From Correlation to Causation: Causal AI for Root Cause Analysis in Semiconductor Manufacturing

As semiconductor processes approach atomic-scale precision, even minor process deviations can cascade into yield loss and costly downtime. Conventional fault detection and classification (FDC) systems excel at flagging anomalies but fall short at explaining them, leaving engineers to manually sift through thousands of variables to find the true root cause. This talk presents a causal AI platform that changes that equation.
By combining causal graph learning, counterfactual simulation, and LLM-assisted diagnostics, the system identifies why faults occur, not just when, from heterogeneous data sources including equipment logs, metrology records, MES events, and maintenance histories. Unlike black-box ML models, the causal approach produces interpretable, engineer-readable explanations that integrate naturally into existing investigation workflows.
This talk will cover the architecture of a production-deployed causal AI system for semiconductor fault investigation, real-world lessons from deploying AI in fab and OEM environments, including air-gapped settings and a practical framework for evaluating AI-assisted RCA against domain expert assessments. We will also discuss the open challenges: sparse fault labels, multimodal data alignment, and what it takes to move from a compelling pilot to a system engineers rely on daily.

Key Technologies Covered

• Causal AI for Root Cause Analysis
• Causal Graph Learning for Fault Investigation
• Counterfactual Simulation for "Why" Diagnostics
• LLM-Assisted Diagnostics and Engineer-Readable Explanations
• Augmenting Fault Detection & Classification (FDC) Systems
• Multimodal Data Integration (Equipment Logs, Metrology, MES, Maintenance)
• Interpretable / Explainable AI vs. Black-Box Models
• AI Deployment in Air-Gapped Fab & OEM Environments
• Evaluation Framework for AI-Assisted RCA vs. Domain Experts
• Yield Loss & Downtime Reduction in Semiconductor Manufacturing