Semiconductor device characteristic prediction system

Abstract

To provide a characteristic prediction system for a semiconductor element.SOLUTION: A characteristic prediction system for a semiconductor element has a storage unit, an input unit, a processing unit, and a calculation unit. The processing unit has a function of creating a learning data set from first data stored in the storage unit, a function of creating prediction data from second data supplied from the input unit, a function of converting qualitative data (name of material, or composition formula) into quantitative data (characteristic of element, and composition), and a function of extracting or removing the first data and the second data. The first data includes a process list for first to m-th (m is an integer of 2 or more) semiconductor elements, and characteristics of the first to m-th semiconductor elements. The second data includes a process list for an (m+1)-th semiconductor element. By having a function of performing learning and inference of supervised learning, the calculation unit performs the leaning on the basis of the learning data set and inference of the characteristic of the semiconductor element from the prediction data.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] SUMMARY OF THE INVENTION One aspect of the present invention relates to a system for predicting characteristics of a semiconductor device. Another aspect of the present invention relates to a method for predicting characteristics of a semiconductor device.

[0002] In this specification and elsewhere, a semiconductor element refers to an element that can function by utilizing semiconductor characteristics. Examples include semiconductor elements such as transistors, diodes, light-emitting elements, and light-receiving elements. Another example of a semiconductor element is a passive element such as a capacitor, resistor, or inductor that is formed by a conductive film or an insulating film. Another example of a semiconductor element is a semiconductor device that includes a circuit having a semiconductor element or a passive element. [Background technology]

[0003] In recent years, in fields using artificial intelligence (AI), robotics, and energy fields that handle high power such as power ICs, the development of new semiconductor elements has been progressing to solve issues such as increasing computational loads and power consumption. While the integrated circuits and semiconductor elements used in integrated circuits required by the market are becoming increasingly complex, there is a demand for the early launch of integrated circuits with new functions. However, the process design, device design, and circuit design in the development of semiconductor elements require the knowledge, know-how, and experience of skilled engineers.

[0004] In recent years, methods for optimizing the manufacturing process of semiconductor devices, methods for estimating device characteristics, etc. Patent Document 1 discloses a method for calculating image feature amounts from an SEM image of a cross-sectional shape pattern of a semiconductor device, and estimating the device characteristics of an evaluation pattern from the correspondence between the image feature amounts and the device characteristics. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Application Laid-Open No. 2007-129059 Summary of the Invention [Problem to be solved by the invention]

[0006] The manufacturing process for semiconductor devices involves many steps before a semiconductor device is completed, and the types of steps and processing conditions vary widely. After a semiconductor device is manufactured through a lengthy process, its electrical characteristics, reliability test results, and other semiconductor device characteristics are measured using measuring equipment. The characteristics of semiconductor devices can be improved by verifying the causal relationship between the manufacturing process and the semiconductor device characteristics one by one through experiments.

[0007] However, it is costly and time-consuming to comprehensively adjust the semiconductor device manufacturing process and investigate the causal relationship with the characteristics of the semiconductor device. Furthermore, it is difficult for humans to understand the vast amount of data. Therefore, optimizing the manufacturing process through experiments requires a great deal of effort.

[0008] In view of the above, an object of one embodiment of the present invention is to provide a system for predicting characteristics of a semiconductor element.Another object of one embodiment of the present invention is to provide a method for predicting characteristics of a semiconductor element.Another object of one embodiment of the present invention is to provide a learning dataset for predicting characteristics of a semiconductor element.

[0009] Note that the description of these problems does not preclude the existence of other problems. Note that one embodiment of the present invention does not necessarily solve all of these problems. Note that problems other than these will become apparent from the description of the specification, drawings, claims, etc., and it is possible to extract other problems from the description of the specification, drawings, claims, etc. [Means for solving the problem]

[0010] One aspect of the present invention is a semiconductor device characteristic prediction system that performs supervised learning based on a training dataset and infers semiconductor device characteristics from prediction data based on the results of the learning. The semiconductor device characteristic prediction system includes a memory unit, an input unit, a processing unit, and a calculation unit. The processing unit has functions of creating a training dataset from first data stored in the memory unit, creating prediction data from second data provided by the input unit, converting qualitative data into quantitative data, and performing extraction or removal on the first data and the second data. The first data includes process lists for first through mth (m is an integer equal to or greater than 2) semiconductor devices and characteristics of the first through mth semiconductor devices. The second data includes a process list for the m+1th semiconductor device. The qualitative data is material names or composition formulas. The quantitative data is elemental properties and compositions. The calculation unit has functions of performing supervised learning and inference.

[0011] In the above-mentioned semiconductor element characteristic prediction system, the characteristic of the element is preferably one or more of atomic number, group, period, electron configuration, atomic weight, atomic radius (covalent bond radius, van der Waals radius, ionic radius, or metallic bond radius), atomic volume, electronegativity, ionization energy, electron affinity, dipole polarizability, melting point of the element, boiling point of the element, lattice constant of the element, density of the element, and thermal conductivity of the element.

[0012] In the semiconductor element characteristic prediction system, the semiconductor element characteristic is preferably a change over time in ΔVsh obtained in a reliability test (+GBT stress test, +DBT stress test, -GBT stress test, +DGBT stress test, +BGBT stress test, or -BGBT stress test). Alternatively, in the semiconductor element characteristic prediction system, the semiconductor element characteristic is preferably an Id-Vg characteristic or an Id-Vd characteristic.

[0013] In the above-described semiconductor device characteristic prediction system, the processing unit preferably has a function of converting the qualitative data into a numerical value using Label Encoding. [Effects of the Invention]

[0014] According to one aspect of the present invention, a system for predicting semiconductor device characteristics can be provided. Also, according to one aspect of the present invention, a method for predicting semiconductor device characteristics can be provided. Also, according to one aspect of the present invention, a training dataset for predicting semiconductor device characteristics can be provided.

[0015] Note that the effects of one embodiment of the present invention are not limited to the effects listed above. The effects listed above do not preclude the existence of other effects. Note that the other effects are effects not mentioned in this section, which will be described below. Effects not mentioned in this section can be derived by a person skilled in the art from the description in the specification, drawings, etc., and can be extracted as appropriate from these descriptions. Note that one embodiment of the present invention has at least one of the effects listed above and / or other effects. Therefore, one embodiment of the present invention may not have the effects listed above in some cases. [Brief explanation of the drawings]

[0016] [Figure 1] 1A and 1B are diagrams illustrating an example of a system for predicting characteristics of a semiconductor device. [Figure 2] FIG. 2 is a flowchart showing an example of a method for predicting characteristics of a semiconductor device. [Figure 3] 3A and 3B are diagrams illustrating the configuration of a neural network. [Figure 4] 4A and 4B are diagrams illustrating the training data set. [Figure 5] Fig. 5A is a diagram illustrating the results obtained in a reliability test of a semiconductor element, and Fig. 5B is a diagram illustrating the Id-Vg characteristics of the semiconductor element. [Figure 6]6A to 6C are diagrams for explaining a method for creating input data. [Figure 7] 7A and 7B are diagrams for explaining a method for creating input data. [Figure 8] FIG. 8 is a diagram illustrating a computer device. DETAILED DESCRIPTION OF THE INVENTION

[0017] The embodiments will be described in detail with reference to the drawings. However, the present invention is not limited to the following description, and it will be readily understood by those skilled in the art that various changes can be made in form and detail without departing from the spirit and scope of the present invention. Therefore, the present invention should not be interpreted as being limited to the description of the embodiments shown below.

[0018] In the configuration of the invention described below, the same parts or parts having similar functions are denoted by the same reference numerals in different drawings, and repeated explanations thereof will be omitted. Furthermore, when referring to similar functions, the same hatch pattern may be used and no particular reference numeral may be assigned.

[0019] Furthermore, for ease of understanding, the position, size, range, etc. of each component shown in the drawings may not represent the actual position, size, range, etc. Therefore, the disclosed invention is not necessarily limited to the position, size, range, etc. disclosed in the drawings.

[0020] It should also be noted that the ordinal numbers "first," "second," and "third" used in this specification and elsewhere are used to avoid confusion of components and are not intended to limit the number.

[0021] (Embodiment 1) In this embodiment, a system for predicting characteristics of a semiconductor element and a method for predicting characteristics of a semiconductor element according to one embodiment of the present invention will be described with reference to FIGS. 1A to 8. FIG.

[0022] A semiconductor element characteristic prediction system according to one embodiment of the present invention is a system capable of predicting the characteristics of a semiconductor element based on information about the semiconductor element. Also, a semiconductor element characteristic prediction method according to one embodiment of the present invention is a method for predicting the characteristics of a semiconductor element using machine learning.

[0023] <Semiconductor device characteristic prediction system> Fig. 1A is a diagram showing the configuration of a characteristic prediction system 100. In other words, Fig. 1A can also be said to be an example of the configuration of a semiconductor element characteristic prediction system, which is one aspect of the present invention.

[0024] The characteristic prediction system 100 may be provided in an information processing device such as a personal computer used by a user, or may be configured such that a processing unit of the characteristic prediction system 100 is provided in a server and the server is accessed and used from a client PC via a network.

[0025] 1A, the characteristic prediction system 100 includes an input unit 101, a processing unit 102, a calculation unit 103, an output unit 104, and a storage unit 105. The input unit 101, the processing unit 102, the calculation unit 103, the output unit 104, and the storage unit 105 may be connected to each other via a transmission path.

[0026] The storage unit 105 stores data relating to each of the multiple semiconductor elements. The information relating to the semiconductor elements includes, for example, a process list for the semiconductor elements, characteristics of the semiconductor elements, and information relating to the shape of the semiconductor elements. Hereinafter, data relating to the process list for the semiconductor elements may be simply referred to as a process list for the semiconductor elements. Data relating to the characteristics of the semiconductor elements may be simply referred to as characteristics of the semiconductor elements. Data relating to the shape of the semiconductor elements may be simply referred to as information relating to the shape of the semiconductor elements.

[0027] In the process list for semiconductor devices, a plurality of processes are set in the order of the manufacturing process of the semiconductor devices, and processing conditions are specified for each process.

[0028] The characteristics of the semiconductor element are the electrical characteristics of the semiconductor element obtained by actual measurement using a measuring device, the results of a reliability test, etc. The data on the characteristics of the semiconductor element are, for example, measurement data of the electrical characteristics of the semiconductor element, data obtained by performing a reliability test, etc.

[0029] Information about the shape of a semiconductor element includes the positions, sizes, ranges, etc. of the constituent elements of the semiconductor element. Data about the shape of a semiconductor element includes, for example, numerical data representing the positions, sizes, ranges, etc. of the constituent elements of the semiconductor element, image data of the semiconductor element and its surroundings, etc. Specifically, such data includes measurement data such as channel length and channel width, observation images obtained by a scanning electron microscope (SEM), observation images obtained by a transmission electron microscope (TEM), etc.

[0030] The storage unit 105 stores at least the process list and characteristics of each of the multiple semiconductor elements. Preferably, an ID is assigned to each process list of the semiconductor elements stored in the storage unit 105. Here, the ID assigned to the process list of the semiconductor element is referred to as a process list ID. The characteristics of the semiconductor elements stored in the storage unit 105 are associated with the process list ID. In other words, the characteristics of the semiconductor elements may be read, written, etc. based on the process list ID.

[0031] Information about the shape of each of the plurality of semiconductor elements may be stored in the storage unit 105. In this case, it is preferable that the information about the shape of the semiconductor elements stored in the storage unit 105 is associated with a process list ID. In this case, reading and writing of the information about the shape of the semiconductor elements may be performed based on the process list ID.

[0032] The process list for the plurality of semiconductor elements and the characteristics of the plurality of semiconductor elements are stored in the storage unit 105 via the input unit 101, a storage medium, communication, etc. Information regarding the shapes of the plurality of semiconductor elements may also be stored in the storage unit 105 via the input unit 101, a storage medium, communication, etc.

[0033] The process list for the plurality of semiconductor elements and the characteristics of the plurality of semiconductor elements are preferably stored in the storage unit 105 as text data. In particular, the characteristics of the plurality of semiconductor elements are preferably stored in the storage unit 105 as numerical data or two-variable data. In this specification and the like, two-variable data refers to a collection of data relating to two variables. Note that the two-variable data may also be a collection of data obtained by extracting data relating to two variables from multi-variable data of three or more variables.

[0034] When the process list for a plurality of semiconductor elements and the characteristics of a plurality of semiconductor elements are image data, they may be stored in storage unit 105 as image data, but it is preferable that the image data be converted into text data before being stored in storage unit 105. Since the data size of text data is smaller than the data size of image data, the load on storage unit 105 can be reduced by converting the image data into text data before storing it in storage unit 105.

[0035] The characteristic prediction system 100 may have an optical character recognition (OCR) function, which allows it to recognize characters included in image data and create text data. For example, the processing unit 102 may have this function. Alternatively, the characteristic prediction system 100 may further have a character recognition unit having this function.

[0036] The storage unit 105 may have a function of storing a trained model (also called an inference model).

[0037] The input unit 101 has a function for a user to input data IN2. The data IN2 is text data or image data. The input unit 101 may be an input device such as a keyboard, a mouse, a touch sensor, a scanner, or a camera. The data IN2 may be stored in the storage unit 105.

[0038] If data IN2 is image data, the characteristic prediction system 100 has the above-mentioned OCR function, which allows it to recognize characters contained in the image data and create text data. For example, if the OCR function is provided in the processing unit 102, data IN2 may remain as image data. Alternatively, if the OCR function is provided in a device other than the processing unit 102 of the characteristic prediction system 100, text data converted from image data may be used as data IN2.

[0039] The processing unit 102 has a function of generating a training dataset DS from data IN1 supplied from the storage unit 105. The training dataset DS is a training dataset for supervised learning. The processing unit 102 also has a function of generating prediction data DI from data IN2 supplied from the input unit 101. The prediction data DI is data for predicting the characteristics of a semiconductor element.

[0040] The data IN1 is a data group used when creating the learning dataset DS. The data group includes some or all of the information about each of the multiple semiconductor elements stored in the storage unit 105.

[0041] Here, some or all of the multiple semiconductor elements are represented as semiconductor elements 30_1 to 30_m (m is an integer equal to or greater than 2). In this case, the process lists for semiconductor elements 30_1 to 30_m are represented as process lists 10_1 to 10_m, respectively. Furthermore, the characteristics of semiconductor elements 30_1 to 30_m measured using a measuring device are represented as characteristics 20_1 to 20_m, respectively. In other words, characteristics 20_1 to 20_m are characteristics measured using a measuring device for semiconductor elements fabricated according to process lists 10_1 to 10_m, respectively.

[0042] Hereinafter, process lists 10_1 to 10_m may be collectively referred to as multiple process lists 10. Furthermore, characteristics 20_1 to 20_m may be collectively referred to as multiple characteristics 20. Furthermore, semiconductor elements 30_1 to 30_m may be collectively referred to as multiple semiconductor elements 30.

[0043] Data IN1 includes, for example, data on process lists 10_1 to 10_m and data on characteristics 20_1 to 20_m. Data IN1 may also include data on information regarding the shape of semiconductor elements associated with the process list IDs of process lists 10_1 to 10_m. Hereinafter, the data on process lists 10_1 to 10_m may be simply referred to as process lists 10_1 to 10_m. The data on characteristics 20_1 to 20_m may be simply referred to as characteristics 20_1 to 20_m.

[0044] Data IN2 is information about a semiconductor element designated by a user for predicting the characteristics of the semiconductor element. Data IN2 includes, for example, a process list designated for predicting the characteristics of the semiconductor element. The process list designated for predicting the characteristics of the semiconductor element is referred to as process list 11. Data IN2 may also include information about the shape of the semiconductor element associated with the process list ID of process list 11.

[0045] The processing unit 102 also has a function of converting qualitative data (also referred to as qualitative data, categorical data, etc.) into numerical values. In other words, the processing unit 102 has a function of converting qualitative data into quantitative data (also referred to as quantitative data, quantity data, etc.). For example, the processing unit 102 preferably implements label encoding, one-hot encoding, target encoding, etc.

[0046] Qualitative data is included in data IN1 and data IN2. Examples of the qualitative data include data related to the equipment, data related to the materials, etc. The quantification of the qualitative data related to the equipment and the qualitative data related to the materials will be described later.

[0047] The calculation unit 103 has a function of performing machine learning. For example, the calculation unit 103 has a function of performing supervised learning based on the learning dataset DS. The calculation unit 103 also has a function of inferring the characteristics of a semiconductor element from the prediction data DI based on the learning results of the supervised learning. By performing supervised learning as machine learning, it is possible to improve the accuracy of inferring the characteristics of a semiconductor element. Note that a trained model may be generated by performing the supervised learning.

[0048] For the supervised learning, it is preferable to use a neural network (particularly, deep learning). As deep learning, it is preferable to use, for example, a convolutional neural network (CNN), a recurrent neural network (RNN), an autoencoder (AE), a variational autoencoder (VAE), a random forest, a support vector machine, gradient boosting, or a generative adversarial network (GAN).

[0049] The output of the calculation unit 103 is used as the characteristics of the semiconductor element. In other words, the output of the neural network is used as the characteristics of the semiconductor element. By using actual measured values ​​as the output of the neural network, the machine learning model can be trained, and then the characteristics of the semiconductor element can be predicted by inputting a process list for any semiconductor element into the neural network.

[0050] In a neural network, a product-sum operation is performed. When the product-sum operation is performed by hardware, the calculation unit 103 preferably has a product-sum operation circuit. The product-sum operation circuit may be a digital circuit or an analog circuit. The product-sum operation may also be performed on software using a program.

[0051] The calculation unit 103 may have a function of performing semi-supervised learning as machine learning. The training data is provided with the characteristics of semiconductor elements as training data (also referred to as training signals, correct labels, etc.), but in order to prepare the training data, it is necessary to actually manufacture semiconductor elements and measure the characteristics of the semiconductor elements. Compared to supervised learning, semi-supervised learning requires a smaller amount of training data to be included in the training dataset, making it possible to perform inference while reducing the time spent on creating training data.

[0052] The output unit 104 has a function of supplying information. The information is the predicted results of the semiconductor device characteristics calculated by the calculation unit 103 or information related to the predicted results. The information is supplied as visual information such as character strings, numerical values, graphs, etc. The output unit 104 can be an output device such as a display. Note that the characteristic prediction system 100 does not necessarily have to include the output unit 104.

[0053] The above constitutes a semiconductor device characteristic prediction system.

[0054] Note that the characteristic prediction system 100 is not limited to the above configuration. For example, as shown in Fig. 1B, the characteristic prediction system 100 may include a storage unit 106 in addition to the input unit 101, the processing unit 102, the calculation unit 103, the output unit 104, and the memory unit 105.

[0055] The storage unit 106 has a function of storing the trained model generated by the calculation unit 103. By including the storage unit 106, the characteristic prediction system 100 can predict the characteristics of a semiconductor device based on the trained model. Therefore, by generating the trained model in advance, it is not necessary to perform supervised learning when predicting the characteristics of a semiconductor device. Therefore, it is possible to shorten the time required to predict the characteristics of a semiconductor device.

[0056] The storage unit 106 is connected to the calculation unit 103 via a transmission path. Note that the storage unit 106 may be connected to each of the input unit 101, the processing unit 102, the output unit 104, and the memory unit 105 via a transmission path.

[0057] The storage unit 106 may be provided within the memory unit 105. The memory unit 105 may also serve as the storage unit 106.

[0058] The above is a description of the configuration of the characteristic prediction system 100. By using the semiconductor element characteristic prediction system, which is one aspect of the present invention, it is possible to predict the characteristics of a semiconductor element from information about the semiconductor element. For example, it is possible to predict the characteristics of a semiconductor element from a process list for the semiconductor element. Furthermore, it is possible to extract, from the process list for the semiconductor element, processes that have a large contribution to the characteristics of the semiconductor element.

[0059] <Method for predicting semiconductor device characteristics> Fig. 2 is a flowchart illustrating the flow of processing executed by the characteristic prediction system 100. In other words, Fig. 2 can also be said to be a flowchart illustrating an example of a method for predicting the characteristics of a semiconductor device, which is an aspect of the present invention.

[0060] The method for predicting the characteristics of a semiconductor device includes steps S001 to S007. Steps S001 to S003 are steps related to learning in supervised learning, and steps S004 to S007 are steps related to inference in supervised learning.

[0061] Step S001 is a step of inputting first data into processing unit 102. The first data corresponds to data IN1 described above. That is, the first data includes information related to semiconductor elements 30_1 to 30_m. Specifically, the first data includes process lists 10_1 to 10_m and characteristics 20_1 to 20_m. Note that the first data may also include information related to the shapes of the semiconductor elements associated with the process list IDs of process lists 10_1 to 10_m, respectively.

[0062] Step S002 is a process of creating a training dataset from the first data. Step S002 is performed by the processing unit 102 shown in Figures 1A and 1B. The training dataset corresponds to the training dataset DS described above.

[0063] Step S002 also includes a step of digitizing qualitative data included in the first data. The qualitative data may be, for example, qualitative data related to equipment, qualitative data related to materials, etc. The data obtained by digitization is included in the training dataset.

[0064] Step S003 is a process of performing supervised learning based on the learning data set. Step S003 is performed by the calculation unit 103 shown in FIGS. 1A and 1B. It is preferable to use a neural network (particularly deep learning) as the algorithm (also called a learning method) for the supervised learning. Note that the supervised learning may generate a trained model for predicting the characteristics of a semiconductor device.

[0065] Step S004 is a step of inputting second data into processing unit 102. The second data corresponds to data IN2 described above. That is, the second data includes information about the semiconductor device designated by the user for predicting the characteristics of the semiconductor device. Specifically, the second data includes process list 11.

[0066] When the semiconductor element is produced according to the process list 11, the second data may include information about the shape of the semiconductor element, characteristics of the semiconductor element, and the like.

[0067] It is preferable that step S004 is performed after steps S003 have been performed, but it may be performed simultaneously with step S001, or may be performed while steps S001 to S003 are being performed.

[0068] Step S005 is a process of creating prediction data for the characteristics of the semiconductor element from the second data. Step S005 is performed by processing unit 102 shown in Figures 1A and 1B. That is, the prediction data for the characteristics of the semiconductor element corresponds to the above-mentioned prediction data DI.

[0069] Step S005 also includes a step of quantifying the qualitative data included in the second data. The qualitative data is, for example, qualitative data related to the device, qualitative data related to the material, etc. The data obtained by quantifying the qualitative data is included in the data for predicting the characteristics of the semiconductor element.

[0070] It is preferable that step S005 is performed after steps up to step S003 have been performed, but it may be performed simultaneously with step S001, or may be performed while steps S001 to S003 are being performed.

[0071] Step S006 is a process of inferring the characteristics of the semiconductor element from the prediction data of the semiconductor element characteristics based on the learning results of the supervised learning performed in step S003. In other words, step S006 is a process of inferring the characteristics of the semiconductor element from the prediction data of the semiconductor element characteristics using the trained model. Step S006 is performed by the calculation unit 103 shown in FIGS. 1A and 1B.

[0072] Step S007 is a step of outputting third data. Step S007 is performed by output unit 104 shown in Figures 1A and 1B. The third data includes the result of the inference or information related to the result of the inference.

[0073] In this way, the characteristics of the semiconductor device can be predicted. Note that, instead of step S007, a step of storing the result of the above inference or information related to the result of the above inference in storage unit 105 shown in FIG. 1A or the like may be performed. Alternatively, step S007 may not be performed.

[0074] The supervised learning process (steps S001 to S003) may be performed each time information about the semiconductor element is stored in the memory unit 105, or may be performed periodically at predetermined times (for example, once a day or once a week).

[0075] The method for predicting the characteristics of a semiconductor element is not limited to the above method. For example, the method for predicting the characteristics of a semiconductor element may include, after step S003, a step of storing the trained model generated in step S003. The trained model is stored in the storage unit 106 shown in FIG. 1B. By generating the trained model in advance, steps S001 to S003 can be omitted when predicting the characteristics of a semiconductor element. This reduces the time required to predict the characteristics of a semiconductor element.

[0076] <<Neural Networks>> Here, we will explain neural networks that can be used for supervised learning.

[0077] As shown in FIG. 3A, a neural network NN can be composed of an input layer IL, an output layer OL, and a hidden layer HL. Each of the input layer IL, output layer OL, and hidden layer HL has one or more neurons (units). The hidden layer HL may have one layer or two or more layers. A neural network with two or more hidden layers HL can also be called a deep neural network (DNN). Learning using a deep neural network can also be called deep learning.

[0078] Input data is input to each neuron in the input layer IL. An output signal from a neuron in the previous or next layer is input to each neuron in the hidden layer HL. An output signal from a neuron in the previous layer is input to each neuron in the output layer OL. Each neuron may be connected to all neurons in the previous or next layer (fully connected), or may be connected to only a portion of the neurons in the previous or next layer.

[0079] Figure 3B shows an example of a neuron operation. It shows neuron N and two neurons in the previous layer that output signals to neuron N. Neuron N receives the output x1 from one neuron in the previous layer and the output x2 from the other neuron in the previous layer. Neuron N then multiplies the output x1 by a weight w1 (x1w1) and the output x2 by a weight w2 (x2w2), calculating the sum x1w1+x2w2. A bias b is then added as necessary to obtain the value a = x1w1+x2w2+b. The value a is then transformed by an activation function h, and neuron N outputs an output signal y = a. The activation function h can be, for example, a sigmoid function, tanh function, softmax function, ReLU function, or threshold function.

[0080] As described above, the computation performed by a neuron includes the sum of the product of the output of a neuron in the previous layer and the weight, i.e., the sum-of-products computation (x1w1+x2w2 as above). This sum-of-products computation may be performed in software using a program, or by hardware. When performing the sum-of-products computation by hardware, a sum-of-products computation circuit can be used. This sum-of-products computation circuit may be a digital circuit or an analog circuit. When an analog circuit is used for the sum-of-products computation circuit, it is possible to reduce the circuit size of the sum-of-products computation circuit or the number of memory accesses, thereby improving processing speed and reducing power consumption.

[0081] The product-sum operation circuit may be configured using transistors containing silicon (such as single crystal silicon) in their channel formation regions (hereinafter also referred to as Si transistors) or transistors containing an oxide semiconductor in their channel formation regions (hereinafter also referred to as OS transistors). OS transistors, in particular, have extremely low off-state current and are therefore suitable as transistors constituting the analog memory of the product-sum operation circuit. The product-sum operation circuit may be configured using both Si transistors and OS transistors.

[0082] When the product-sum operation is performed by hardware, the product-sum operation circuit may be included in the operation unit 103 of the characteristic prediction system 100.

[0083] The above is a description of the neural network. In one embodiment of the present invention, it is preferable to use deep learning. In other words, it is preferable to use a neural network having two or more hidden layers HL.

[0084] The above is a description of an example of a method for predicting the characteristics of a semiconductor device.

[0085] <Details of the method for predicting semiconductor device characteristics> The method for predicting the characteristics of a semiconductor device will be described in detail below with reference to FIGS. 4A to 7B.

[0086] <<Structure of semiconductor elements>> First, the structure of a semiconductor element will be described, and a transistor will be described as an example of the semiconductor element.

[0087] Transistors are classified into various types based on the relative positions and shapes of their components. For example, transistor structures are classified into bottom-gate and top-gate structures based on the relative positions of the substrate, gate, and channel formation region. A transistor structure in which the gate is provided between the channel formation region and the substrate is called a bottom-gate structure. On the other hand, a transistor structure in which the channel formation region is provided between the gate and the substrate is called a top-gate structure.

[0088] Furthermore, transistor structures are classified into bottom-contact and top-contact structures depending on the connection points between the source and drain and the semiconductor layer that forms the channel. A transistor structure in which the source and drain and the semiconductor layer that forms the channel are connected on the substrate side is called a bottom-contact structure. A transistor structure in which the source and drain and the semiconductor layer that forms the channel are connected on the opposite side from the substrate is called a top-contact structure.

[0089] That is, transistor structures are classified into BGBC (bottom gate bottom contact) type structure, BGTC (bottom gate top contact) type structure, TGTC (top gate top contact) type structure, and TGBC (top gate bottom contact) type structure.

[0090] In addition to the four transistor structures mentioned above, there are also other types of transistor structures, such as a dual-gate structure in which the gates are located above and below the semiconductor layer, and a TGSA (Top-Gate Self-Align) structure in which the source and drain are formed self-aligned to the gate pattern.

[0091] The semiconductor elements 30_1 to 30_m preferably have the same or similar structures. For example, if the semiconductor elements 30_1 to 30_m are transistors, the structures of the semiconductor elements 30_1 to 30_m are preferably BGBC type structures, BGTC type structures, TGTC type structures, TGBC type structures, dual gate type structures, or TGSA type structures. Identical structures of the semiconductor elements can improve the accuracy of predicting the characteristics of the semiconductor elements.

[0092] The semiconductor elements 30_1 to 30_m may have different structures. When the semiconductor elements 30_1 to 30_m are transistors, for example, some of the structures of the semiconductor elements 30_1 to 30_m may have a TGTC structure, and the rest may have a TGSA structure. Combining multiple structures enables highly versatile prediction of the characteristics of semiconductor elements.

[0093] The above is a description of the structure of the semiconductor element.

[0094] <<Semiconductor element characteristics>> Next, the characteristics of the semiconductor element will be described.

[0095] In this specification and elsewhere, the characteristics of a semiconductor element refer to the electrical characteristics of the semiconductor element, such as the drain current (Id)-gate voltage (Vg) characteristics, the drain current (Id)-drain voltage (Vd) characteristics, and the capacitance (C)-gate voltage (V) characteristics.

[0096] The characteristics of the semiconductor element may also be the results obtained by a reliability test, such as the change in on-current (Ion) over time (also referred to as the stress time dependency of Ion), the change in ΔVsh over time (also referred to as the stress time dependency of ΔVsh), etc.

[0097] ΔVsh is the amount of change in the shift voltage (Vsh), which is defined as the Vg at which the tangent to the maximum slope of the drain current (Id)-gate voltage (Vg) curve of a transistor intersects with the line where Id = 1 pA.

[0098] Reliability tests include +GBT (Gate Bias Temperature) stress tests, +DBT (Drain Bias Temperature) stress tests, -GBT stress tests, +DGBT (Drain Gate Bias Temperature) stress tests, +BGBT (Back Gate Bias Temperature) stress tests, and -BGBT stress tests.

[0099] Since reliability tests may require long-term measurements, it takes time to obtain the results of the reliability tests. Furthermore, the measurement equipment is occupied during the measurement period. Therefore, by using a semiconductor element characteristic prediction system according to one embodiment of the present invention, the results of the reliability tests can be predicted. Therefore, by determining whether or not to perform the reliability tests based on the predicted results, it is possible to omit some of the reliability tests. Alternatively, it is possible to determine the priority of the reliability tests. This allows for effective use of the measurement equipment.

[0100] In this specification and elsewhere, the characteristics of a semiconductor element also include characteristic values ​​calculated from the measurement results of the electrical characteristics of the semiconductor element. Examples of such characteristic values ​​include threshold voltage (Vth), Vsh, subthreshold swing (S value), Ion, and field-effect mobility (μFE). Here, the subthreshold swing (S value) refers to the amount of change in gate voltage in the subthreshold region that changes the drain current by one order of magnitude at a constant drain voltage. Hereinafter, characteristic values ​​calculated from the measurement results of the electrical characteristics of a semiconductor element may be referred to as characteristic values ​​of the semiconductor element, or simply as characteristic values.

[0101] The characteristics of a semiconductor element also include temperature characteristics. Examples of the temperature characteristics include the temperature characteristics of a threshold voltage and the temperature dependence of capacitance characteristics. Since the temperature characteristics need to be measured at a plurality of different temperatures, it takes time to evaluate the temperature characteristics. By using the characteristic prediction system for a semiconductor element according to one embodiment of the present invention, it is possible to predict the temperature characteristics without manufacturing the semiconductor element and performing measurements to evaluate the temperature characteristics.

[0102] When the characteristics of the semiconductor element are characteristic values, the characteristics of the semiconductor element are stored as numerical data in the storage unit 105. When the characteristics of the semiconductor element are electrical characteristics or temperature characteristics, the characteristics of the semiconductor element are stored as binary data in the storage unit 105. In other words, the characteristics of the semiconductor element stored in the storage unit 105 are quantified.

[0103] For example, if the characteristic of the semiconductor element is a change over time in ΔVsh obtained in a reliability test, a collection of data relating to time and ΔVsh is stored in the storage unit 105. Also, for example, if the characteristic of the semiconductor element is an Id-Vg characteristic, a collection of data relating to Vg and Id is stored in the storage unit 105.

[0104] The above is a description of the characteristics of the semiconductor element.

[0105] <<Learning Dataset>> Here, we will explain the training dataset for supervised learning.

[0106] 4A and 4B are diagrams showing the configuration of a training dataset 50. The training dataset 50 corresponds to the training dataset DS generated by the processing unit 102. The training dataset 50 includes training data 51_1 to training data 51_m. Training data 51_i (i is an integer between 1 and m) includes input data 52_i and teacher data 53_i. The training data 51_i includes information related to the semiconductor element 30_i.

[0107] The training dataset 50 is generated from the data IN1 input to the processing unit 102 shown in Figures 1A and 1B. Therefore, the training dataset 50 is generated by extracting, processing, converting, selecting, removing, etc., the data included in the data IN1.

[0108] In this embodiment, the training data is information about the semiconductor element, which is information about the semiconductor element, and therefore the target of prediction in this embodiment is the characteristics of the semiconductor element.

[0109] Furthermore, in this embodiment, the input data is preferably created from a process list for a semiconductor element, which is information related to the semiconductor device. That is, the input data preferably includes a part of the process list for a semiconductor element, which is information related to the semiconductor device. The characteristics of the semiconductor element to be predicted are affected by the type of semiconductor material used in the layer in which the channel is formed, the type of conductive material used in the layer that functions as the gate electrode, the type of insulating material used in the layer that functions as the gate insulating film, the film thickness of each of these layers, and the film formation conditions for each of these layers. Note that the type of material used in the layers, the film thickness of the layers, the film formation conditions for the layers, etc. are included in the process list for the semiconductor element. Therefore, the input data is preferably created from the process list for the semiconductor element.

[0110] The data contained in a training dataset for supervised learning is preferably quantitative data. In other words, the data is preferably quantified. Compared to a training dataset containing non-numeric data (qualitative data), quantification of the data contained in the training dataset can prevent the machine learning model from becoming too complex.

[0111] 4A, input data 52_1 to input data 52_m are generated from process lists 10_1 to 10_m, respectively. Teacher data 53_1 to teacher data 53_m are generated from characteristics 20_1 to 20_m, respectively.

[0112] 4B, input data 52_1 to input data 52_m may be generated from process lists 10_1 to 10_m and information on the shapes of semiconductor elements 30_1 to 30_m, respectively. Adding information on the shapes of semiconductor elements 30_1 to 30_m to input data 52_1 to input data 52_m, respectively, can improve the prediction accuracy of the characteristics of the semiconductor elements.

[0113] It is preferable that the number of steps in each of the process lists 10_1 to 10_m is the same. This makes it easier to create a learning data set or prediction data. When creating a learning data set or prediction data, the process list is selected. For example, a portion of the process list is extracted, or another portion of the process list is removed. Therefore, the number of steps in each of the process lists 10_1 to 10_m may be different.

[0114] As described above, the characteristics 20_1 to 20_m are digitized data, and therefore can be included in the teacher data 53_1 to 53_m, respectively, without any particular conversion.

[0115] When the characteristics 20_1 to 20_m are two-variable data, one or more characteristic points may be extracted from the two-variable data and included in the teacher data 53_1 to 53_m, respectively. Alternatively, multiple points may be extracted from the two-variable data so that the values ​​of one of the two variables are equally spaced, and included in the teacher data 53_1 to 53_m, respectively.

[0116] 5A is a diagram illustrating the results obtained in a reliability test of a semiconductor device. In FIG. 5A, the horizontal axis represents the elapsed time (also called stress time) [h] from the start of measurement, and the vertical axis represents ΔVsh [mV]. For example, the values ​​of ΔVsh from time A1 to time A10 and at time A1 to time A10 can be extracted and used as training data.

[0117] Note that some or all of the times A1 to A10 may have a characteristic value of ΔVsh. Alternatively, the times A1 to A10 may be equally spaced. Alternatively, some of the times A1 to A10 may be a first interval, and the remaining times A1 to A10 may be a second interval different from the first interval.

[0118] Furthermore, the number of pairs of extracted times and values ​​of ΔVsh at those times is not limited to 10, but may be 1 to 9 or 11 or more.

[0119] FIG. 5B is a diagram illustrating the Id-Vg characteristics of a semiconductor element. In the Id-Vg curve, the drain current value at a gate voltage of 0V is one of the characteristic points. For example, let the gate voltage be voltage B4. Voltages B1 to B3 are designated as voltages lower than voltage B4. Voltages B5 to B10 are designated as voltages higher than voltage B4. For example, voltages B1 to B10 and the drain current values ​​at voltages B1 to B10 can be extracted and used as training data. Alternatively, one of voltages B1 to B10, excluding voltage B4, can be set to 0V.

[0120] Furthermore, the number of pairs of extracted voltages and values ​​of drain currents at those voltages is not limited to 10, but may be 1 to 9 or 11 or more.

[0121] <<How to create input data>> Here, a method for creating the input data 52_1 to 52_m shown in FIG. 4A will be described.

[0122] First, a method for creating the input data 52_1 will be described. Here, an example in which the input data 52_1 is created from the process list 10_1 will be described with reference to Figs. 6A to 6C.

[0123] In the process list, a plurality of processes are set in the order of the manufacturing process of the semiconductor device, such as film formation, cleaning, resist application, exposure, development, processing, heat treatment, inspection, and substrate transfer.

[0124] Furthermore, processing conditions are specified for each of the multiple processes set in the process list. For example, processing conditions for a film formation process include the equipment, material, film thickness, temperature, pressure, power, flow rate, etc. The processing conditions for the film formation process may affect the characteristics of semiconductor devices. Furthermore, even in processes other than film formation, the processing conditions, the presence or absence of a process, the order of processes, etc. may affect the characteristics of semiconductor devices.

[0125] The processing conditions are a mixture of qualitative and quantitative data, and are expressed in various scales. To express the similarity of the features in each process, for example, qualitative data on materials can be converted into quantitative data such as physical properties for each material, and the resulting set can be used as the feature.

[0126] Here, it is assumed that the process list 10_1 has n processes (n is an integer of 2 or more). For example, in the process list 10_1 shown in FIG. 6A, the first process is a substrate transfer process, the jth process (j is an integer of 2 or more and (n-4) or less) is a film formation process, the (j+1)th process is a processing process, the (j+2)th process is a film formation process, the (j+3)th process is a heat treatment process, and the nth process is a substrate transfer process. Note that the numbers shown in FIGS. 6A and 6B are process numbers.

[0127] The processing conditions specified in the jth process (film formation process) are referred to as Condition 1 through Condition p (p is an integer of 2 or greater). The processing conditions specified in the (j+1)th process (processing process) are referred to as Condition 1 through Condition q (q is an integer of 1 or greater). The processing conditions specified in the (j+2)th process (film formation process) are referred to as Condition 1 through Condition r (r is an integer of 2 or greater). The processing conditions specified in the (j+3)th process (heat treatment process) are referred to as Condition 1 through Condition s (s is an integer of 1 or greater).

[0128] First, some processes are extracted from the n processes included in the process list 10_1. The extracted processes are, for example, processes that are estimated to have a large contribution from the characteristics of semiconductor elements. Also, for example, processes that have many changes in conditions. By extracting some of the processes included in the process list 10_1, the number of parameters in machine learning can be reduced. In other words, the number of neurons included in the input layer in supervised learning using a neural network can be reduced. This optimizes the number of hidden layers and the number of neurons in the hidden layers, reducing the amount of calculation or calculation time for learning or inference. It may also be possible to prevent overfitting.

[0129] For example, if the processing conditions specified in the jth process (film formation process) are often changed among the process lists 10_1 to 10_m, the jth process (film formation process) may be extracted from the process list 10_1. Also, if the (j+3)th process (heat treatment process) is estimated to have a large contribution from the characteristics of the semiconductor element, the (j+3)th process (heat treatment process) may be extracted from the process list 10_1.

[0130] Alternatively, some processes different from the above may be removed from the n processes included in the process list 10_1. The removed processes are, for example, processes estimated to have little contribution from semiconductor characteristics. Also, for example, processes whose processing conditions have not changed. By removing some processes different from the above, the number of parameters in machine learning can be reduced. In other words, the number of neurons included in the input layer in supervised learning using a neural network can be reduced. This optimizes the number of hidden layers and the number of neurons in the hidden layers, reducing the amount of calculation or calculation time for learning or inference. It may also be possible to prevent overfitting.

[0131] For example, it is estimated that the substrate transfer processes (the first process and the nth process) do not affect the characteristics of the semiconductor device. Therefore, it is advisable to remove the substrate transfer processes (the first process and the nth process) from the process list 10_1. Also, for example, if the processing conditions specified in the (j+1)th process (processing process) and the processing conditions specified in the (j+2)th process (film formation process) are the same among the process lists 10_1 to 10_m, it is advisable to remove the (j+1)th process (processing process) and the (j+2)th process (film formation process) from the process list 10_1.

[0132] As a result of the above, some processes are extracted from the process list 10_1. Alternatively, some processes different from the above are removed from the process list 10_1. FIG. 6B shows an example in which the jth process, the (j+3)th process, etc. are extracted. Note that the example shown in FIG. 6B also applies to a case in which the 1st process, the (j+1)th process, the (j+2)th process, the nth process, etc. are removed.

[0133] Next, data of the processing conditions included in the process list 10_1 from which some processes have been extracted or in the process list 10_1 from which some processes different from the extracted some processes have been removed are quantified.

[0134] As described above, for example, the processing conditions for a film formation process include the equipment, material, film thickness, temperature, pressure, power, flow rate, etc. Since the film thickness, temperature, pressure, power, flow rate, etc. are set as values, these processing conditions are digitized data. Therefore, these processing conditions can be included in the input data 52_1 without any special conversion.

[0135] It is preferable that the set values ​​of each processing condition be in the same unit. By using the same unit, the amount of data included in the training dataset 50 can be reduced. This reduces the time spent on data transmission and reception, training, inference, etc.

[0136] Equipment data may be included in the process list as qualitative data, such as the equipment name (including abbreviations and names), the method used by the equipment, etc.

[0137] Furthermore, data on materials may be included in the process list as qualitative data, such as the name of the material (including abbreviations and names), composition formula, etc.

[0138] As described above, it is preferable that the data included in the training data set used in supervised learning be quantified, and therefore it is preferable to quantify the qualitative data included in the process list.

[0139] [Quantification of qualitative data related to the device] Here, the digitization of qualitative data relating to the equipment will be described, taking an example where the equipment name is input as qualitative data relating to the equipment in the condition 1 of the film formation process.

[0140] Film formation equipment includes equipment capable of forming films using the chemical vapor deposition (CVD) method (sometimes called CVD equipment), equipment capable of forming films using the sputtering method (sometimes called sputtering equipment), and equipment capable of forming films using the atomic layer deposition (ALD) method (sometimes called ALD equipment).

[0141] CVD methods can be classified into plasma-enhanced CVD (PECVD), which uses plasma; thermal CVD (TCVD), which uses heat; and photo-CVD (Photo-CVD), which uses light. Therefore, different CVD systems may be used depending on the CVD method. In other words, multiple CVD systems may be available. The same applies to sputtering systems, ALD systems, etc.

[0142] The data related to the equipment (here, the equipment name) input as a processing condition is qualitative data. Therefore, it is preferable to use Label Encoding to digitize the qualitative data related to the equipment. For example, it is good to manage the equipment name with an ID. It is good to assign an ID different from the process list ID to each equipment name. Here, the ID assigned to the equipment name is referred to as the equipment ID.

[0143] FIG. 7A shows an example of a correspondence table between device names and device IDs. For example, if the device name is CVD1, the device ID is 1. If the device name is CVD2, the device ID is 2. If the device name is SP1, the device ID is 3. By converting the device name into a device ID, the device name can be treated as numerical data.

[0144] The correspondence table may be stored in the storage unit 105. Furthermore, each time an available device is increased, a new device name and a device ID associated with the new device name may be added to the correspondence table via the input unit 101, a storage medium, communication, or the like.

[0145] Although the method of quantifying qualitative data about a device using Label Encoding has been described, the method of quantifying qualitative data about a device is not limited to this. One-hot encoding (also called 1 of K encoding) may also be used to quantify qualitative data about a device.

[0146] For example, if the number of devices that can be used in a film formation process is t (t is an integer greater than or equal to 1), the device names can be expressed as a t-dimensional one-hot vector. If the number of devices that can be used in each process is not large, they can be expressed as a low-dimensional vector. This allows for a reduction in the amount of calculation or time required for learning or inference.

[0147] Furthermore, as a method for converting qualitative data relating to a device into a numerical value, other methods such as Target Encoding may also be used.

[0148] The above is an explanation of how qualitative data relating to the device is quantified.

[0149] [Quantification of qualitative data on materials] Here, we will explain how to quantify qualitative data related to a material. Note that we will explain using an example where the name of the material (including abbreviations, names, etc.) is input as qualitative data related to the material in Condition 2 of the film formation process. Furthermore, we will assume that the material is an inorganic material.

[0150] The crystal structure, film quality, etc. of materials used in semiconductor devices change depending on processing conditions. Furthermore, they also change depending on the material used as the film to be formed and the roughness of the surface to be formed. Therefore, when using a database or the like to convert qualitative data about the material's name into quantitative data about the material's physical properties (crystal structure, density, dielectric constant, etc.), the accuracy of predicting the characteristics of semiconductor devices may decrease. Therefore, in this embodiment, qualitative data about the material (here, the material's name) is converted into constituent elements and composition.

[0151] First, the name of the material is converted into a composition formula. For example, if "silicon oxide" is entered as condition 2 of the film formation process, it should be converted into "SiO2." The conversion from the material name to the composition formula may be performed using a concept dictionary or database, or may be performed using a previously prepared correspondence table between material names and composition formulas.

[0152] Next, the composition formula is converted into the constituent elements and composition. For example, if a material is composed of elements M1, M2, M3, and M4 and has a composition of M1:M2:M3:M4=w:x:y:z, the composition formula is converted into "M1, M2, M3, M4, w:x:y:z" or "M1, w, M2, x, M3, y, M4, z."

[0153] It is preferable to standardize the composition. For example, the composition may be standardized so that w+x+y+z=1 is satisfied. This makes it possible to distinguish between materials that have the same combination of constituent elements but different compositions.

[0154] If a material is composed of one element, M2, M3, M4, x, y, and z should be written as zero. Similarly, if a material is composed of two elements, M3, M4, y, and z should be written as zero. Furthermore, if a material is composed of three elements, M4 and z should be written as zero.

[0155] Specifically, if the composition formula is "SiO2", it will be converted to "Si, O, 0, 0, 0.333:0.667: 0: 0" or "Si, 0.333, O, 0.667, 0, 0, 0, 0".

[0156] In the above, the number of elements and composition are described so that they can be applied to materials with four or fewer constituent elements, but this is not limited to this. For example, the number of elements and composition may be described so that they can be applied to materials with five or more constituent elements. Alternatively, for example, if the material used in a semiconductor device has three or fewer constituent elements, the composition formula may be converted to "M1, M2, M3, w:x:y" or "M1, w, M2, x, M3, y." This reduces the number of parameters in machine learning. This optimizes the number of hidden layers and the number of neurons in the hidden layers, reducing the amount of calculation or time required for learning or inference.

[0157] Next, the element is converted into its properties, such as atomic number, group, period, electron configuration, atomic weight, atomic radius, atomic volume, electronegativity, ionization energy, electron affinity, dipole polarizability, melting point of the element, boiling point of the element, lattice constant of the element, density of the element, and thermal conductivity of the element. The atomic radius may be one or more of the covalent bond radius, van der Waals radius, ionic radius, and metallic bond radius.

[0158] In particular, it is preferable to select the atomic number or electron configuration and electronegativity as the characteristics of the element to be converted. When a material is composed of a single element, the characteristics of the material are likely to be reflected in the atomic number and electronegativity. When a material is composed of two or more elements, the electronegativity is likely to be reflected in the bonding type between different elements. For example, between elements with similar electronegativity, covalent or metallic bonding will predominate. On the other hand, between elements with significantly different electronegativities, ionic bonding will predominate.

[0159] 7B shows an example of a correspondence table between elements and their properties. In FIG. 7B, the properties of elements include atomic number, electron configuration, electronegativity, melting point (K) of a simple substance, etc. Note that the conversion from an element to its properties may be performed using a database or a previously created correspondence table between elements and their properties.

[0160] Specifically, when selecting atomic number and electronegativity as the element properties, "Si" is converted to "14, 1.90", and "O" is converted to "8, 3.44".

[0161] As a result, "silicon oxide" entered as condition 2 for the film formation process can be converted to "14, 1.90, 8, 3.44, 0, 0, 0, 0, 0.333:0.667:0:0" or "14, 1.90, 0.333, 8, 3.44, 0.667, 0, 0, 0, 0, 0, 0." This allows qualitative data about materials to be quantified.

[0162] From the above, it is possible to predict the characteristics of semiconductor elements from a first-principles perspective. Furthermore, even when a material that has never been used before is used in a semiconductor element, it is possible to predict the characteristics of the semiconductor element without sacrificing accuracy. Furthermore, even when a material that is not listed in a database or the like is used in a semiconductor element, it is possible to predict the characteristics of the semiconductor element without sacrificing accuracy.

[0163] The name of the material may be directly converted into the constituent elements and composition without going through the composition formula.

[0164] This concludes the explanation of how to quantify qualitative data related to materials.

[0165] As a result, input data 52_1 consisting of digitized data as shown in Fig. 6C can be created. Specifically, the input data 52_1 includes data on the processing conditions shown in Fig. 6C. The input data 52_1 may also include a process number.

[0166] The order of the step of selecting (extracting or removing) the process list and the step of converting qualitative data into quantitative data is not limited to the above. For example, after converting qualitative data into quantitative data, the input data 52_1 may be created from the process list 10_1 by selecting (extracting or removing) the process list.

[0167] In this manner, training data 51_1 including digitized data can be generated. Note that training data 51_2 to training data 51_m have the same configuration as training data 51_1. That is, training data 51_2 to training data 51_m can be generated by the above method.

[0168] 4A illustrates a case where input data 52_1 to input data 52_m are generated from process lists 10_1 to 10_m, respectively, but this is not limiting. For example, as illustrated in FIG. 4B, input data 52_1 to input data 52_m may be generated from process lists 10_1 to 10_m and information on the shapes of semiconductor elements 30_1 to 30_m, respectively.

[0169] In addition, input data 52_1 to input data 52_m may be created from process list 10_1 to process list 10_m and first characteristics of semiconductor elements 30_1 to 30_m, respectively, and teacher data 53_1 to teacher data 53_m may be created from second characteristics of semiconductor elements 30_1 to 30_m, respectively.

[0170] In the above, the first characteristic and the second characteristic are different. For example, the first characteristic may be a characteristic value of the semiconductor element, and the second characteristic may be the result of a reliability test of the semiconductor element. There are many factors that affect the reliability of a semiconductor element, and these factors are intricately intertwined, making it difficult to predict based on experience. Therefore, the reliability of a semiconductor element is suitable as a target for estimation. Furthermore, the characteristic value of a semiconductor element indirectly includes information such as the manufacturing process of the semiconductor element. Therefore, by adding the characteristic value of the semiconductor element to the input data, this information is provided to supervised learning, thereby improving the accuracy of prediction of the characteristics of the semiconductor element.

[0171] The training dataset 50 may be composed of only data on semiconductor elements having the same or similar structure. In other words, the training dataset 50 may be created for each structure of the semiconductor element. This can improve the accuracy of prediction of the characteristics of the semiconductor element. The training dataset 50 may also be composed of data on semiconductor elements regardless of their structure. This allows for highly versatile prediction of the characteristics of semiconductor elements.

[0172] This concludes the explanation of the training dataset. By training a machine learning model using input data and training data, it is possible to predict the characteristics of semiconductor devices.

[0173] <<Data for predicting the characteristics of semiconductor elements>> Here, the prediction data for the characteristics of the semiconductor device will be described.

[0174] The prediction data for the characteristics of the semiconductor element is generated from the data IN2 input to the processing unit 102 shown in Figures 1A and 1B. Therefore, the prediction data for the characteristics of the semiconductor element is generated by extracting, processing, converting, selecting, removing, etc. the data included in the data IN2.

[0175] The data IN2 includes at least information about the semiconductor device, and may also include characteristics of the semiconductor device.

[0176] The prediction data for the characteristics of the semiconductor elements may have the same configuration as the input data for the learning data described above. For example, if the input data 52_1 to 52_m are created from the process lists 10_1 to 10_m, respectively, the prediction data for the characteristics of the semiconductor elements may be created from the process list 11. Also, if the input data 52_1 to 52_m are created from the process lists 10_1 to 10_m and the characteristics of the semiconductor elements 30_1 to 30_m, respectively, the prediction data for the characteristics of the semiconductor elements may be created from the process list 11 and the characteristics of the semiconductor elements associated with the process list ID of the process list 11.

[0177] The above is an explanation of the data for predicting the characteristics of semiconductor elements.

[0178] According to one embodiment of the present invention, the characteristics of a semiconductor element can be predicted without using the physical properties of materials contained in the semiconductor element. Furthermore, by using past experimental data, the optimization of the structure of a semiconductor element can be accelerated by virtual screening. Even if a person views the data and finds it not interpolative, it may be considered interpolative due to the nonlinear or high-order representation of the machine learning model. Furthermore, by extracting and examining fragments of the representation obtained by the machine learning model, it is possible to discover rules that were previously unnoticed.

[0179] <Computer equipment> In this section, a computer device having a semiconductor element characteristic prediction system, which is one embodiment of the present invention, will be described with reference to FIG.

[0180] 8 is a diagram illustrating a computer device having a semiconductor device characteristic prediction system. The computer device 1000 includes an arithmetic unit 1001, a memory 1002, an input / output interface 1003, a communication device 1004, and a storage device 1005. The computer device 1000 is electrically connected to a display device 1006a and a keyboard 1006b via the input / output interface 1003.

[0181] The computer device 1000 may be an information processing device such as a personal computer used by a user. In this case, the arithmetic device 1001 includes the processing unit 102 and the arithmetic unit 103 shown in FIGS. 1A and 1B. The storage 1005 includes the memory unit 105 and / or the storage unit 106 shown in FIGS. 1A and 1B. The display device 1006a corresponds to the output unit 104 shown in FIGS. 1A and 1B. The keyboard 1006b corresponds to the input unit 101 shown in FIGS. 1A and 1B.

[0182] The trained model may be stored in memory 1002 or in storage 1005.

[0183] The computer device 1000 may also be connected to a database 1011, a remote computer 1012, and a remote computer 1013 via a network. The computer device 1000 is electrically connected to a network interface 1007 via a communication device 1004. The network interface 1007 is also electrically connected to the database 1011, the remote computer 1012, and the remote computer 1013 via a network.

[0184] Here, the network includes a local area network (LAN) and the Internet. The network can use either or both of wired and wireless communication. When wireless communication is used in the network, various communication methods can be used, such as short-range communication methods such as Wi-Fi (registered trademark) and Bluetooth (registered trademark), as well as communication methods conforming to the third generation mobile communication system (3G), LTE (sometimes called 3.9G), fourth generation mobile communication system (4G), or fifth generation mobile communication system (5G).

[0185] As described above, the processing unit of the semiconductor device characteristic prediction system may be provided in a server, and may be accessed and used by a client PC via a network. For example, the computer device 1000 may be regarded as the client PC, and the remote computer 1012 and / or the remote computer 1013 may be regarded as the server.

[0186] In this case, the processing unit 102 and the calculation unit 103 shown in Figures 1A and 1B are provided in the remote computer 1012 and / or the remote computer 1013. That is, the calculation device included in the remote computer 1012 and / or the remote computer 1013 includes the processing unit 102 and the calculation unit 103. In addition, the database 1011 includes the memory unit 105 and / or the storage unit 106 shown in Figures 1A and 1B.

[0187] As described above, one aspect of the present invention can provide a semiconductor device characteristic prediction system. Another aspect of the present invention can provide a semiconductor device characteristic prediction method. Another aspect of the present invention can provide a training dataset for predicting semiconductor device characteristics.

[0188] This embodiment can be carried out by combining parts thereof as appropriate. [Explanation of symbols]

[0189] IN1: data, IN2: data, 10: multiple process lists, 10_m: process list, 10_1: process list, 11: process list, 20: multiple characteristics, 20_m: characteristics, 20_1: characteristics, 30: multiple semiconductor elements, 30_i: semiconductor element, 30_m: semiconductor element, 30_1: semiconductor element, 50: training data set, 51_i: training data, 51_m: training data, 51_1: training data, 51_2: training data, 52_i: input data, 52_m: input data, 52_1: input data, 53_i: training data, 53_m : Teacher data, 53_1: Teacher data, 100: Property prediction system, 101: Input unit, 102: Processing unit, 103: Calculation unit, 104: Output unit, 105: Memory unit, 106: Storage unit, 1000: Computer device, 1001: Calculation device, 1002: Memory, 1003: Input / output interface, 1004: Communication device, 1005: Storage, 1006a: Display device, 1006b: Keyboard, 1007: Network interface, 1011: Database, 1012: Remote computer, 1013: Remote computer

Claims

1. A semiconductor device characteristic prediction system that performs supervised learning based on a learning data set, and infers characteristics of the semiconductor device from prediction data based on a result of the learning, the semiconductor device characteristic prediction system includes a storage unit, an input unit, a processing unit, and an arithmetic unit; the processing unit has a function of creating the learning dataset from first data stored in the storage unit, a function of creating the prediction data from second data supplied from the input unit, a function of converting qualitative data into quantitative data, and a function of extracting or removing from the first data and the second data, the first data includes m sets (m is an integer of 2 or more) of a process list for a first semiconductor device and characteristics of the first semiconductor device manufactured in accordance with the process list for the first semiconductor device; the second data includes a process list for a second semiconductor device; The qualitative data is the name or composition formula of the material, The quantitative data is elemental properties and compositions; The calculation unit has a function of performing learning and inference in the supervised learning. A system for predicting semiconductor device characteristics.

2. In claim 1, The properties of the element are one or more of atomic number, group, period, electron configuration, atomic weight, atomic radius, atomic volume, electronegativity, ionization energy, electron affinity, dipole polarizability, melting point of the element, boiling point of the element, lattice constant of the element, density of the element, and thermal conductivity of the element. A system for predicting semiconductor device characteristics.

3. In claim 2, The atomic radius is one of a covalent radius, a van der Waals radius, an ionic radius, or a metallic radius. A system for predicting semiconductor device characteristics.

4. In any one of claims 1 to 3, The characteristic of the semiconductor element is a change over time in ΔVsh obtained in any one of a +GBT stress test, a +DBT stress test, a −GBT stress test, a +DGBT stress test, a +BGBT stress test, and a −BGBT stress test. A system for predicting semiconductor device characteristics.

5. In any one of claims 1 to 3, The characteristics of the semiconductor element are Id-Vg characteristics or Id-Vd characteristics. A system for predicting semiconductor device characteristics.

6. In any one of claims 1 to 5, The processing unit has a function of converting the qualitative data into a numerical value using Label Encoding. A system for predicting semiconductor device characteristics.

Images