What would it take for an AI system to discover a scientific theory — not just summarize existing work, but propose new principles about the world? This thesis explores the conceptual and technical foundations of building AI systems capable of scientific discovery, inspired by the idea of an AI system that could, in principle, invent something as profound as relativity theory.

The thesis focuses on conceptual design, system architecture, and experimental prototyping — not just fine-tuning. You will explore how LLMs can be combined with structure, constraints, reasoning modules, simulations, or external tools to autonomously generate and test hypotheses. The scientific domain (e.g., physics, chemistry, biology) is flexible and can be chosen according to your interests.

This is an ambitious research-oriented thesis with strong publication potential.

What are the tasks?

Understand LLM capabilities & limitations in scientific reasoning

• Review literature on LLMs in science (reasoning, hypothesis generation, automated discovery).
• Analyze how far current LLMs can go in reasoning and where they break.
• Identify a scientific domain and a concrete discovery challenge (e.g., rediscovering a known law, proposing hypotheses from data).

Design the architecture of an “AI Scientist” system

• Propose how to combine an LLM with additional components, such as:
  ◦ simulation environments (e.g., physics engines)
  ◦ symbolic reasoning or equation solvers
  ◦ retrieval or domain-specific knowledge bases
  ◦ planning or multi-agent frameworks
• Develop a pipeline for hypothesis generation, testing, refinement, and reporting.
• Decide on levels of autonomy vs. user guidance.

Implement a prototype and run discovery experiments

• Build a minimal working system that:
  ◦ generates hypotheses,
  ◦ tests them using simulations or data,
  ◦ updates or refines them.
• Evaluate whether the system can:
  ◦ rediscover known principles (baseline validation),
  ◦ propose plausible new hypotheses.
• Analyze outcomes and failure modes.

Reflect on scientific validity & future potential

• Discuss gaps between LLM-driven discovery and human scientific reasoning.
• Outline what is needed to scale such systems to complex scientific domains.

What prerequisites do you need?

• Strong interest in scientific reasoning, discovery processes, or philosophy of science.
• Good programming skills in Python; experience with machine learning is helpful.
• Very good English skills.

Why this thesis is special

• High-impact vision: If successful, even in a small domain, results can inspire new research directions.
• Scientific depth: Combines AI, reasoning, science, and systems design.
• Strong publication potential: Novel, conceptual, and timely.