Circuit design method, and a system and program for said method.

Abstract

This invention provides methods, systems, and programs that enable efficient circuit design. [Solution] An analog circuit design method using a large-scale language model, comprising the steps of: the large-scale language model receiving natural language input including information related to circuit specifications; the large-scale language model selecting knowledge information based at least on the input; the large-scale language model generating information necessary for circuit simulation based at least on at least a portion of the selected knowledge information; the circuit simulator executing a circuit simulation and generating information on the simulation results based at least on the necessary information; and the large-scale language model generating circuit design evaluation information based at least on at least a portion of the simulation results.

Description

Technical Field

[0001] The present disclosure relates to a circuit design method, as well as a system and a program for the method. Specifically, the present disclosure relates to an analog circuit design method, as well as a system and a program for the method.

Background Art

[0002] In recent years, the development of large language models (LLMs) has been remarkable, and their applications are expanding in all industrial fields. For example, Mitsubishi Electric Technical Report (https: / / www.giho.mitsubishielectric.co.jp / giho / pdf / 2025 / 2509.pdf) discloses generating HDL code that meets the standards by using RAG to search for past HDL codes, which are the company's design assets, and providing them to generative AI.

Prior Art Documents

Non-Patent Documents

[0003]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] The above non-patent document discloses the use of generative AI and RAG for the design of digital circuits. Circuit design includes digital circuit design and analog circuit design. Although both have their own difficulties, more expertise is required for analog circuit design. The reason for this is that the theoretical values often do not match the behavior of the actual device, and a deep understanding of physical phenomena and intuition are required.

[0005] Therefore, analog circuit design is difficult for beginners in circuit design. Thus, an object of the present disclosure is to provide a means that enables efficient circuit design even for those who are not proficient in circuit design. [Means for solving the problem]

[0006] To achieve the above objectives, this disclosure encompasses, in one aspect, the following inventions. (Invention 1) A method for designing analog circuits using a large-scale language model, The aforementioned method, The process includes the large-scale language model receiving natural language input, including information related to the circuit specifications, The process includes a step in which the large-scale language model selects knowledge information based at least on the input, The process involves the large-scale language model generating information necessary for circuit simulation based on at least a portion of the selected knowledge information, The process includes: • Based at least the necessary information, the circuit simulator performs a circuit simulation and generates information on the simulation results; The process involves the large-scale language model generating circuit design evaluation information based on at least a portion of the simulation results, Includes, Here, the aforementioned knowledge information is • Circuit template information that defines the circuit configuration, • Design rule information that defines design constraints, • Analysis control information that defines the circuit analysis conditions, including, method. (Invention 2) The method of Invention 1, wherein the step selected by the large-scale language model is: • Presenting multiple candidates as part of multiple circuit template information. • Presenting multiple options as part of multiple design rule information, and / or • Presenting multiple candidates as part of multiple analysis and control information. Includes, The process by which the large-scale language model makes a selection further includes selecting one candidate from among the multiple candidates in response to user input. method. (Invention 3) A method according to invention 1 or 2, The aforementioned knowledge information is stored in a vector database in an embedded state. The process selected by the aforementioned large-scale language model is: The aforementioned natural language input is embedded, Performing a search in the vector database based at least on the embedded results, Extracting multiple candidates that fall within a predetermined similarity range, Methods that include... (Invention 4) A method according to any one of inventions 1 to 3, The circuit template information includes information on calculation formulas related to circuit characteristics, The process by which the large-scale language model generates the information necessary for circuit simulation is as follows: Extracting the information of the calculation formula from the circuit template information, The component constants are calculated in reverse based at least on the information in the aforementioned calculation formula, Based at least on the aforementioned inverse calculation results, circuit diagram information is generated, Methods that include... (Invention 5) A method according to any one of inventions 1 to 4, wherein the step of the large-scale language model generating information necessary for circuit simulation is: The information necessary for the aforementioned circuit simulation includes the control block, The circuit simulator generates a file in an executable format, Methods that include... (Invention 6) A method according to any one of inventions 1 to 5, The process of generating information from the simulation results is: Performing an analysis by SPICE based at least on the circuit diagram information, Outputting DC / AC / transient response waveform data, A method comprising. (Invention 7) The method according to Invention 6, wherein the DC / AC / transient response waveform data is output in an image format and / or a numerical format. (Invention 8) The method according to any one of Inventions 1 to 7, wherein the design evaluation information includes information on circuit constant calculation basis and information on circuit analysis results. (Invention 9) The method according to any one of Inventions 1 to 8, further comprising a step in which the large language model generates a design improvement plan and / or an additional analysis plan based at least on the circuit design evaluation information. (Invention 10) The method according to Invention 9, wherein the step in which the large language model generates a design improvement plan and / or an additional analysis plan Calculating performance indicators based at least on at least a part of the simulation results, Generating a design improvement plan based at least on the performance indicators, A method comprising. (Invention 11) The method according to Invention 9 or 10, wherein the design improvement plan includes candidates for circuit constants. (Invention 12) The method according to any one of Inventions 1 to 11, which is executed through the same interface. (Invention 13) A system for executing the method according to any one of Inventions 1 to 12, the system including a device provided with a processor, The device is provided with an interface for requesting processing to the large language model existing inside or outside the system, The device is provided with an interface for requesting processing to the circuit simulator existing inside or outside the system. system. (Invention 14) A program for performing a method according to any one of inventions 1 to 13 by instructing a device equipped with a processor. [Effects of the Invention]

[0007] In one aspect, the above invention involves a large-scale language model selecting knowledge information. This selected knowledge information includes circuit template information, design rule information, and analysis control information. Based on at least a portion of this selected knowledge information, information necessary for circuit simulation is generated, and by performing the circuit simulation based on at least this necessary information, it is possible to prevent results that would require rework. For example, it is possible to prevent the breakdown of circuit topology and deviations from design rules. Therefore, circuit design can be performed efficiently. [Brief explanation of the drawing]

[0008] [Figure 1] An information processing device according to one embodiment of the present disclosure is shown. [Figure 2] The present disclosure describes a system in one embodiment. The system may include a plurality of terminals and at least one server. The terminals and the server may be connected via a network. [Figure 3] The method disclosed herein in one embodiment is shown. [Figure 4] The content of the knowledge information (specifically, circuit template information) in one embodiment of this disclosure is shown below. [Figure 5] The content of the knowledge information (specifically, design rule information) in one embodiment of this disclosure is shown below. [Figure 6] The content of the knowledge information (specifically, analysis and control information) in one embodiment of this disclosure is shown below. [Figure 7] In one embodiment, an example is shown in which knowledge information is stored in a vector database in an embedded state. [Figure 8]In one embodiment, an example of the information required for circuit simulation is shown. It includes the circuit diagram (top) and control information (bottom, the part from "CONTROL_AUTO_MERGE" onwards). The content is created in response to the input prompt, "Create an RC bandpass filter with a cutoff low frequency of 500Hz and a high frequency of 2kHz." [Figure 9] In one embodiment, the results of a circuit simulation are shown. Specifically, DC / AC / transient response waveform data are represented in the form of a plotted image. [Figure 10] In one embodiment, the contents of a report generated through a large-scale language model are shown. [Figure 11] In one embodiment, the contents of a report generated through a large-scale language model are shown. [Figure 12] In one embodiment, the contents of a report generated through a large-scale language model are shown. [Figure 13] In one embodiment, the contents of a report generated through a large-scale language model are shown. [Modes for carrying out the invention]

[0009] The following describes specific embodiments for carrying out the invention. The following description is intended to facilitate understanding of the invention and is not intended to limit the scope of the present invention.

[0010] 1. Overview In one embodiment, the present disclosure relates to a circuit design method, as well as a program and system for performing the method.

[0011] Circuit design includes analog circuit design and digital circuit design, but this disclosure focuses on analog circuit design methods.

[0012] Examples of analog circuits, though not limited to them, may include one or more of the following: filter / frequency processing systems, amplification / signal conditioning / amplifier systems, comparison / judgment / waveform generation systems, power supply / bias / reference systems, sensor / interface systems, arithmetic / analog calculation circuits, communication / signal systems, noise / stabilization / protection systems, etc. Further specific examples may include one or more of the following: high-pass filters, low-pass filters, band-pass filters, notch filters, comparators, window comparators, Schmitt triggers, voltage regulators, constant current circuits, current-to-voltage conversion circuits (IV conversion), differential / integral circuits, voltage followers, non-inverting / inverting amplifier circuits, differential amplifier circuits, instrumentation amplifiers, and summing amplifiers.

[0013] 2. Execution Environment The above circuit design method may be carried out by executing a program. The environment for executing the program and method is not particularly limited, and a typical information processing device (also called a computing device) can be used. The information processing device (100) may typically include a processor (110), memory (120), non-temporary storage medium (130), and a communication module (140), as shown in Figure 1.

[0014] The information processing device (100) includes, but is not limited to, the following: a server, a personal computer, a tablet device, a smartphone, a smartwatch, smart glasses, etc.

[0015] The program is stored in a non-temporary storage medium (130, e.g., HDD, SSD, etc.), loaded into memory (120, e.g., RAM, etc.) as needed, and executed by a processor (110, e.g., CPU, NPU, GPU, TPU (Tensor Processing Unit), LPU (Language Processing Unit), VPU, AI SoC, FPGA, CGRA, DSP, quantization computer processor, etc.). If necessary, the program can connect to a network via a communication module (140) to send and receive information.

[0016] In one embodiment, the program may be installed as application software on an information processing device (100) and executed by the information processing device (100).

[0017] In another embodiment, the number of information processing devices (100) is not limited to one, and multiple information processing devices (100) may be used as needed. In that case, the functions of the program may be distributed among multiple information processing devices (100).

[0018] Alternatively, as shown in Figure 2, a system (200) configuration may be adopted in which a server (210) and a terminal (220) are interconnected via a network. In this system (200), the terminal (220) may receive input from a user and send at least a portion of the received input to the server (210). The server (210) may receive the input information sent from the terminal (220), process the information, and send a portion of the output to the terminal (220). The terminal (220) may then receive the output information sent from the server (210) and display it on the terminal (220).

[0019] Therefore, in another aspect, this disclosure also relates to an information processing device including the program of this disclosure, and a system including said information processing device. In yet another aspect, this disclosure relates to terminals and / or servers constituting the system of this disclosure. The internal configuration of the terminals and servers may be the same as that of the information processing device shown in Figure 1. In yet another aspect, this disclosure relates to a computer-readable non-temporary storage medium (e.g., HDD, SSD, flash memory, optical disk, etc.) storing the program.

[0020] The information processing device described above may be connected to a display or the like as appropriate. The information processing device can then transmit signals to the display for displaying calculation results from a processor or the like on the display.

[0021] The system configuration is not limited to a specific form and may be any form as long as the circuit design method of this disclosure can be implemented. For example, in order to implement the circuit design method of this disclosure, at least the following hardware is required. User terminal • Device equipped with a circuit simulator • Devices equipped with large-scale language models

[0022] These three components may be the same device or they may be independent devices. For example, the user terminal may be one independent device, while the circuit simulator and large-scale language model are provided by another independent device. Alternatively, multiple devices may work together to operate virtually as a single large-scale language model.

[0023] The three components described above may exist inside or outside the system in the following patterns. Pattern 1: User terminal, circuit simulator, and large-scale language model within the system Pattern 2: User terminals within the system, circuit simulator and large-scale language model outside the system. Pattern 3: User terminals within the system, circuit simulator, and a large-scale language model outside the system. Pattern 4: User terminals within the system, large-scale language model, circuit simulator outside the system.

[0024] In addition, in Pattern 1, a single information processing device may serve as the user terminal, circuit simulator, and large-scale language model.

[0025] Components within the system and components outside the system may be connected to each other via a network (e.g., the Internet). Furthermore, components within the system may be connected to each other via a network (e.g., the Internet, LAN, WAN, etc.).

[0026] Furthermore, the three components described above may be connected to each other by additional hardware. For example, a server (e.g., a web server) may be located within or outside the system. This server may be configured to send and receive information with the large-scale language model and / or circuit simulator through an interface (e.g., an API) provided by the server. Alternatively, the user terminal may be equipped with an interface for sending and receiving information with the large-scale language model and / or circuit simulator.

[0027] The large-scale language model is not particularly limited and may be any large-scale language model used in this field. For example, a large-scale language model may be one or more of the following: Gemini, GPT, Claude, Llama, Phi, OpenAI, Mistral, Grok, DeepSeek, Qwen. Alternatively, the large-scale language model may be built on-premises. It is preferable to build the large-scale language model on-premises from the standpoint of hallucination and confidentiality management. For convenience, the term "large-scale language model" is used, but in this specification, the term "large-scale language model" also includes "small-scale language models." This is because the only difference between the two is the scale of the parameters, and the objectives of this disclosure can be achieved even with a "small-scale language model."

[0028] The circuit simulator is not particularly limited and any circuit simulator used in this field may be used. For example, the circuit simulator may be one or more of the following: SPICE (e.g., LTspice, PSpice, etc.), Keysight ADS, SIMetrix / SIMPLIS, Ansys Icepak / SIwave, HSPICE, Spectre, Eldo, ngspice, Xyce.

[0029] 3. Overview of Circuit Design Method In one embodiment, this disclosure relates to an analog circuit design method using a large-scale language model. The method includes at least the following steps (Figure 3). • A process in which a large-scale language model receives natural language input, including information related to circuit specifications. • The process by which a large-scale language model selects knowledge information based at least on the input. • A process in which a large-scale language model generates information necessary for circuit simulation based on at least a portion of the selected knowledge information. • The process involves the circuit simulator performing a circuit simulation based on at least the necessary information, and then generating information about the simulation results. A process in which a large-scale language model generates circuit design evaluation information based on at least a portion of the simulation results.

[0030] Here, the knowledge information includes the following: • Circuit template information that defines the circuit configuration. • Design rule information that defines design constraints, • Analysis and control information that defines the circuit analysis conditions.

[0031] These pieces of knowledge may exist in independent forms or as an integrated whole.

[0032] The process by which a large-scale language model selects knowledge information is not simply limited to narrowing down multiple pieces of knowledge information to one. For example, the process by which a large-scale language model selects knowledge information may include narrowing down multiple pieces of knowledge information to a few options (for example, by searching for knowledge information) and presenting those options to the user. In this case, the knowledge information may be further narrowed down to a specific one based on information from the user specifying a particular option. Details will be described later.

[0033] The following sections will explain the details of each process, as well as processes other than those mentioned above.

[0034] 3-1. The process by which a large-scale language model receives natural language input, including information related to circuit specifications. The user inputs information about the circuit specifications by operating a user terminal. The input is ultimately sent to a large-scale language model. The large-scale language model may receive input from the user terminal through a dedicated interface (e.g., an API), or it may receive input from the user terminal via a server.

[0035] The input may be in natural language format. For example, an input prompt may be in the following natural language format: "Create an RC bandpass filter with a cutoff frequency of 500Hz for the low frequency and 2kHz for the high frequency." "Design a circuit that uses a 5V operational amplifier to amplify the microphone signal 100 times." "Design a circuit that turns on an LED when the brightness drops below 10 lux."

[0036] In the above input prompt, you may also specify the component. For example, the following input prompt is also possible: "Create a low-pass filter with a cutoff frequency of 5kHz, but use a 10Kohm resistor."

[0037] The format of the above input prompt is not particularly limited, but the content of the input prompt may include at least information regarding the circuit specifications.

[0038] 3-2. The process by which a large-scale language model selects knowledge information based at least on the input. After receiving natural language input, the large-scale language model can select knowledge information based at least on the input. As mentioned above, the knowledge information includes at least circuit template information, design rule information, and analysis and control information.

[0039] 3-2-1. Circuit Template Information The format of the circuit template information defining the circuit configuration is not particularly limited and may be in binary file format or text file format. For example, Figure 4 shows an example of a circuit template for a bandpass filter.<R1_HP_VALUE><C1_HP_VALUE><R2_LP_VALUE><C2_LP_VALUE><RLOAD_VALUE> The following parts are replaceable: R1 and R2 represent the values ​​of the resistors used in the filter circuit, and C1 and C2 represent the values ​​of the capacitances of the capacitors used in the filter circuit. R1 and C1 represent the values ​​for the high-pass filter, while R2 and C2 represent the values ​​for the low-pass filter.

[0040] The circuit template information may include multiple files, each defining a circuit diagram corresponding to one of the various circuits described above. Therefore, the content shown in Figure 4 is only the content of one of several files, and there may be multiple files depending on the type of circuit. Finally, at least one candidate file may be selected from among these multiple files. As shown in Figure 4, the circuit template information may also include calculation formulas related to the circuit. Alternatively, information on calculation formulas related to the circuit may exist separately and independently of the circuit template information. If the circuit template information includes calculation formulas related to the circuit, for example, as shown in Figure 4, a field called "##Formula" may be created to define the calculation formulas. In Figure 4, calculation formulas for the high-pass and low-pass cutoff frequencies are defined, respectively.

[0041] The existence of circuit templates helps prevent circuit topology breakdowns. It also avoids the effort of designing circuits from scratch. Furthermore, because circuit templates include information about calculation formulas, the basis for those calculations can be verified later.

[0042] 3-2-2. Design Rule Information Design rule information includes definitions of various conventions in circuit design. It is crucial for avoiding design results that deviate from design constraints. The specific content of design constraints is not limited, but may include one or more of the following, for example: definition of component sequences, acceptable deviations from user-defined goals, design rules for each type of circuit, rules for violations, and definitions of actual components preferred for use. Design rules may also include information defined as tacit knowledge based on corporate culture (e.g., types of usable components, types of components preferred for use, etc.).

[0043] The design rule information may be in binary file format or text file format. An example is shown in Figure 5. For example, Figure 5 describes the design rules for a low-pass filter and a high-pass filter. The types of capacitors and resistors to be selected are also specified. Furthermore, a tolerance of ±20% for the user-specified goal is defined.

[0044] For example, when designing a filter circuit, if you simply design the circuit based on the circuit template information (and the calculation formulas contained within it), you may end up designing resistors and capacitors with values ​​that do not actually exist as components (for example, a resistor component with a resistance of 10Ω exists, but a resistor component with an odd resistance value of 10.5Ω does not actually exist).

[0045] Furthermore, due to the unique circumstances of each manufacturer, there may be manufacturer-specific design rules. In such cases, by considering not only circuit template information but also design rule information, efficient circuit design becomes possible, and the likelihood of avoiding a situation where the design is dependent on a specific individual (i.e., a situation where only skilled engineers can design) can be increased.

[0046] In Figure 5, one file is shown as a design rule, but there may be multiple files corresponding to multiple design rules. Furthermore, at least one candidate file may be selected from among these multiple files. In some cases, there may be at least one file among the multiple files corresponding to multiple design rules that can be selected as the default.

[0047] 3-2-3. Analysis and Control Information The analysis control information defines the circuit analysis conditions. The circuit analysis conditions may also define the perspective from which the circuit simulator, described later, performs the circuit analysis. An example is shown in Figure 6. In Figure 6, DC Sweep (or DC analysis), AC Sweep (or AC analysis), and TRAN analysis (or transient analysis) are defined.

[0048] The specific details of the circuit analysis conditions are not limited, but may include one or more of the following, for example: OP (Operating Point analysis), NOISE (Noise analysis), TF (Transfer Function analysis), STEP (Parameter sweep), FOUR (Fourier analysis), SENS (Sensitivity analysis), TEMP (Temperature analysis), MC (Monte Carlo analysis), DCSweep, ACAnalysis, TRAN, PZ, STB, CornerAnalysis, AgingAnalysis, DistortionAnalysis, IM, PhaseNoiseAnalysis, Jitter, PowerDisipation, Efficiency, Electro-ThermalCo-Simulation, HarmonicBalance, PeriodicSteady-State, SphereOperatingArea.

[0049] In some cases, the analysis control information may be integrated with the circuit template information. For example, if the content to be analyzed is determined by the selected circuit, the analysis control information may be integrated with the circuit template information. On the other hand, if there are multiple analysis candidates from the selected circuit, the analysis control information may exist independently of the circuit template information.

[0050] Similar to the design rule information described above, Figure 6 shows one file as analysis and control information, but multiple files may exist to correspond to multiple analysis and control information. Furthermore, at least one candidate file may be selected from among these multiple files. In some cases, there may be at least one file among the multiple files corresponding to multiple analysis and control information that can be selected as the default.

[0051] As described above, three types of information have been explained as knowledge information, but knowledge information is not limited to these three types of information; it may include other types of knowledge information as well.

[0052] 3-2-4. Storage Formats of Knowledge Information The form in which knowledge information is stored is not particularly limited. In one embodiment, the knowledge information may be stored in a vector database in an embedded state. In this case, the steps selected by the large-scale language model may include the following: Embedding natural language input, Performing a search in a vector database based at least on the embedded results, Extracting multiple candidates that fall within a predetermined similarity range.

[0053] An example is shown in Figure 7. The left-hand block shows the state in which the three types of knowledge information described above are stored in separate folders. The contents of this knowledge information (the contents of the documents in which the knowledge information is written) can then be converted into numerical data using the Sentence-Embedding vectorization module. The results can then be saved to another folder, as shown in the right-hand block. The data format is not particularly limited (for example, it can be binary or in JSON format).

[0054] Therefore, in one embodiment, the above method may further include the step of embedding the knowledge information and then storing it in a vector database. Embedding may include chunking the target information and vectorizing it. The method of vectorization is not particularly limited, but known methods in the art may be used. For example, Python libraries such as Sentence-Transformers (SBERT), OpenAI Embedding API, Google Universal Sentence Encoder, BERT, E5, GTE, Word2Vec, and Code embedding may be used.

[0055] 3-2-5. Selection of Knowledge Information After storing the data in the vector database, natural language input from the user is similarly embedded. Then, information conceptually similar to the natural language input is extracted from the vector database. In a preferred embodiment, it is not necessary to extract only one piece of information from the vector database from the beginning; multiple candidate pieces of information may be extracted. However, since it is not practical to increase the number of candidates indefinitely, only those that fall within a predetermined similarity range may be extracted. For example, the similarity range may be based on cosine similarity. Furthermore, the number of candidates may be limited to any number, such as the top 10, top 5, or top 3 in terms of similarity.

[0056] Therefore, large-scale language models are Select at least some (for example, one, two, or three) of the circuit template information. Select at least some (for example, one, two, or three) of the multiple design rule information. From multiple analysis and control information, select at least some (for example, one, two, or three). It may be configured to do so.

[0057] For example, by deliberately not limiting each piece of knowledge information to just one, you can give the user the opportunity to make a final choice. Also, when vectorizing, there may be knowledge information with similar numerical values, and when extracting content from a vector database that is similar to the vectorized content of natural language input, there is a possibility of extracting inappropriate content. For example, low-pass filters and high-pass filters have a high degree of similarity (close numerical scores), so even if a low-pass filter is requested in natural language, a high-pass filter may be selected as a candidate. By listing multiple options rather than limiting it to one, you can leave the final decision to the user.

[0058] Therefore, in the method of one embodiment of the present disclosure, the steps selected by the large-scale language model are • Presenting multiple candidates as part of multiple circuit template information. • Presenting multiple options as part of multiple design rule information, and / or • Presenting multiple candidates as part of multiple analysis and control information. This may include the following. Furthermore, in the method of the embodiment, the step selected by the large-scale language model may further include selecting one candidate from among several candidates in response to user input.

[0059] For example, the interface may be configured to present three candidate pieces of knowledge information and prompt the user to input 1 through 3. Alternatively, a GUI (e.g., a website) may be configured to prompt the user to select a specific item (e.g., through radio buttons, dropdown menus, etc.).

[0060] 3-3. The process by which a large-scale language model generates information necessary for circuit simulation based on at least a portion of the selected knowledge information. After the selection of knowledge information has been performed, the large-scale language model generates the information necessary for circuit simulation, based at least on some of the selected knowledge information.

[0061] As mentioned above, users may further refine the selected knowledge information, so it is not necessary to base the results on all of the selected knowledge information. Furthermore, in addition to knowledge information, the information necessary for circuit simulation can be generated based on other factors as well. Typically, this may be based on the natural language input described above.

[0062] For example, a large-scale language model understands the essential rules in the selected design rules and, while adhering to them, generates schematic information by referring to the circuit template and its associated calculation formulas. For instance, it generates schematic information by appropriately embedding values ​​in the embeddable parts of the circuit template. In this process, it is preferable to generate schematic information while maintaining the circuit topology in the circuit template.

[0063] Figure 8 shows an example of the information required for circuit simulation. This information includes, for example, information defining the circuit diagram and control information (also called control blocks). The part defining the circuit diagram is based on the circuit template information shown in Figure 4 (especially the part after #SPICE TEMPLATE in Figure 4). The control information is based on the analysis control information shown in Figure 6. Here, the two elements—the information defining the circuit diagram and the control information—may be integrated into a single file, or they may be separate.

[0064] However, it is preferable that the circuit simulator (e.g., SPICE) used in subsequent steps is in an executable format. Therefore, in the method of one embodiment of this disclosure, the step of a large-scale language model generating the information necessary for circuit simulation is: The information required for circuit simulation includes control blocks, The circuit simulator generates a file in an executable format, It may include.

[0065] As mentioned above, the part of the information that defines the circuit diagram is based on the circuit template information shown in Figure 4. Here, when creating the information that defines the circuit diagram from the circuit template information, certain calculations may be performed. For example, the circuit template information may include information on calculation formulas related to the circuit characteristics contained in the information. In the example in Figure 4, the following calculation formulas are included. These are formulas for determining the cutoff frequency. fc_low = 1 / (2π·R1·C1) fc_high = 1 / (2π·R²·C²)

[0066] The above explains an example of a filter circuit, but the calculation formula will differ depending on the characteristics of the circuit. For example, in the case of an amplifier circuit, Gain = 1 + R f / R l The circuit template information may include information about such equations. For example, in the case of a resonant circuit, the circuit template information may include information about an equation such as f = 1 / (2π√(LC)). Examples of combinations of circuit types and corresponding calculation formulas are listed below, but are not limited to these. Voltage divider circuit: vout = vion x R2 / (R1 + R2) Synthesis circuit R = R1 + R2 RC low-pass filter fc = 1 / (2πRC) RL filter fc = R / (2πL) Schmitt trigger vth = Vref x (1+R1 / R2) Integrator circuit Vout = -(1 / RC) S Vin dt Differential circuit Vout = -RC (dVin / dt)

[0067] Furthermore, a calculation formula may be extracted from the circuit template information, and the component constants may be calculated in reverse based at least on the information in the calculation formula. For example, in the case of the filter circuit example above, the cutoff frequency is determined by the input prompt, so the values ​​of R1 and C1, and R2 and C2 in the above formula may be determined based on this.

[0068] Then, once the values ​​of R1, C1, R2, and C2 are determined, the circuit diagram information is generated. For example, the replaceable parts in the circuit template may be replaced with the values ​​of R1, C1, R2, and C2.

[0069] Therefore, in the method of one embodiment of the present disclosure, the circuit template information may include information on calculation formulas relating to circuit characteristics, and the step of a large-scale language model generating information necessary for circuit simulation is, Extracting calculation formula information from circuit template information, This involves at least calculating the component constants based on the information in the calculation formula, To generate circuit diagram information based at least on the reverse calculation results, It may include.

[0070] 3-4. The process by which the circuit simulator performs a circuit simulation based on at least the necessary information and generates information on the simulation results. After the information necessary for circuit simulation has been generated, the circuit simulator executes the simulation. For example, the circuit simulator reads a file with the contents shown in Figure 8 and executes the circuit simulation according to its contents.

[0071] A circuit simulator (e.g., SPICE) generates information from the simulation results. These simulation results may include the results of DC Sweep (or DC analysis), AC Sweep (or AC analysis), and TRAN analysis (or transient analysis) (e.g., DC / AC / transient response waveform data).

[0072] Therefore, in the method of one embodiment of the present disclosure, the step of generating simulation result information may include performing SPICE analysis based at least on circuit diagram information and outputting DC / AC / transient response waveform data.

[0073] The format of this data is not particularly limited and may be one or more of the following: text file format, CSV file format, spreadsheet software-specific format (e.g., Excel's XLSX), image format (e.g., GIF, JPEG, PNG, etc.).

[0074] Therefore, in the method of one embodiment of the present disclosure, DC / AC / transient response waveform data may be output in image format (see, for example, Figure 9) and / or numerical format.

[0075] Here, some of the output data may be cut (for example, to reduce the load on a large language model).

[0076] 3-5. A process in which a large-scale language model generates circuit design evaluation information based on at least a portion of the simulation results. After the simulation is performed, the large-scale language model generates circuit design evaluation information based on at least a portion of the simulation results. Here, the design evaluation information may include information on the basis for calculating circuit constants and information on the circuit analysis results.

[0077] The following explanation will refer to Figures 10 to 13.

[0078] The information used to calculate the circuit constants may include, for example, information on how the circuit constants are calculated by working backward from the calculation formula included in the circuit template information and the circuit specification information included in the natural language input. For example, in the filter circuit example described above, it may include information on how the resistance value of the resistor and the capacitance of the capacitor were selected from the specific cutoff frequency (for example, the cutoff frequency entered by the user in the input prompt) and the calculation formula fc_high = 1 / (2π·R2·C2) (see, for example, Figure 13).

[0079] The circuit analysis results may include information such as whether there is a difference between the circuit specifications information included in the natural language input and the simulation results (or whether that difference is within the tolerance range specified in the design rules) (see, for example, Figure 13).

[0080] In a preferred embodiment, when the large-scale language model generates circuit design evaluation information, it may also be based on factors other than the simulation results, such as chat history information between the user and the large-scale language model (including natural language input, including information related to the circuit specifications), and information necessary for circuit simulation (for example, information defining the circuit diagram and control information).

[0081] A large-scale language model can perform the following actions, for example, based on some of the simulation results and other information: • Extract the following information from chat history data and output it to a report (see Figure 10, for example). User input in natural language (e.g., user prompts), The knowledge information that was ultimately selected (for example, the information of the single circuit template that was ultimately selected, the design rule information, and the analysis and control information), Information on the basis for calculating circuit constants • The system receives information from circuit simulator simulation results, or excerpts thereof, as input, analyzes it, and outputs the results in a report (see, for example, Figure 11). • The system receives and verifies the contents of the generated report and outputs the following information (see, for example, Figure 13). Summary Comparison of the target characteristics of a circuit diagram designed based on user requirements and the resulting characteristics obtained from circuit simulation using a circuit simulator.

[0082] By using large-scale language models, even engineers without extensive circuit design experience can efficiently design circuits. Furthermore, by using knowledge information, the possibility of the large-scale language model designing circuits that do not meet the requirements (in extreme cases, the possibility of hallucination) can be reduced. By utilizing circuit template information within the knowledge information, efficient circuit design can be achieved. Moreover, by designing circuits while maintaining the circuit topology, it is possible to avoid breakdowns in the circuit topology. In addition, when the knowledge information is kept independent of each other, it becomes possible to clearly define the boundaries of responsibility, reducing the possibility of the large-scale language model becoming confused.

[0083] Furthermore, outputting a report that includes the calculation basis and the difference between the theoretical calculation and the simulation results, as described above, makes verification easier.

[0084] 3-6. A process in which a large-scale language model generates design improvement proposals and / or additional analysis proposals based at least on circuit design evaluation information. In a preferred embodiment, though not required, the large-scale language model may further generate design improvement proposals and / or additional analysis proposals based at least on circuit design evaluation information (see, for example, Figures 11 and 12).

[0085] By obtaining these design improvement proposals and / or additional analysis proposals, even engineers who are not highly skilled in circuit design can perform circuit design efficiently.

[0086] Furthermore, in a preferred embodiment, the step of a large-scale language model generating design improvement proposals and / or additional analysis proposals is: Calculating performance metrics based on at least a portion of the simulation results, To generate design improvement proposals based at least on performance indicators, It may include.

[0087] For example, in the case of the filter circuit described above, the simulation results may include information on the cutoff frequency based on those results. If there is a difference between this cutoff frequency performance indicator and the cutoff frequency performance indicator required by the user, or if that difference exceeds an acceptable range, the large-scale language model may generate an improved circuit design proposal to bridge that difference.

[0088] In this regard, the design improvement proposal may include candidate circuit constants. For example, in the example of the filter circuit described above, alternative values ​​for the resistance and capacitance of the resistors and capacitors that make up the filter circuit may be shown.

[0089] 4. Others By following the process described above, even inexperienced engineers can efficiently design circuits.

[0090] In a preferred embodiment, the series of processes described above may be performed through the same interface. For example, in another embodiment, it is possible to connect directly to the large-scale language model from the user terminal to exchange input prompts, and then run the circuit simulation by launching separate circuit simulation software. However, in this case, the user must operate by switching between the interfaces of the large-scale language model and the circuit simulation, which may reduce efficiency and increase the likelihood of errors.

[0091] Therefore, if it becomes possible to interact with large-scale language models and perform circuit simulations on the same interface (for example, the same prompt input screen), efficiency will be improved and the possibility of errors will be reduced.

[0092] Therefore, in one embodiment, a particular program may be configured to provide a user interface while simultaneously being able to send and receive information with a large-scale language model and a circuit simulator through a separate API.

[0093] In one example, the process described above can be broadly divided into three parts. A. Circuit design -> Schematic diagram Circuit B simulation (SPICE) -> Results Evaluation and review of C circuit -> Comparison of 1 (desktop) and 2 (shim)

[0094] A can include steps 1 through 3 of the process shown in Figure 3. B can include step 4 of the process shown in Figure 3. C can include step 5 of the process shown in Figure 3.

[0095] These steps A through C can be completed using the same interface, for example, through input prompts, command operations, and web screen operations. Furthermore, the user can verify the results in each phase A through C. For example, in A, the designed circuit diagram and control information can be viewed. In B, the results of the circuit simulation can be viewed. In C, a comparison between the theoretical calculations in A and the actual results based on B can be viewed. This allows the user to check for errors in each phase (intermediate steps).

[0096] The above describes specific embodiments of the invention. The above embodiments are merely examples, and the present invention is not limited to these embodiments. For example, the technical features disclosed in one of the above embodiments can be applied to other embodiments. Also, unless otherwise specified, for a particular method, it is possible to swap the order of some steps with those of other steps, and further steps may be added between two specific steps. The scope of the present invention is defined by the claims.

Claims

1. A method for designing analog circuits using a large-scale language model, The above method involves a computer, - A process of having the large-scale language model receive natural language input, including information related to circuit specifications, - A step of causing the large-scale language model to select knowledge information based at least on the input, - A step of causing the large-scale language model to generate information necessary for circuit simulation based on at least a portion of the selected knowledge information, - A step of having the circuit simulator perform a circuit simulation based on the necessary information mentioned above, and generating information of the simulation results, - A step of generating circuit design evaluation information in the large-scale language model based on at least a portion of the simulation results, This includes performing the following: Here, the aforementioned knowledge information is • Circuit template information that defines the circuit configuration, • Design rule information that defines design constraints, - Analysis control information that defines the circuit analysis conditions, Includes, The circuit design evaluation information includes information on the basis for calculating circuit constants and information on the results of circuit analysis. method.

2. The method according to claim 1, wherein the step of causing the large-scale language model to make a selection is: - Presenting multiple candidates as part of multiple circuit template information. - Presenting multiple candidates as part of multiple design rule information, and / or, - Presenting multiple candidates as part of multiple analysis and control information, Includes, The process of having the large-scale language model make a selection further includes selecting one candidate from the multiple candidates in response to user input. method.

3. The method according to claim 1, The aforementioned knowledge information is stored in a vector database in an embedded state. The process of having the aforementioned large-scale language model make a selection is: The aforementioned natural language input is embedded, Performing a search in the vector database based at least on the embedded results, Extracting multiple candidates that fall within a predetermined similarity range, Methods that include...

4. The method according to claim 1, The circuit template information includes information on calculation formulas related to circuit characteristics, The process of generating the information necessary for circuit simulation in the large-scale language model is as follows: Extracting the information of the calculation formula from the circuit template information, The component constants are calculated in reverse based at least on the information in the aforementioned calculation formula, Based at least on the aforementioned inverse calculation results, circuit diagram information is generated, Methods that include...

5. The method according to claim 1, wherein the step of generating information necessary for circuit simulation in the large-scale language model is: The information necessary for the aforementioned circuit simulation includes the control block, The circuit simulator generates a file in an executable format, Methods that include...

6. The method according to claim 4, The process of generating information from the simulation results is: Performing SPICE analysis based at least on the circuit diagram information, Outputting DC / AC / transient response waveform data, Methods that include...

7. A method according to claim 6, wherein the DC / AC / transient response waveform data is output in image format and / or numerical format.

8. A method according to claim 1, wherein the method further comprises the step of causing the large-scale language model to generate design improvement proposals and / or additional analysis proposals based at least on the circuit design evaluation information.

9. The method according to claim 8, wherein the step of causing the large-scale language model to generate design improvement proposals and / or additional analysis proposals is: The performance index is calculated based on at least a portion of the aforementioned simulation results, To generate design improvement proposals based at least on the aforementioned performance indicators, Methods that include...

10. A method according to claim 8, wherein the design improvement proposal includes candidate circuit constants.

11. A method according to claim 1, the method being performed through the same interface.

12. A system for performing the method according to any one of claims 1 to 11, wherein the system includes a device comprising a processor, The device includes an interface for requesting processing from the large-scale language model located within or outside the system. The device includes an interface for requesting processing from the circuit simulator located within or outside the system. system.

13. A program for causing the processor of a device equipped with a processor to perform the method according to any one of claims 1 to 11.

Images