Method of generating circuit model and manufacturing integrated circuit using the same

By generating a hybrid circuit model using machine learning and physical rule-based models, the method addresses the challenge of accurately predicting semiconductor device characteristics, enhancing simulation accuracy and reducing costs in integrated circuit manufacturing.

KR103046416B1Active Publication Date: 2026-11-10SAMSUNG ELECTRONICS CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
KR1020200135524
Authority / Receiving Office
KR · KR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2020-10-19
Publication Date
2026-11-10
Estimated Expiration
2040-10-19

AI Technical Summary

Technical Problem

Existing methods struggle to accurately predict the electrical characteristics of semiconductor devices in integrated circuits due to complex interactions during design and manufacturing, necessitating improved simulation techniques to enhance accuracy and reduce experimental costs.

Method used

A method involving generating a circuit model through classifying characteristic data of semiconductor devices, preprocessing the data, and extracting machine learning models to create a hybrid model that combines machine learning and physical rule-based models, ensuring consistency and accuracy in predicting electrical characteristics.

Benefits of technology

The method generates a high-performance circuit model that accurately predicts and optimizes the characteristics of integrated circuits, minimizing the disadvantages of machine learning and physical rule-based models while maximizing their advantages.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 112020110538161-PAT00001_ABST
    Figure 112020110538161-PAT00001_ABST
Patent Text Reader

Abstract

A method for generating a circuit model used to simulate an integrated circuit is disclosed. The method comprises the steps of: classifying characteristic data of a target semiconductor device according to measurement requirements to generate first characteristic element data and second characteristic element data; preprocessing the first characteristic element data and the second characteristic element data, respectively, to generate first target data and second target data; extracting a first machine learning model using the first target data and extracting a second machine learning model using the second target data; and generating a circuit model using the first machine learning model and the second machine learning model.
Need to check novelty before this filing date? Find Prior Art

Description

Technology Field

[0001] The technical concept of the present disclosure relates to a method for manufacturing an integrated circuit, which involves generating a circuit model, analyzing the characteristics of a semiconductor device using the generated circuit model, and designing an integrated circuit. Background Technology

[0002] As semiconductors become more highly integrated and miniaturized, factors at each stage of designing and manufacturing semiconductor devices interact in a complex manner, resulting in various unintended electrical characteristics in the devices. Consequently, there is a growing demand in the semiconductor industry for Technology Computer Aided Design (TCAD) process-device simulation environments based on physical simulation to overcome the limitations of semiconductor processes and devices, and to improve understanding of phenomena and reduce experimental costs. Furthermore, to provide accurate product specifications for semiconductor devices, it is necessary to predict and simulate the characteristics of the devices.

[0003] Modeling techniques can be used to estimate objects or phenomena that have causal relationships, and models generated through modeling techniques can be used to predict or optimize objects or phenomena. For example, machine learning models can be generated by training (or learning) based on vast sample data, and rule-based models can be generated by at least one rule defined based on physical laws, etc. The problem to be solved

[0004] The problem that the technical concept of the present disclosure aims to solve is a method for manufacturing an integrated circuit that can more accurately predict the electrical characteristics of a semiconductor device included in the integrated circuit by generating a circuit model with improved consistency with the characteristics of the semiconductor device. means of solving the problem

[0005] A method for generating a circuit model used to simulate an integrated circuit according to the technical concept of the present disclosure may include: a step of generating first characteristic element data and second characteristic element data by classifying characteristic data of a target semiconductor device according to measurement requirements; a step of generating first target data and second target data by preprocessing the first characteristic element data and second characteristic element data, respectively; a step of extracting a first machine learning model using the first target data and extracting a second machine learning model using the second target data; and a step of generating a circuit model using the first machine learning model and the second machine learning model.

[0006] A method for manufacturing an integrated circuit according to the technical concept of the present disclosure may include the steps of: classifying characteristic data of a target semiconductor device according to measurement requirements to generate first characteristic element data and second characteristic element data; preprocessing the first characteristic element data and second characteristic element data, respectively, to generate first target data and second target data; extracting a first machine learning model using the first target data; and generating a circuit model corresponding to the target semiconductor device using the first machine learning model.

[0007] A method for manufacturing an integrated circuit according to the technical concept of the present disclosure may include the steps of: classifying characteristic data of a target semiconductor device according to measurement requirements into characteristic element data based on mutually independent characteristic elements of the target semiconductor device; preprocessing characteristic element data corresponding to each of a plurality of characteristic elements to generate target data; extracting machine learning models corresponding to each of the plurality of characteristic elements using the target data; and generating a circuit model corresponding to the target semiconductor device using the machine learning models. Effects of the invention

[0008] According to the method for generating a circuit model and the method for manufacturing an integrated circuit including the same according to an exemplary embodiment of the present disclosure, a circuit model having higher performance can be generated, and accordingly, the accuracy of predicting and optimizing the characteristics of an integrated circuit by modeling can be improved.

[0009] According to the method for generating a circuit model and the method for manufacturing an integrated circuit including the same according to an exemplary embodiment of the present disclosure, a hybrid model can be provided in which the disadvantages of machine learning models and physical rule-based models are minimized and the advantages are maximized.

[0010] The effects obtainable from the exemplary embodiments of the present disclosure are not limited to those mentioned above, and other unmentioned effects can be clearly derived and understood by those skilled in the art to which the exemplary embodiments of the present disclosure belong from the description below. That is, unintended effects resulting from the implementation of the exemplary embodiments of the present disclosure can also be derived by those skilled in the art from the exemplary embodiments of the present disclosure. Brief explanation of the drawing

[0011] FIG. 1 is a flowchart illustrating a method for manufacturing an integrated circuit according to an exemplary embodiment of the present disclosure. Figure 2 is a diagram illustrating an example of an operation to generate a circuit model by modeling the characteristics of the semiconductor device of Figure 1. Figure 3 is a diagram illustrating an example of the characteristic element classification operation of Figure 2. Figure 4 is a diagram illustrating an example of the preprocessing operation of Figure 2. FIG. 5 is a flowchart illustrating a method for manufacturing an integrated circuit according to an exemplary embodiment of the present disclosure. FIG. 6 is a diagram illustrating an example of step S231 of FIG. 5. FIG. 7 is a flowchart illustrating an example of step S233 of FIG. 5. FIG. 8 is a diagram illustrating an example of a machine learning operation of FIG. 2 and an example of step S23 of FIG. 5. FIG. 9 is a flowchart illustrating a method for manufacturing an integrated circuit according to an exemplary embodiment of the present disclosure. FIG. 10 is a diagram illustrating an example of an operation to implement a machine learning model according to steps S242 to S244 of FIG. 9 as a circuit model in a simulation tool. Figure 11 is a diagram illustrating an example of the model compression operation of Figure 10, and is a diagram illustrating the case where the machine learning model is a neural network model. Figure 12 is a diagram illustrating an example of the model merging operation of Figure 10. FIG. 13 is a block diagram showing a computing system including a memory for storing a program according to an exemplary embodiment of the present disclosure. FIG. 14 is a block diagram illustrating a computer system accessing a storage medium storing a program according to an exemplary embodiment of the present disclosure. FIG. 15 is a flowchart illustrating a method for manufacturing an integrated circuit according to an exemplary embodiment of the present disclosure. Specific details for implementing the invention

[0012] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings.

[0013] FIG. 1 is a flowchart illustrating a method for manufacturing an integrated circuit according to an exemplary embodiment of the present disclosure. Specifically, the flowchart of FIG. 1 illustrates a method for designing an integrated circuit by performing simulation of the integrated circuit using a circuit model.

[0014] Referring to FIG. 1, in step S10, an operation to acquire characteristic data of a semiconductor device according to measurement requirements can be performed. For example, characteristic data of a semiconductor device can be acquired through a method of measuring electrical characteristics using a Test Elements Group (TEG). For example, characteristic data of a semiconductor device can be acquired by fabricating a TEG on a wafer and measuring electrical characteristics according to measurement requirements from the wafer having the TEG.

[0015] In step S20, an operation to generate a circuit model by modeling the characteristics of a semiconductor device may be performed. The circuit model generated in step S20 may be implemented in any computing system (e.g., 170 in FIG. 13) to model an object or phenomenon. For example, the circuit model may be implemented in a stand-alone computing system or in distributed computing systems capable of communicating with each other through a network, etc. Additionally, the circuit model may include a part implemented by a processor that executes a program containing a series of instructions, or a part implemented by logic hardware designed by logic synthesis. In this specification, a processor may refer to any hardware-implemented data processing device that includes a physically structured circuit to execute predefined operations, including instructions included in a program and / or operations expressed in code. For example, the data processing device may include a microprocessor, a CPU (central processing unit), a GPU (graphics processing unit), a NPU (neural processing unit), a processor core, a multi-core processor, a multiprocessor, an ASIC (application-specific integrated circuit), an ASIP (application-specific instruction-set processor), and a FPGA (field programmable gate array). In an exemplary embodiment, step S20 may include steps S21 through S24.

[0016] In step S21, an operation to classify characteristic data of a semiconductor device based on multiple independent characteristic elements may be performed, and characteristic element data may be generated accordingly. One characteristic of a semiconductor device may be composed of multiple characteristic elements, and characteristic data for said one characteristic may be classified into different characteristic element data corresponding to each of the multiple characteristic elements. For example, a resistor may be composed of a intrinsic resistance and a variable resistance, and characteristic data for said resistor may be classified into characteristic element data corresponding to the intrinsic resistance and characteristic element data for the variable resistance.

[0017] In step S22, an operation to generate target data by preprocessing the feature element data of each of the classified multiple feature elements may be performed. Subsequently, in step S23, a preprocessing operation to improve the accuracy of the machine learning model extracted may be performed. A detailed explanation of step S22 will be provided later in FIG. 2.

[0018] In step S23, machine learning models can be extracted using preprocessed feature element data. For example, in step S23, regression models corresponding to each of the different preprocessed feature element data can be selected, and machine learning models can be extracted by extracting the model parameters of each regression model through machine learning. A detailed explanation of step S23 will be provided later in Fig. 5.

[0019] In step S24, an operation to generate a circuit model using the extracted machine learning models may be performed. The circuit model may be programmed to be implemented in a simulation tool for simulating an integrated circuit. In this case, the simulation tool may be an Electronic Design Automation (EDA) tool, for example, a Simulation Program with Integrated Circuit Emphasis (SPICE) tool, and the extracted machine learning model may be included in a Process Design Kit (PDK) used by the tool.

[0020] In step S30, an operation to simulate an integrated circuit using the circuit model generated in step S20 can be performed. As illustrated in FIG. 15, after step S30, an operation to manufacture an integrated circuit by a semiconductor process can be performed based on the results of simulating the integrated circuit.

[0021] The method for manufacturing an integrated circuit according to the present disclosure can generate a circuit model used to simulate the electrical characteristics of the integrated circuit through machine learning. Accordingly, since a more accurate circuit model corresponding to the range of process requirements between different process requirements for manufacturing semiconductor devices is provided, the consistency of the circuit model is improved, and the electrical characteristics of the semiconductor devices included in the integrated circuit can be predicted more accurately.

[0022] FIG. 2 is a diagram illustrating an example of an operation to generate a circuit model by modeling the characteristics of a semiconductor device of FIG. 1. FIG. 3 is a diagram illustrating an example of a characteristic element classification operation of FIG. 2. FIG. 4 is a diagram illustrating an example of a preprocessing operation of FIG. 2.

[0023] Referring to FIG. 2, a characteristic element classification operation (121) can be performed using characteristic data (Y) of a target semiconductor device that changes according to changes in measurement requirements (X). One characteristic of the target semiconductor can be classified into first to nth characteristic elements, and the characteristic data (Y) can be classified into first to nth characteristic element data (Y1~Yn) corresponding to each of the first to nth characteristic elements. Each of the first to nth characteristic elements may be independent and may not affect each other.

[0024] For example, n may be a natural number greater than or equal to 2, but if it is difficult to classify one characteristic of the target semiconductor into multiple independent characteristic elements, n may be 1. As the characteristic data (Y) is classified into the first to nth characteristic element data (Y1 to Yn), each of the first to nth characteristic elements can be modeled individually, and the consistency of the modeling can be improved.

[0025] The measurement requirements (X) may include process requirements and operation requirements of the target semiconductor device. The process requirements may include technology requirements (T1, T2, …) that vary depending on the design tool used to design the target semiconductor device, and design requirements (D1, D2, …) based on the physical / structural characteristics of the target semiconductor device. For example, if the target semiconductor device is a transistor, the design requirements may include the length of the gate line, the width of the active region, the channel length, the device width, the doping profile, the oxide thickness, the oxide dielectric constant, the channel length modulation constant, the temperature at which the target semiconductor device operates, etc.

[0026] The operating requirements may include environmental requirements in which the target semiconductor device operates, and the operating requirements may include voltages (V1, V2, …) input to the target semiconductor device or currents (I1, I2, …) input to the target semiconductor device. For example, if the target semiconductor device is a transistor, the operating requirements may be the body-source voltage, drain-source voltage, and gate-source voltage of the transistor, etc.

[0027] Referring to FIGS. 2 and 3, for example, if the target semiconductor device for modeling is a capacitor which is a passive device, the capacitance values ​​can be obtained as characteristic data (Y) in step S10 of FIG. 1. The capacitance can be classified into area capacitance, which is a characteristic of the surface of the capacitor plate following the general formula for capacitance, and fingering capacitance, which is a characteristic of the edge of the capacitor plate. That is, area capacitance can be classified as a first characteristic element, and fingering capacitance can be classified as a second characteristic element. The characteristic data (Y) containing information about the capacitance can be classified into first characteristic element data (Y1) containing information about the area capacitance and second characteristic element data (Y2) containing information about the fingering capacitance.

[0028] Alternatively, for example, if the target semiconductor device for modeling is a resistor which is a passive component, the resistance values ​​can be obtained as characteristic data (Y) in step S10 of FIG. 1. The resistance may be classified such that the intrinsic resistance is classified as a first characteristic element and the variable resistance is classified as a second characteristic element. Accordingly, the characteristic data containing information about the resistance values ​​can be classified into first characteristic element data (Y1) containing information about the intrinsic resistance values ​​and second characteristic element data (Y2) containing information about the variable resistance values.

[0029] Alternatively, for example, if the target semiconductor device for modeling is a bipolar transistor which is an active device, information regarding the magnitude of the current flowing through the transistor can be obtained as characteristic data (Y) in step S10 of FIG. 1. The current can be classified into hole current and electron current, and the characteristic data (Y) can be classified into first characteristic element data (Y1) containing information about the hole current and second characteristic element data (Y2) containing information about the electron current.

[0030] Alternatively, for example, if the target semiconductor device for modeling is a unipolar transistor which is an active device, the current flowing through the transistor can be obtained as characteristic data (Y) in step S10 of FIG. 1. The current can be classified into transport current, body current, and tunneling current, and the characteristic data (Y) can be classified into first characteristic element data (Y1) containing information about the transport current, second characteristic element data (Y2) containing information about the body current, and third characteristic element data (Y3) containing information about the tunneling current.

[0031] Referring to FIGS. 2 and 4, a preprocessing operation (122) for the first to nth characteristic element data (Y1 to Yn) can be performed to improve the consistency of the circuit model. In the preprocessing operation (122), the first to nth characteristic element data (Y1 to Yn) corresponding to each of the measurement requirements (X) can be corrected by considering the parasitic components included in each of the first to nth characteristic element data (Y1 to Yn), and the preprocessed first to nth characteristic element data can be generated as the first to nth target data (YT1 to YTn). The first to nth characteristic element values ​​(Y1 to Yn) can be corrected by considering the parasitic resistance (Rpara) and parasitic capacitance (Cpara) occurring in the target semiconductor device.

[0032] For example, when voltages (V11~V1k) applied to a target semiconductor device change as measurement requirements (X), information regarding currents (I11~I1k) flowing through the target semiconductor device corresponding to each different voltage (V11~V1k) may be included in the first characteristic element data (Y1) of the characteristic data (Y). In this case, k may be a natural number greater than 3, but the present disclosure is not limited thereto, and k may be various numbers. As the preprocessing operation (122) is performed, additional currents (I(Rpara(V11))~I(Rpara(V1k))) caused by parasitic resistance (Rpara) are added to the currents (I11~I1k), thereby correcting the first characteristic element data (Y1) into the first target data (YT1).

[0033] Alternatively, for example, when voltages (V11~V1k) applied to the target semiconductor device as measurement requirements (X) change, information regarding the capacitances (C11~C1k) of the target semiconductor device corresponding to each different voltage (V11~V1k) may be included in the first characteristic element data (Y1) of the characteristic data (Y). As the preprocessing operation (122) is performed, the first characteristic element data (Y1) may be corrected into the first target data (YT1) by correcting using parasitic capacitances (Cpara(V11)~Cpara(V1k)) corresponding to each of the capacitances (C11~C1k).

[0034] Referring again to FIG. 2, a machine learning operation (123) for extracting first to i-th machine learning models (MM1 to MMi) using at least a portion of the first to n-th target data (YT1 to YTn) may be performed. Extracting the first to i-th machine learning models (MM1 to MMi) may include extracting first to i-th model parameters corresponding to each of the first to i-th machine learning models (MM1 to MMi). In an exemplary embodiment, i may be a natural number equal to n. Or, in an exemplary embodiment, i may be a natural number smaller than n, and the target data among the first to n-th target data (YT1 to YTn) for which the machine learning operation (123) is not performed may be used to generate the physical rule-based model of FIG. 10.

[0035] The machine learning operation (123) can be performed in various ways, and the machine learning models selected in the machine learning operation (123) can be trained in various ways. For example, the machine learning model may include an artificial neural network model, and the first to i machine learning models (MM1 to MMi) may be extracted by correcting the weights of the neural network model based on values ​​backpropagated from errors.

[0036] FIG. 5 is a flowchart illustrating a method for manufacturing an integrated circuit according to an exemplary embodiment of the present disclosure. FIG. 6 is a diagram illustrating an example of step S231 of FIG. 5. Step S23 of FIG. 5 may be an example of step S23 of FIG. 1 and may include steps S231 through S233.

[0037] Referring to FIG. 5, in step S231, the operation of selecting regression models corresponding to each of a plurality of characteristic elements can be performed. That is, regression models corresponding to each of different target data can be selected.

[0038] For example, regression models may include linear regression models, polynomial regression models, neural network models such as multilayer perceptrons (fully connected neural networks), and machine learning models such as deep learning models and reinforcement learning models. Neural network models may include, for example, convolutional neural network models and recurrent neural network models.

[0039] Referring to FIGS. 5 and 6, in step S231, a regression model that is most optimized to express the relationship between the measurement requirements (X) and the first to nth target data (YT1 to YTn) may be selected. For example, a linear regression model may be selected as the regression model most optimized for the first target data (YT1), a multilayer perceptron model may be selected as the regression analysis hypothesis most optimized for the second target data (YT2), a polynomial regression model may be selected as the hypothesis most optimized for the third target data (YT3), and a convolutional neural network model may be selected as the hypothesis most optimized for the nth target data (YTn).

[0040] In step S232, an operation to generate predicted characteristic data of a target semiconductor device based on the predicted characteristics of the target semiconductor device using selected regression models may be performed. In step S233, an operation to train a machine learning model based on the error between the predicted characteristic data and the target data may be performed. For example, an operation to correct the model parameters of the machine learning model to reduce the error between the predicted characteristic data and the target may be performed repeatedly. The machine learning model may be trained in various ways. For example, in step S233, automatic parameter optimization (backward propagation) through machine learning may be performed.

[0041] FIG. 7 is a flowchart illustrating an example of step S233 of FIG. 5. Step S233 may include steps S233_1 and S233_2.

[0042] Referring to FIG. 7, in step S233_1, an operation to calculate a loss function based on target data and predicted feature data may be performed. The loss function may be defined to evaluate the predicted feature data and may be referred to as a cost function. The loss function may be defined to increase as the predicted feature data moves further away from the target data; that is, the loss function may be defined to increase as the errors in the predicted feature data increase.

[0043] In step S233_2, an operation to correct the loss function based on physical guide rules may be performed. For example, the machine learning model may correct the loss function so that it satisfies the law of conservation of charge and the law of conservation of energy. Additionally, the loss function may be corrected for regularization, or the loss function may be corrected to minimize the error between the derivative value of the target data and the derivative value of the predicted feature data, or the loss function may be corrected to minimize the error between the integral value of the target data and the integral value of the predicted feature data.

[0044] In step S233_3, an operation to train a machine learning model so that the result of the loss function decreases can be performed.

[0045] Figure 8 is a diagram illustrating the machine learning operation of Figure 2 and step S23 of Figure 5.

[0046] Referring to FIG. 8, in the machine learning operation (123), a first predicted characteristic data (YP1) according to measurement requirements (X) can be generated by using a first regression model (RM1) corresponding to the first target data (YT1) of the first characteristic element. Additionally, in the machine learning operation (123), a second predicted characteristic data (YPi) according to measurement requirements (X) can be generated by using a second regression model (RMi) corresponding to the second target data (YTi) of the second characteristic element.

[0047] A first loss function (LF1) can be calculated based on the error between the first target data (YT1) and the first predicted characteristic data (YP1), and a second loss function (LFi) can be calculated based on the error between the second target data (YTi) and the second predicted characteristic data (YPi). At this time, the first loss function (LF1) can be corrected based on the first physical guide rule (PR1), and the second loss function (LFi) can be corrected based on the second physical guide rule (PRi).

[0048] Based on the backpropagated values ​​from the calculated first loss function (LF1) and the i-th loss function (LFi), respectively, the first model parameters of the first regression model (RM1) and the i-th model parameters of the i-th regression model (RMi) may be corrected. The first model parameters of the first regression model (RM1) and the i-th model parameters of the i-th regression model (RMi) may be corrected so that the values ​​of the first loss function (LF1) and the i-th loss function (LFi), respectively, become smaller. Finally, the first machine learning model (MM1) and the i-th machine learning model (MMi) according to the corrected first model parameters of the first regression model (RM1) and the i-th model parameters of the i-th regression model may be extracted as the result of the machine learning operation (123).

[0049] FIG. 9 is a flowchart illustrating a method for manufacturing an integrated circuit according to an exemplary embodiment of the present disclosure. Step S24 of FIG. 9 may be an example of Step S24 of FIG. 1 and may include steps S241 to S245.

[0050] Referring to FIG. 9, in step S241, an operation to determine whether the machine learning model extracted in step S23 satisfies a reference requirement may be performed. For example, it may be determined whether the inference time of the machine learning model exceeds a reference time, or whether the storage capacity for storing the machine learning model exceeds a reference value. If the inference time of the machine learning model exceeds a reference time, or if the storage capacity for storing the machine learning model as a circuit model exceeds a reference value, the load on simulating an integrated circuit using the machine learning model increases; therefore, steps S241 and S242 may be performed to prevent an excessive load from occurring in the simulation operation.

[0051] If the machine learning model does not satisfy the standard requirements, in step S242, a compression operation of the machine learning model may be performed to reduce the complexity of the machine learning model. A compression operation that can reduce the complexity of the machine learning model while maintaining the consistency of the machine learning model may be performed. Step S242 will be described in detail later in FIGS. 10 and FIGS. 11.

[0052] After the machine learning model satisfies the standard requirements or the model parameters are compressed, in step S243, an operation to construct an equivalent circuit corresponding to the target semiconductor device using the machine learning model may be performed. In step S244, an operation to modify the machine learning model so that the equivalent circuit satisfies physical laws may be performed. Steps S243 and S244 will be described later in FIG. 12.

[0053] In step S245, the operation of programming a machine learning model as a circuit model to correspond to a simulation program can be performed. For example, a machine learning model finally configured in the form of an equivalent circuit can be programmed into a circuit model in the form of an equivalent circuit using a model API (Application Program Interface) provided by commercial EDA software, thereby enabling circuit analysis in EDA.

[0054] FIG. 10 is a diagram illustrating an example of an operation in which a machine learning model according to steps S242 to S244 of FIG. 9 is implemented as a circuit model in a simulation tool. FIG. 11 is a diagram illustrating an example of a model compression operation of FIG. 10. FIG. 12 is a diagram illustrating an example of a model merging operation of FIG. 10.

[0055] Referring to FIG. 10, a model compression operation (242) may be performed to compress at least some of the first to i-th machine learning models (MM1 to MMi). For example, a compression operation (242) may be performed on machine learning models among the first to i-th machine learning models (MM1 to MMi) that do not satisfy the criteria requirements. FIG. 10 describes a case where all of the first to i-th machine learning models (MM1 to MMi) are compressed through the model compression operation (242) and the first to i-th compressed models (CM1 to CMi) are generated, but the present disclosure is not limited thereto. The model compression operation (242) may include an operation to reduce the number of model parameters of at least some of the machine learning models among the first to i-th machine learning models (MM1 to MMi).

[0056] Referring to FIG. 11, the model compression operation (242) may include a compression operation (242_1) that compresses according to a compression policy and an operation (242_2) that evaluates the compression policy. FIG. 11 describes the case where the machine learning model is a neural network model for convenience of explanation, but the present disclosure is not limited thereto.

[0057] A neural network model may include multiple layers. Each layer of a neural network model may include multiple nodes referred to as neurons. Each node or neuron may represent a computational unit having one or more inputs and outputs. Each input from multiple nodes of a layer may be supplied from each node of an adjacent layer. Similarly, the outputs of multiple nodes of a layer may be supplied to multiple nodes of an adjacent layer. Each node of a specific layer of a neural network model may be connected to at least some nodes of an adjacent layer.

[0058] A rate of reduction in the number of model parameters of a machine learning model (e.g., a pruning rate) can be determined according to the compression policy, and, for example, the compression policy can be determined through reinforcement learning. If the machine learning model is a neural network model, the compression policy can be determined from the state of each of the multiple layers, and whether to perform a compression operation can be determined sequentially for each of the multiple layers. The model compression operation (242) can determine whether to perform a compression operation (242_1) for the j-1 layer, then determine whether to perform a compression operation (242_1) for the j-th layer, and subsequently determine whether to perform a compression operation (242_1) for the last layer. For a specific layer where the performance of the compression operation (242_1) is determined, the number of nodes connected to the nodes of adjacent layers can be reduced. For example, the state of each of the multiple layers may include the layer index, the number of neurons, the model lightweighting index, the lightweighting budget used so far, the remaining lightweighting budget, whether compression was performed in the previous step, etc.

[0059] The operation (242_2) for evaluating the compression policy may evaluate the consistency of the compressed compression model based on the result of performing the compression operation (242_1), and may include an operation to change the compression policy value based on the evaluation result. For example, if the operation (242_2) for evaluating the compression policy determines that the consistency of the compressed compression model is less than the reference consistency according to the compression operation (242_1), the operation may include an operation to change the compression policy so that the ratio of reducing the number of model parameters is reduced.

[0060] Referring again to FIG. 10, a model merging operation (243) can be performed to merge the first to i-th compression models (CM1~CMi) and the first to m-th physical rule-based models (PM1~PMm), and to construct an equivalent circuit (EC) corresponding to a target semiconductor device using the merged model. For example, the sum of i and m may be the same as n in FIG. 2.

[0061] For example, machine learning models can be generated by training (or learning) based on vast sample data, and physical rule-based models can be generated by at least one rule defined based on physical laws, etc. Machine learning models and physical rule-based models may have different characteristics, and accordingly, they may have different fields of application as well as different advantages and disadvantages. Accordingly, in the method for manufacturing an integrated circuit according to the present disclosure, hybrid models including the first to i-th machine learning models (MM1 to MMi) and the first to m-th physical rule-based models (PM1 to PMm) are used in modeling the characteristics of a target semiconductor device, so that the disadvantages of machine learning models and physical rule-based models can be minimized and the advantages can be maximized.

[0062] However, unlike what is shown in FIG. 10, the circuit model may not include physical rule-based models, or it may include only machine learning models. Depending on the characteristics of the target semiconductor device, the target semiconductor device may be modeled to include only machine learning models, or the target semiconductor device may be modeled to include only physical rule-based models, or the target semiconductor device may be modeled as a hybrid model including both machine learning models and physical rule-based models.

[0063] A model compensation operation (244) can be performed to compensate the equivalent circuit (EC) so that it satisfies physical laws. The model compensation operation (244) may include an operation to modify at least one of the first to i compression models (CM1~CMi) and the first to m physical rule-based models (PM1~PMm) that configure the equivalent circuit (EC) so that the equivalent circuit (EC) satisfies Kirchhoff's current law and Kirchhoff's voltage law. A circuit model (RSCM) modified by the model compensation operation (244) may be generated.

[0064] Referring to FIG. 12, a transistor, which is a target semiconductor device, can be modeled as an equivalent circuit (EC) composed of first to fifth compression models (CM1 to CM5), first to third physical rule-based models (PM1 to PM3), and first to fourth ports. The first to fourth ports may be input / output ports of the target semiconductor device.

[0065] The first port may correspond to the gate of the transistor, the second port to the source of the transistor, the third port to the body of the transistor, and the fourth port to the drain of the transistor. For example, the current between the body and drain of the transistor may be modeled by a first compression model (CM1), the current between the body and channel of the transistor may be modeled by a second compression model (CM2), the current between the drain and source of the transistor may be modeled by a third compression model (CM3), the voltage between the gate and drain of the transistor may be modeled by a fourth compression model (CM4), and the voltage between the gate and source of the transistor may be modeled by a fifth compression model (CM5). Alternatively, for example, the voltage between the drain and body of the transistor may be modeled by a first physical rule-based model (PM1), the voltage between the source and body of the transistor may be modeled by a second physical rule-based model (PM2), and the current between the body and source of the transistor may be modeled by a third physical rule-based model (PM3).

[0066] An equivalent circuit (EC) can be configured to satisfy Kirchhoff's current law and Kirchhoff's voltage law based on each of the first to fourth ports. For example, the first to fifth compression models (CM1 to CM5) and the first to third physical rule-based models (PM1 to PM3) can be corrected so that the sum of the currents input to the first to fourth ports is zero. Or, for example, the first to fifth compression models (CM1 to CM5) and the first to third physical rule-based models (PM1 to PM3) can be corrected so that the sum of the voltages between each of the first to fourth ports is zero.

[0067] FIG. 13 is a block diagram illustrating a computing system including a memory for storing a program according to an exemplary embodiment of the present disclosure. At least some of the steps of generating a circuit model by modeling the characteristics of a target semiconductor device may be performed in the computing system (170). In exemplary embodiments, the computing system (170) may be referred to as a system for modeling the characteristics of a target semiconductor device.

[0068] The computing system (170) may be a fixed computing system, such as a desktop computer, workstation, server, etc., or a portable computing system, such as a laptop computer, etc. As illustrated in FIG. 13, the computing system (170) may include a processor (171), input / output devices (172), a network interface (173), RAM (random access memory) (174), ROM (read only memory) (175), and a storage device (176). The processor (171), input / output devices (172), network interface (173), RAM (174), ROM (175), and storage device (176) may be connected to a bus (177) and may communicate with each other through the bus (177).

[0069] The processor (171) may be referred to as a processing unit and may include at least one core capable of executing any instruction set (e.g., IA-32 (Intel Architecture-32), 64-bit extended IA-32, x86-64, PowerPC, Sparc, MIPS, ARM, IA-64, etc.), such as a microprocessor, application processor (AP), digital signal processor (DSP), or graphic processing unit (GPU). For example, the processor (171) may access memory, i.e., RAM (174) or ROM (175), through a bus (177) and execute instructions stored in RAM (174) or ROM (175).

[0070] RAM (174) may store a program (174_1) or at least a part thereof for performing a method of modeling the characteristics of a target semiconductor device and generating a circuit model, and the program (174_1) may enable the processor (171) to perform at least some of the steps included in the method of generating a circuit model. That is, the program (174_1) may include a plurality of instructions executable by the processor (171), and the plurality of instructions included in the program (174_1) may enable the processor (171) to perform at least some of the steps included in the aforementioned method.

[0071] The storage device (176) may not lose stored data even if the power supplied to the computing system (170) is cut off. For example, the storage device (176) may include a non-volatile memory device and may include a storage medium such as a magnetic tape, an optical disk, or a magnetic disk. Additionally, the storage device (176) may be detachable from the computing system (170). The storage device (176) may store a program (174_1) according to an exemplary embodiment of the present disclosure, and the program (174_1) or at least a part thereof may be loaded from the storage device (176) into RAM (174) before the program (174_1) is executed by the processor (171). Alternatively, the storage device (176) may store a file written in a programming language, and the program (174_1) or at least a part thereof generated from the file by a compiler, etc. may be loaded into RAM (174). Additionally, as illustrated in FIG. 13, the storage device (176) can store a database (176_1), and the database (176_1) may contain information necessary to perform a method for a hybrid model, for example, sample data.

[0072] The storage device (176) may store data to be processed by the processor (171) or data processed by the processor (171). That is, the processor (171) may generate data by processing the data stored in the storage device (176) according to the program (174_1), and may also store the generated data in the storage device (176).

[0073] The input / output devices (172) may include input devices such as a keyboard and a pointing device, and output devices such as a display device and a printer. For example, through the input / output devices (172), the user may trigger the execution of the program (174_1) by the processor (171), input learning data, and check result data.

[0074] The network interface (173) may provide access to a network outside the computing system (170). For example, the network may include a number of computing systems and communication links, and the communication links may include wired links, optical links, wireless links, or any other type of link.

[0075] FIG. 14 is a block diagram illustrating a computer system accessing a storage medium storing a program according to an exemplary embodiment of the present disclosure. At least some of the steps of modeling the characteristics of a target semiconductor device and generating a circuit model may be performed by the computer system (182). The computer system (182) may access a computer-readable medium (184) and execute a program (184_1) stored in the computer-readable medium (184). In an exemplary embodiment, the computer system (182) and the computer-readable medium (184) may be collectively referred to as a system for modeling the characteristics of a target semiconductor device.

[0076] The computer system (182) may include at least one computer subsystem, and the program (184_1) may include at least one component executed by the at least one computer subsystem. For example, the at least one component may include machine learning models described above with reference to the drawings, and may include a model trainer that trains or modifies machine learning models. The computer-readable medium (184) may include a non-volatile memory device similar to the storage device (176) of FIG. 13, and may include a storage medium such as a magnetic tape, an optical disc, or a magnetic disc. Additionally, the computer-readable medium (184) may be detachable from the computer system (182).

[0077] FIG. 15 is a flowchart illustrating a method for manufacturing an integrated circuit according to an exemplary embodiment of the present disclosure. Steps S40 and S50 of FIG. 15 may be performed after step S30 of FIG. 1.

[0078] In step S40, results of simulating the characteristics of an integrated circuit can be obtained. The simulation operation can be performed using a circuit model generated by performing a modeling operation of a target semiconductor device described in FIGS. 1 to 13 (e.g., S20 of FIG. 1). A circuit model with improved consistency can be generated through machine learning operations, and since consistency is further improved by merging the machine learning model with a physical rule-based model based on existing physical formulas, the accuracy of the simulation results regarding the characteristics of the integrated circuit can be improved.

[0079] In step S50, based on the simulation results, an operation to manufacture an integrated circuit including a target semiconductor device by a semiconductor process may be performed. For example, an integrated circuit may be manufactured by a semiconductor process applying the process parameters finally adjusted in step S50. The semiconductor process may include a front-end-of-line (FEOL) process and a back-end-of-line (BEOL) process using masks fabricated based on the integrated circuit. For example, FEOL may include steps such as planarizing and cleaning the wafer, forming trenches, forming wells, forming gate lines, and forming sources and drains. Additionally, BEOL may include steps such as silicidating gate, source, and drain regions, adding a dielectric, planarizing, forming holes, adding a metal layer, forming vias, and forming a passivation layer. The integrated circuit manufactured in step S50 can have characteristics similar to the simulation results obtained in step S40 due to the high consistency of the circuit model obtained in step S24 of Fig. 1. Therefore, the time and cost for manufacturing an integrated circuit with good characteristics can be reduced, and an integrated circuit with better characteristics can be manufactured.

[0080] As described above, exemplary embodiments have been disclosed in the drawings and specification. Although specific terms have been used to describe the embodiments in this specification, they are used only for the purpose of explaining the technical concept of the present disclosure and are not intended to limit the meaning or the scope of the present disclosure as defined in the claims. Therefore, those skilled in the art will understand that various modifications and equivalent alternative embodiments are possible therefrom.

Claims

Claim 1 A method for generating a circuit model used to simulate an integrated circuit, which is performed by a computing system comprising a processor and a memory that stores a program executed by said processor, comprising the steps of: classifying characteristic data of a target semiconductor device according to measurement requirements based on mutually independent first characteristic elements and second characteristic elements, and generating first characteristic element data corresponding to said first characteristic elements and second characteristic element data corresponding to said second characteristic elements; preprocessing said first characteristic element data and said second characteristic element data, respectively, to generate first target data and second target data; generating a first machine learning model using said first target data and extracting a second machine learning model using said second target data; and generating said circuit model using said first machine learning model and said second machine learning model. Claim 2 A method for generating a circuit model according to claim 1, wherein the step of generating the first target data and the second target data comprises the step of preprocessing the first characteristic element data and the second characteristic element data, respectively, based on at least one of the parasitic capacitance and parasitic resistance of the target semiconductor device. Claim 3 A method for generating a circuit model according to claim 1, wherein the step of extracting the first machine learning model and the second machine learning model comprises: selecting a first regression model corresponding to the first target data and selecting a second regression model corresponding to the second target data; generating first prediction characteristic data of the target semiconductor device using the first regression model and generating second prediction characteristic data of the target semiconductor device using the second regression model; and training the first machine learning model based on the error between the first prediction characteristic data and the first target data, and training the second machine learning model based on the error between the second prediction characteristic data and the second target data. Claim 4 A method for generating a circuit model according to claim 3, wherein the step of training the first machine learning model and the second machine learning model comprises: calculating a first loss function based on the first prediction feature data and the first target data, and calculating a second loss function based on the second prediction feature data and the second target data; correcting the first loss function to reflect a first physical guide rule and correcting the second loss function to reflect a second physical guide rule; and training the first machine learning model so that the result value of the first loss function is reduced, and training the second machine learning model so that the result value of the second loss function is reduced. Claim 5 A method for generating a circuit model according to claim 1, wherein the step of generating the circuit model using the first machine learning model and the second machine learning model further comprises the step of compressing the first machine learning model such that the number of model parameters of the first machine learning model is reduced when the inference time of the first machine learning model exceeds a reference time. Claim 6 A method for generating a circuit model according to claim 1, wherein the step of generating the circuit model using the first machine learning model and the second machine learning model further comprises the step of compressing the first machine learning model such that the number of model parameters of the first machine learning model is reduced when the storage capacity for storing the first machine learning model exceeds a reference value. Claim 7 A method for manufacturing an integrated circuit performed by a computing system comprising a processor and a memory for storing a program executed by said processor, comprising: a step of classifying characteristic data of a target semiconductor device according to measurement requirements based on mutually independent first characteristic elements and second characteristic elements, and generating first characteristic element data corresponding to said first characteristic element and second characteristic element data corresponding to said second characteristic element; a step of preprocessing said first characteristic element data and said second characteristic element data respectively to generate first target data and second target data; a step of extracting a first machine learning model using said first target data; and a step of generating a circuit model corresponding to said target semiconductor device using said first machine learning model. Claim 8 A method for manufacturing an integrated circuit according to claim 7, wherein the step of generating the first target data and the second target data comprises the step of preprocessing the first characteristic element data and the second characteristic element data, respectively, based on at least one of the parasitic capacitance and parasitic resistance of the target semiconductor device. Claim 9 A method for manufacturing an integrated circuit according to claim 7, wherein the measurement requirements include process requirements for the target semiconductor device and operation requirements for the environment in which the target semiconductor device operates. Claim 10 A method for manufacturing an integrated circuit according to claim 7, wherein the step of extracting the first machine learning model comprises: selecting a first regression model corresponding to the first target data; generating first prediction characteristic data of the target semiconductor device using the first regression model; and training the first machine learning model based on the error between the first prediction characteristic data and the first target data. Claim 11 A method for manufacturing an integrated circuit according to claim 10, wherein the step of training the first machine learning model comprises: a step of calculating a first loss function based on the first prediction feature data and the first target data; a step of correcting the first loss function to reflect a physical guide rule; and a step of training the first machine learning model so that the result value of the first loss function is reduced. Claim 12 A method for manufacturing an integrated circuit according to claim 7, wherein the step of generating a circuit model using the first machine learning model comprises the step of compressing the first machine learning model such that the number of model parameters of the first machine learning model is reduced according to a compression policy. Claim 13 A method for manufacturing an integrated circuit according to claim 12, wherein the step of generating a circuit model using the first machine learning model comprises the step of changing the compression policy according to the result of evaluating the consistency of the compressed first machine learning model. Claim 14 A method for manufacturing an integrated circuit according to claim 7, wherein the step of generating a circuit model using the first machine learning model comprises: a step of extracting a physical rule-based model corresponding to the second target data; and a step of generating the circuit model using the first machine learning model and the physical rule-based model. Claim 15 A method for manufacturing an integrated circuit according to claim 7, wherein the step of generating a circuit model using the first machine learning model comprises: a step of configuring an equivalent circuit corresponding to the target semiconductor device using the first machine learning model; and a step of modifying the first machine learning model so that the equivalent circuit satisfies physical laws. Claim 16 A method for manufacturing an integrated circuit performed by a computing system comprising a processor and a memory for storing a program executed by said processor, the method comprising: a step of classifying characteristic data of a target semiconductor device according to measurement requirements into characteristic element data based on a plurality of independent characteristic elements of said target semiconductor device; a step of preprocessing characteristic element data corresponding to each of said plurality of characteristic elements to generate target data; a step of extracting machine learning models corresponding to each of said plurality of characteristic elements using said target data; and a step of generating a circuit model corresponding to said target semiconductor device using said machine learning models. Claim 17 A method for manufacturing an integrated circuit according to claim 16, wherein the step of extracting machine learning models comprises: selecting regression models corresponding to the target data; generating predicted characteristic data of the target semiconductor device using the regression models; and training the machine learning models based on the error between the predicted characteristic data and the target data. Claim 18 A method for manufacturing an integrated circuit according to claim 17, wherein the step of training the machine learning models comprises: a step of calculating loss functions based on the prediction feature data and the target data; and a step of training the machine learning models such that the result value of the loss functions is reduced. Claim 19 A method for manufacturing an integrated circuit according to claim 16, wherein the step of generating a circuit model using the machine learning models comprises the step of compressing the machine learning models such that the number of model parameters of the machine learning models is reduced. Claim 20 A method for manufacturing an integrated circuit according to claim 16, wherein the step of generating a circuit model using the machine learning models comprises the step of constructing an equivalent circuit corresponding to the target semiconductor device using the machine learning models and the step of modifying the machine learning models so that the equivalent circuit satisfies physical laws.

Citation Information

Patent Citations

  • Data-mining-based knowledge extraction and visualization of analog / mixed-signal / custom digital circuit design flow

    US7707533B2

  • Semiconductor circuit design method and semiconductor circuit manufacturing method

    US8024674B2

  • Method and system for enhancing the yield in semiconductor manufacturing

    US8090464B2

  • Computer aided design apparatus, computer aided design program, computer aided design method for a semiconductor device and method of manufacturing a semiconductor circuit based on characteristic value and simulation parameter

    US8935146B2