SemiAI Delivers Invited Talk on "Agentic AI for Closed-Loop Semiconductor Manufacturing" at KISM 2026

Taekwon Jee, CEO of SemiAI, a semiconductor manufacturing AI company, delivered an invited talk on November 12 at KISM 2026 BUSAN, the 5th Korean International Semiconductor Conference & Exhibition on Manufacturing Technology, at the Paradise Hotel Busan.

KISM 2026, held November 9 to 13 under the theme "Beyond Scaling: Smart Manufacturing for AI Era," featured Jee in the "Advanced Metrology & Inspection, Process Diagnostics & Control, and Yield Management V" session. He presented "Agentic AI for Closed-Loop Semiconductor Manufacturing: Autonomous Diagnostics, Virtual Metrology, Process Control, and Yield Optimization." The work was co-authored with Jenghun Suh, Jungsuk Kim, Minjeong Jin and Jihoon Jung.

Jee explained that sub-nanometer process control is no longer only a hardware problem but a problem of information and decision-making. In most fabs, diagnostics, metrology, process control and yield management are split across different systems and teams, so coupled decisions are handed off manually.

The talk introduced building blocks based on papers published from 2024 to 2026. PRISM, a domain-knowledge fusion ML model for overlay prediction, achieved about 3x lower RMSE than linear regression in a bimodal deformation stress test. LLM-based overlay issue classification gave the correct diagnosis in 4,760 of 5,000 cases (95%). A human-in-the-loop root-cause analysis (RCA) agent using Causal-RAG achieved 91% Top-5 root-cause retrieval accuracy.

Jee then presented a closed-loop framework that groups these building blocks into five agents: Data, Diagnostics, Prediction, Control and Yield. The agents share evidence and uncertainty with one another as peers, and a deterministic orchestration layer manages validation gates, engineer approvals and audit trails. Autonomy proceeds in stages: observe (L0), recommend (L1), supervised action (L2), bounded autonomy (L3) and self-optimizing fab (L4, a research goal). The current stage is L1, in which agents recommend with evidence and engineers decide. L2 is in development. Evaluation covers four axes: diagnostic validity, prediction calibration, control stability and measured yield impact.

"AI will not replace engineers. It will close the loop around them," Jee said.

KISM 2026 was organized by The Korean Society of Semiconductor & Display Technology (KSDT). Founded in 2024, SemiAI is based in Seoul and Menlo Park, U.S. Part of this work was supported by the Technology Development Program funded by the Ministry of SMEs and Startups (MSS), Korea.