Predicting Deposition Efficiency Across Diverse Cold Spray Process Parameters Using Machine Learning

Abstract Cold spray (CS) is an additive manufacturing process that is highly complex due to the many process parameters involved in the fabrication process. The efficiency of the process can be reasonably assessed and quantified through the metric of deposition efficiency (DE), denoting the ratio of the powder material successfully deposited to the total powder flowing through the nozzle. There is an industrial need to predict DE because it affects the powder usage and production cost. Machine learning (ML) has been proven to be a viable method to predict properties of additively manufactured parts as it can handle large datasets with numerous variables and is therefore well-suited to model the complex CS process. A large training dataset is needed to ensure that the ML model can be universally applied to the problem at hand. In this work, two datasets with different dimensionality and data quantities were developed, with data collected from experiments reported in the literature, and from newly obtained experimental data. These datasets were then used to train and develop ML models that can be applied to a wide range of CS spray scenarios, including a high number of variable spray parameters and large parameter ranges and high powder and substrate material flexibility. Four ML algorithms were selected for training, including K-nearest neighbors, random forest, gradient boosting, and neural network. The most accurate predictions of the DE were achieved with neural network and gradient boosting algorithms, with a root-mean-squared error under 6% DE on unseen data. An analysis of the performance using the learning curve concept revealed that the performance of most models could be further improved by collecting more training data. Shapley values and prediction maps emphasize the significant impact of gas temperature on DE, showcasing nonmonotonic changes with other CS process parameters.