Interface State Density Prediction between an Insulator and a Semiconductor by Gaussian Process Regression Models for...

Area of Science:

• Materials Science
• Semiconductor Physics
• Chemical Engineering

Background:

• Optimizing laboratory-scale processes requires effective use of limited data.
• Process development involves frequent modifications, creating related but distinct datasets.
• Accurate prediction of insulator-semiconductor interface quality is crucial for GaN power devices.

Purpose of the Study:

• To predict the interface state density (D_it) for insulator/Gallium Nitride (GaN) interfaces.
• To evaluate model-building strategies for D_it prediction, especially with limited and varied datasets.
• To assess the impact of ultraviolet (UV)/Ozone gas treatment on interface quality.

Main Methods:

• Retrospective evaluation of various model-building approaches for D_it prediction.
• Utilizing data from untreated interfaces to improve models for treated interfaces.
• Employing automatic relevance vector-based Gaussian process regression with optimized hyperparameters.

Main Results:

• Data from untreated interfaces significantly improved D_it prediction for UV/O3-treated GaN interfaces.
• This performance improvement was not observed for Silicon (Si) interfaces.
• Gaussian process regression demonstrated high predictive performance and reliable uncertainty estimation.

Conclusions:

• Leveraging related, albeit different, datasets enhances predictive model accuracy for GaN interfaces.
• Gaussian process regression is a suitable method for predicting interface quality and quantifying prediction uncertainty.
• Hyperparameter optimization via cross-validation is key for robust Gaussian process models in process development.