Generative design algorithms usually generate solutions in a single pass. Feedback from evaluation, for example from structural or manufacturing simulations, is rarely used to specifically improve flawed solutions. In addition, rule-based approaches are often restricted to clearly defined boundary conditions.

The institute has developed a rule-based generative design algorithm for bent sheet metal parts that generates manufacturable design concepts. However, the algorithm only operates within certain geometric boundary conditions and discards solutions that violate its rules.

AI agents based on Large Language Models can call tools autonomously, interpret results and refine solutions iteratively. Combined with simulation feedback and design knowledge, they can orchestrate generative algorithms and extend their scope of application.

Tasks

The goal of this thesis is to develop an orchestration agent that uses the existing algorithm as a tool and iteratively improves sheet metal parts based on simulation feedback.

Possible tasks include:

• Design of an agent architecture that uses the algorithm or individual subfunctions as tools via a Python API
• Integration of existing FEM and manufacturing simulations as feedback sources for the agent
• Integration of design methods and guidelines as a knowledge base
• Development of strategies for the iterative improvement of flawed or previously discarded parts
• Extension of the generation to geometries beyond the current boundary conditions
• Evaluation of the agent regarding solution quality, improvement rate and comparison with the algorithm without the agent

The specific focus of the thesis can be adapted to your interests and field of study.

Profile

• Studies in mechanical engineering, mechatronics, computer science or a related field
• Interest in AI agents, generative design or simulation
• Programming skills in Python

If you are interested, feel free to contact me: christoph.wittig∂kit.edu