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

Area of Science:

• Materials Science
• Chemical Engineering
• Artificial Intelligence

Background:

• Achieving nanoscale surface finish on silicon carbide (SiC) via chemical mechanical polishing (CMP) demands high material removal rates (MRR) and subnanometer surface roughness (Ra).
• Empirical optimization of CMP processes is challenging due to complex, coupled tribochemical effects.
• Existing methods often struggle to efficiently identify optimal processing parameters for SiC.

Purpose of the Study:

• To develop an interpretable machine learning (ML) framework for optimizing SiC CMP.
• To identify the key relationships between processing variables and CMP outcomes (MRR and Ra).
• To establish actionable process windows for achieving high MRR and low Ra simultaneously.

Main Methods:

• A designed set of 50 CMP experiments using cerium oxide (CeO2)/hydrogen peroxide (H2O2) slurry.
• Training and benchmarking of six baseline and two enhanced ML models, identifying a multilayer perceptron ensemble with polynomial feature expansion (5-MLP-PFE) as the top performer.
• Utilizing partial dependence, SHAP, and RadViz analyses to interpret model predictions and understand variable importance.

Main Results:

• The 5-MLP-PFE model achieved high prediction accuracy (R² ≈ 0.94 for MRR, 0.91 for Ra).
• Mechanical parameters (pressure, speeds) primarily influence MRR, while slurry chemistry (H2O2, CeO2, pH) dictates the achievable Ra limit.
• Experimental validation confirmed ML-recommended settings yielded Ra ≤ 0.13 nm with MRR up to 230.8 nm/h.

Conclusions:

• The developed ML framework provides a reproducible and interpretable method for SiC CMP optimization.
• Actionable process windows were identified, enabling inverse design of CMP settings for desired outcomes.
• The ML approach significantly shortens formulation search time and is transferable to other semiconductor polishing processes.