This work addresses a novel technique for selecting the best process parameters for the 4H–SiC epitaxial layer in a horizontal hot-wall chemical vapor reactor using a transient multi-physical (thermal-fluid-chemical) simulation model and combined with a machine-learning model. An experiment was performed to validate the feasibility of the numerical model. Secondly, a single-factor analysis was conducted to investigate the effects of process parameters, including the deposition temperature, inlet-flow volume, rotational speed of the susceptor, and cavity pressure, on the quality of the 4H–SiC epitaxial layer. Finally, a machine learning algorithm, the ant colony optimization-back propagation neural network (ACO–BPNN), was employed to develop the input/output model and optimize process parameters for obtaining a high-quality epitaxial layer and reducing the optimization cycle and costs. Notably, the optimized process was validated by real experiments, where the error between calculation and experiment is 4.03 % for deposition rate and 0.49 % for coefficient of variation, respectively. The results highlight the model as reliable and lay the foundation for the CVD growth of the 4H–SiC epitaxial layer.

@en

4H-SiC is widely used in power electronics owing to its superior physical properties. However, temperature-induced stresses compromise the reliability of 4H-SiC power devices in high-temperature applications, warranting precise, and nondestructive stress characterization responsive to temperature variations. Herein, a temperature-dependent predictive model is proposed for analyzing the Raman shift–stress in 4H-SiC. The 4H-SiC epitaxial samples prepared via chemical vapor deposition are characterized using in situ variable-temperature Raman spectroscopy, resulting in a temperature correction factor of approximately −0.021 cm−1 K−1, which is integrated into the conventional Raman shift–stress relationship to assess stress variations induced by temperature variations. The elastic modulus tensor of 4H-SiC at various temperatures determined using molecular dynamics simulations indicates a linear reduction in modulus with increasing temperature. This variable temperature modulus is incorporated into the Raman shift–stress relationship. Furthermore, a finite element method is used for model simplification to perform stress calculations in three axial directions. The experimental results confirm the consistency between calculated and experimental values with a 10% error range under the uniaxial stress condition. The study findings provide valuable insights into assessing stress evolution in 4H-SiC under temperature variations based on Raman spectroscopy, thereby advancing the application of spectroscopic techniques in material stress detection.@en