Design method for generating electronic component package through multi-mode large language model

Abstract

The invention belongs to the field of electronic design automation, and provides a design method for generating electronic component packaging through a multi-mode large language model. The method aims at automatically processing and analyzing data tables in different formats by using a multi-modal large language model, and the problem of data table type isomerism is solved. Through an automated process, the MLLM automatically generates a package based on the data table. The MLLM is guided to generate a more accurate result through a designed query process and a data set for providing context examples. In the package generation process, visual prompts and thinking chains are introduced, so that the model can describe the bonding pad more accurately according to a given index, and the accuracy of package generation is improved. And finally, performing evaluation processing on the packaging quality of the generated electronic component, thereby realizing quantitative evaluation on the quality of the electronic component.

Description

Technical Field

[0001] The invention belongs to the field of electronic design automation and provides a design method for generating electronic component packaging using a multi-modal large language model. Background Art

[0002] A complete and accurate PCB component library is the foundation of PCB design, providing designers with a set of predefined electronic components for circuit schematics and PCB layout. Such libraries provide a comprehensive set of resources, including datasheets, schematics, 2D packages, and 3D models.

[0003] Traditionally, companies have manually created footprints based on datasheets to build component libraries. However, this manual process is laborious and error-prone. In addition, datasheet formats vary widely from manufacturer to manufacturer, complicating the implementation of a standardized design process. The varying requirements of various electronic design automation (EDA) software further add to the challenge, making it difficult to reuse library resources across design platforms, thereby increasing development costs. With new components emerging on the market every year, a more efficient method for developing component libraries is essential to keep pace with technological advancements and market demands.

[0004] The heterogeneous format of datasheets and the lack of large-scale datasets make it difficult for traditional AI methods to capture component design details. However, multimodal large language models (MLLM) and hint engineering techniques provide a promising solution to extract key information from datasheets and generate accurate design files. Thanks to its extensive knowledge, MLLM can meet challenges without retraining. On the other hand, as electronic designs become more complex, AI-assisted PCB design points out an opportunity to significantly shorten the design cycle while ensuring high design quality. Automatic component library construction is a necessary foundation for MLLM to understand and create PCB designs.

[0005] Large language models (LLMs) with billions of parameters, such as GPT, Gemini, and LLama, have demonstrated impressive capabilities in handling complex tasks in a variety of domains, including code generation, mathematics, science, and electronic design automation (EDA). Advanced MLLMs, such as GPT-4 and Qwen-VL, extend their versatility by providing more flexible and user-friendly interfaces, enabling them to process and understand additional modalities, such as vision. Their ability to handle multimodal inputs enables MLLMs to interpret component specifications in various formats, including lists, tables, and images, making them well suited for processing component datasheets. However, PCB design has a low tolerance for errors in component libraries. Directly using MLLMs for component design generation may lead to unsatisfactory results, such as incorrect pin counts and inaccurate dimensions of created shapes. This demonstrates the importance of specific query flows and datasets that provide contextual examples that guide MLLMs to generate more accurate results.

[0006] In general, the defects of the prior art are as follows:

[0007] 1. The process of manually creating a package is time-consuming and error-prone. Engineers need to extract detailed information from the datasheet and then manually create a package that is compatible with the EDA software. This process is not only inefficient, but also prone to errors due to human factors.

[0008] 2. Heterogeneity of data table formats. The data table formats provided by different manufacturers vary greatly, which brings challenges to the construction of automated component libraries. Traditional AI methods have difficulty processing such heterogeneous data, which affects the accuracy and efficiency of the automation process.

[0009] 3. Lack of large-scale datasets. Due to the diversity and complexity of electronic components, it is a challenge to build a large dataset containing data on various components. The lack of such a dataset limits the ability of traditional AI methods to capture component design details.

[0010] 4. Limitations of using MLLM directly for component design generation. Although MLLM has potential in processing multimodal inputs, using MLLM directly for component design generation may lead to unsatisfactory results, such as incorrect pin counts and inaccurate dimensions of the generated package shapes. Summary of the invention

[0011] This paper proposes a design method for electronic component package generation using a multimodal large language model (MLLM), aiming to solve the problems of low efficiency and error-proneness in the traditional manual package creation process, as well as the challenges brought by data table format heterogeneity and lack of large-scale data sets. At the specific technical implementation level, this paper solves the following key problems:

[0012] Data heterogeneity processing: By using a multimodal large language model (MLLM), it can process and parse data tables of different formats, thus solving the problem of data table format heterogeneity. MLLM can understand and process component specifications in various formats including lists, tables, and images.

[0013] Efficient and automated generation: In traditional methods, the packaging of electronic components needs to be created manually, which is inefficient and error-prone. The present invention uses MLLM to automatically generate packaging based on data tables through an automated process, greatly improving efficiency.

[0014] Improved precision and accuracy: Directly using MLLM for component design generation may lead to unsatisfactory results, such as incorrect number of pins and inaccurate size of the generated package shape. The present invention guides MLLM to generate more accurate results by designing a specific query process and providing a data set of contextual examples.

[0015] Learning and optimization of logical and spatial position parameters: In the package generation process, by introducing visual cues and thought chain CoT, the model can more accurately describe the pads based on the given index, improving the accuracy of package generation.

[0016] In general, the present invention mainly solves key problems such as data heterogeneity processing, automatic generation, precision and accuracy improvement, logic and spatial parameter learning optimization, and quality assessment quantification at the technical implementation level.

[0017] The present invention provides a design method for generating electronic component packaging using a multimodal large language model, comprising the following steps:

[0018] Step 1: Collect and preprocess data: Collect data sheets containing package details from the Internet and electronic design automation software. The package information includes pad size, pad position, and component body outline.

[0019] Step 2: Define the conditional text generation problem for encapsulation generation;

[0020] Step 3: Develop a package generation design process based on a multimodal large language model;

[0021] Step 4: Implement the package generation process by extracting, contextualizing, enhancing visual cues, and guiding thought chains from the input data table;

[0022] Step 5: By evaluating the quality of the packaging of the generated electronic components, a quantitative evaluation of the quality is achieved.

[0023] In the above scheme, step 1 includes:

[0024] Step 1.1, collect <data sheet, package> data pairs from the Internet and EDA software;

[0025] Step 1.2. Select a diverse subset of components based on the number of pads and component type.

[0026] In the above scheme, step 2 includes:

[0027] Step 2.1, define the component package generation task as a conditional text generation problem;

[0028] Step 2.2: Automatically generate a package based on the provided component data sheet.

[0029] In the above scheme, step 3 includes:

[0030] Step 3.1: Use the dataset Where (x j ,y j ) is the jth sample of the dataset, a {input constraint, output text} pair, and M is the total number of samples;

[0031] Step 3.2: Provide k samples As an ICL template, where x j and j They are constraint and target files respectively;

[0032] Step 3.3: Based on constraint x test , the goal is to generate high-quality component library files y test ;

[0033] Step 3.4, design the prompt P to integrate the input constraints, ICL examples, task description and other relevant information;

[0034] Step 3.5: Introduce CoT into the prompt to enhance the model’s logical processing capability of input information.

[0035] Step 3.6: Integrate the prompt P into the MLLM, where P = [R; x1; y1; ...; x k ;y k ;x test ], R is the task description, x1 to x k is the constraint of the ICL example, y1 to y k is the corresponding target file, x test is the constraint of the current input;

[0036] Step 3.7: Process the integrated prompt P through MLLM to generate a component library file y that meets the constraints test .

[0037] In the above scheme, step 4 includes:

[0038] Step 4.1, extract the package information from the input data sheet, including the size and position of each pad;

[0039] Step 4.2: Use a small amount of context to learn ICL examples to help MLLM understand the constraints of pad location and size;

[0040] Step 4.3: Introduce visual cues VP for package generation to enhance SFgen's ability to understand pad size and location;

[0041] Step 4.4, respond to the input by generating a preliminary package file that can be used for visualization by the EDA software;

[0042] Step 4.5: Introduce the modified prompts, check the encapsulation step by step and correct the errors;

[0043] Step 4.6, include the thought chain CoT in the prompt to guide the model to more accurately describe the pad according to the given index, improving the accuracy of the package generation;

[0044] In the above scheme, step 5 includes:

[0045] Step 5.1, a set of evaluation metrics are defined to quantify the quality of the generated footprint, including ACCN, ACCT, Dif A and Dif P ;

[0046] Step 5.1.1. Calculate the accuracy index ACCN, which is used to measure the ratio of the number of correctly identified pads to the total number of pads in the generated package:

[0047]

[0049] Step 5.1.2. Calculate the accuracy index ACCT, which is used to measure the ratio of the number of correctly identified shape types to the total number in the generated footprint:

[0050]

[0051] Step 5.1.3. Calculate the difference index Dif A , which measures the average difference between the generated pad size and the actual size;

[0052]

[0053] Step 5.1.4. Calculate the difference index Dif P , which measures the average difference between the pad positions in the generated footprint and the actual positions;

[0054]

[0055] Area represents the pad area, Pos represents the pad position, i represents the i-th component, and the subscripts gen and GT represent the generated and true values, respectively.

[0056] This paper proposes a design method for electronic component package generation using a multimodal large language model (MLLM), aiming to solve the inefficiency and error-proneness of the traditional manual package creation process, as well as the challenges brought by the heterogeneity of data table formats and the lack of large-scale data sets. In terms of specific technical implementation, this paper solves the following key problems:

[0057] Data heterogeneity processing: By using a multimodal large language model (MLLM), it can process and parse data tables of different formats, thus solving the problem of data table format heterogeneity. MLLM can understand and process component specifications in various formats including lists, tables, and images.

[0058] Efficient and automated generation: In traditional methods, the packaging of electronic components needs to be created manually, which is inefficient and error-prone. The present invention uses MLLM to automatically generate packaging based on data tables through an automated process, greatly improving efficiency.

[0059] Improved precision and accuracy: Directly using MLLM for component design generation may lead to unsatisfactory results, such as incorrect number of pins and inaccurate size of the generated package shape. The present invention guides MLLM to generate more accurate results by designing a specific query process and providing a data set of contextual examples.

[0060] Learning and optimization of logical and spatial position parameters: In the package generation process, by introducing visual cues and thought chain CoT, the model can more accurately describe the pads based on the given index, improving the accuracy of package generation.

[0061] Quantitative quality assessment: The quality of the generated electronic component packaging is assessed to achieve quantitative quality assessment, including accuracy indicators ACCN, ACCT and difference indicator Dif A 、Dif P .

[0062] In general, the present invention mainly solves key problems such as data heterogeneity processing, automatic generation, precision and accuracy improvement, logic and spatial parameter learning optimization, and quality assessment quantification at the technical implementation level. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 :Comparison between traditional design method and SFgen of this application;

[0064] Figure 2 :An embodiment of the encapsulation generation process;

[0065] Figure 3 : Explanation of the visual cue (VP): Without VP, the large language model (LLM) answered four pads incorrectly; with VP, the large language model (LLM) answered six pads correctly.

[0066] Figure 4 : Example of package generation prompt;

[0067] Figure 5 : Visualization results of the generated package

[0068] Figure 6 : Visualization results of encapsulation ablation studies. DETAILED DESCRIPTION

[0069] The following is a detailed description of the embodiments of the present invention. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, modifications or equivalent substitutions made to the present invention should all be included in the scope of the claims of the present invention.

[0070] In addition, in order to better illustrate the present invention, numerous specific details are given in the following specific embodiments. It will be understood by those skilled in the art that the present invention can also be implemented without these specific details.

[0071] The present invention provides a design method for generating electronic component packaging using a multimodal large language model, comprising the following steps:

[0072] Step 1: Collect and preprocess data: Collect data sheets containing package details from the Internet and electronic design automation software. The package information includes pad size, pad position, and component body outline.

[0073] Step 2: Define the conditional text generation problem for encapsulation generation;

[0074] Step 3: Develop a package generation design process based on a multimodal large language model;

[0075] Step 4: Implement the package generation process by extracting, contextualizing, enhancing visual cues, and guiding thought chains from the input data table;

[0076] Step 5: By evaluating the quality of the packaging of the generated electronic components, a quantitative evaluation of the quality is achieved.

[0077] In the above scheme, step 1 includes:

[0078] Step 1.1, collect <data sheet, package> data pairs from the Internet and EDA software;

[0079] Step 1.2. Select a diverse subset of components based on the number of pads and component type.

[0080] In the above scheme, step 2 includes:

[0081] Step 2.1, define the component package generation task as a conditional text generation problem;

[0082] Step 2.2: Automatically generate a package based on the provided component data sheet.

[0083] In the above scheme, step 3 includes:

[0084] Step 3.1: Use the dataset Where (x j ,y j ) is the jth sample of the dataset, a {input constraint, output text} pair, and M is the total number of samples;

[0085] Step 3.2: Provide k samples As an ICL template, where x j and j They are constraint and target files respectively;

[0086] Step 3.3: Based on constraint x test , the goal is to generate high-quality component library files y test ;

[0087] Step 3.4, design the prompt P to integrate the input constraints, ICL examples, task description and other relevant information;

[0088] Step 3.5: Introduce CoT into the prompt to enhance the model’s logical processing capability of input information.

[0089] Step 3.6: Integrate the prompt P into the MLLM, where P = [R; x1; y1; ...; x k ;y k ;x test ], R is the task description, x1 to x k is the constraint of the ICL example, y1 to y k is the corresponding target file, x test is the constraint of the current input;

[0090] Step 3.7: Process the integrated prompt P through MLLM to generate a component library file y that meets the constraints test .

[0091] In the above scheme, step 4 includes:

[0092] Step 4.1, extract the package information from the input data sheet, including the size and position of each pad;

[0093] Step 4.2: Use a small amount of context to learn ICL examples to help MLLM understand the constraints of pad location and size;

[0094] Step 4.3: Introduce visual cues VP for package generation to enhance SFgen's ability to understand pad size and location;

[0095] Step 4.4. Respond to the input by generating a preliminary package file that can be used by the EDA

[0096] Software is used for visualization;

[0097] Step 4.5: Introduce the modified prompts, check the encapsulation step by step and correct the errors;

[0098] Step 4.6, include the thought chain CoT in the prompt to guide the model to more accurately describe the pad according to the given index, improving the accuracy of the package generation;

[0099] In the above scheme, step 5 includes:

[0100] Step 5.1, a set of evaluation metrics are defined to quantify the quality of the generated footprint, including

[0101] ACCN, ACCT, Dif A and Difp;

[0102] Step 5.1.1. Calculate the accuracy index ACCN, which is used to measure the ratio of the number of correctly identified pads to the total number of pads in the generated package:

[0103]

[0104] Step 5.1.2. Calculate the accuracy index ACCT, which is used to measure the ratio of the number of correctly identified shape types to the total number in the generated footprint:

[0105]

[0106] Step 5.1.3. Calculate the difference index Dif A , which measures the average difference between the generated pad size and the actual size;

[0107]

[0108] Step 5.1.4. Calculate the difference index Dif P , which measures the average difference between the pad positions in the generated footprint and the actual positions;

[0109] represents the pad area, Pos represents the pad position, i represents the i-th component, and the subscripts gen and GT represent the generated and true values, respectively.

[0110] experiment

[0111] The present invention implements the SFgen framework and conducts experiments using the state-of-the-art MLLM, GPT4. All experiments are performed via API calls. Visualization is done with KiCad 8.0.5. The complete prompt is as follows Figure 4 shown.

[0112] Packaging generation results

[0113] Four indicators ACCN, ACCT, Diff A and Diff P To measure the quality of the generated footprints, Table 2 summarizes the results. We grouped the components into several families. ACCN reached 100% indicating that all packages were generated with the correct number of pads. ACCT was 99.55% indicating that almost all pad types were correctly identified. The average values ​​of DiffA and DiffP were calculated to be 2.82% and 6.12%, respectively, indicating that the generated footprints are similar to GT in terms of their area and spatial layout.

[0114]

[0115] Figure 5 Some examples of package visualization results are shown. The generated footprints consistently reflect the pad shape, size, and spacing, demonstrating SFgen’s ability to ensure design accuracy. Since the generated files are text-based and editable, minor modifications can be made through the EDA platform to make the generated footprint conform to the desired metallization pattern.

[0116] Generated Ablation Studies

[0117] We conduct a series of package generation ablation studies to evaluate the effectiveness of our VP, CoT, and modified hinting techniques. Figure 6 Sample visualization results under different experimental settings are shown, and the average footprint generation results are summarized in Table 3. Since reference examples are essential for SFgen to generate meaningful content, ICL is assumed for all test cases.

[0118] The addition of CoT can significantly improve the spatial distribution and configuration of elements in CAD design. Figure 6 As shown, incorporating CoT helps SFgen generate a footprint with a more accurate number of pads and positions that are more aligned with GT than the baseline. Figure 6(a) shows that, in the absence of VP, CoT alone cannot guarantee the correct number of pads in other cases. Table 3 shows that applying CoT produces 93.06% pad number accuracy and 92.94% pad type accuracy, which are significantly higher than the baseline's 22.85% and 14.64% results.

[0119]

[0120] Similar to CoT, VP enhances the number and location of pads in the generated footprint compared to the baseline. This highlights the fact that SFgen benefits significantly from the contextual information annotated on the image. The combination of VP with CoT produces stronger results for all components, where accurate pad counts are generated and only minor modifications are required. Table 3 shows that applying VP produces 100% pad count accuracy and 98.65% pad type accuracy, which is better than CoT. This is likely due to the additional information provided by VP. Moreover, combining VP with CoT further improves the accuracy, reaching 98.82% for pad type, while the number of pads remains at 100%.

[0121] The footprint generated using VP+CoT was then re-imported into SFgen for further optimization. Figure 6 As shown in Table 3, the proposed hint modification procedure aligns the design elements to the GT more accurately by correcting the subtle differences step by step. The modifications have resolved the single inaccuracy and improved the integrity of the overall layout. As shown in Table 3, with the hint modification technique, SFgen achieves 100% pad number accuracy and 99.55% pad type accuracy. It also minimizes the differences in pad area and position.

Claims

1. A design method for generating electronic component packaging using a multimodal large language model, characterized in that: The following steps are involved: Step 1: Collect and preprocess data: Collect data sheets containing package details from the Internet and electronic design automation software. The package information includes pad size, pad position, and component body outline. Step 2: Define the conditional text generation problem for encapsulation generation; Step 3: Develop a package generation design process based on a multimodal large language model; Step 4: Implement the package generation process by extracting, contextualizing, enhancing visual cues, and guiding thought chains from the input data table; Step 5: By evaluating the quality of the packaging of the generated electronic components, a quantitative evaluation of the quality is achieved.

2. The method for designing electronic component packaging using a multimodal large language model according to claim 1, characterized in that: Step 1 includes: Step 1.1, collect <data sheet, package> data pairs from the Internet and EDA software; Step 1.

2. Select a diverse subset of components based on the number of pads and component type.

3. The design method for generating electronic component packaging using a multimodal large language model according to claim 1, characterized in that: Step 2 includes: Step 2.1, define the component package generation task as a conditional text generation problem; Step 2.2: Automatically generate a package based on the provided component data sheet.

4. The method for designing electronic component packaging using a multimodal large language model according to claim 1, characterized in that: Step 3 includes: Step 3.1: Use the dataset Where (x j ,y j ) is the jth sample of the dataset, a {input constraint, output text} pair, and M is the total number of samples; Step 3.2: Provide k samples As an ICL template, where x j and j They are constraint and target files respectively; Step 3.3: Based on constraint x test , the goal is to generate high-quality component library files y test ; Step 3.4, design the prompt P to integrate the input constraints, ICL examples, task description and other relevant information; Step 3.5: Introduce CoT into the prompt to enhance the model’s logical processing capability of input information. Step 3.6: Integrate the prompt P into the MLLM, where P = [R; x1; y1; ...; x k ;y k ;x test ], R is the task description, x1 to x k is the constraint of the ICL example, y1 to y k is the corresponding target file, x test is the constraint of the current input; Step 3.7: Process the integrated prompt P through MLLM to generate a component library file y that meets the constraints test .

5. The method for designing electronic component packaging using a multimodal large language model according to claim 1, characterized in that: Step 4 includes: Step 4.1, extract the package information from the input data sheet, including the size and position of each pad; Step 4.2: Use a small amount of context to learn ICL examples to help MLLM understand the constraints of pad location and size; Step 4.3: Introduce visual cues VP for package generation to enhance SFgen's ability to understand pad size and location; Step 4.4, respond to the input by generating a preliminary package file that can be used for visualization by the EDA software; Step 4.5: Introduce the modified prompts, check the encapsulation step by step and correct the errors; Step 4.

6. Include the chain of thought CoT in the prompt to guide the model to more accurately describe the pad according to the given index, improving the accuracy of package generation.

6. The method for designing electronic component packaging using a multimodal large language model according to claim 1, characterized in that: Step 5 includes: Step 5.1, a set of evaluation metrics are defined to quantify the quality of the generated footprint, including ACCN, ACCT, Dif A and Dif P ; Step 5.1.

1. Calculate the accuracy index ACCN, which is used to measure the ratio of the number of correctly identified pads to the total number of pads in the generated package: Step 5.1.

2. Calculate the accuracy index ACCT, which is used to measure the ratio of the number of correctly identified shape types to the total number in the generated footprint: Step 5.1.

3. Calculate the difference index Dif A , which measures the average difference between the generated pad size and the actual size; Step 5.1.

4. Calculate the difference index Dif P , which measures the average difference between the pad positions in the generated footprint and the actual positions; represents the pad area, Pos represents the pad position, i represents the i-th component, and the subscripts gen and GT represent the generated and true values, respectively.

Images