Hybrid ML for a Methane-Coupling Plasma Reactor

Problem

Predicting the performance of a non-oxidative methane coupling DBD reactor normally takes costly experimental trials, and only 24 experiments were available.

Approach

• Benchmarked Linear Regression, Random Forest, XGBoost and Gaussian Process Regression
• One-hot feature engineering and leave-one-out cross-validation for small-data generalization
• Feature-importance analysis and a study of extrapolation and confounding limits

Result

Gaussian Process Regression reached cross-validated R² up to 0.996 across six outputs (CH₄ conversion, H₂ and C₂–C₄ selectivities), pointing to Bayesian Optimization for experiment-guided tuning.