Modeling high-temperature creep behavior of steel using long short-term memories and convolutional neural networks

The time-series nature of creep strain under varying temperatures makes it a dynamic problem, requiring treatment as a multivariate time series dataset where each sequence is influenced by the prior one. This research introduces an innovative approach to predicting the creep strain curve of steel under high-temperature conditions by utilizing Long Short-Term Memory (LSTM) and Convolutional Neural Network-LSTM (CNN-LSTM) models. The models work with and without incorporating steel features into the model and require a three-dimensional input array. The dataset comprises 101 experimental creep-strain curves compiled from previously published studies, covering Q345B, Q460GJ, Q550, Q690, and Q890 structural steels tested at temperatures ranging from 300°C to 800°C under various stress levels. The dataset was preprocessed in two ways: one included steel type, temperature, and stress level, and the other excluded these features. Bayesian optimization algorithms were employed to optimize the model structures, significantly enhancing performance. The LSTM model performed best with an R2 score of 0.96 on the test set, in contrast to the poorer performance of the CNN-LSTM models. Moreover, the models could predict the curve without including the steel features and conditions. The results demonstrate that the proposed data-driven framework can accurately predict creep strain from historical creep data, providing a promising basis for structural fire analysis and condition-monitoring applications.