Predictive modeling for the mean diameter of carbon nanotubes produced by methane decomposition

The catalytic decomposition of methane represents an efficient route for carbon nanotube (CNT) synthesis, yet the quantitative prediction of CNT mean diameter under multivariate processing conditions remains challenging. This study aims to develop and compare a suite of data‑driven models capable of capturing the nonlinear correlations among compositional and operational parameters governing CNT growth. A dataset of 66 experimental entries from peer‑reviewed sources was compiled, including 16 catalysts and reaction‑related variables numerically encoded and normalized before modeling. Neural networks (ANN and CNN), kernel, linear, and nine ensemble regression algorithms were optimized via systematic hyperparameter tuning and evaluated using R², MSE, and MRD% metrics across training, testing, and validation stages. Outlier analysis confirmed the statistical integrity and balanced diversity of the dataset, supporting robust learning and generalization. Among the models, Gradient Boosting and Random Forest achieved nearly perfect predictive fidelity (R² ≈ 0.999, MSE < 0.35, MRD ≈ 2%), followed by CatBoost, Gaussian Process, and CNN with high yet slightly variable accuracy. Linear regressors yielded moderate fits with limited nonlinear adaptability, while instance‑based (KNN) and structural (DT) models exhibited partial overfitting. SHAP interpretability of CatBoost output revealed Mo, CeO₂, and reduction parameters as the most influential variables controlling nanotube diameter, linking model predictions to catalytic chemistry. The integrated modeling strategy provides a transparent and highly accurate pathway for predictive CNT morphological control, establishing ensemble‑based algorithms as effective tools for catalytic nanomaterial optimization and design.