Evaluation of neural network attribution methods for microstructural representation of aero-engine material

Abstract

Efficiency and accuracy of evaluation for the key to microstructure parameters in aero-engine materials is crucial for understanding the properties and performance. However, it is still a tough research topic to identify the important microstructural variables using conventional methods. In the present work, attribution (variable importance evaluation) methods based on neural networks have been systematically sorted out with a comprehensive understanding of the strengths and limitations. Microstructures of nickel base single-crystal alloys is taken as an example for discuss the variable importance evaluation methods, namely forward stepwise, backward stepwise, and partial derivative. Suggestions are provided for future application in material science and solid mechanics.