AI & Technology

Machine Learning Optimised CMP

Integrating machine learning with CMP process engineering to replace expensive trial-and-error approaches with data-driven precision.

Machine Learning Framework

Our proprietary framework benchmarks 8 models across 50+ experiments to find optimal CMP parameters �balancing high material removal rate (MRR > 140 μm/h) with ultra-smooth surfaces (Ra < 0.13 nm).

Research Background

SiC is a core third-generation semiconductor material with outstanding properties for power electronics and RF devices. CMP is the only effective method to achieve nanometre-level ultra-smooth SiC wafer surfaces. Traditional methods rely on costly trial-and-error �our ML approach cuts development cycles from months to days.

ML Pipeline

01

Experimental Design

50+ experiments across 8 key CMP parameters: CeOâ‚?concentration, Hâ‚‚Oâ‚?concentration, slurry pH, polishing pressure, head RPM, platen RPM, slurry flow rate, polishing time

02

Data Preprocessing

Min-Max normalisation to eliminate dimensional effects, outlier detection via box plots, Pearson correlation analysis to guide model selection

03

Model Training & Optimisation

Comparative analysis of 8 architectures: Linear Regression, Random Forest, XGBoost, MLP, BNN, Gaussian Process, 5-MLP-PFE, Bayesian Optimisation

04

Interpretability & Validation

SHAP analysis for parameter importance, new experimental validation, surface characterisation, and FEM simulation of 8-inch SiC wafers

Model Performance Comparison

5-MLP-PFE
R² = 0.95
Bayesian Opt.
R² = 0.92
XGBoost
R² = 0.87
Random Forest
R² = 0.84
Gaussian GP
R² = 0.81
BNN
R² = 0.78
MLP
R² = 0.75
Linear Reg.
R² = 0.69

FEM Wafer Simulation

Finite element modelling of 8-inch SiC wafers validates stress distribution, equivalent deformation, and wafer warp �confirming ML predictions with physics-based simulation.

Equivalent StressEquivalent DeformationWafer WarpMechanical Strength

Key Innovation: 5-MLP-PFE

Our best-performing model combines Polynomial Feature Expansion with ensemble MLP learning.

Polynomial Feature Expansion

Adds second-order polynomial interaction terms to capture complex nonlinear coupling effects between CMP parameters �revealing relationships invisible to standard models.

5-Member MLP Ensemble

Integrates 5 independently trained MLP models through collective decision-making, improving prediction accuracy and stability while preventing overfitting on small datasets.

Apply AI to your CMP process

Let our ML framework optimise your polishing parameters �achieving results in days that would take months of manual experimentation.

Get Started