Machine Learning-Driven Optimization of Silicon Carbide Chemical Mechanical Polishing with Surface Roughness Constraints

Nanoscale surface finishing of silicon carbide (SiC) via chemical mechanical polishing (CMP) requires simultaneously achieving a high material removal rate (MRR) and subnanometer surface roughness (Ra), yet empirical tuning struggles to capture coupled tribochemical effects. Here, we develop an interpretable machine-learning (ML) framework trained on a designed set of 50 polishing experiments with a CeO₂ (cerium oxide)/H₂O₂ (hydrogen peroxide) slurry to learn the multidimensional relationships between eight processing variables and CMP outcomes. We benchmark six baseline and two enhanced models and identify a five-member multilayer perceptron ensemble with polynomial feature expansion (5-MLP-PFE) as the top performer (<i>R</i>² ≈ 0.94 for MRR; 0.91 for Ra). Partial dependence, Shapley additive explanations (SHAP), and radial visualization (RadViz) analyses show that mechanical parameters (pressure, head speed, platen speed) principally control MRR, whereas slurry chemistry (H₂O₂, CeO₂, pH) sets the attainable Ra limit. The combined analyses delineate a practical high-MRR/low-Ra window─head speed 70-75 rpm, pressure 4.8-5.2 psi, H₂O₂ 4-5 vol %, CeO₂ 2-3 vol %, and pH 9.5-10.5─used to inverse-design process settings. Experimental validation of four ML-recommended formulations achieved Ra ≤ 0.13 nm with MRR up to 230.8 nm h^-1, and complementary X-ray photoelectron spectroscopy (XPS)/scanning electron microscope (SEM) characterization and finite-element modeling corroborated the predicted chemical and mechanical mechanisms. This reproducible, interpretable ML approach shortens formulation search and provides actionable process windows for SiC CMP, and the framework is transferable to other wide-bandgap semiconductor polishing processes.