Optimizing and Understanding amorphous SiN Thin Film Deposition for GaN-HEMTs using Machine Learning and Causal Discovery

Machine learning–based material optimization often proceeds as a black box process, making it difficult to explain the rationale behind the optimized process conditions. This study used correlation analysis and causal discovery to determine amorphous silicon nitride (a-SiN) deposition parameters for GaN high-electron-mobility transistor (HEMT) devices, comparing their interpretability. To simultaneously optimize the breakdown voltage, hysteresis voltage difference, and sheet resistance, experiments were conducted by varying the SiH4 flow rate, applied radiofrequency (RF) power, postannealing temperature, and annealing time. A genetic algorithm (GA) was used to optimize a Gaussian process (GP) model trained on 10 datasets. The yielded optimal conditions were as follows: SiH4 flow rate of 1.45–1.65 SCCM, a RF power of 58–62 W, postannealing temperature of 700°C, and annealing time of 1–5 min. The mechanism underlying the optimal conditions was elucidated by analyzing relationships among independent variables, objective functions, and intermediate variables using correlation analysis and causal discovery with DirectLiNGAM. Both methods identified consistent intermediate factors, with causal discovery providing a quantitative hypothesis confirming the correlation analysis and improving interpretability.