Data-Driven Inversion of Semiconductor Epitaxial Layer Parameters Adaptive Signal Processing and Robust M-Estimation

Abstract

Addressing the challenges of nonlinear noise interference and dynamic parameter estimation in measuring silicon carbide epitaxial layer thickness during semiconductor manufacturing, this study establishes a deep data analysis system integrating adaptive signal processing and robust optimization algorithms. During feature extraction, a hierarchical preprocessing strategy based on signal-to-noise ratio evaluation is proposed. By integrating empirical mode decomposition, wavelet transform, and adaptive Savitzky-Golay filtering, high-fidelity denoising of multi-source spectral data is achieved. To address the complex search space resulting from coupled physical parameters, an auto-correlation-based method for estimating interference periods was designed. The Levenberg-Marquardt iterative algorithm was employed to achieve global optimization of the constrained nonlinear objective function. Furthermore, to enhance robustness against outliers, the model incorporates the Huber M-estimator for anomaly detection and weight correction, while Bootstrap resampling quantifies nonparametric confidence intervals. Experimental results demonstrate the algorithmic system's exceptional convergence and stability when processing complex spectral data, achieving a relative error of only 0.96%. This provides an efficient algorithmic model for automating and intelligentizing precision measurements in semiconductor manufacturing.