Wafer thin film property real-time prediction device and method

Abstract

A wafer thin film property real-time prediction device according to an embodiment comprises a memory for storing at least one command, and a processor, wherein: as the at least one command is executed by the processor, process data that define a deposition process of a wafer is acquired; as the acquired process data is provided to a pre-trained thin film property prediction model, thin film property for each point in the wafer is predicted in real time on a cycle basis; the point of the wafer, for which the thin film property is predicted, includes an edge, a center, and a middle of the wafer, the middle being a region between the edge and the center; and the thin film properties respectively predicted in the edge, the center, and the middle have a predetermined co-relation therebetween.

Description

Device and method for real-time prediction of wafer thin film characteristics

[0001] The present disclosure relates to an apparatus and method for predicting wafer thin film characteristics in real time.

[0002] Generally, in semiconductor manufacturing processes, Statistical Process Control (SPC) is utilized to monitor process variables (e.g., temperature, pressure, gas flow rate), and corrective action is taken only after the data is analyzed if an anomaly is detected. However, since this approach analyzes data after the process is completed, it is currently difficult to detect and respond to anomalies in real time.

[0003] In addition, conventional SPC methods often estimated the overall state of the wafer by measuring thickness or sheet resistance at specific points (e.g., sampling at only 13 points). However, the state estimated in this way does not accurately reflect the uniformity across the wafer. In particular, there are limitations in effectively managing skew between the center, edge, and the middle between the center and edge.

[0004] Furthermore, given the vast scale of data generated in the semiconductor manufacturing process, there is a problem in that existing SPC methods cannot fully utilize this data. For example, while the semiconductor manufacturing process collects data in real time from numerous sensors, existing SPC methods analyze this data using simple statistical methods, failing to properly reflect the complex interactions between process variables. Consequently, adopting existing SPC methods results in lower prediction accuracy and longer response times after anomalies occur, which can ultimately lead to a decline in yield and production efficiency.

[0005] The problem to be solved according to one embodiment includes predicting the thin film property at each point of the wafer in real time.

[0006] In addition, the aforementioned task may include providing the aforementioned prediction without actual measurement results regarding the thin film characteristics of the wafer.

[0007] Additionally, the aforementioned task may include ensuring that the deposition process is immediately stopped when an abnormality is detected during the wafer deposition process.

[0008] However, the aforementioned problems are not limited to those mentioned above, and other problems not mentioned but intended to be solved will be clearly understood by those skilled in the art to which the present invention pertains from the description below.

[0009] A wafer thin film property real-time prediction device according to a first embodiment includes a memory storing at least one instruction; and a processor. By executing the at least one instruction by the processor, process data defining a deposition process of a wafer is obtained, and the obtained process data is provided to a pre-trained thin film property prediction model. Consequently, the thin film property for each point within the wafer is predicted in real-time on a cycle basis. The points within the wafer where the thin film property is predicted include the edge of the wafer, the center, and the middle, which is the region between the edge and the center. The thin film properties predicted at each of the edge, the center, and the middle each have a predetermined correlation with one another.

[0010] In addition, the process data may include at least one of the wafer temperature, pressure, stable time, stage / shower head gap, gas flow rate, source time, purge time, reactant time, and number of cycles.

[0011] In addition, at least some of the thin film characteristics predicted at each of the above edges, the above centers, and the above intermediates can be post-corrected dependently on the above correlation.

[0012] In addition, the above-mentioned thin film characteristic prediction model may be trained by augmenting pre-prepared training data using a domain knowledge-based data augmentation technique and utilizing the augmented results.

[0013] In addition, the domain knowledge used in the above data augmentation technique may include at least one of Atomic Layer Deposition (ALD) reaction kinetics, growth mechanism, and precursor.

[0014] In addition, the training data prepared in advance includes measurement results for thin film characteristics at each point of a predetermined wafer for each cycle, and the augmented result may be obtained by dividing the measurement results by cycle and then augmenting each of the divided results.

[0015] In addition, real-time prediction of the thin film characteristics of the wafer can be performed without actual measurement results of the thin film characteristics of the wafer.

[0016] In addition, the thin film characteristics predicted for each of the above points may include at least one of physical characteristics, electrical characteristics, and optical characteristics. In this case, the physical characteristics may include deposition thickness, the electrical characteristics may include at least one of sheet resistance, capacitance, and leakage current, and the optical characteristics may include at least one of reflectance and refractive index.

[0017] In addition, the deposition process on the wafer currently in progress can be controlled to be stopped based on the predicted result by executing at least one of the above instructions by the processor.

[0018] A real-time wafer thin film characteristic prediction method according to a second embodiment is performed by a real-time wafer thin film characteristic prediction device and includes the step of obtaining process data defining a deposition process of a wafer; and the step of providing the obtained process data to a pre-trained thin film characteristic prediction model to predict thin film characteristics for each point within the wafer in real-time on a cycle basis, wherein the points within the wafer where thin film characteristics are predicted include the edge, the center, and the middle region between the edge and the center of the wafer, and the thin film characteristics predicted at each of the edge, the center, and the middle may have a certain co-relation with each other.

[0019] In addition, the process data may include at least one of the wafer temperature, pressure, stable time, stage / shower head gap, gas flow rate, source time, purge time, reactant time, and number of cycles.

[0020] In addition, at least some of the thin film characteristics predicted at each of the above edges, the above centers, and the above intermediates can be post-corrected dependently on the above correlation.

[0021] In addition, the above-mentioned thin film characteristic prediction model may be trained by augmenting pre-prepared training data using a domain knowledge-based data augmentation technique and utilizing the augmented results.

[0022] In addition, the domain knowledge used in the above data augmentation technique may include at least one of Atomic Layer Deposition (ALD) reaction kinetics, growth mechanism, and precursor, and additionally may include Physical Vapor Deposition, Chemical Vapor Deposition, Sol-Gel Deposition, spin coating, or spray pyrolysis, but is not limited thereto.

[0023] In addition, the training data prepared in advance includes measurement results for thin film characteristics at each point of a predetermined wafer for each cycle, and the augmented result may be obtained by dividing the measurement results by cycle and then augmenting each of the divided results.

[0024] In addition, real-time prediction of the thin film characteristics of the wafer can be performed without actual measurement results of the thin film characteristics of the wafer.

[0025] In addition, the thin film characteristics predicted for each of the above points may include at least one of physical characteristics, electrical characteristics, and optical characteristics. In this case, the physical characteristics may include deposition thickness, the electrical characteristics may include at least one of sheet resistance, capacitance, and leakage current, and the optical characteristics may include at least one of reflectance and refractive index.

[0026] In addition, based on the above predicted results, it may further include a step of controlling the deposition process on the currently ongoing wafer to be stopped.

[0027] A non-transient computer-readable recording medium according to a third embodiment stores a computer program programmed to perform each step included in the method described above.

[0028] A computer program according to the fourth embodiment is stored on a non-transient computer-readable recording medium and may be programmed to perform each step included in the method described above.

[0029] According to one embodiment, at the end of each cycle in the wafer deposition process, thin film characteristics such as deposition thickness or sheet resistance in that cycle can be predicted in real time without actual measurement of the wafer. Therefore, if an anomaly is detected in a specific cycle, the process can be stopped immediately, thereby minimizing wafer loss in the event of an anomaly.

[0030] In addition, unlike conventional methods, thin film characteristics can be predicted for all desired points on the wafer, such as the edge, center, or middle, as well as for each of multiple points (e.g., 49 points). This allows for verification of process uniformity, which enables the optimization of process data such as process parameters and allows for the analysis of local variations (edge ​​effects) on the wafer. Furthermore, since it is possible to analyze which regions consistently exhibit thin film characteristics, issues with process equipment or errors in process settings causing these issues can be detected early, which can lead to improved yield.

[0031] FIG. 1 is a diagram illustrating the concept of a wafer thin film characteristic real-time prediction device according to one embodiment.

[0032] Figure 2 is a conceptual block diagram showing the configuration of a real-time wafer thin film characteristic prediction device illustrated in Figure 1.

[0033] Figure 3 is a conceptual diagram of the architecture of a deep learning model.

[0034] FIG. 4 is a flowchart of a method for real-time prediction of wafer thin film characteristics according to one embodiment.

[0035] Figure 5 is a diagram showing the accuracy of the thin film characteristics predicted according to one embodiment.

[0036] The advantages and features of the present invention and the methods for achieving them will become clear by referring to the embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below but can be implemented in various different forms. These embodiments are provided merely to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the invention, and the present invention is defined only by the scope of the claims.

[0037] In describing the embodiments of the present invention, specific descriptions of known functions or configurations will be omitted if it is determined that such detailed descriptions could unnecessarily obscure the essence of the invention. Furthermore, the terms described below are defined in consideration of their functions in the embodiments of the present invention, and these definitions may vary depending on the intentions or practices of the user or operator. Therefore, such definitions should be based on the content throughout this specification.

[0038] FIG. 1 is a diagram illustrating the concept of a wafer thin film property real-time prediction device according to one embodiment. Referring to FIG. 1, a thin film property prediction device (100) (hereinafter referred to as the 'prediction device (100)') is provided. This prediction device (100) is provided to predict the thin film properties of a wafer in real time.

[0039] Specifically, when the prediction device (100) receives process data for a wafer, it is configured to use this data to output thin film characteristics on the wafer.

[0040] Here, the process data provided to the prediction device (100) includes at least one of temperature, pressure, stable time, stage / shower head gap, gas flow rate, source time, and purge time as shown in FIG. 1, and may include reactant time and number of deposition cycles, although not shown in FIG. 1. At this time, the gas flow rate may include Carrier Ar, Dilution Ar (Source), Reactant, Dilution Ar (Reactant), or Backside Ar, etc.

[0041] There are various ways in which process data is acquired. For example, the aforementioned temperature or pressure may be values ​​detected by sensors, and the gas flow rate, precursor injection time, or distance between the substrate and the showerhead may be obtained from values ​​designed for the process and actually input into the process, but are not limited thereto.

[0042] Meanwhile, such process data is acquired in the form of log data whenever each cycle of the deposition process ends. For example, process data for 3 cycles can be acquired when 3 cycles are completed. Therefore, if the deposition process consists of 100 cycles, process data for the corresponding cycle can be acquired when each of the 100 cycles is completed. At this time, the process data may have the same value for each cycle depending on its type, but it may also have different values.

[0043] The aforementioned process data may be provided to the prediction device (100) on a cycle-by-cycle basis whenever each cycle ends, that is, a value corresponding to the ended cycle. For example, process data for one cycle may be provided to the prediction device (100) when one cycle ends, and process data for two cycles may be provided to the prediction device (100) when two cycles end.

[0044] Next, let us look at the output. The prediction device (100) is configured to output thin film characteristics as illustrated in FIG. 1. These thin film characteristics may include at least one of physical characteristics, electrical characteristics, and optical characteristics. In this case, the physical characteristics may include deposition thickness, the electrical characteristics may include at least one of sheet resistance, capacitance, and leakage current, and the optical characteristics may include at least one of reflectance and refractive index, but are not limited thereto.

[0045] At this time, these thin film characteristics are predicted on a point-by-point basis. For example, thin film characteristics can be predicted on a point-by-point basis for each of the edge, center, and middle regions between the edge and the center of the wafer, or they can be predicted for each of the 49 points (the number of points is merely an example).

[0046] Meanwhile, thin film characteristics can be predicted on a cycle basis. For example, thin film characteristics for each cycle can be predicted on a cycle basis, such as thin film characteristics for 1 cycle to thin film characteristics for 100 cycles. This corresponds to process data being provided to the prediction device (100) on a cycle basis. Specifically, for example, thin film characteristics for 1 cycle can be predicted for process data for 1 cycle, and thin film characteristics for 2 cycles can be predicted for process data for 2 cycles.

[0047] Based on this, thin film characteristics such as deposition thickness or sheet resistance at the end of each cycle in the wafer deposition process can be predicted in real time without actual measurements of the wafer. Therefore, since the process can be stopped immediately if an anomaly is detected in a specific cycle, wafer loss in the event of an anomaly can be minimized.

[0048] In addition, unlike conventional methods, thin film characteristics can be predicted for all desired points on the wafer, such as the edge, center, or middle, as well as for each of multiple points (e.g., 49 points). This allows for verification of process uniformity, which enables the optimization of process data such as process parameters and allows for the analysis of local variations (edge ​​effects) on the wafer. Furthermore, since it is possible to analyze which regions consistently exhibit thin film characteristics, issues with process equipment or errors in process settings causing these issues can be detected early, which can lead to improved yield.

[0049] Below, we will examine this prediction device (100) in more detail.

[0050] FIG. 2 is an exemplary configuration diagram of a prediction device (100) according to one embodiment. Referring to FIG. 2, the prediction device (100) includes a communication unit (110), a memory (120), and a processor (130). However, the configuration diagram shown in FIG. 2 is merely exemplary, and the concept of the present invention is not limited by the configuration diagram shown in FIG. 2. For example, the prediction device (100) may include at least one configuration not shown in FIG. 2, or may not include at least one of the configurations shown in FIG. 2.

[0051] The communication unit (110) can be implemented by various types of wired or wireless communication modules. The prediction device (100) can communicate with an external terminal or server, such as a deposition equipment that performs a wafer deposition process or a manager's terminal that manages the wafer process, through this communication unit (110).

[0052] Memory (120) can be implemented by a medium that stores information. Such a medium may be at least one type of storage medium among flash memory type, hard disk type, multimedia card micro type, card type memory (e.g., SD or XD memory), RAM (Random Access Memory, RAM), SRAM (Static Random Access Memory), ROM (Read-Only Memory, ROM), EEPROM (Electrically Erasable Programmable Read-Only Memory), PROM (Programmable Read-Only Memory), magnetic memory, magnetic disk, and optical disk, but is not limited thereto.

[0053] Various types of information can be stored in this memory (120). For example, process data obtained by the prediction device (100) from the deposition equipment via the communication unit (110) can be stored in the memory (120) on a cycle-by-cycle basis.

[0054] Meanwhile, various types of modules or models may be implemented in the memory (120). When these modules or models are executed by the processor (130) described later, the intended function is performed. At this time, at least one of these modules or models may be implemented based on rules or an artificial intelligence network. Below, we will examine the artificial intelligence network, etc.

[0055] In this specification, a model or module may refer to any form of computer program that operates based on a network function, an artificial neural network, and / or a neural network. Throughout this specification, the terms model, neural network, network function, and neural network may be used interchangeably. A neural network is formed in which one or more nodes are interconnected through one or more links to form input and output node relationships within the neural network. The characteristics of a neural network may be determined by the number of nodes and links within the neural network, the relationships between the nodes and links, and the values ​​of the weights assigned to each link. A neural network may be composed of a set of one or more nodes. A subset of nodes constituting a neural network may form a layer.

[0056] Artificial neural networks or neural networks may include machine learning-based or deep learning-based models. For example, they may include decision tree-based models such as Randomforest, LightBGM, XGBoost, CatBoost, Linear Regression, Ridge Regression, Deep learning models such as MLP, CNN, RNNs, or Machine Learning models such as Bayesian Regression and Gaussian Process Regression, but are not limited thereto.

[0057] Here, deep learning is also referred to as deep neural networks, which we will examine below.

[0058] A deep neural network (DNN) may refer to a neural network that includes multiple hidden layers in addition to an input layer and an output layer, and as illustrated exemplarily in FIG. 3, the intermediate hidden layer in a deep neural network is composed of one or more, preferably two or more.

[0059] These deep neural networks may include convolutional neural networks (CNN), vision transformers, recurrent neural networks (RNN), Long Short Term Memory (LSTM) networks, Generative Pre-trained Transformer (GPT), auto encoders, Generative Adversarial Networks (GAN), restricted Boltzmann machines (RBM), deep belief networks (DBN), Q networks, U networks, Siamese networks, Generative Adversarial Networks (GAN), transformers, etc.

[0060] Alternatively, according to an embodiment, the deep neural network may be a model trained using a transfer learning method. Here, transfer learning refers to a learning method in which a large amount of unlabeled training data is pre-trained using a semi-supervised or self-learning method to obtain a pre-trained model (or base part) having a first task, and a target model is implemented by training labeled training data using a supervised learning method to fine-tune the pre-trained model to suit a second task. Examples of models trained using such a transfer learning method include BERT (Bidirectional Encoder Representations from Transformers), but are not limited thereto.

[0061] Meanwhile, the model disclosed in one embodiment may borrow at least a part of a transformer. The transformer may be composed of an encoder that encodes embedded data and a decoder that decodes the encoded data. The transformer may have a structure that receives a series of data and outputs a series of data of different types after undergoing encoding and decoding steps. In one embodiment, the series of data may be processed into a form that the transformer can compute. The process of processing the series of data into a form that the transformer can compute may include an embedding process. Expressions such as data token, embedding vector, embedding token, etc., may refer to data embedded in a form that the transformer can process.

[0062] To encode and decode a series of data, the encoders and decoders within the transformer can be processed using an attention algorithm. An attention algorithm can refer to an algorithm that calculates the similarity between one or more keys for a given query, applies this similarity to the values ​​corresponding to each key, and then calculates an attention value by performing a weighted sum of the similarity-applied values.

[0063] Various types of attention algorithms can be classified depending on how the query, key, and value are configured. For example, if attention is calculated by setting the query, key, and value identically, this can be referred to as a self-attention algorithm. If attention is calculated by reducing the dimensionality of embedding vectors to process a series of input data in parallel and determining an individual attention head for each partitioned embedding vector, this can be referred to as a multi-head attention algorithm.

[0064] In one embodiment, the transformer may be composed of modules that perform a plurality of multi-head self-attention algorithms or multi-head encoder-decoder algorithms. In one embodiment, the transformer may also include additional components other than attention algorithms, such as embeddings, normalization, and softmax. A method for constructing the transformer using an attention algorithm may include the method disclosed in Vaswani et al., Attention Is All You Need, 2017 NIPS, which is incorporated herein by reference.

[0065] A transformer can be applied to various data domains, such as embedded natural language, segmented image data, and audio waveforms, to convert a series of input data into a series of output data. To convert data with various data domains into a series of data that can be input to the transformer, the transformer can embed the data. The transformer can process additional data that represents the relative positional or phase relationships between the series of input data. Alternatively, the series of input data may be embedded by additionally reflecting vectors that represent the relative positional or phase relationships between the input data. In one example, the relative positional relationships between the series of input data may include, but are not limited to, word order within a natural language sentence, the relative positional relationships of each segmented image, and the temporal order of segmented audio waveforms. The process of adding information that represents the relative positional or phase relationships between the series of input data may be referred to as positional encoding.

[0066] Referring again to FIG. 2, let's look at the processor (130).

[0067] A processor (130) according to one embodiment can perform technical features according to embodiments of the present disclosure to be described below by executing at least one instruction stored in memory (120). In one embodiment, the processor (130) may be composed of at least one core and may include a processor for data analysis and / or processing, such as a central processing unit (CPU), a general purpose graphics processing unit (GPGPU), or a tensor processing unit (TPU) of the prediction device (100).

[0068] Specifically, this processor (130) can train a neural network or model designed in a machine learning or deep learning manner. To this end, the processor (130) can perform calculations for training the neural network, such as processing input data for training, extracting features from input data, calculating errors, and updating the weights of the neural network using backpropagation.

[0069] Additionally, the processor (130) may perform inference for a specific purpose using a model implemented in an artificial neural network manner.

[0070] Hereinafter, we will examine the various operations or functions that the prediction device (100) can perform by executing at least one instruction stored in memory (120) by the processor (130).

[0071] First, the processor (130) can control the communication unit (110). Through this, the prediction device (100) can obtain information by communicating with a manager terminal or a deposition device through the communication unit (110).

[0072] Additionally, the processor (130) can read the aforementioned data or instructions stored in the memory (120) and write new data or instructions to the memory (120). Additionally, the processor (130) can modify or delete data or instructions that have already been written. Below, we will examine this in more detail.

[0073] By executing at least one instruction stored in memory (120) by the processor (130), the prediction device (100) obtains process data that defines the deposition process of the wafer. Here, "define" means that the process data reflects what specifications the deposition process will or has been performed under, as well as what the values ​​of various parameters were during or after the deposition process. Meanwhile, since it has already been explained what types of process data exist, how each type is acquired or obtained, and that process data is acquired on a cycle-by-cycle basis whenever each cycle ends, the explanation regarding this will be omitted.

[0074] Meanwhile, by executing at least one instruction stored in memory (120) by the processor (130), the prediction device (100) can pre-process the previously obtained process data.

[0075] In the preprocessing stage, a process of highlighting can be performed so that a specific machine learning or deep learning model can better understand the process data or effectively capture its features. For this reason, one of the operations performed during the preprocessing stage can be referred to as feature extraction. For instance, since temperature, pressure, or settling time in the process data are numerical values, a preprocessing process such as min-max scaling or normalization can be performed to convert them into values ​​within a specific range.

[0076] In addition, missing values ​​or outliers included in the process data may be handled during the preprocessing stage. There are various processing methods; for example, in the case of missing values ​​or outliers, the rows containing the corresponding values ​​may be removed or replaced with the average value.

[0077] Next, by executing at least one instruction stored in memory (120) by the processor (130), the process data described above is provided to the thin film characteristic prediction model in the prediction device (100). At this time, as described above, the data may be provided on a cycle-by-cycle basis whenever each cycle of the thin film process ends, or alternatively, it may be provided on a cycle-by-cycle basis after all cycles have ended.

[0078] Meanwhile, the aforementioned thin film characteristic prediction model is implemented in the form of a machine learning or deep learning model. Specifically, this thin film characteristic prediction model is stored in the memory (120) of the prediction device (100) shown in FIG. 2 and is executed by a processor (130).

[0079] In addition, models for predicting thin film characteristics can be implemented in various forms. For example, they can be implemented as decision tree-based models such as Randomforest, LightBGM, XGBoost, CatBoost, Linear Regression, and Ridge Regression; deep learning models such as MLP, CNN, and RNN; or machine learning models such as Bayesian Regression and Gaussian Process Regression; or alternatively, they can be implemented as rule-based models or modules. Specific methodologies for training models for predicting thin film characteristics will be described later.

[0080] As illustrated in FIG. 1, the process data received by the prediction device (100) on a cycle-by-cycle basis is provided to a thin film characteristic prediction model on a cycle-by-cycle basis. Accordingly, the thin film characteristic prediction model predicts the thin film characteristics within the corresponding wafer on a cycle-by-cycle basis, and consequently, the predicted result is provided externally by the prediction device (100).

[0081] Let us examine the thin film properties, which are predicted values, in more detail below. Thin film properties may include at least one of physical properties, electrical properties, and optical properties. In this case, physical properties may include deposition thickness; electrical properties may include at least one of sheet resistance, capacitance, and leakage current; and optical properties may include at least one of reflectance and refractive index, but are not limited thereto.

[0082] At this time, these thin film characteristics are predicted point by point. For example, they can be predicted point by point for each of the edge, center, and middle regions between the edge and the center of the wafer, or they can be predicted for each of the 49 points (the number of points is merely an example).

[0083] Meanwhile, thin film characteristics can be predicted on a cycle basis. For example, thin film characteristics for each cycle can be predicted on a cycle basis, such as thin film characteristics for 1 cycle to thin film characteristics for 100 cycles. This corresponds to process data being provided to the prediction device (100) on a cycle basis. Specifically, to look at an example, when process data for 1 cycle is provided to the prediction device (100), thin film characteristics for 1 cycle can be predicted, and when process data for 2 cycles is provided to the prediction device (100), thin film characteristics for 2 cycles can be predicted.

[0084] Based on this, one of the predicted thin film characteristics, 'thin film thickness per cycle,' is expressed as GPC (Growth Per Cycle). Therefore, the total thickness of the wafer when thin film deposition is completed can be expressed by the following mathematical formula 1, but is not limited thereto.

[0085]

[0086] [Mathematical Formula 1]

[0087] Total Thickness = ∑ GPC(Growth per Cycle) = Numbers of Cycles * GPC

[0088]

[0089] Next, another predicted thin film characteristic, 'sheet resistance per cycle,' is expressed as Thickness Per Cycle. Therefore, the sheet resistance of the wafer when thin film deposition is completed, i.e., resistance per unit area (Sheet Resistance), can be expressed by the following mathematical formula 2, but is not limited thereto.

[0090]

[0091] [Mathematical Formula 2]

[0092] Sheet Resistance = Total Sheet Resistance * Total Thickness / Thickness per Cycle

[0093]

[0094] Here, Total Sheet Resistance refers to the sheet resistance of the entire layer, Total Thickness refers to the total thickness of the wafer, and thickness per Cycle refers to the thickness deposited in one cycle.

[0095] Meanwhile, as previously mentioned, these thin film characteristics are predicted on a point-by-point basis. Here, if each point includes, for example, the edge, middle, and center, a certain correlation exists between them. First, the deposition thicknesses of the edge, middle, and center, respectively, generally exhibit the following characteristics.

[0096]

[0097] - Center: Generally, the thickness is the most uniform and tends to be close to the target value.

[0098] - Middle: Can vary depending on gas flow, temperature distribution, etc. in the deposition process.

[0099] - Edge: Thickness and sheet resistance may vary significantly due to non-uniform gas flow, changes in deposition rate, and the influence of the wafer holder.

[0100]

[0101] Since the aforementioned characteristics may follow specific mathematical patterns such as polynomials, Gaussian distributions, or sine functions, if mathematical patterns reflecting the characteristics of each point are utilized, the correlation between the thin film characteristics of each point can be derived, and an example of such a correlation may be described as follows.

[0102]

[0103] - It exhibits a pattern where the median is the most stable and large deviations occur at the edges.

[0104] - Changes in thickness at the edges are linearly or non-linearly related to changes in thickness in the center and middle regions.

[0105] - If the film thickness tends to thin at the edges, it is highly likely that this pattern will appear repeatedly.

[0106]

[0107] In addition, by executing at least one instruction stored in memory (120) by the processor (130), the prediction device (100) can post-process the predicted thin film characteristics by point by taking into account this correlation. The post-processing method is diverse, and one or more of the following methods may be employed, but are not limited thereto.

[0108]

[0109] (1) Interpolation

[0110] A method for predicting the value of a specific point based on data measured at multiple points on a wafer.

[0111] - Linear interpolation: Correction based on the assumption of a simple linear relationship.

[0112] - Polynomial fitting: Applying a polynomial model when data exhibits a curved shape.

[0113] (2) Regression Analysis

[0114] Learn measurement data from specific points (edge, center, middle) to derive a relationship equation, and correct based on the predicted value.

[0115] - Polynomial Regression: Correction based on the assumption that the wafer's sheet resistance or thickness follows a specific function pattern.

[0116] - Gaussian Process Regression (GPR): Predicts changes on the wafer surface by smoothing them.

[0117] (3) Application of machine learning models

[0118] A method of learning the spatial variation pattern of a wafer using deep learning or machine learning techniques, and then correcting the value at a specific point.

[0119] - Complex patterns can also be learned by utilizing non-linear models such as Random Forest and XGBoost.

[0120] - It is also possible to learn non-linear characteristics by utilizing neural network-based models.

[0121] (4) Physical model-based correction

[0122] Apply a theoretical model utilizing process conditions (temperature distribution, gas flow rate, plasma distribution, etc.).

[0123] - Predict and correct thin film thickness and resistance distribution using CFD (Computational Fluid Dynamics) simulations.

[0124]

[0125] Next, by executing at least one instruction stored in memory (120) by the processor (130), the prediction device (100) can control the deposition process for the wafer currently in progress to be stopped. Specifically, if the characteristic value predicted by the thin film characteristic prediction model for a specific cycle deviates from a predetermined threshold range, the prediction device (100) can detect this immediately. In addition, the prediction device (100) can control the deposition device through the communication unit (110) to stop the deposition process for the wafer or to contact the terminal of the manager of the deposition process.

[0126] As described above, according to one embodiment, thin film characteristics such as deposition thickness or sheet resistance at the end of each cycle in the wafer deposition process can be predicted in real time without actual measurement of the wafer. Therefore, if an anomaly is detected in a specific cycle, the process can be stopped immediately, thereby minimizing wafer loss in the event of an anomaly.

[0127] In addition, unlike conventional methods, thin film characteristics can be predicted for all desired points on the wafer, such as the edge, center, or middle, as well as for each of multiple points (e.g., 49 points). This allows for verification of process uniformity, which enables the optimization of process data such as process parameters and allows for the analysis of local variations (edge ​​effects) on the wafer. Furthermore, since it is possible to analyze which regions consistently exhibit thin film characteristics, issues with process equipment or errors in process settings causing these issues can be detected early, which can lead to improved yield.

[0128] Hereinafter, let us look at a method for real-time prediction of wafer thin film characteristics according to one embodiment, which can be performed by such a prediction device (100).

[0129] FIG. 4 illustrates an exemplary flowchart of a method for real-time prediction of wafer thin film characteristics according to one embodiment. Here, such a flowchart is merely exemplary and the scope of the invention is not limited thereto. For example, depending on the embodiment, each step may be performed in a different order than that shown in FIG. 4, or at least one step not shown in FIG. 4 may be additionally performed, or at least one of the steps shown in FIG. 4 may not be performed.

[0130] Referring to FIG. 4, a step (S100) of obtaining process data defining the deposition process of a wafer is performed.

[0131] In addition, the above-mentioned process data is provided to a pre-trained thin film characteristic prediction model, thereby performing a step (S200) of predicting the thin film characteristics for each point within the wafer in real-time on a cycle basis.

[0132] Here, this method is performed by the prediction device (100) as described above, and the description of the prediction device (100) is referenced, and no redundant description is provided below.

[0133] Meanwhile, FIG. 5 is a diagram showing the accuracy of the thin film characteristics predicted according to one embodiment. Referring to FIG. 5, it can be seen that the prediction accuracy of the deposition thickness and sheet resistance is high for the center, middle, edge, and all regions.

[0134] Below, we will examine the implementation method of the wafer thin film characteristic prediction model described earlier—that is, the specific methods for training and verifying it.

[0135] First, the model for predicting wafer thin film characteristics (hereinafter referred to as the "prediction model") may employ various machine learning models or deep learning models. For example, Decision tree-based models such as Randomforest, LightBGM, XGBoost, CatBoost, Linear Regression, Ridge Regression, etc., Deep learning models such as MLP, CNN, RNNs, etc., or Machine Learning models such as Bayesian Regression, Gaussian Process Regression, etc., may be included as candidates for adoption, but are not limited thereto.

[0136] Furthermore, prediction models can be trained using a supervised learning method. For this to work, a sufficient amount of labeled training data is required for training, and a sufficient amount of training data is also needed for validation and testing.

[0137] However, the reality is that it is difficult to secure training data for processes such as thin film deposition. Therefore, in one embodiment, a method is performed to augment a small number of training data to secure a sufficient amount of training data, and then utilize this to implement or train a prediction model with desired performance. Let us examine this in detail below.

[0138] First, for each Design of Experiment (DOE), process data for, for example, two wafers and data on thin film characteristics per cycle are obtained as a result of the completion of the thin film process. Of course, values ​​for the thin film characteristics per cycle are prepared for each point on the corresponding wafer. Furthermore, this data covers the process from the first cycle to the last cycle, where the last cycle could be, for example, 100 cycles. Since this data pertains to a small number of wafers, such as two, it is not difficult to obtain.

[0139] Next, the data obtained in this way is divided into detailed steps and cycles. Here, four detailed steps, such as the precursor exposure step, the purge step, the reactant exposure step, and the purge step, constitute one cycle. In addition, for example, if there is data from cycle 1 to cycle 100 for the first wafer, it is divided into 100 parts, and if there is data from cycle 1 to cycle 500 for the second wafer, it is divided into 500 parts.

[0140] Next, features are extracted from the data separated by cycle in this way.

[0141] As a result, the data obtained from just two wafers is augmented by the sum of the number of cycles performed on each wafer, for example, to 500 (200 + 300) in the above example. In this respect, classifying the number of data in cycle units in this way can be referred to as a data enhanced algorithm.

[0142] Next, this augmented data is effectively for one cycle. For example, it is like data for the 1st cycle, data for the 2nd cycle, ... data for the 50th cycle, ... data for the 100th cycle.

[0143] However, the deposition process is performed over multiple cycles. Therefore, the cycle-specific data augmented in this way is augmented from one cycle to a predetermined number of cycles, for example, up to 200 cycles. For instance, data for the first cycle can be augmented to data from cycle 1 to cycle 200, data for the second cycle can be augmented to data from cycle 1 to cycle 200, ... data for the hundredth cycle can be augmented to data from cycle 1 to cycle 200. When performing such augmentation, not only process data but also estimated values ​​for thin film characteristics are required. However, since the estimated values ​​cannot be determined arbitrarily, a domain-knowledge based data augmentation technique is applied in one embodiment. That is, by utilizing various domain knowledge required for the semiconductor manufacturing process, even if only a value for a single cycle exists, it is augmented to a value ranging from one cycle to hundreds or thousands of cycles. The domain knowledge used at this time may include at least one of Atomic Layer Deposition (ALD) reaction kinetics, growth mechanism, and precursor, and additionally may include Physical Vapor Deposition, Chemical Vapor Deposition, Sol-Gel Deposition, spin coating, or spray pyrolysis, but is not limited thereto. That is, since the data used for training must be reliable in order to improve the performance of the model, in one embodiment, such augmentation may be performed by utilizing domain knowledge.

[0144] Finally, the augmented training data learned in this way is divided into training, validation, and test. For example, the ratio of training vs. validation vs. test can be divided as 50% : 30% : 20%, but is not limited to this.

[0145] Accordingly, in one embodiment, data for at least several hundred to several thousand wafers can be obtained using only data for two wafers per DOE, for example, and supervised learning can be performed by utilizing this as training data. Specifically, process data may be provided as training input data, and thin film characteristics by point for each cycle may be provided as training ground data. During the training process, back-propagation may be performed so that the difference between the result obtained by providing training input data to the model and the training ground data is minimized, and as a result, the parameters of the model may be optimized. Of course, the training process is not limited to this, and various embodiments may be applied.

[0146] As described above, according to one embodiment, thin film characteristics such as deposition thickness or sheet resistance at the end of each cycle in the wafer deposition process can be predicted in real time without actual measurement of the wafer. Therefore, if an anomaly is detected in a specific cycle, the process can be stopped immediately, thereby minimizing wafer loss in the event of an anomaly.

[0147] In addition, unlike conventional methods, thin film characteristics can be predicted for all desired points on the wafer, such as the edge, center, or middle, as well as for each of multiple points (e.g., 49 points). This allows for verification of process uniformity, which enables the optimization of process data such as process parameters and allows for the analysis of local variations (edge ​​effects) on the wafer. Furthermore, since it is possible to analyze which regions consistently exhibit thin film characteristics, issues with process equipment or errors in process settings causing these issues can be detected early, which can lead to improved yield.

[0148] Meanwhile, the method according to the various embodiments described above may be implemented in the form of a computer program stored on a computer-readable recording medium programmed to perform each step of the method, and may also be implemented in the form of a computer-readable recording medium storing a computer program programmed to perform each step of the method.

[0149] The above description is merely an illustrative explanation of the technical concept of the present invention, and those skilled in the art to which the present invention pertains will be able to make various modifications and variations within the scope of the essential quality of the present invention. Accordingly, the embodiments disclosed in the present invention are intended to explain, not limit, the technical concept of the present invention, and the scope of the technical concept of the present invention is not limited by such embodiments. The scope of protection of the present invention shall be interpreted by the claims below, and all technical concepts within the equivalent scope shall be interpreted as being included within the scope of rights of the present invention.

Claims

1. Memory storing at least one instruction; and Includes a processor, By executing the above at least one instruction by the processor, Process data defining the wafer deposition process is secured, and The above-mentioned process data is provided to a pre-trained thin film property prediction model, thereby enabling real-time prediction of thin film properties for each point within the wafer on a cycle basis. At points within the wafer where the above thin film characteristics are predicted, The wafer includes an edge, a center, and a middle region between the edge and the center, wherein the predicted thin film characteristics at each of the edge, the center, and the middle have a certain co-relation with each other. Wafer thin film characteristic real-time prediction device.

2. In Paragraph 1, The above process data is, at least one of wafer temperature, pressure, stable time, stage / shower head gap, gas flow rate, source time, purge time, reactant time, and number of deposition cycles. Wafer thin film characteristic real-time prediction device.

3. In Paragraph 1, At least some of the thin film characteristics predicted at each of the above edges, the above centers, and the above intermediates are, Post-corrected dependently on the above relevance Wafer thin film characteristic real-time prediction device.

4. In Paragraph 1, The above-mentioned thin film characteristic prediction model is, After augmenting pre-prepared training data using domain knowledge-based data augmentation techniques, the learned data utilizing the augmented results Wafer thin film characteristic real-time prediction device.

5. In Paragraph 4, The domain knowledge used in the above data augmentation technique includes, at least one of ALD (Atomic Layer Deposition) reaction kinetics, growth mechanism, and precursor is included Wafer thin film characteristic real-time prediction device.

6. In Paragraph 4, The training data prepared in advance above includes, Measurement results for thin film characteristics at each point of a given wafer are included by cycle, and The above augmented result is, The above measurement result obtained by dividing it by cycle and then augmenting each of the divided results Wafer thin film characteristic real-time prediction device.

7. In Paragraph 1, The thin film characteristics predicted for each of the above points include: at least one of deposition thickness and sheet resistance is included Wafer thin film characteristic real-time prediction device.

8. In Paragraph 1, By executing the above at least one instruction by the processor, Based on the above predicted results, the deposition process on the currently ongoing wafer is controlled to be stopped. Wafer thin film characteristic real-time prediction device.

9. A method for real-time prediction of wafer thin film characteristics performed by a wafer thin film characteristic real-time prediction device, A step of obtaining process data that defines the deposition process of a wafer; and The above-mentioned process data is provided to a pre-trained model for predicting thin film characteristics, thereby including a step of predicting thin film characteristics for each point within the wafer in real-time on a cycle basis. At points within the wafer where the above thin film characteristics are predicted, The wafer includes an edge, a center, and a middle region between the edge and the center, wherein the predicted thin film characteristics at each of the edge, the center, and the middle have a certain co-relation with each other. Real-time prediction method for wafer thin film characteristics.

10. In Paragraph 9, The above process data is, at least one of wafer temperature, pressure, stable time, stage / shower head gap, gas flow rate, source time, purge time, reactant time, and number of deposition cycles. Real-time prediction method for wafer thin film characteristics.

11. In Paragraph 9, At least some of the thin film characteristics predicted at each of the above edges, the above centers, and the above intermediates are, Post-corrected dependently on the above relevance Real-time prediction method for wafer thin film characteristics.

12. In Paragraph 9, The above-mentioned thin film characteristic prediction model is, After augmenting pre-prepared training data using domain knowledge-based data augmentation techniques, the learned data utilizing the augmented results Real-time prediction method for wafer thin film characteristics.

13. In Paragraph 12, The domain knowledge used in the above data augmentation technique includes, at least one of ALD (Atomic Layer Deposition) reaction kinetics, growth mechanism, and precursor is included Real-time prediction method for wafer thin film characteristics.

14. In Paragraph 12, The training data prepared in advance above includes, Measurement results for thin film characteristics at each point of a given wafer are included by cycle, and The above augmented result is, The above measurement result obtained by dividing it by cycle and then augmenting each of the divided results Real-time prediction method for wafer thin film characteristics.

15. In Paragraph 9, The thin film characteristics predicted for each of the above points include: at least one of deposition thickness and sheet resistance is included Real-time prediction method for wafer thin film characteristics.

16. A non-transient computer-readable recording medium storing at least one computer-executable instruction, wherein, when the at least one instruction is executed by a processor, A step of obtaining process data that defines the deposition process of a wafer; and The above-mentioned process data is provided to a pre-trained model for predicting thin film characteristics, thereby including a step of predicting thin film characteristics for each point within the wafer in real-time on a cycle basis. At points within the wafer where the above thin film characteristics are predicted, A method wherein the processor performs a method comprising including the edge, center, and middle region between the edge and the center of a wafer, wherein the predicted thin film characteristics at each of the edge, the center, and the middle have a predetermined co-relation with each other. Non-transient computer-readable recording medium.

17. In Paragraph 16, The above process data is, at least one of wafer temperature, pressure, stable time, stage / shower head gap, gas flow rate, source time, purge time, reactant time, and number of deposition cycles. Non-transient computer-readable recording medium.

18. In Paragraph 16, At least some of the thin film characteristics predicted at each of the above edges, the above centers, and the above intermediates are, Post-corrected dependently on the above relevance Non-transient computer-readable recording medium.

19. In Paragraph 16, The above-mentioned thin film characteristic prediction model is, After augmenting pre-prepared training data using domain knowledge-based data augmentation techniques, the learned data utilizing the augmented results Non-transient computer-readable recording medium.

20. In Paragraph 16, By executing the above at least one instruction by the processor, Based on the above predicted results, the deposition process on the currently ongoing wafer is controlled to be stopped. Non-transient computer-readable recording medium.

Images