Hacking Claude Code for Scientific Research (AI-Driven Material Science) | Dr. Peter Felfer

Описание: AI-Driven Knowledge Extraction: How Materials Science Extends the Limits of LLMs
(Prof. Dr. Peter Felfer, Institute of General Materials Properties, Friedrich-Alexander-Universität Erlangen–Nürnberg)

The explosive growth of scientific literature poses an increasing challenge for researchers. In many fields—particularly highly empirical domains such as materials science—it has become virtually impossible to maintain a comprehensive overview of the entire knowledge base. Our case study, the effects of hydrogen on materials, already encompasses an estimated 15,000 publications, with numbers rising rapidly. This is a critical field for the energy transition, where all existing knowledge must be considered, as no universal first principles fully capture the complexity involved.

Processing such vast amounts of data exceeds the capabilities of conventional large language models (LLMs), since the information cannot fit into a single context window. To address this scalability challenge, we present our solution: a Hierarchical Retrieval-Augmented Generation (RAG) system. This system leverages domain-specific expertise in materials science to systematically structure and preprocess the extensive body of literature. Instead of handling all documents at once, it applies a multi-stage AI-driven preprocessing and retrieval pipeline, combined with metadata, ensuring that only the most relevant information is fed into the LLM’s context window.

This hierarchical approach not only overcomes the context window limitations of LLMs in research settings, but also provides a transferable blueprint for industry applications. Wherever large, heterogeneous data silos exist—such as in legal compliance, pharmacovigilance (drug safety monitoring), or the management of technical documentation and corporate knowledge—this system can transform LLMs into tools capable of delivering precise, evidence-based answers from extensive data archives.

In this talk, we demonstrate how a hierarchical RAG approach enables the practical use of LLMs in scenarios that were previously infeasible due to sheer data scale.