WO2024071517A1

WO2024071517A1 - Neural compact modeling method and computing device

Info

Publication number: WO2024071517A1
Application number: PCT/KR2022/019333
Authority: WO
Inventors: 차예슬; 박정환; 황현석
Original assignee: 주식회사 알세미
Priority date: 2022-09-27
Filing date: 2022-12-01
Publication date: 2024-04-04
Also published as: KR102579974B1

Abstract

A neural compact modeling method performed by a processor is disclosed. The neural compact modeling method comprises the steps of: extracting a second vector for a second voltage dataset of a source device according to a first vector, which is extracted by applying, to an encoder, a first voltage dataset of the source device and a first current; comparing a second current, which is output by applying the second vector to a decoder, with a ground true to train the encoder and the decoder; generating a third vector by using a third voltage dataset of a target device, a third current, and the trained encoder; and estimating a fourth current for a fourth voltage dataset of the target device by using the third vector, the trained encoder, and the trained decoder.

Description

Neural compact modeling method and computing device

The present invention relates to a neural compact modeling method and computing device, and more specifically, to a neural compact modeling method and computing device using meta-learning techniques.

One semiconductor chip is implemented with multiple transistors. In the process of designing a single semiconductor chip, a netlist is created that describes the connectivity between transistors, capacitors, and resistors. The netlist and compact model are used in a simulator such as SPICE (Simulation Program for Integrated Circuit Emphasis) to simulate the operation of each of the transistors. The compact model is simple mathematical descriptions of the behavior of circuit elements that make up one semiconductor chip.

The existing semiconductor device modeling method (No. 10-2285516) required a large data set to train a neural network. Therefore, it takes a lot of time and money to secure a large data set.

The technical problem to be achieved by the present invention is to provide a neural compact modeling method and computing device that can perform neural compact modeling of a semiconductor device that is a target device using a limited amount of data of the target device.

The neural compact modeling method performed by a processor according to an embodiment of the present invention is a second voltage data set of the source device according to a first vector extracted by applying the first voltage data set and the first current of the source device to the encoder. Extracting a second vector for, applying the second vector and the second voltage data set to a decoder and comparing the second current and ground true output to train the encoder and the decoder, the target device generating a third vector using a third voltage data set, a third current, and the trained encoder, and generating a third vector of the target device using the third vector, the trained encoder, and the trained decoder. and estimating a fourth current for the four voltage data set.

The encoder includes an initial vector generation module that generates an initial vector of the first vector or the second vector, and updates the first vector by comparing the first voltage data set, the first vector, and the first current. It includes an updater module that outputs a value, and an attention module that compares the first voltage data set and the second voltage data set and outputs a second vector updated from the updated value of the first vector.

Extracting the second vector includes generating a first initial vector by applying the first voltage data set to the first neural network of the encoder, and applying the second voltage data set to the first neural network of the encoder. Generating a second initial vector, applying the first current, the first voltage data set, and the first initial vector to a second neural network of the encoder to output an update value of the first vector, Setting the update value of the second initial vector as the update value of the first vector and outputting the update value of the second vector, combining the update value of the first vector and the first voltage data set with the third input of the encoder updating the first initial vector and outputting the updated first vector by applying it to a neural network, and converting the first voltage data set, the second voltage data set, and the updated values of the second vector into the first voltage data set of the encoder. 3. It includes the step of updating the second initial vector by applying it to a neural network and outputting the updated second vector.

The neural compact modeling method includes a first output operation of applying the first current, the first voltage data set, and the updated first vector to a second neural network of the encoder to output a re-updated value of the first vector. performing a second output operation of setting the re-update value of the updated second vector to the re-update value of the first vector and outputting the re-update value of the second vector, Applying the re-updated vector value and the first voltage data set to a third neural network of the encoder to re-update the updated first vector and performing a third output operation of outputting the re-updated first vector, and applying the re-updated values of the first voltage data set, the second voltage data set, and the second vector to a third neural network of the encoder to re-update the updated second vector and generate the re-updated second vector. It may further include performing a fourth output operation to output.

In the neural compact modeling method, the first output operation, the second output operation, the third output operation, and the fourth output operation are repeatedly performed for an arbitrary number of times.

The second vector is expressed by compressing data about the width, length, and temperature of the source device.

The second vector is a function that maps the second voltage data set to data on the width, length, and temperature of the source device.

The second voltage data set is drain-source voltage, gate-source voltage, and bulk-source voltage.

In the process of extracting the second vector, the encoder parameters and the decoder parameters are not updated.

Generating the third vector includes generating a third initial vector by applying the third voltage data set to the first neural network of the trained encoder, and applying the fourth voltage data set to the first neural network of the trained encoder. Generating a fourth initial vector by applying it to a neural network; applying the third current, the third voltage data set, and the third initial vector to a second neural network of the trained encoder to generate an update value of the third vector Outputting, setting the update value of the fourth initial vector as the update value of the third vector and outputting the update value of the fourth vector, and the update value of the third vector and the third voltage data set. Applying to the third neural network of the trained encoder to update the third initial vector and outputting the updated third vector.

The step of estimating the fourth current includes outputting an updated fourth vector using the third voltage data set, the fourth voltage data set, and the third vector, and the updated fourth vector using the third voltage data set, the fourth voltage data set, and the third vector. and estimating a fourth current for the fourth voltage data set of the target device by applying it to a trained decoder.

A computing device according to an embodiment of the present invention includes a memory that stores instructions, and a processor that executes the instructions.

The instructions extract a second vector for the second voltage data set of the source device according to the first vector extracted by applying the first voltage data set and the first current of the source device to the encoder, and the second vector and The encoder and the decoder are trained by comparing the second current output by applying the second voltage data set to the decoder and the ground true, and the third voltage data set, the third current, and the trained encoder of the target device are used. A third vector is generated using the third vector, the trained encoder, and the trained decoder to estimate a fourth current for the fourth voltage data set of the target device.

The neural compact modeling method and computing device according to an embodiment of the present invention has the effect of generating an accurate neural compact model of a semiconductor device that is a target device even with a limited amount of data of the target device by using a meta-learning technique.

In order to more fully understand the drawings cited in the detailed description of the present invention, a detailed description of each drawing is provided.

Figure 1 shows a block diagram of a computing device for performing a neural compact modeling method according to an embodiment of the present invention.

Figure 2 shows a block diagram of a neural network for performing a neural compact modeling method according to an embodiment of the present invention.

Figure 3 shows a flowchart for explaining the neural compact modeling method according to an embodiment of the present invention.

Figure 4 shows an internal block diagram of the encoder of the neural network for the source device shown in Figure 2.

FIG. 5 shows a flowchart for explaining in detail the operation of extracting the second vector shown in FIG. 3.

FIG. 6 shows an internal block diagram of the encoder of the neural network for the target device shown in FIG. 2.

FIG. 7 shows a flowchart for explaining in detail the operation of extracting the third vector shown in FIG. 3.

FIG. 8 shows a flowchart for explaining in detail the operation of estimating the fourth current shown in FIG. 3.

Specific structural or functional descriptions of the embodiments according to the concept of the present invention disclosed in this specification are merely illustrative for the purpose of explaining the embodiments according to the concept of the present invention, and the embodiments according to the concept of the present invention are It may be implemented in various forms and is not limited to the embodiments described herein.

Since the embodiments according to the concept of the present invention can make various changes and have various forms, the embodiments will be illustrated in the drawings and described in detail in this specification. However, this is not intended to limit the embodiments according to the concept of the present invention to specific disclosed forms, and includes all changes, equivalents, or substitutes included in the spirit and technical scope of the present invention.

Terms such as first or second may be used to describe various components, but the components should not be limited by the terms. The above terms are used only for the purpose of distinguishing one component from another component, for example, without departing from the scope of rights according to the concept of the present invention, a first component may be named a second component, and similarly The second component may also be named the first component.

When a component is said to be "connected" or "connected" to another component, it is understood that it may be directly connected to or connected to the other component, but that other components may exist in between. It should be. On the other hand, when it is mentioned that a component is “directly connected” or “directly connected” to another component, it should be understood that there are no other components in between. Other expressions that describe the relationship between components, such as "between" and "immediately between" or "neighboring" and "directly adjacent to" should be interpreted similarly.

The terms used in this specification are merely used to describe specific embodiments and are not intended to limit the invention. Singular expressions include plural expressions unless the context clearly dictates otherwise. As used herein, “comprises.” Or “to have.” Terms such as are intended to designate the presence of the described feature, number, step, operation, component, part, or combination thereof, but are not intended to indicate the presence of one or more other features, numbers, steps, operations, components, parts, or combination thereof. It should be understood that it does not exclude in advance the existence or possibility of addition of things.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person of ordinary skill in the technical field to which the present invention pertains. Terms as defined in commonly used dictionaries should be interpreted as having meanings consistent with the meanings they have in the context of the related technology, and unless clearly defined in this specification, should not be interpreted in an idealized or overly formal sense. No.

Hereinafter, the present invention will be described in detail by explaining preferred embodiments of the present invention with reference to the accompanying drawings.

Referring to FIG. 1, the computing device 10 for performing the neural compact modeling method can generate an accurate neural compact model of the semiconductor device that is the target device even with a limited amount of data from the target device by using a meta-learning technique. . Computing device 10 may be an electronic device such as a server, computer, laptop, tablet PC, or personal PC. Computing device 10 includes a processor 11 and memory 13. The processor 10 executes instructions implementing the neural compact modeling method. The memory 20 stores instructions implementing the neural compact modeling method. Hereinafter, a specific neural compact modeling method is disclosed. The neural compact modeling method refers to creating a compact model using a neural network.

Referring to Figures 1 and 2, the neural network 100 is a neural network for the source device. The source device refers to a semiconductor device that has a lot of data on voltage data sets (21, 25) and drain currents (23, 27).

Neural network 200 is a neural network for the target device. The target device refers to a semiconductor device corresponding to the drain current 77 of which the voltage data set 75 is desired. The semiconductor device may be a transistor.

The purpose of the present invention is to predict the drain current of the target device using a neural compact modeling method. That is, the purpose of the present invention is to generate a compact model for the target device using a neural compact modeling method. In creating a compact model for the target device, the source device's voltage data sets (21, 25) and drain currents (23, 27) are used. Neural compact modeling means that compact modeling is performed using a neural network.

When compared to the data numbers of the voltage data sets (71, 75) and drain currents (73, 77) of the target device with the data numbers of the voltage data sets (21, 25) and drain currents (23, 27) of the source device. , relatively few.

Therefore, in the present invention, the neural network 100 is first trained using the

voltage data sets

21 and 25 and the

drain currents

23 and 27 of the source device to create a compact model for the target device with limited data. .

Neural network 100 includes an encoder 20 and a decoder 60.

Referring to FIGS. 1 to 3, the processor 11 applies the first voltage data set 21 and the first current 23 of the source device to the encoder 20 to extract the first vector (31 in FIG. 4) ), the updated second vector 51 for the second voltage data set 25 of the source device is extracted (S100). The extraction refers to the output. The first current 23 refers to the drain current of the source device. That is, the processor 11 applies the first voltage data set 21 and the first current 23 of the source device to the encoder 20 to extract the first vector (31 in FIG. 4) and the first vector (31 in FIG. 4) of the source device. The updated second vector 51 is extracted by comparing the second vector 51 generated by applying the second voltage data set 25 to the encoder 20. In other words, the processor 11 applies the first voltage data set 21 and the first current 23 of the source device to the encoder 20 to extract the first vector (31 in FIG. 4) and the first vector of the source device. An updated second vector 51 is extracted from the second vector 51 generated by applying the first voltage data set 21 and the second voltage data set 25 of the source device to the encoder 20. Applying the first vector (31 in FIG. 4) to the encoder 20 means that the updated second vector 51 is extracted using the first vector (31 in FIG. 4). The updated second vector 51 can use Equation 2 below.

The first vector 31 is a function that maps the first voltage data set 21 to data about the width, length, and temperature of the source device. The second vector 51 is a function that maps the second voltage data set 25 to data on the width, length, and temperature of the source device.

The processor 11 applies the second vector 51 and the second voltage data set 25 to the decoder 60 and compares the second current 27 and ground true output to the encoder 20 and the decoder 60. ) train (S200). The second current 27 refers to the drain current of the source device. The encoder 20 and decoder 60 are trained so that the difference between the second current 27 and ground true is minimized.

The processor 11 generates the third vector 91 using the third voltage data set 71, the third current 73, and the trained encoder 80 of the target device (S300). That is, the processor 11 generates the third vector 91 by applying the third voltage data set 71 and the third current 73 of the target device to the trained encoder 80.

The processor 11 uses the third vector (91 in FIG. 6), the trained encoder 80, and the trained decoder 130 to calculate the fourth voltage data set 75 of the target device. Estimate the current 77 (S400). The processor 11 applies the third vector (91 in FIG. 6), the third voltage data set 71, and the fourth voltage data set 75 to the trained encoder 80 to generate an updated fourth vector ( 121), and apply the updated fourth vector 121 and the fourth voltage data set 75 to the trained decoder 130 to obtain the fourth voltage data set 75 of the target device. 4Assume current (77). Applying the third vector (91 in FIG. 6) to the trained encoder 80 means that the updated fourth vector 121 is generated using the third vector (91 in FIG. 6). The updated fourth vector 121 can be generated using Equation 2 below.

Referring to FIGS. 1 to 4 , the encoder 20 includes an initial vector generation module 30, an updater module 40, and an attention module 50.

The initial vector generation module 30, the updater module 40, and the attention module 50 are each implemented with multilayer perception (MLP).

FIG. 5 shows a flowchart for explaining in detail the operation (S100) of extracting the second vector shown in FIG. 3.

Referring to FIGS. 1 to 5, the initial vector generation module 30 receives the first voltage data set 21 of the source device or the second voltage data set 25 of the source device and generates the first vector 31. The initial vector of or the initial vector of the second vector 51 is generated.

The first voltage data set 21 includes drain-source voltage (V _DS ), gate-source voltage (V _GS ), and bulk-source voltage (V _BS ) for the source device. The second voltage data set 25 includes drain-source voltage (V _DS ), gate-source voltage (V _GS ), and bulk-source voltage (V _BS ) for the source device. The first voltage data set 21 and the first current 23 may be referred to as a support set. The second voltage data set 25 and the second current 27 may be referred to as a query set. A set of current and voltage data for the width, length, and temperature of the source device is defined as a task. Multiple tasks may be used. Each task can be divided into a support set and a query set.

The first vector 31 or the second vector 51 is a vector expressed by compressing data about the width, length, and temperature of the source device. The first vector 31 or the second vector 51 is not a fixed value, but is in the form of a function whose value changes depending on the voltage. Therefore, by using the form of a function, current characteristics that vary in various voltage regions (e.g., linear region, saturation region, and cut-off region) can be more appropriately measured. It can be reflected clearly.

The initial vector of the first vector 31 and the initial vector of the second vector 51 mean that the first vector 31 or the second vector 51 is initialized. That is, the processor 11 generates the first initial vector 31 by applying the first voltage data set 21 to the first neural network of the encoder 20 (S110). The processor 11 generates a second initial vector 51 by applying the second voltage data set 25 to the first neural network of the encoder 20 (S115). The first neural network refers to the initial vector generation module 30. The first initial vector 31 refers to the vector in which the first vector 31 is initialized. The second initial vector 51 refers to the vector in which the second vector 51 is initialized.

The updater module 40 compares the first voltage data set 21, the first vector 31, and the first current 23 and outputs an updated value of the first vector 31 (S120). Comparing the first voltage data set 21, the first vector 31, and the first current 23 refers to the first voltage data set 21, the first vector 31, and the first current 23. ) is applied to the neural network to output the updated value of the first vector (31). The neural network refers to the updater module 40. The update value of the first vector 31 means a specific value of the function. That is, the processor 11 applies the first initial vector of the first current 23, the first voltage data set 21, and the first vector 31 to the second neural network of the encoder 20 to generate the first vector. Outputs the update value of (31). The second neural network refers to the updater module 40.

The attention module 50 compares the first voltage data set 21 and the second voltage data set 25 and outputs a second vector 25 updated from the updated value of the first vector 21.

Hereinafter, the operations of the attention module 50 will be described in detail.

The processor 11 sets the update value of the second initial vector 25 to the update value of the first vector 21 and outputs the update value of the second vector 25 (S125). That is, the processor 11 sets the update value of the first vector 21 to the update value of the second initial vector 25 and outputs the update value of the second vector 25. The update value of the second vector 25 is set to the update value of the first vector 21.

The second initial vector 25 refers to the vector from which the second vector 25 was initialized. The output of the update value of the second vector 25 may be performed by the attention module 50.

The processor 11 applies the updated value of the first vector 21 and the first voltage data set 21 to the third neural network of the merchant coder 20 to update the first initial vector 31 and update the first initial vector 31. Output vector 31 (S130). The third neural network refers to the attention module 50. The updated first vector 31 can be expressed as Equation 1 below.

[Equation 1]

Updated support representation = initial support representation + alpha * ATTENTION(querys=support V, keys=support V, values=support representation initial update value)

In Equation 1, the updated support representation represents the updated first vector 31, the initial support representation represents the first initial vector 31 where the first vector 31 is an initialized vector, and alpha represents the learning rate. (learning rate), the ATTENTION represents MLP (Multilayer perception), the queries, the keys, and the value represent vectors in the attention of a neural network, and the support V represents the first voltage data set 21. The support representation initial update value represents the update value of the first vector 31.

The processor 11 applies the updated values of the first voltage data set 21, the second voltage data set 25, and the second vector 51 to the third neural network of the encoder 20 to create a second initial vector ( 51) is updated and the updated second vector 51 is output (S135). In FIG. 3, it is shown that the first voltage data set 21 and the second voltage data set 25 are not input to the attention module 50, but in reality, the first voltage data set 21 is input to the attention module 50. It should be understood that is input. The updated second vector 51 can be expressed as Equation 2 below.

[Equation 2]

Updated query representation = initial query representation + alpha * ATTENTION(querys=query V, keys=support V, values=query representation initial update value)

In Equation 2, the updated query representation represents the updated second vector 51, the initial query representation represents the second initial vector 51 where the second vector 51 is an initialized vector, and alpha represents the learning rate. (learning rate), the ATTENTION represents MLP (Multilayer perception), the queries, the keys, and the value represent vectors in the attention of a neural network, and the query V represents the second voltage data set 25. The support V represents the first voltage data set (21), and the query representation initial update value represents the update value of the second vector (51).

The processor 11 applies the first current 23, the first voltage data set 21, and the updated first vector 31 to the second neural network 40 of the encoder 20 to generate a first vector ( The first output operation to output the re-updated value of 31) is performed (S140).

The processor 11 performs a second output operation of setting the re-update value of the updated second vector 51 to the re-update value of the first vector 31 and outputting the re-update value of the second vector 51. Perform (S145). That is, the processor 11 sets the re-update value of the first vector 31 to the updated re-update value of the second vector 51 and outputs the re-update value of the second vector 51. Perform the action. The updated re-update value of the second vector 51 is set to the re-update value of the first vector 31.

The output of the update value of the second vector 25 may be performed by the attention module 50.

The processor 11 applies the re-updated value of the first vector 31 and the first voltage data set 21 to the third neural network 50 of the encoder 20 to regenerate the updated first vector 31. A third output operation is performed to output the updated and re-updated first vector 31 (S150).

The processor 11 applies the re-updated values of the first voltage data set 21, the second voltage data set 25, and the second vector 51 to the third neural network 50 of the encoder 20 to A fourth output operation is performed to re-update the updated second vector 51 and output the re-updated second vector 51 (S155). That is, the re-updated second vector 51 is extracted.

The processor 11 repeatedly performs the first output operation (S140), the second output operation (S145), the third output operation (S150), and the fourth output operation (S155) for an arbitrary number of times. . That is, the processor 11 performs the first output operation (S140), the second output operation (S145), the third output operation (S150), and the fourth output operation (S155) N times (N is a natural number). ) Determine whether or not it was performed repeatedly. When the processor 11 determines that the operations have been performed repeatedly more than N times, it terminates the operations. Therefore, even if there is no current value for the second voltage data set 25, the second vector 51 for the second voltage data set 25 can be extracted. The arbitrary number of times may be set arbitrarily.

In the process of extracting the second vector 51 (S110 to S155), the parameters (weights and bias) of the encoder 20 and the parameters of the decoder 60 are not updated. . That is, the operation of extracting the second vector 51 is not performed in backpropagation in the training stage of the neural network 100.

After extracting the second vector 51 for the second voltage data set 25 of the source device (S100), the processor 11 converts the second vector 51 and the second voltage data set 25 into a decoder ( 60), the encoder 20 and the decoder 60 are trained by comparing the second current 27 output and the ground true (S200). Backpropagation is performed. The parameters (weights and bias) of the encoder 20 and decoder 60 are updated.

After the encoder 20 and the decoder 60 are trained (S200), the processor 11 uses the third voltage data set 71, the third current 73, and the trained encoder 80 of the target device. A third vector (91 in FIG. 6) is generated (S300).

Neural network 200 includes a trained encoder 80 and a trained decoder 130. The trained encoder 80 and trained decoder 130 refer to the encoder 20 and decoder 60 trained in the neural network 100. That is, the encoder 20 and decoder 60 whose parameters are updated through training in the neural network 100 are the trained encoder 80 and the trained decoder 130.

Referring to FIGS. 2 and 6 , the encoder 80 includes an initial vector generation module 90, an updater module 110, and an attention module 120.

The initial vector generation module 90, the updater module 110, and the attention module 120 are each implemented with multilayer perception (MLP). The initial vector generation module 90, the updater module 110, and the attention module 120 are the initial vector generation module 30, the updater module 40, and the attention module 50 through training of the neural network 100. This is a module with updated parameters.

FIG. 7 shows a flowchart for explaining in detail the operation (S300) of extracting the third vector shown in FIG. 3.

1 to 7, the processor 11 generates a third initial vector 91 by applying the third voltage data set 71 to the first neural network of the trained encoder 80 (S310). . The first neural network represents the trained initial vector generation module 90.

The processor 11 generates a fourth initial vector 121 by applying the fourth voltage data set 75 to the first neural network 90 of the trained encoder 80 (S315).

The third initial vector 91 refers to the vector in which the third vector 91 is initialized. The fourth initial vector 121 refers to the vector in which the fourth vector 121 is initialized.

The third vector 91 or the fourth vector 121 is a vector expressed by compressing data about the width, length, and temperature of the target device. The third vector 91 or the fourth vector 121 is not a fixed value, but is a function whose value changes depending on the voltage.

The third voltage data set 71 includes drain-source voltage (V _DS ), gate-source voltage (V _GS ), and bulk-source voltage (V _BS ) for the target device. The fourth voltage data set 75 includes drain-source voltage (V _DS ), gate-source voltage (V _GS ), and bulk-source voltage (V _BS ) for the target device.

The third voltage data set 71 and the third current 73 may be referred to as a support set. The third current 73 refers to the drain current for the target device.

The fourth voltage data set 75 and the fourth current 77 may be referred to as a query set. The fourth current 77 refers to the drain current for the target device.

The amount of data in the third voltage data set 71 and third current 73 for the target device is limited. The data amount of the third voltage data set 71 and the third current 73 for the target device is less than the data amount of the first voltage data set 21 and the first current 23 for the source device.

The processor 11 applies the third current 73, the third voltage data set 71, and the third initial vector 91 to the second neural network of the trained encoder 80 to generate a third vector 91. Output the updated value (S320). The second neural network represents the trained updater module 110. The update value of the third vector 91 means a specific value of the function.

The processor 11 sets the update value of the fourth initial vector 121 to the update value of the third vector 91 and outputs the update value of the fourth vector 121 (S325).

That is, the processor 11 sets the update value of the third vector 91 to the update value of the fourth initial vector 121 and outputs the update value of the fourth vector 121. The update value of the fourth vector 121 is set to the update value of the third vector 91.

The output of the update value of the fourth vector 121 can be performed by the trained attention module 120. The update value of the fourth vector 121 means a specific value of the function.

The processor 11 updates the third initial vector 91 by applying the updated value of the third vector 91 and the third voltage data set 71 to the third neural network 120 of the trained encoder 80. The updated third vector 91 is output (S330). Equation 1 above can be applied to the operation (S330) of outputting the updated third vector 91.

At this time, in Equation 1, the updated support representation represents the updated third vector 91, the initial support representation represents the third initial vector 91, where the third vector 91 is an initialized vector, and the alpha is the learning rate, the ATTENTION is MLP (Multilayer perception), the queries, the keys, and the value are vectors in the attention of the neural network, and the support V is the third voltage data set 71. , the support representation initial update value represents the update value of the third vector 91.

FIG. 8 shows a flowchart for explaining in detail the operation (S400) of estimating the fourth current shown in FIG. 3.

Referring to FIGS. 1 to 8 , the operation S400 of estimating the fourth current includes an operation S410 of outputting a fourth vector and an operation S440 of estimating the fourth current.

The processor 11 outputs the updated fourth vector 121 using the third voltage data set 71, the fourth voltage data set 75, and the third vector 91 (S410).

Equation 2 above may be applied to the operation (S410) of outputting the updated fourth vector 121. At this time, in Equation 2, the updated query representation represents the updated fourth vector 121, the initial query representation represents the fourth initial vector 121 in which the fourth vector 121 is an initialized vector, and the alpha is the learning rate, the ATTENTION is MLP (Multilayer perception), the queries, the keys, and the value are vectors in the attention of the neural network, and the query V is the fourth voltage data set 75. , the support V represents the third voltage data set 71, and the query representation initial update value represents the update value of the fourth vector 121.

The processor 11 applies the third voltage data set 71 and the third current 73 to the trained encoder 80 and applies the fourth voltage data set 75 of the target device according to the extracted third vector 91. ) outputs the updated fourth vector 121. That is, the processor 11 applies the third voltage data set 71 and the third current 73 of the target device to the trained encoder 80 to extract the third vector 91 and the third vector 91 of the target device. The updated fourth vector 121 is extracted by comparing the fourth vector 121 generated by applying the 4-voltage data set 75 to the trained encoder 80. In other words, the processor 11 applies the third voltage data set 71 and the third current 73 of the target device to the trained encoder 80 to obtain the extracted third vector 91 and the third vector 91 of the target device. The updated fourth vector 121 is extracted from the voltage data set 71 and the fourth vector 121 generated by applying the fourth voltage data set 75 of the target device to the trained encoder 80. . Applying the third vector 91 to the trained encoder 80 means that the updated fourth vector 121 is extracted using the third vector 91. The updated third vector 91 is generated using Equation 2 above.

Hereinafter, detailed operations for outputting the fourth vector 121 will be described.

The operation (S410) of outputting the fourth vector 121 includes the fifth output operation (S415), the sixth output operation (S420), the seventh output operation (S425), the eighth output operation (S430), and the judgment operation ( S435).

The processor 11 applies the third current 73, the third voltage data set 71, and the updated third vector 91 to the second neural network 110 of the trained encoder 80 to generate a third A fifth output operation is performed to output the re-updated value of the vector 91 (S415). The second neural network 110 is a trained updater module.

The processor 11 performs a sixth output operation of setting the re-update value of the updated fourth vector 121 as the re-update value of the third vector 91 and outputting the re-update value of the fourth vector 121. Perform (S420). That is, the processor 11 sets the re-update value of the third vector 91 to the updated re-update value of the fourth vector 121 and outputs the re-update value of the fourth vector 121. Perform the action. The re-update value of the fourth vector 121 is set to the re-update value of the third vector 91.

The output of the update value of the fourth vector 121 may be performed in the attention module 120.

The processor 11 applies the re-updated value of the third vector 91 and the third voltage data set 71 to the third neural network 120 of the trained encoder 80 to generate the updated third vector 91. A seventh output operation is performed to re-update and output the re-updated third vector 91 (S425).

The processor 11 applies the re-updated values of the third voltage data set 71, the fourth voltage data set 75, and the fourth vector 121 to the third neural network 120 of the trained encoder 80. Then, the updated fourth vector 121 is re-updated and an eighth output operation is performed to output the re-updated fourth vector 121 (S430). That is, the re-updated fourth vector 121 is extracted.

The processor 11 repeatedly performs the fifth output operation (S415), the sixth output operation (S420), the seventh output operation (S425), and the eighth output operation (S430) for an arbitrary number of times. . That is, the processor 11 performs the fifth output operation (S415), the sixth output operation (S420), the seventh output operation (S425), and the eighth output operation (S430) M times (M is a natural number). ) Determine whether the above was performed repeatedly (S435). When the processor 11 determines that the operations have been performed repeatedly M times or more, it ends. The arbitrary number of times may be set arbitrarily.

The processor 11 applies the updated fourth vector 121 and the fourth voltage data set 75 to the trained decoder 130 to generate a fourth signal for the fourth voltage data set 75 of the target device. Estimate the current 77 (S440).

There is no separate training process for the neural network 200. In the case of conventional transfer learning, fine tuning is required. In the case of registered patent No. 10-2285516, referring to paragraph number [0085], the parameters of the second neural network 175 are updated. Fine tuning is required for the second neural network 175 to perform a backpropagation operation. However, in the case of the present invention, separate fine tuning is not performed on the neural network 200. Therefore, the drain current of the target device can be estimated faster than transfer learning.

The table below shows compact modeling methods using three different techniques. Random initialization refers to a compact modeling method using a single neural network. In other words, random initialization is a method in which a neural network is randomly initialized and trained using limited data from the target device. Transfer learning refers to a compact modeling method according to registered patent No. 10-2285516. Meta-learning refers to a compact modeling method according to the present invention.

In Table 1 below, the pre-training time means the training time of the neural network 100 in the present invention. In the case of registered patent No. 10-2285516, it refers to the training time of the neural network 160. Adaptation time refers to the time until the current state of the target device is ready to be estimated using the neural network 200 in the present invention. In the case of registered patent No. 10-2285516, it refers to the time when the backpropagation operation is completed for the neural network 175 or the neural network 185. In the present invention, the backpropagation operation of the neural network 200 is not performed.

Referring to Table 1 below, it can be seen that the present invention not only requires a short adaptation time but also has a low error compared to transfer learning.

	Random initializationRandom initialization	Tramsfer LearningTramsfer Learning	Meta learningMeta-learning
Pretraining TimePretraining Time	N/AN/A	646 sec.646sec.	17 hours17 hours
Adapation Time(per W/L/T)Adaptation Time(per W/L/T)	538 sec 538 seconds	186 sec186 seconds	1 sec1 sec
Relative Linear Error(%)Relative Linear Error(%)	22.922.9	4.34.3	2.32.3
Relative Log Error(%)Relative Log Error(%)	1.561.56	0.400.40	0.110.11

The present invention can increase the accuracy of a neural compact model or ANN (Artificial Neural Network)-based compact model by using meta-learning. This is useful when the amount of data on the target device is limited.

The present invention has been described with reference to an embodiment shown in the drawings, but this is merely illustrative, and those skilled in the art will understand that various modifications and other equivalent embodiments are possible therefrom. Therefore, the true scope of technical protection of the present invention should be determined by the technical spirit of the attached registration claims.

Claims

In the neural compact modeling method performed by a processor,

extracting a second vector for a second voltage data set of the source device according to the first vector extracted by applying the first voltage data set and the first current of the source device to an encoder;

training the encoder and the decoder by comparing the second current and ground true output by applying the second vector and the second voltage data set to the decoder;

generating a third vector using a third voltage data set of a target device, a third current, and the trained encoder; and

A neural compact modeling method comprising estimating a fourth current for a fourth voltage data set of the target device using the third vector, the trained encoder, and the trained decoder.
The method of claim 1, wherein the encoder:

an initial vector generation module that generates an initial vector of the first vector or the second vector;

an updater module that compares the first voltage data set, the first vector, and the first current and outputs an updated value of the first vector; and

A neural compact modeling method including an attention module that compares the first voltage data set and the second voltage data set and outputs an updated second vector from the update value of the first vector.
The method of claim 1, wherein extracting the second vector comprises:

generating a first initial vector by applying the first voltage data set to a first neural network of the encoder;

generating a second initial vector by applying the second voltage data set to a first neural network of the encoder;

applying the first current, the first voltage data set, and the first initial vector to a second neural network of the encoder to output an update value of the first vector;

Setting the update value of the second initial vector as the update value of the first vector and outputting the update value of the second vector;

applying the updated value of the first vector and the first voltage data set to a third neural network of the encoder to update the first initial vector and output the updated first vector; and

Applying the first voltage data set, the second voltage data set, and the updated values of the second vector to a third neural network of the encoder to update the second initial vector and output the updated second vector. Neural compact modeling method including.
The neural compact modeling method of claim 3,

performing a first output operation of applying the first current, the first voltage data set, and the updated first vector to a second neural network of the encoder to output a re-updated value of the first vector;

performing a second output operation of setting the re-update value of the updated second vector as the re-update value of the first vector and outputting the re-update value of the second vector;

Applying the re-updated value of the first vector and the first voltage data set to the third neural network of the encoder to re-update the updated first vector and perform a third output operation to output the re-updated first vector steps; and

The first voltage data set, the second voltage data set, and the re-updated value of the second vector are applied to the third neural network of the encoder to re-update the updated second vector and generate the re-updated second vector. A neural compact modeling method further comprising performing a fourth output operation.
The method of claim 4, wherein the neural compact modeling method:

A neural compact modeling method in which the first output operation, the second output operation, the third output operation, and the fourth output operation are repeatedly performed for an arbitrary number of times.
The method of claim 1, wherein the second vector is:

A neural compact modeling method expressed by compressing data on the width, length, and temperature of the source device.
The method of claim 1, wherein the second vector is:

A neural compact modeling method that is a function that maps the second voltage data set to data about the width, length, and temperature of the source device.
The method of claim 1, wherein the second voltage data set is:

Neural compact modeling method: drain-source voltage, gate-source voltage, and bulk-source voltage.
The neural compact modeling method of claim 1, wherein the encoder parameters and the decoder parameters are not updated during the process of extracting the second vector.
The method of claim 1, wherein generating the third vector comprises:

generating a third initial vector by applying the third voltage data set to the first neural network of the trained encoder;

generating a fourth initial vector by applying the fourth voltage data set to a first neural network of the trained encoder;

applying the third current, the third voltage data set, and the third initial vector to a second neural network of the trained encoder to output an update value of the third vector;

Setting the update value of the fourth initial vector as the update value of the third vector and outputting the update value of the fourth vector; and

A neural compact modeling method comprising applying the update value of the third vector and the third voltage data set to a third neural network of the trained encoder to update the third initial vector and output the updated third vector. .
The method of claim 10, wherein the step of estimating the fourth current is,

outputting an updated fourth vector using the third voltage data set, the fourth voltage data set, and the third vector; and

A neural compact modeling method comprising estimating a fourth current for the fourth voltage data set of the target device by applying the updated fourth vector to the trained decoder.
memory to store instructions; and

It includes a processor that executes the instructions,

The above commands are:

Applying the first voltage data set and the first current of the source device to the encoder to extract a second vector for the second voltage data set of the source device according to the extracted first vector,

Train the encoder and the decoder by applying the second vector and the second voltage data set to the decoder and comparing the second current and ground true output,

Generate a third vector using the third voltage data set of the target device, the third current, and the trained encoder,

A computing device configured to estimate a fourth current for a fourth voltage data set of the target device using the third vector, the trained encoder, and the trained decoder.
The method of claim 12, wherein the encoder:

an initial vector generation module that generates an initial vector of the first vector or the second vector;

an updater module that compares the first voltage data set, the first vector, and the first current and outputs an updated value of the first vector; and

A computing device comprising an attention module that compares the first voltage data set and the second voltage data set and outputs a second vector updated from the update value of the first vector.
The method of claim 12, wherein the second vector is:

A computing device expressed by compressing data about the width, length, and temperature of the source device.
The method of claim 12, wherein the second vector is:

A computing device that maps the second set of voltage data to data for width, length, and temperature of the source device.
The method of claim 12, wherein the second voltage data set is:

A computing device that is a drain-to-source voltage, a gate-to-source voltage, and a bulk-to-source voltage.
The computing device of claim 12, wherein the encoder parameters and the decoder parameters are not updated during the process of extracting the second vector.
The method of claim 12, wherein the instructions for generating the third vector are:

Applying the third voltage data set to the first neural network of the trained encoder to generate a third initial vector,

Applying the fourth voltage data set to the first neural network of the trained encoder to generate a fourth initial vector,

Applying the third current, the third voltage data set, and the third initial vector to a second neural network of the trained encoder to output an update value of the third vector,

Setting the update value of the fourth initial vector to the update value of the third vector and outputting the update value of the fourth vector,

A computing device implemented to update the third initial vector and output an updated third vector by applying the update value of the third vector and the third voltage data set to a third neural network of the trained encoder.
The method of claim 18, wherein the instructions for estimating the fourth current are:

Outputting an updated fourth vector using the third voltage data set, the fourth voltage data set, and the third vector,

A computing device configured to estimate a fourth current for the fourth voltage data set of the target device by applying the updated fourth vector to the trained decoder.