AVS 72 Session AIML2-FrM: AI/ML in Semiconductor Processing & Manufacturing

We developed a framework based on physics-informed process informatics for supercritical fluid deposition (SCFD) that enables efficient optimization of high deposition rates and conformality in high-aspect-ratio (HAR) structures. As semiconductor devices transition toward three-dimensional integration, high-throughput conformal deposition in HAR features such as through-silicon vias is increasingly required. SCFD, which involves chemical reactions of metal organics in supercritical CO₂(scCO₂) to allow thin film deposition, is promising because the high diffusivity and solubility of precursors in scCO₂ enable concentrated precursor transport deep into microstructures. However, fluid properties and reaction kinetics are highly sensitive to process conditions, and their relationships with deposition behavior have not been quantitatively established. Consequently, conventional process development has relied largely on empirical trial-and-error.

To address this issue, we propose a causal decomposition framework that separates process-condition from physicochemical-response effects, introducing physicochemical parameters as intermediate variables between process conditions and deposition outcomes. Specifically, the SCFD process is mapped onto a causal chain: process conditions → physicochemical parameters → deposition outcomes. The physicochemical parameters are linked to deposition outcomes through mass balance analysis of precursors in HAR features, clarifying their relationship with deposition rate and conformality. To model the relationship between process conditions and physicochemical parameters, we applied Physics-Informed Bayesian Optimization (PIBO). In PIBO, a prior physical model is incorporated as the mean function of Gaussian process regression. The predictive model is sequentially updated using newly acquired data to recommend subsequent process conditions in a closed-loop manner. In this contribution, PIBO was evaluated through sequential sampling from synthetic process-response surfaces designed to emulate SCFD behavior and compared with conventional Bayesian Optimization.

Rather than relying solely on data-driven learning, the proposed framework integrates physically describable behavior with data-driven residual learning. This hybrid approach enables optimal deposition conditions to be derived from sparse datasets while maintaining physical consistency and extrapolation reliability, and can be extended to different target structures without re-optimizing the entire process from scratch. This framework provides a scalable and physically interpretable strategy for optimizing next-generation semiconductor thin-film deposition processes.

The rapid expansion of AI services has significantly increased demand for advanced semiconductor chips, accelerating the need for equipment strategies that maximize yield, process stability, and tool availability. In fabs, even a single instance of unscheduled downtime can cause substantial losses in production throughput. However, conventional rule-based process management systems such as Fault Detection and Classification (FDC) and Statistical Process Control (SPC) have inherent limitations. They require extensive manual effort to define process-specific rules across diverse process conditions and hardware configurations. In addition, high-cost of metrology equipment typically limits wafer sampling to approximately 1% of mass production wafers, making real-time quality assurance difficult.

To address these challenges, Wonik IPS has developed an intelligent monitoring framework that combines AI-based equipment health diagnosis with wafer quality prediction. The core of the framework is a Condition-Based Maintenance (CBM) model. In deposition equipment, complex physical interactions among sensors often make direct analysis of raw sensor data unreliable. Wonik’s CBM model overcomes this limitation by integrating process log data with a sensor relationship network that captures process role and physical location of each sensor, thereby enabling equipment health diagnosis from the system level down to individual components. Unlike conventional Time-Based Maintenance (TBM), CBM provides a more effective approach for maintaining optimal equipment conditions by reducing downtime and enabling early fault diagnosis.

Wafer quality prediction is further enhanced by integrating CBM-based equipment health information with real-time Virtual Metrology (VM). Wonik’s VM solution adopts a hybrid architecture that combines a neural network optimized for subtle pattern recognition and a physics-based model capable of robust inference based on physical principles. Using process log data collected from Wonik’s deposition equipment, GEMINI HQ, thickness prediction results demonstrated strong performance even under data-constrained conditions, achieving average error rates of approximately 1% for the neural network model and 5% for the physics-based model.

Going forward, Wonik plans to integrate CBM and VM into a unified predictive framework to further advance autonomous process control. By combining equipment data, health diagnosis, and recipe conditions, this data-driven optimization strategy is expected to enhance real-time monitoring across all process stages and contribute to the development of an intelligent semiconductor equipment ecosystem.