TCAD Structure Input File Generation Using Large Language Model

In this paper, we study the possibility of using Large Language Models (LLMs) to create Technology Computer-Aided-Design (TCAD) structure generation input files. LLMs are machine learning models trained on vast amounts of text data from the Web and are designed to understand, generate, and interact with humans through natural languages such as English. However, unlike programming languages with abundant training examples on the Web, TCAD examples are scarce. In this work, by using TCAD Sentaurus Structure Editor (SSE) as an example, 7000 nanowire data are generated to fine-tune open-source models (Llama 2 and 3) to obtain chatbots that can generate an SSE input file for a nanowire with 18 parameters with English as an instruction. Structures are then created using SSE to verify the correctness of the input files. It shows that it is possible to create a chatbot even with limited resources. In-context one-shot learning is also studied. It shows that with large commercial models, in-context one-shot learning is sufficient to generate the desired SSE input file.