Predicting the hardness of diamond-like carbon coatings using machine learning and generative adversarial networks

Producing diamond-like carbon (DLC) coatings with increased hardness remains an inspiration for achieving diamond-like properties. A multivariable parametric analysis may necessitate 50+ experiments to establish optimum plasma dynamics with a combination of electric, magnetic, kinetic, and thermal energies that build a DLC coating of a specific hardness. Overall, this places a strain on resources and the climate. This research aims to predict the hardness of DLC coatings as a function of bias voltage (0 to 140V) and annealing temperature as a direct and two-stage heat treatment. In addition, this work investigates the critical features for estimating the hardness of DLC coatings. The conditional tabular generative adversarial networks (CTGANs) model is used to expand the small experimental data of DLC coatings, and a large dataset is obtained from the optimal CTGANs model. A range of 15 machine learning models are used to predict the hardness of DLC, and their efficacy is measured using the six well-known error-based performance measures. The data-driven modelling reflects that top-performing models, including SVR, XGBoost, LightGBM, CatBoost, ANNs, and FNNs, achieved exceptional predictive accuracy (~99.9%). Furthermore, the significance of each explanatory variable is indicated using the Shapley additive explanations (SHAP) model, which identified bias voltages of 40 V and 120 V and second-stage heat treatment as critical factors influencing hardness. Therefore, it is concluded that the implementation of the second stage potentially increasing the hardness and the bias levels of 40 and 120V may decrease the hardness of the DLC coating. The approach demonstrates a novel and efficient data-driven strategy for optimizing DLC coating processes, offering significant potential to accelerate material design in applications where experimental trials are resource-intensive.