IDP Project: natural language assistant of an open-source multi-physics library and simulator

Institut

Lehrstuhl für Aerodynamik (TUM-ED)

Typ

Semesterarbeit /

Inhalt

theoretisch /  

Beschreibung

Project Overview and Goal
Generally, the learning curve for open-source library is quite steep because the lack of documentation, especially when the library is large and complicate, for instant the multi-physics libraries. Nowadays, the advanced AI tools open a new way for decrease the steepness of learning curve by fine-tuning a baseline LLM model to an expert on a specific open-source
library.
This project aims to build up a continuous fine-tuning workflow for the open-source multiphysics library SPHinXsys and its python simulator SPHinXsim. The focus will be on
exploring variable approaches by which the model can be trained directly based on the Github repositories.
 

Key Highlights
• Parameter-Efficient Fine-Tuning (PEFT), a technique like LoRA (Low-Rank Adaptation) will be used.
• No code-generation is required, and the simulator already has been using LLM for structured JSON generation.
• The integrated platform will be built on python3 for easy access.
• Continuous fine‑tuning (retrain on new development of the libraries).
 

References: 

To explore SPHinXsys and SPHinXsim, check out from the repository:
github.com/Xiangyu-Hu

Möglicher Beginn

sofort

Kontakt

PD Dr. -Ing. habil. Xiangyu Hu
Raum: MW1636
Tel.: +49.89.289.16152
xiangyu.hutum.de

Ausschreibung