Determination of mechanical properties of physical vapor deposition tool coatings using machine learning

The wear resistance of physical vapor deposition (PVD) coatings is heavily influenced by their elastic and plastic properties. These properties serve as essential inputs for finite element method (FEM) simulations of the thermomechanical load experienced by the coating during the cutting process to predict tool wear, including the elastic modulus for the characterization of elastic properties and parameters of the Ludwik-Hollomon model for plastic behavior. In this study, machine learning models are developed to directly map load-depth curves from nanoindentation to elastic modulus and Ludwik-Hollomon parameters of the coating. A FEM simulation model of nanoindentation is employed to generate a dataset comprising load-depth curves from a wide range of input mechanical properties. For each definition of mechanical properties, simulations of nanoindentation at two different indentation forces are run to generate the dataset. Several machine learning models including support vector regression (SVR), multilayer perceptron (MLP), long short-term memory (LSTM) and gated recurrent unit (GRU) are then trained, validated and compared using this dataset. The inputs to these models consist of simulated load-depth curves, with the target being mechanical properties of coatings. Among these machine learning models, SVR achieves the best accuracy for predicting elastic modulus and GRU achieves the best accuracy for predicting plastic properties. Ultimately, a hybrid model combining SVR and GRU is used to predict mechanical properties of TiAlCrN coatings using experimental load-depth curves. FEM simulations using the predicted mechanical properties show good alignment with nanoindentation experiments at two different forces. The determined properties can serve as input parameters for FEM models simulating thermomechanical load during the cutting process. • Mechanical properties determination of coatings using machine learning • Comparison between several machine learning models • Precise parameters determination using SVR and GRU • Feature engineering showing useful features for prediction • Approach validated using experimental data