 ICACT20250028 Slide.01
[Big Slide]
|
Chrome  Click!! |
.jpg) |
Hello. I am very pleased to present my research on Generative AI-based Optimized Code Generation for Semiconductor Equipment Interfaces today. |
| ICACT20250028 Slide.02
[Big Slide]
|
Chrome Click!! |
.jpg) |
Today's presentation will proceed in the following order. |
| ICACT20250028 Slide.03
[Big Slide]
|
Chrome Click!! |
.jpg) |
- |
| ICACT20250028 Slide.04
[Big Slide]
|
Chrome Click!! |
.jpg) |
The semiconductor industry demands a high level of automation and unmanned operation, making equipment control based on the SECS/GEM standard essential. However, due to a shortage of skilled developers and high turnover rates, developing equipment control software using the SECS/GEM standard is becoming increasingly challenging.
In this study, we propose a novel approach that leverages Generative AI to overcome the limitations of traditional code generation methods. Existing SECS/GEM-based code generation tools provide predefined template codes, but adapting these templates to real-world applications requires significant developer intervention. In contrast, the Generative AI-based optimized code generation method proposed in this study learns from existing code data and effectively generates code that closely aligns with the developer's intended functionality.
This approach reduces the time required for code modification and refinement, enabling efficient semiconductor equipment control software development with fewer personnel.
We experimentally evaluated the performance of the proposed AI model to validate its effectiveness, demonstrating the feasibility of our approach. We believe that this research marks a significant turning point in enhancing productivity and advancing automation in the semiconductor industry.
|
| ICACT20250028 Slide.05
[Big Slide]
|
Chrome Click!! |
.jpg) |
In this study, we first review the SECS/GEM standard and existing code generation techniques before proposing an automatic code generation method for semiconductor equipment control using Generative AI.
We validated the effectiveness of our approach using evaluation metrics commonly employed in LLM research and will share the results of our experimental evaluation. The experimental findings confirm the significance of the Generative AI-based approach.
We expect this research to contribute to the automation and optimization of semiconductor equipment software development. In the final part of this presentation, we will discuss the implications of our study and outline directions for future research.
|
| ICACT20250028 Slide.06
[Big Slide]
|
Chrome Click!! |
.jpg) |
- |
| ICACT20250028 Slide.07
[Big Slide]
|
Chrome Click!! |
.jpg) |
Standardized protocols are essential for seamless communication between semiconductor equipment and host systems. SEMI, a global organization with over 2,000 member companies, defines the SECS/GEM standard to facilitate such communication.
The SECS/GEM standard systematically standardizes equipment control and data collection, serving as a key component in enabling automation and smart factory implementation in semiconductor manufacturing processes. |
| ICACT20250028 Slide.08
[Big Slide]
|
Chrome Click!! |
.jpg) |
The core components of smart factories and factory automation are manufacturing equipment, the Equipment Interface Server (EIS), and the Manufacturing Execution System (MES).
Manufacturing equipment includes Programmable Logic Controllers (PLCs), which communicate with the Equipment Interface Server (EIS) to transmit and receive data. However, traditional interface methods are often inefficient, requiring additional controllers to ensure smooth communication and efficient equipment control.
While the introduction of these additional controllers increases initial costs, they play a crucial role in enhancing equipment control capabilities and supporting the SECS/GEM standard. As a result, they enable more precise process management and data exchange, ultimately contributing to the optimization of semiconductor manufacturing processes and the implementation of smart factories.
|
| ICACT20250028 Slide.09
[Big Slide]
|
Chrome Click!! |
.jpg) |
The automatic generation of SECS/GEM standard code has been utilized since its early adoption. By selecting the appropriate Stream and Function according to the standard, a code template is automatically generated, which developers then customize to align with the equipment specifications and operational scenarios.
This automated code generation approach is a key factor in improving productivity in semiconductor equipment interface development. However, traditional methods still have limitations, as they require a significant amount of manual adjustments by developers.
|
| ICACT20250028 Slide.10
[Big Slide]
|
Chrome Click!! |
.jpg) |
In code generation using generative AI, large language models (LLMs) play a crucial role. LLMs are pre-trained on extensive language datasets through upstream tasks and then fine-tuned on downstream tasks to address specific problems.
However, LLMs present key risks, such as hallucination and limited controllability. To mitigate these challenges, various technical approaches are employed, as outlined below.
|
| ICACT20250028 Slide.11
[Big Slide]
|
Chrome Click!! |
.jpg) |
- |
| ICACT20250028 Slide.12
[Big Slide]
|
Chrome Click!! |
.jpg) |
In this paper, we propose an automatic SECS/GEM code generation method using LLMs. Existing SECS/GEM codes, which serve as control data for semiconductor equipment, are stored in a vector database. When a user requests code generation through a prompt, the system utilizes the Retrieval-Augmented Generation (RAG) approach to automatically generate the desired SECS/GEM code based on the stored data.
Fine-tuning is essential for handling downstream tasks, but it requires significant computational resources. To address this, we opted for the relatively lightweight RAG technique, offering an efficient and cost-effective solution.
To facilitate code generation, the necessary design documents are converted into PDF format and loaded into a Lang Chain-based service. These documents are then stored in a vector database, enabling efficient retrieval. This approach not only conserves computational resources but also provides code generation based on design documents. |
| ICACT20250028 Slide.13
[Big Slide]
|
Chrome Click!! |
.jpg) |
When a user generates a query, the system first searches the vector store for relevant information. The retrieved results are then combined with the query and sent to the LLM service, which generates the final response. This process plays a crucial role in efficient code generation.
The following figure illustrates the overall architecture of the SECS/GEM code generator. This architecture visually explains how data retrieval and LLM-based code generation occur at each stage.
|
| ICACT20250028 Slide.14
[Big Slide]
|
Chrome Click!! |
.jpg) |
This flowchart illustrates the process of generating equipment interface source code based on user requests |
| ICACT20250028 Slide.15
[Big Slide]
|
Chrome Click!! |
.jpg) |
- |
| ICACT20250028 Slide.16
[Big Slide]
|
Chrome Click!! |
.jpg) |
This experiment demonstrates how the process of automatically converting SECS messages into XML-formatted code is applied in actual system development. The experiment was conducted in the following steps:
The code for converting SECS messages into XML was prepared and converted into a PDF format, after which the information was stored in a Vector Database (Vector DB).
Upon a user's request, the system retrieves relevant information from the Vector DB and uses a prepared prompt to invoke the LLM, which automatically generates the XML code for the SECS messages.
|
| ICACT20250028 Slide.17
[Big Slide]
|
Chrome Click!! |
.jpg) |
The Chain of Thought (CoT) prompt used in this experiment is as follows:
CoT is a technique that systematically outlines the problem-solving process, helping the model to generate more accurate and consistent results through logical reasoning.
This method ensures that even in complex code generation tasks, a clear logical flow is maintained step by step.
By using this prompt, the model can clearly understand each step before performing a task, allowing it to generate optimized code based on this understanding.
|
| ICACT20250028 Slide.18
[Big Slide]
|
Chrome Click!! |
.jpg) |
On the left, the SECS message set data is displayed, while on the right, the XML code automatically generated using the method proposed in this paper is shown.
This screen allows for a comparison of how SECS messages are converted into XML code. The message data on the left represents the raw data in its original form, while the XML code on the right is the automatically generated output, providing an important example of efficient code transformation.
This approach suggests that automated code generation can deliver accurate results while saving time and costs. |
| ICACT20250028 Slide.19
[Big Slide]
|
Chrome Click!! |
.jpg) |
ROUGE is a metric designed to evaluate the performance of text summarization models. It is particularly useful for assessing the performance of natural language generation models used in automatic text summarization or machine translation.
ROUGE calculates performance scores by comparing the summaries or translations generated by the model to human-generated reference texts. This process serves as an important benchmark for evaluating how accurately and naturally the model conveys key information.
Therefore, in this paper, the ROUGE metric was used to evaluate the performance of the SECS/GEM code generation model proposed, verifying the quality and accuracy of the automatically generated code.
|
| ICACT20250028 Slide.20
[Big Slide]
|
Chrome Click!! |
.jpg) |
This experiment was conducted on four different datasets, and the validity of the generated code was verified by comparing it with pre-prepared reference code.
To assess the similarity of the code, ROUGE-1, ROUGE-2, and ROUGE-L metrics were used, allowing us to quantitatively verify how closely the code generated by the model matches the actual code.
The table summarizes the ROUGE-1, ROUGE-2, and ROUGE-L values obtained from each experiment, along with their average values. This helps confirm that the method proposed in this paper ensures consistent performance.
The table below shows the evaluation method for the ROUGE values. A score of 0.5 or higher indicates satisfactory performance. |
| ICACT20250028 Slide.21
[Big Slide]
|
Chrome Click!! |
.jpg) |
In this study, we conducted experiments to validate the effectiveness of the proposed method.
The experiment was carried out in the process of converting SECS messages written by developers into XML code, where completed XML example code was pre-stored in a database. Based on this, a prompt template was designed to account for considerations when automatically converting SECS messages into XML code.
The system operates as follows:
First, when the user inputs a sentence requesting code generation, relevant information is retrieved from the stored database. Then, the system proceeds by using the configured prompt template and generative AI to generate the requested code.
As mentioned earlier, the similarity of the results was evaluated using the ROUGE-1, ROUGE-2, and ROUGE-L metrics.
Based on the predefined evaluation criteria, the experimental results demonstrated a high level of similarity, confirming the effectiveness of the proposed method.
|
| ICACT20250028 Slide.22
[Big Slide]
|
Chrome Click!! |
.jpg) |
- |
| ICACT20250028 Slide.23
[Big Slide]
|
Chrome Click!! |
.jpg) |
This study proposes a method for automatically generating the necessary code for semiconductor equipment interface development using generative AI. Through experimentation, we confirmed that the system demonstrates relatively effective performance.
The proposed system is capable of generating code that adheres to specific standards, highlighting the potential for using generative AI in no-code and low-code environments.
Future research will focus on advancing information generation technologies based on specialized documentation, with the goal of developing autonomous technologies that can replace tasks traditionally performed by human experts. |
| ICACT20250028 Slide.24
[Big Slide]
|
Chrome Click!! |
.jpg) |
- |
