Exploring LLMs for Causal Discovery in Cyber-Physical Systems

Causality plays a fundamental role in both human reasoning and complex system analysis. As Cyber-Physical
Systems (CPS) become increasingly complex, understanding causal relationships between system events is
essential for tasks such as anomaly detection and fault diagnosis. This paper explores the potential of Large
Language Models (LLMs) to support causal discovery in CPS. In particular, the capabilities of LLMs in assisting
domain experts to identify system states and their causal relations are investigated. We propose a hybrid workflow
that integrates LLM-generated suggestions with domain expert validation, aiming to improve the efficiency of
causal analysis. Our evaluation is conducted on a real-world smart grid use case and compares LLM-generated
causal relations with domain expert-validated ground truth. The results indicate that, while LLMs can propose
relevant causal structures, their effectiveness varies depending on the complexity of temporal and topological
relationships between system states. Although these models do not replace human domain expertise, they can
serve as a valuable tool for supporting causal discovery in a hybrid workflow. Future research should focus on
refining LLM capabilities and expanding their application across different CPS domains. Investigating different
LLMs, causal models, and larger datasets may provide deeper insights into their potential for causal discovery.