Machine learning-based prediction for time series damage evolution of Ni-based superalloy microstructures

Abstract

The Directionally Solidified (DS) superalloy DZ125 developed time-series related microstructure damage with serve time. The conventional grey prediction model (GM (1,1)), the fractional order accumulation grey prediction model (FAGM (1,1)), and the long short-term memory (LSTM) neural network prediction model were used to forecast the values of the time-related damage variables. The damage variables of the superalloy microstructure evolution were obtained through experimental observations of the evolution characteristics of the γ' precipitates and γ matrix. The microstructure damage variables were predicted by three machine learning models, and the fatigue damage evolution mechanism of the superalloy was analyzed. It was found that the root mean square error of the three prediction models was 0.072, 0.001, and 0.008, respectively. Finally, a model to predict fatigue life was established based on the Chaboche fatigue damage theory, and the fatigue life was obtained using the damage variables predicted by machine learning. The results revealed that the fractional order cumulative grey model and the long short-term memory neural network prediction model were both more accurate, with the conventional grey model being better for exponentially small sample time series data. • Damage variable was predicted with time-series small sample using grey model. • The prediction accuracy of the GM (1,1), FAGM (1,1), and LSTM were compared. • Fatigue life prediction model of DZ125 superalloy containing damage was developed.