Application of explainable artificial intelligence for prediction and feature analysis of carbon diffusivity in austenite

Carbon diffusivity in austenite is an important value in heat treatment, such as annealing and normalizing, for controlling the overall properties of steels. However, it is still a challenge to improve the model to more accurately predict the carbon diffusivity and analyze the features. Here, machine learning methods were employed to precisely predict carbon diffusivity and provide specific insights into prediction mechanisms of features. Shapley additive explanation method was employed to analyze the feature mechanisms. A total of 263 utilizable datasets were collected from the literature and analyzed, and three outliers were discarded. We then searched for hyperparameters using fivefold cross-validation and a grid search. Random forest regression (RFR) was selected based on the determination coefficient. The RFR was validated with an empirical equation using training, testing, and additional datasets. The importance and mechanisms of input features such as temperature, carbon concentration, Cr, Ni, Si, Al, Mn, and Mo were discussed using the calculation results of the Shapley value. Temperature had the greatest influence on carbon diffusivity, followed by carbon concentration, Cr, Si, Mo, Ni, Al, and Mn. Temperature, carbon concentration, Mo, Ni, and Mn increased carbon diffusivity but Cr, Si, and Al decreased carbon diffusivity.