Smart Predictions: Machine Learning in Constructing Sm-Fe-V Phase Diagram

The equilibrium Sm-Fe-V phase diagram is re-investigated by using machine learning algorithms to predict unknown regions and to select the next experiment from the most uncertain region in the predicted data. The least confidence score is applied to the random forest, k-nearest neighbor (KNN), and Neural Network-Multi-Layer Perceptron (NN-MLP) to identify the uncertain region. The uncertainty score of the NN-MLP is high at the potential phase boundaries, while in the random forest, the high uncertainty is in various regions of the phase diagram. Therefore, the first four compositions were selected from the boundary region to synthesize experimentally. The obtained results were incorporated into the dataset, and upon completing the initial iteration, it was revealed that the 1:12 and liquid phase equilibrium can exist in the V-lean region with a composition of Sm 12 Fe 77 V 11 .