Experimental and Neural Networks Analysis on Elevated-Temperature Mechanical Properties of Structural Steels

Zhenli Wu,Zhaoqian Liu,Lingzhi Li,Zhoudao Lu

DOI: https://doi.org/10.1016/j.mtcomm.2022.104092

IF: 3.8

2022-01-01

Materials Today Communications

Abstract:This study presents the possibility of using artificial neural network (ANN) for describing and predicting the high temperature mechanical behavior of structural steel. For this reason, a series of steady-state tensile tests were performed on two representative structural steels Q345B and Q460GJ in temperature range of 20 degrees C to 800 degrees C, in the forms of steel bars with diameters of 8, 10 and 12 mm. The high-temperature mechanical properties were evaluated and compared, revealing the necessity of a general and reliable prediction model on the elevated-temperature mechanical properties of different structural steels. Furthermore, the experimental results could be well predicted by using the Ramberg-Osgood model only in a limited temperature range. Therefore, the application of back-propagation neuron network (BPNN) was proposed to predict the yield stress and ultimate stress. In order to model flow property, a long short-term memory recurrent neural network (LSTM-RNN) was first adopted in strength of mechanics. Two preprocessing methodologies including one-hot encoding and polynomial feature function were used in the models. The satisfactory agreements indicate that the trained BPNN and LSTM-RNN models are efficient and accurate in predicting the mechanical properties of structural steels.