Proactive Yield Maximization in Photolithography via Human-in-the-Loop AI on an On-Premise Big Data Platform
Yield loss in advanced semiconductor photolithography has become increasingly critical as device features continue to scale into the nanometer regime. Conventional reactive process control methodologies, such as Statistical Process Control (SPC), have proven insufficient for managing the growing complexity and data volume inherent in modern fabrication environments, frequently leading to elevated false alarm rates and delayed detection of process excursions. To address these limitations, a novel framework is proposed that transitions from reactive fault correction toward proactive yield optimization through the integration of Artificial Intelligence (AI) with a Human-in-the-Loop (HITL) protocol. The framework is designed to operate entirely on a secure, high-performance, on-premise big data platform, ensuring data sovereignty and low-latency performance.
The proposed framework comprises three core technological components: (1) a Causal-RAG (Retrieval-Augmented Generation) system for real-time root cause analysis of defect excursions; (2) a Process-Aware Generative AI model for predictive identification of patterning hotspots prior to wafer processing; and (3) a Semiconductor-Specific Multi-Modal AI Agent serving as an intelligent decision support system for integrated process and metrology control. Two diagrams are provided to illustrate the framework: the first depicts the high-level data flow and AI-HITL interaction loop, while the second presents a detailed architectural blueprint of the on-premise big data and MLOps platform.
Experimental validation demonstrates the effectiveness of the proposed approach. The Causal-RAG system achieved a Top-5 accuracy of 91% in retrieving and ranking relevant causal factors from historical defect databases. The Process-Aware Generative AI model attained a coefficient of determination (R²) exceeding 0.9 with a prediction accuracy of 0.2 nm for hotspot forecasting. The Semiconductor-Specific Multi-Modal AI Agent is presented as a conceptual architecture, with its design and anticipated operational workflow outlined for future implementation. These findings suggest that AI-driven proactive yield management offers a viable approach for addressing the challenges of advanced photolithography processes.
