TCAD RAG Pipeline: Automated Sentaurus Script Generation

A Retrieval-Augmented Generation (RAG) pipeline that converts natural language prompts into executable Synopsys Sentaurus TCAD scripts (SDE and SDevice) using vector search and Google's Gemini large language model.

---

Overview

Writing Sentaurus scripts manually is time-consuming and requires detailed knowledge of syntax, device physics models, and simulation workflows.

This project automates the process by combining:

• Retrieval-Augmented Generation (RAG) to retrieve relevant sections from Sentaurus manuals and example scripts
• FAISS for vector similarity search
• Sentence Transformers for semantic embeddings
• Google Gemini for script generation
• Streamlit for an interactive web interface

Users can describe a device in plain English, and the system generates complete, runnable TCAD scripts.

---

Example Prompt

Generate a 35nm NMOSFET with LDD and halo doping and an Id-Vg sweep.

Output

• SDE script defining geometry, contacts, doping, and mesh
• SDevice script containing physics models, solver settings, and bias sweeps

---

Key Features

• Natural language to Sentaurus script generation
• Hybrid retrieval from:
  • Sentaurus SDE and SDevice manuals
  • Example device scripts
• FAISS vector database for fast semantic search
• Prompt engineering to enforce script structure
• Built-in validation for required sections
• Downloadable .cmd files
• Streamlit web application

---

Architecture

User Prompt
    ↓
Embedding Model (all-MiniLM-L6-v2)
    ↓
FAISS Similarity Search
    ↓
Retrieve:
  • Manual Context
  • Example Scripts
    ↓
Prompt Construction
    ↓
Google Gemini
    ↓
Generated:
  • SDE Script
  • SDevice Script
    ↓
Validation + Download