What is SEMIKONG
SEMIKONG is a large language model (LLM) customized specifically for the semiconductor industry, jointly launched by Aitomatic, FPT Software, and Tokyo Electron Limited. Based on in-depth domain knowledge, it addresses unique challenges in semiconductor manufacturing and design, such as complex physics and chemistry problems. SEMIKONG integrates expert knowledge and optimizes the pretraining process to provide a foundation model capable of expert-level understanding of etching issues. Compared with general-purpose LLMs, SemiKong demonstrates superior performance in semiconductor manufacturing tasks, laying the foundation for the development of company- or tool-specific proprietary models and driving further research and application of domain-specific AI models.
Main Features of SEMIKONG
- Understanding specialized knowledge in the semiconductor field: Understanding complex issues in semiconductor manufacturing and design, especially in the etching field.
- Optimizing manufacturing processes: Based on learning from large amounts of semiconductor-related data, assisting in optimizing semiconductor manufacturing processes, such as parameter optimization, anomaly detection, and predictive maintenance.
- Assisting IC design: SEMIKONG can assist with integrated circuit (IC) design tasks, including design rule checking, layout generation, and design space exploration.
- Improving the performance of AI solutions: Based on fine-tuning pretrained large language models, improving the performance of AI-driven semiconductor manufacturing tasks.
- Expert knowledge integration: Introducing a framework that integrates expert knowledge to advance the evaluation process of domain-specific AI models.
Technical Principles of SEMIKONG
- Data curation: The development of SEMIKONG begins with the curation of large-scale, high-quality semiconductor domain-specific text datasets, including technical books, papers, and patents.
- Pretraining and fine-tuning: The model is trained using pretraining and fine-tuning methods. The pretraining stage uses domain-specific data to enhance the model's domain knowledge, and the fine-tuning stage enables the model to perform specific tasks.
- Domain ontology: Collaborating with semiconductor experts to build a systematic ontology of semiconductor manufacturing processes, helping AI researchers develop domain-specific AI models more effectively.
- Expert feedback loop: Based on an expert feedback loop, using experts to evaluate answers generated by the model, producing evaluation criteria and high-quality benchmarks.
- Model quantization and adaptation: After pretraining and fine-tuning, model quantization and adaptation processing are carried out to facilitate model deployment.
SEMIKONG Project Address
- Project official website：semikong.ai
- GitHub repository：https://github.com/aitomatic/semikong
- HuggingFace model library：https://huggingface.co/pentagoniac
- arXiv technical paper：https://arxiv.org/pdf/2411.13802
Application Scenarios of SEMIKONG
- Process parameter optimization: Adjusting parameters in semiconductor manufacturing processes to improve production efficiency and product quality.
- Anomaly detection system: Assisting in identifying anomalies in manufacturing processes, reducing defects and improving reliability.
- IC design assistance tools: Helping engineers comply with design rules in integrated circuit design and generate optimized layouts.
- Expert decision support: Serving as an expert system to provide solutions and decision support for complex technical issues.
- Technical education platform: Used as an educational tool to help learners understand the complex concepts of semiconductor manufacturing and design.