Machine Learning Empowers Silicon Carbide Crystal Growth: From Data-Driven to Process Optimization

This article explores the application of machine learning technology in the Physical Vapor Transport (PVT) process of Silicon Carbide (SiC). By comparing models such as gradient boosting, neural networks, and K-nearest neighbors, it optimizes crystal growth rate prediction and process parameters. The study found that the gradient boosting model has the highest prediction accuracy, and key parameters like temperature gradient and gas phase composition have a significant impact on growth, providing a data-driven optimization framework for semiconductor material preparation.

As a third-generation semiconductor material, Silicon Carbide (SiC) is widely used in power electronics and other fields due to its wide bandgap and high thermal conductivity. However, the preparation of high-quality single crystals is an industrial bottleneck. The mainstream PVT process involves multi-physical field coupling with complex parameters. Traditional empirical trial-and-error and physical simulation are difficult to grasp nonlinear relationships, restricting development.

Data-driven methods bring a new paradigm to materials science. The study constructs a dataset covering parameters such as temperature, pressure, and gas phase composition, and compares three types of models: gradient boosting (ensemble learning with regularization to prevent overfitting), neural networks (nonlinear fitting with optimized architecture), and K-nearest neighbors (instance-based, benchmark reference). Data preprocessing uses stratified sampling and normalization, and training uses cross-validation and grid search to optimize hyperparameters.

Machine learning models outperform traditional linear regression, with the gradient boosting model having the highest accuracy. Feature importance analysis shows that temperature and its gradient are the primary influencing factors; gas phase transport, seed crystal characteristics, and pressure control also significantly affectaffect growth rate, providing a clear direction for process optimization.

The predictive framework can accelerate process development (reduce the number of experiments), lower R&D costs, improve crystal quality, and support intelligent manufacturing. In the future, with data accumulation and deep learning development, it is expected to play a greater role in real-time monitoring and defect prediction.

The current model's generalization ability is limited by specific equipment conditions, and the 'black box' nature nature leads to limited physicalY physical interpretability. Subsequentsequent recommendations: expand data scope, introduceroduce hybrid modeling with physical constraints, develop online learning to adapt to process drift, and explore the application of reinforcement learning in autonomous optimization.