Surface quality prediction and validation of Electro-Fenton chemical mechanical polishing for single crystal SiC based on neural network models

• Predictive models are developed to analyse the relationship between process parameters and both MRR and surface roughness Ra using three neural networks. • A new prediction and validation method of surface quality in EF-CMP effectively address the challenge of predicting the nonlinear mapping relationship between multiple process parameters and individual evaluation indicators. • Based on SHAP value analysis, this study is the first to quantitatively identify the most critical process parameters influencing SiC MRR and surface roughness Ra in the EF-CMP process. To enhance the Electro-Fenton Chemical Mechanical Polishing (EF-CMP) process for single-crystal SiC, this study employs factorial analysis to investigate the influence of process parameters on material removal rate (MRR) and surface roughness Ra. Predictive models are developed using Radial Basis Function (RBF), Back Propagation (BP), and Convolutional Neural Networks (CNN) to explore the relationships between process parameters, MRR, and surface roughness Ra. A comparative analysis of these three neural network models was conducted to identify the one with the highest prediction accuracy and overall performance. The results demonstrate that all three models can effectively predict the relationship between process parameters, MRR, and surface roughness Ra in EF-CMP. However, the BP neural network outperforms both the RBF and CNN models regarding prediction accuracy and overall performance. The BP neural network achieved a regression coefficient (R²) of 0.98824 when fitting the training data for both MRR and surface roughness Ra. During test data validation, the BP model produced a root mean square error (RMSE) of 101.19 nm/h for MRR and 0.03 nm for surface roughness Ra. Based on SHAP value analysis, most influential process parameter affecting the MRR is the abrasive size, while the most critical factor impacting surface quality is the H₂O₂ concentration in the EF-CMP process. The neural network-based approach for predicting surface quality in EF-CMP effectively addresses the challenge of forecasting the nonlinear relationship between process parameters and individual performance metrics (MRR or surface roughness Ra). This method contributes to optimising and refining the EF-CMP process for single-crystal SiC.