in each epoch, the discriminant network 220 and the generation network 210 are alternately learned as many as the number of batches.

Wow

Denotes the loss of the discriminant network 220 and the generation network 210 in the epoch (i), and l denotes the epoch time at which the learning of LSGANs is stable. When learning LSGANs, the vibrations of d_loss and g_loss are severe in the initial epoch, so to reduce the error of the c(t) _URC value, the epoch after the point l when the learning is stabilized is selected. In order to reduce the total loss of the learning unit of the present invention and balance d_loss and g_loss, the average value obtained by multiplying the sum of d_loss and g_loss by the absolute difference between d_loss and g_loss and dividing the epoch number ml is defined as c(t) _URC . As the value of c(t) _URC decreases, the difference in loss between the generation network 210 and the discrimination network 220 decreases, and it can be seen that the state is not biased toward either the generation network or the discrimination network.

For solution(t), c(t) _URC is expensive because it can be observed only after learning is completed. Therefore, efficient sampling is required in the solution space. BO assumes that c(t) _URC has a Gaussian process prior. Initially, the BO configures the learning unit 200 by randomly selecting one solution, and evaluates c(t) _URC through the training information 14. The BO with observed data re-estimates c(t) _URC by calculating the posterior, and the next point to be sampled in the solution space through an acquisition function based on the updated c(t) _URC . Choose After that, sampling is repeated until the condition of completion of learning is satisfied to find the best solution (t ^* ).

Referring back to FIG. 2, accordingly, the control unit 300 according to an embodiment of the present invention receives the first process control information u ^k output from the generation network 210, and each of the parts for polishing the wafer. Control information (polishing time and pressure value for polishing) (y ^k ) is input to the controller 300, and at this time, the controller 300 may transmit control information (y ^k ) transmitted to the parts to the update unit 400. That is, whenever the wafer process is completed, the control information (control value) of the process may be stored in the update unit 400, and the thickness value (output value) for each region of the wafer polished based on the control information It may be stored in the update unit 400 until the use time of the part.

Accordingly, the update unit 400 according to an embodiment of the present invention may update the stored control information (run data) by transmitting the stored control information (run data) to the input unit 100 for subsequent learning by the learning unit 200.

Next, a process control device that controls a process in real time in an on-site wafer polishing process using the process control information generation unit learned through the process control information generation device described with reference to FIGS. 2 to 5 will be described. 6 is a block diagram schematically showing the configuration of a process control apparatus according to an embodiment of the present invention.

The process control apparatus 600 according to the embodiment of the present invention as shown in FIG. 6 includes an input unit 610, a process control information generation unit 620, a control unit 630, and an update unit 640. Can be configured. The input unit 610, the control unit 630, and the update unit 640 of the process control device 600 are substantially related to the technical operation of the input unit 100, the control unit 300, and the update unit 400 shown in FIG. Since it is similar, the process control information generation unit 620 and the control unit 630 will be mainly described below.

The process control information generation unit 620 according to an embodiment of the present invention receives target scale information related to the scale of regions belonging to the wafer to be polished, and pressure setting information and equipment of at least one equipment for polishing the wafer. It is possible to generate a process control feature value for on-site process control including time setting information related to the polishing time of.

In addition, the controller 630 may control a process condition of the equipment based on the process control feature value generated from the process control information generation unit 620.

The process control information generation unit 620 is set with filter coefficients of layers determined with an optimized weight through the learning process of the process control information generating apparatus of the present invention, receives real-time input values and generates process control feature values. , The control unit 630 may control the equipment based on the generated process control feature value with a pressure value for polishing and a polishing time.

Here, input values input to the process control information generation unit 620 may be the same as input information (values) input to the generation network of the process control information generation apparatus described above.

In addition, the process control information generation unit 620 of the present invention generates and outputs a process control feature value in the form of a feature vector or a feature matrix before being decoded with a pressure value for polishing, which is process control information, and a polishing time, and the controller 630 By decoding the process control feature value, the pressure value and the polishing time may be obtained, and the equipment may be controlled using the obtained pressure value and the polishing time. As another example, the process control feature value is generated to control the process. It may be implemented to decode the feature value and output the pressure value and the polishing time to the controller.

The performance of the R2R controller is measured based on a control error. Here, the control error refers to the absolute difference between the output of the learning unit and the actual output for each run. In the present invention, (SD) and mean absolute error (MAE) were calculated for each control variable to evaluate the performance of the models as shown in FIG. 7. In FIG. 7, p denotes control variables, n denotes test data size, MAE denotes how accurately the trained model predicts the actual value, and denotes the prediction accuracy and variability of the trained model. It is an indicator. Both indicators indicate that the lower the value is, the higher the performance of the control model is.

FIG. 8 is a result of an experiment to improve prediction performance by optimizing a hyper-parameter to BO for each of the two comparison models in FIG. 7. NNR2R(A) uses two neural networks, one as a VM model, and the other as an R2R controller. NNR2R(A) learns a VM model using the latest process data selected by a sliding window after each process is finished, and learns an R2R controller based on the learned VM model. Therefore, the performance of the R2R controller of NNR2R(A) largely depends on the performance of the VM model. In the present invention, in order to improve the performance of the VM model, the CNN for extracting the features of the process change applied to the learning unit is also applied to the VM model.

로 설정하였다. 여기서

은 학습 에폭의 수를 말하며 20회를 설정했고, l은 NNR2R의 학습이 안정화되는 에폭 시점을 의미하며 5회를 설정했다.

와

는 epoch i에서 hyper-parameters solution(t)을 사용한 VM과 R2R controller의 각각 mean squared error를 뜻한다. 본 발명의 학습부(B)의 경우, NNR2R(A)과 동일하게 BO 연속 횟수는 100회, m=20, l=5로 설정했다. 도8은 최적화 실험 결과를 보여준다. 도8과 같은 실시예에서 학습주기(learning rate)는 0.000001부터 0.01사이이며, CNN 특징맵의 개수는 4개에서 300개 사이이고, 각 FC 레이어(fully connected layer)의 은닉 노드의 개수는 10개에서 400개 사이이며, 풀링 레이어의 타입은 average pooling과 max pooling이다.In hyper-parameter optimization, the number of consecutive BOs is 100, and the network function of NNR2R(A) is

Was set to. here

Is the number of learning epochs, and 20 is set, and l is the epoch time when NNR2R learning is stabilized, and 5 times is set.

Wow

Means each mean squared error of VM and R2R controller using hyper-parameters solution(t) in epoch i. In the case of the learning unit (B) of the present invention, similarly to the NNR2R(A), the number of consecutive BOs was set to 100 times, m=20, l=5. 8 shows the results of the optimization experiment. In the embodiment shown in FIG. 8, the learning rate is between 0.000001 and 0.01, the number of CNN feature maps is between 4 and 300, and the number of hidden nodes in each FC layer is 10. It is between 400 and the pooling layer types are average pooling and max pooling.

9 is a diagram showing the loss of the generation network for each learning epoch and the loss of the discriminant network after optimizing with the algorithm (a) of randomly finding hyper-parameters of the learning unit and the algorithm (b) of the BO method. Fig. 9(a) shows a phenomenon in which the loss of the generation network and the discriminant network continuously progresses during learning and does not converge. This phenomenon is a typical problem that occurs in the learning process of GANs. Since the discriminant network and the generation network learn alternately, they cannot converge to the global solution in the process of deceiving each other. Therefore, the generation network is not suitable for use as an R2R controller because it cannot learn all the distribution of actual data. 9B is a graph of the loss of the generation network and the discriminant network after optimizing the hyper-parameters to BO, and shows that the generation network and the discriminant network are stably learned without biasing the loss to either side. This stable loss distribution secures the reproducibility of the R2R controller and enables fast learning of new data.

10 is a diagram showing an experiment result for hyper-parameter optimization. The performance result of the learning unit after applying the hyper-parameter optimization algorithm to 1,450 experimental data is shown. In FIG. 10, "Random" and "BO" are performance measurements when the random values and BO results of FIG. 8 are respectively used for hyper-parameters. Looking at the results of eight control parameters, it was confirmed that the performance improved significantly after hyper-parameter optimization. In terms of MAE, each parameter showed improved performance by about 26.4% and up to 63.8%, and in terms of +SD, there was a performance improvement from about 31.2% to 80.3%. From the results of FIGS. 8 and 9, it can be seen that the hyper-parameter solution found by the BO enables stable learning that balances the capabilities of the generating network and the discriminant network of the learning unit. In addition, it can be seen that the generation network applying the BO result can be used as an R2R controller with low control error and variability.

)의 절대 제어 오차의 합과 압력값들(

)의 절대 제어 오차의 합이며, 오른쪽 y축은 부품들의 정규화된 수명을 나타내며 1의 값에 도달하면 교체에 의해 0에서부터 증가함을 확인할 수 있다. 11 is a result of expressing a control error of an experimental wafer process in a run sequence plot. This plot contains one process shift caused by replacement of CMP equipment components. The y-axis on the left in (a) and (b) of Fig. 11 are the polishing times (

) Of the sum of absolute control errors and pressure values (

) Is the sum of the absolute control errors, and the y-axis on the right represents the normalized life of the parts, and when it reaches the value of 1, it can be seen that it increases from 0 by replacement.

Even if all the components constituting the embodiments of the present invention described above are described as being combined into one or operating in combination, the present invention is not necessarily limited to these embodiments. That is, within the scope of the object of the present invention, all the constituent elements may be selectively combined and operated in one or more. In addition, although all of the components can be implemented as one independent hardware, a program module that performs some or all functions combined in one or more hardware by selectively combining some or all of the components. It may be implemented as a computer program having In addition, such a computer program is stored in a computer readable media such as a USB memory, a CD disk, a flash memory, etc., and is read and executed by a computer, thereby implementing an embodiment of the present invention. The recording medium of the computer program may include a magnetic recording medium or an optical recording medium.

The above description is merely illustrative of the technical idea of the present invention, and those of ordinary skill in the technical field to which the present invention belongs can make various modifications, changes, and substitutions within the scope not departing from the essential characteristics of the present invention. will be. Accordingly, the embodiments disclosed in the present invention and the accompanying drawings are not intended to limit the technical idea of the present invention, but to explain, and the scope of the technical idea of the present invention is not limited by these embodiments and the accompanying drawings. . The scope of protection of the present invention should be interpreted by the following claims, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the present invention.

Claims (17)

An input unit for receiving target scale information related to the scale of regions belonging to the substrate to be manufactured and process parameter information for at least one device for manufacturing the substrate;
A generation network that generates first process control information necessary for a process for manufacturing the substrate by using a plurality of layers based on the target scale information and the process parameter information; And
A process control information generating apparatus required for a substrate manufacturing process comprising a discrimination network for deriving an evaluation result for the first process control information by learning based on the first process control information and second process control information, which is training data. . The method of claim 1,
The scale information is thickness information of each area belonging to the substrate,
The process parameter information is information on the cumulative use time of the equipment,
The first process control information and the second process control information includes a polishing time and a polishing pressure value required in a single operation process. The method of claim 1, wherein the generation network,
A first feature map is generated using previously stored process control history information related to the process control of the substrate, and the first feature map is subjected to convolutional filtering using filters of different scales. The apparatus for generating process control information necessary for a substrate manufacturing process, further comprising: a drift characteristic information generation unit that generates information. The method of claim 3, wherein the generation network,
A FC feature information generator configured to receive the drift feature information, the scale information, and the process parameter information and perform convolution filtering using at least one convolution layer to generate FC feature information; and
A process control information generating apparatus required for a substrate manufacturing process, characterized in that decoding the FC characteristic information and outputting the first process control information. The method of claim 3, wherein the drift characteristic information generation unit,
A second feature map generator configured to generate a second feature map in consideration of a correlation between the process parameter information in the same process by filtering the first feature map using a first filter; And
By filtering the first feature map using a second filter having a scale different from that of the first filter, a third feature map is generated based on the same process parameter information in consideration of correlation between successive process processes. The apparatus for generating process control information required for a substrate manufacturing process, further comprising: a feature map generator. The method of claim 1,
And an update unit for receiving the second process control information, process control history information in a previous process of manufacturing the substrate, and scale history information for each area of the substrate manufactured in the previous process and transmitting the received information to the input unit. An apparatus for generating process control information necessary for a substrate manufacturing process. The method of claim 1,
Wherein the generation network changes a filter coefficient of at least one of the plurality of layers in consideration of an evaluation result derived from the discrimination network. The method of claim 1,
And the generation network and the discrimination network use a least square loss function based on a least squares method. The method of claim 1,
The generation network and the discrimination network change and set filter coefficients to be changed based on a Bayesian optimization model. Receiving target scale information related to the scale of regions belonging to the substrate to be manufactured and process parameter information for at least one equipment for manufacturing the substrate;
Generating first process control information necessary for a process for manufacturing the substrate by using a plurality of layers based on the target scale information and the process parameter information; And
Deriving an evaluation result for the first process control information by learning based on the first process control information and second process control information, which is training data, and generating process control information necessary for a substrate manufacturing process. The method of claim 10,
The scale information is thickness information of each area belonging to the substrate,
The process parameter information is information on the cumulative use time of the equipment,
The first process control information and the second process control information include a polishing time and a polishing pressure value required in a single operation process. The method of claim 10, wherein the generation network,
A first feature map is generated using previously stored process control history information related to the process control of the substrate, and the first feature map is subjected to convolutional filtering using filters of different scales. Generating information; And
Receiving the drift characteristic information, the scale information, and the process parameter information and performing convolution filtering using at least one convolution layer to generate FC characteristic information for the first process control information; further comprising ,
The method for generating process control information required for a substrate manufacturing process, characterized in that the FC characteristic information is decoded and the first process control information is output. Receive target scale information related to the scale of regions belonging to the substrate to be polished, and perform on-site process control including pressure setting information of at least one equipment for polishing the substrate and time setting information related to the polishing time of the equipment. A process control information generation unit that generates a process control feature value for; And
Process control device comprising; a control unit for controlling the process condition of the equipment based on the process control characteristic value. The method of claim 13,
The control unit transfers the process conditions of the controlled equipment and the accumulated use time of the equipment to the process control information generation unit,
The target scale information is information on a target thickness of each region belonging to the substrate. The method of claim 13, wherein the process control information generation unit,
A first feature map is generated using previously stored process control history information related to the process control of the substrate, and the first feature map is subjected to convolutional filtering using filters of different scales. A drift feature information generation unit that generates information; And
An FC feature information generator configured to receive the drift feature information, the scale information, and the process parameter information and perform convolution filtering using at least one convolution layer to generate FC feature information for the first process control information; Including more,
The process control device, characterized in that for generating the process control information by decoding the FC characteristic information. The method of claim 15, wherein the drift characteristic information generation unit,
A second feature map generator configured to generate a second feature map in consideration of a correlation between the process parameter information in the same operation process by filtering the first feature map using a first filter; And
By filtering the first feature map using a second filter having a scale different from that of the first filter, a third feature map that considers the relationship between successive operation processes based on the same factory parameter information is generated. The process control device further comprising a; feature map generator. A computer program stored in a computer-readable recording medium in order to execute the process control information generating method according to any one of claims 10 to 12 on a computer.

Images