Estimation of stress–strain constitutive model for ultra-high performance concrete after high temperature with an deep neural network based method

•The residual compressive behavior and stress–strain curve relationships of ultra-high performance concrete (UHPC) materials after high temperatures is studied.•A Deep Neural Network model is constructed to predict the complete residual stress–strain response of UHPC materials after high temperature.•Mechanism of explosive spalling of concrete under high temperature is reviewed. In order to better study and analyze the residual compressive behavior and stress–strain curve relationships of ultra-high performance concrete (UHPC) materials after high temperatures, high-temperature tests and uniaxial compression tests were carried out on 144 specimens in the temperature range of 20 °C-900 °C in this paper. The effects of different curing regimes and cooling regimes on the mechanical properties, apparent characteristic and stress–strain curves of UHPC after experiencing different temperatures were investigated respectively. As the target temperature increased, the stress–strain curves of all UHPC specimens gradually tended to be flatten, the peak point of the curve shifted to the right, the peak strain increased sharply, and the elastic modulus and proportional limits gradually decreased. The compressive strength and modulus of toughness of the UHPC specimens both increased and then decreased as the target temperature increased, with a critical temperature of 300 °C and 500 °C respectively. The Poisson's ratio of the UHPC specimens decreased and then increased as the target temperature increased, with a critical temperature of 500 °C. In addition, based on the experimental results, a Deep Neural Network (DNN) model with four hidden layers and 100 neurons in each layer was constructed to predict the complete residual stress–strain response of UHPC materials after high temperature. The training and validation of the DNN model was completed by pre-processing and segmenting the original dataset with the help of Pytorch deep learning libraries. The DNN model predicts stress–strain curves that are in excellent agreement with the UHPC test curves.