Turning POSCO DX's Engineering Tacit Knowledge into AI-Usable Knowledge Assets

The core of manufacturing sites, PLC (Programmable Logic Controller) control design know-how, often relies on the experience of senior engineers not found in manuals—that is, Tacit Knowledge. POSCO DX is pursuing AI-based design automation to solve the problem of this knowledge being lost upon retirement.

As a result of verification, approximately 80% (28) of the 35 identified design rules turned out to be UDK (Untrained Domain Knowledge) that general-purpose LLMs had not learned. In particular, 60% of these were classified as the company's own tacit knowledge that was not even documented, confirming that while AI can provide general answers, it has limitations in deriving the optimal solution for the field.

To address this, POSCO DX systematized knowledge by classifying it along three axes (whether AI has learned it, the nature of the knowledge, and the scope of the knowledge). Going beyond simple text search, the company is using a Knowledge Graph based on Amazon Bedrock and Graphiti to structure UDK, and is verifying through a PoC an architecture that allows AI to reference the accurate design rules at the point of need.