Zhen Fan

Sim2Know: Digital Twins for Action Recognition

Challenge: Collecting labeled real-world data for manufacturing is resource-intensive.

Method: Built a high-fidelity Digital Twin of a metal AM system to generate synthetic human action data. Trained a V-JEPA model using transfer learning with a mix of synthetic and real data.

Submitted to CIRP Annals 2025.

MetalMind: Human-Centric Knowledge System

Challenge: Expert knowledge is often "implicit" or locked in unstructured PDF manuals.

Method: Developed an automated pipeline using LLMs to extract entities and build a structured Knowledge Graph (3,102 Nodes). Implemented a Hybrid Retrieval strategy (Vector + Graph) for better context.

Submitted to npj Advanced Manufacturing.