Abstract: Machine learning identification of dominant descriptors in cold spraying of similar and dissimilar metals
Cold spray (CS) is a solid-state deposition and manufacturing process where high-velocity particle impact forms layer-by-layer deposits/coatings on a substrate. Here, using machine learning in conjunction with molecular dynamics, an approach for identifying the primary descriptors that underlie impact bonding in the cold spray process of similar and dissimilar metals, is developed. High-throughput molecular dynamics (MD) simulations are performed to quantify the effects of cold spray process variables, i.e., particle/substrate material combinations, particle size, and particle velocity, on deformation regime and bonding strength. The final dataset is obtained through data featurization, resulting in numerical features that include physical, mechanical, thermal, atomic, and acoustical properties of the involved materials. A three-step feature selection process is subsequently employed, involving filter-based feature selection using Pearson correlation, evaluation of machine learning regression algorithms, and wrapper-based feature selection, to identify the most important descriptors for the targeted property. The results show that the mass, density, and impact velocity of the particle, as well as the speed of sound and Young’s modulus of the substrate, are the most effective descriptors in predicting the CS penetration depth. The most dominant descriptors for impact bonding energy (IBE) are found to be the velocity, Young’s modulus, speed of sound, and mass of the particle, as well as the bulk modulus and speed of sound in the substrate. The proposed method sets the stage for machine learning based optimization of the cold spray process variables to manufacture components with tailored performance-metrics.