Predicting the service condition of K492M superalloy by high-throughput stress rupture tests and artificial neural network machine learning

Predicting the service condition of cast superalloys for industrial gas turbines and aero-engines is the principal problem to be solved. However, it has long been a formidable issue due to the lack of extensive datasets and a valid model for machine learning (ML). In this work, high-throughput stress rupture tests were conducted for K492M superalloy at temperatures ranging from 650 to 800°C under stress levels between 130 and 950 MPa over durations from 50 to 130 h. A quantitative database was constructed via the quantitative characterization of microstructural descriptors including the volume fraction and size of γ' phase, as well as the elemental segregation coefficient. A dual-output full connected neural network (FCNN) model, which consists of four hidden layers with neuron configurations of a 2-1-3-3 set, was formulated to predict the service temperature and stress. The relationship among temperature, stress, service duration and microstructure of K492M superalloy was established by ML. As a result, we successfully predicted the service temperatures and stresses with the average absolute error of 10°C and 15 MPa, respectively, deviating from the practical values. This work provides a robust framework for predicting the equivalent service conditions and has practical implications for evaluating the service life of cast equiaxed superalloys.