An autonomous LLM agent built using the Model Context Protocol (MCP) that optimizes symbolic regression on real industrial datasets from particle technology research. This project was done as Master's Thesis in Institute of Particle Technology within TU Braunschweig
Traditional symbolic regression requires manual parameter tuning — a time-consuming trial-and-error process with no guarantee of physical consistency.
An agentic loop where the LLM:
- Evaluates regression outputs against physics constraints
- Detects overfitting and unphysical terms
- Autonomously adjusts PySR parameters & search grammars
- Iterates until convergence on a physically meaningful equation
Python · PySR · MCP · Gemini 2.5 Flash · PyTorch · Scikit-learn
- Successfully derived physically consistent equations from 6 industrial targets
- Reduced manual iteration cycles significantly
- Composite physics scoring across 7 evaluation dimensions