Predicting creep‐fatigue and thermomechanical fatigue life of Ni‐base superalloys using a probabilistic physics‐guided neural network

Predicting the life of thermomechanical fatigue (TMF) is challenging because there are several parameters describing the mechanical and thermal cycles including dwell periods within the cycle that can impact life. The relationships between these TMF history parameters and fatigue life are not always clear. This paper explores the use of a neural network (NN) with a probabilistic physics‐guided architecture to learn these relationships and predict the cycles to failure for a wide range of possible creep‐fatigue and thermomechanical fatigue histories. Using inputs of strain range, maximum and minimum temperature, the phasing between the thermal and mechanical cycles, cycling frequency, and dwell time in either tension or compression, the model predicts both the mean fatigue life and its confidence intervals (CI). The model is demonstrated from a comprehensive set of data on single‐crystal and directionally solidified Ni‐base superalloys extracted from the literature. The model is evaluated in several ways to determine the success of learning the relationship between the applied TMF histories and cycles to failure. Highlights A TMF dataset on Ni‐base superalloys is mined from open literature. A probabilistic physics‐guided neural network (PPgNN) is developed. Physics‐guided constraints are applied to improve the training process and predictions. Both mean fatigue life and the confidence intervals are predicted.