Feature Importance and Growth Rate Prediction in SiC PVT Processes through Advanced Machine Learning Models
Amir Reza Ansari Dezfoli
Abstract
Amir Reza Ansari Dezfoli
Abstract
Silicon carbide is a key wide-bandgap semiconductor material for next-generation power electronics, yet the Physical Vapor Transport (PVT) method used for bulk crystal growth remains constrained by complex thermal-chemical interactions and low growth rates. This study develops a data-driven framework to overcome these limitations by integrating a curated dataset of available experimental observations extracted from literature with machine learning (ML) models. The dataset includes seed temperature and equilibrium partial pressures of , and , with growth rate as the target. Feature scaling, normalization, and stratified sampling were applied to prepare training and test sets. Three modeling approaches—regularized linear regression, artificial neural networks, and k-nearest neighbors-were compared, with Gradient Boosting incorporated as a benchmark for advanced non-linear prediction. Results show that Gradient Boosting achieved the highest predictive accuracy ( ), outperforming and ). Feature selection revealed that gas-phase species, particularly exert slightly greater influence than temperature, emphasizing the coupled role of vapor chemistry and thermal fields. The findings demonstrate that machine learning provides a powerful alternative to traditional physics-based models, enabling more reliable prediction of SiC PVT growth rates and offering guidance for optimizing process design and control.
A significance statement is not available in the OpenAlex record.
A contribution statement is not available in the OpenAlex record.
Method details are not available in the OpenAlex metadata.
Findings are not separately available in the OpenAlex metadata.
Limitations are not available in the OpenAlex metadata.
Application details are not available in the OpenAlex metadata.
Silicon carbide is a key wide-bandgap semiconductor material for next-generation power electronics, yet the Physical Vapor Transport (PVT) method used for bulk crystal growth remains constrained by complex thermal-chemical interactions and low growth rates. This study develops a data-driven framework to overcome these limitations by integrating a curated dataset of available experimental observations extracted from literature with machine learning (ML) models. The dataset includes seed temperature and equilibrium partial pressures of , and , with growth rate as the target. Feature scaling, normalization, and stratified sampling were applied to prepare training and test sets. Three modeling approaches—regularized linear regression, artificial neural networks, and k-nearest neighbors-were compared, with Gradient Boosting incorporated as a benchmark for advanced non-linear prediction. Results show that Gradient Boosting achieved the highest predictive accuracy ( ), outperforming and ). Feature selection revealed that gas-phase species, particularly exert slightly greater influence than temperature, emphasizing the coupled role of vapor chemistry and thermal fields. The findings demonstrate that machine learning provides a powerful alternative to traditional physics-based models, enabling more reliable prediction of SiC PVT growth rates and offering guidance for optimizing process design and control.
Key concepts: Machine learning, Artificial intelligence, Gradient boosting, Feature selection, Boosting (machine learning), Computer science, Benchmark (surveying), Silicon carbide