AI-Driven Automation in the Czochralski Process: A Step Towards Sustainable Silicon Manufacturing

Abstract

The transition to renewable energy relies heavily on high-quality monocrystalline silicon for photovoltaics, yet its production via the Czochralski (Cz) process remains resource-intensive and yield-limited. Structural defects during crystallization account for significant material waste, challenging the sustainability and cost-efficiency of silicon manufacturing. The Cz process is the only industrially relevant method for producing monocrystalline silicon, enabling the growth of large single crystals in defined orientations [1]. Despite its technological maturity, increasing productivity while reducing crystallization costs remains a major challenge. The primary yield- limiting factor is the loss of the monocrystalline structure, triggered by various still poorly understood phenomena [2]. Visual monitoring by human operators should be replaced by automated image recognition systems to detect disturbances early and enable timely countermeasures. Logging these anomalies enables subsequent process analysis. Our goal is to integrate process monitoring and analysis into a unified tool that supports continuous evaluation and optimization through machine learning. This work presents a modular and extensible software framework for the automated preprocessing, analysis, and training of machine learning models using real-world manufacturing data. The framework integrates domain expertise with data- driven approaches to support anomaly detection in complex production processes. A central focus is placed on autoencoders [3], which are used to identify deviations in process parameter trajectories, providing a promising tool for quality assurance. The developed Python modules enable flexible handling and visualization of time-series data and serve as a foundation for further research and industrial deployment. While the current implementation does not yet support fully automated anomaly detection in production environments, it offers valuable insights into the practical application of unsupervised learning in manufacturing.