Prediction of impurity concentrations in AlN single crystals by absorption at 230 nm using random forest regression

Abstract

This study introduces a rapid and non-destructive impurity characterization method using UV absorption spectroscopy that is calibrated against secondary ion mass spectrometry (SIMS) data. A random forest regression model was evaluated for carbon, oxygen, and silicon impurity prediction based on absorption spectra. AlN boules were grown using the seeded PVT method with tungsten crucibles, processed into wafers, and characterized. A matrix of 37 samples with varying impurity concentrations in the range 1 × 10¹⁷ to 5 × 10¹⁹ cm⁻³ was created using element-specific doping methods. SIMS and absorption spectroscopy data revealed characteristic absorption patterns for different impurities. Absorption at 230 nm, which is a crucial wavelength for UVC-LEDs, correlated well with the overall impurity concentration. The random forest model predicted impurity concentrations accurately when similar training data were available, but high prediction errors occurred for unique impurity profiles. To improve prediction accuracy, a more extensive sample series and/or more complex AI tools are required.