Leveraging Transfer Learning to Overcome Data Limitations in Czochralski Crystal Growth

TL;DR

This paper introduces a transfer learning framework that improves predictive modeling for Czochralski crystal growth across different materials, addressing data scarcity issues in machine learning applications.

Contribution

It presents novel transfer learning strategies tailored for crystal growth modeling, enabling effective adaptation of models between similar and different materials.

Findings

Transfer learning improves prediction accuracy with limited data.

Models trained on silicon data can be adapted to germanium and GaAs.

The approach is robust across varying material similarities.

Abstract

The Czochralski (Cz) method is a widely used process for growing high-quality single crystals, critical for applications in semiconductors, optics, and advanced materials. Achieving optimal growth conditions requires precise control of process and furnace design parameters. Still, data scarcity -- especially for new materials -- limits the application of machine learning (ML) in predictive modeling and optimization. This study proposes a transfer learning approach to overcome this limitation by adapting ML models trained on a higher data volume of one source material (Si) to a lower data volume of another target material (Ge and GaAs). The materials were deliberately selected to assess the robustness of the transfer learning approach in handling varying data similarity, with Cz-Ge being similar to Cz-Si, and GaAs grown via the liquid encapsulated Czochralski method (LEC), which differs from Cz-Si. We explore various transfer learning strategies, including Warm Start, Merged Training, and Hyperparameters Transfer, and evaluate multiple ML architectures across two different materials. Our results demonstrate that transfer learning significantly enhances predictive accuracy with minimal data, providing a practical framework for optimizing Cz growth parameters across diverse materials.