Predicting the mechanical properties of pristine and defective carbon nanotubes using a random forest model.

Data-driven models have lately emerged as a faster and less time-consuming method for computing material properties than computationally expensive conventional molecular dynamics and density functional theory-based simulations. Here, we developed a random forest (RF) model for comprehensively predicting mechanical properties such as stress and Poisson's ratio under varying strain and ultimate tensile strain of pristine and defective carbon nanotubes (CNTs). The variations in stress and Poisson's ratio with the strain of CNTs with a 0.4-2 nm diameter range were calculated by classical molecular dynamics simulations and characterized using parameters extracted from fitting polynomial equations. The fitting parameters and ultimate tensile strength showed distinct dependency on chiral indices, chiral angles, radii, and the presence of defects in CNTs, which constituted the target dataset. The dataset features were selected through principal component analysis, and the correlation with targets was scrutinized. We performed a comparative analysis of different machine learning algorithms for predicting mechanical properties, revealing the RF model as the best-performing algorithm. The RMSE for the stress-strain curve had a maximum value of 0.013 and 0.0143 for pristine and defective CNTs, respectively, while the correlation coefficients were ≫ 0.99 for all CNTs, showcasing the excellent predictive power of the model. The model made excellent predictions of properties for CNTs with diameters >2 nm, which is beyond the training dataset range, demonstrating the robustness of the model as a substitute for MD simulation. The insight gained from this study will benefit the research of nanocomposites, nanoelectronics, and nanomechanical systems incorporating CNTs.