Machine learning assisted CALPHAD framework for thermodynamic analysis of CVD SiOxNy thin films

Abstract

A machine learning assisted CALPHAD framework is applied in this study to thermodynamically analyze the chemical vapor deposition (CVD) process for silicon oxynitride films. Among the various machine learning algorithms evaluated, Random Forest (RF) was identified as the optimal model due to its superior accuracy and generalization performance. The study identified that only 5 % data of the original dataset is required to effectively train the RF model. The best-trained RF model can excellently reproduce results from CALPHAD. SHAP analysis was performed to quantify the contribution of input features to the performance of machine learning model. The results revealed that NH₃/N₂O and NH₃/SiCl₄ ratios have the most significant influence on the mole fractions of SiO₂ and Si₂N₂O, while the NH₃/N₂O ratio is the dominant factor affecting the mole fraction of Si₃N₄ in the deposit. Notably, the ML-assisted CALPHAD framework demonstrated a 20-fold increase in analysis efficiency compared to traditional CALPHAD calculations.