Accelerated design of multicomponent metallic glasses using machine learning

The present study examines the role of important elemental, thermodynamic, structural and kinetic attributes in amorphous phase formation and proposes a near fool-proof design strategy for multicomponent metallic glasses (MMGs) using machine learning (ML) approach. The feature space was optimized using feature engineering and incorporating the scientific fundamentals of glass formation as the 'veto' method. The incorporation of the characteristic transformation temperatures to the feature space allowed viewing the glass formation phenomenon from previously unexplored dimensions. A multilayer perceptron neural network (MLPNN) with error backpropagation was used to classify MMGs and crystalline multicomponent alloys (CMAs). The trained model performed reasonably well based on various scoring metrics with a cross-validation accuracy of 90.35%. Further, several new MMGs were designed, synthesized and examined for their glass-forming ability (GFA). The analysis showed good agreement between the experimental results and model predictions, validating the efficacy of machine learning approach in steering the development of MMGs in future.