Prediction Model of Material Dynamic Mechanical Properties Embedded with Physical Mechanism Neural Network

High-speed tensile test is an important experimental tool for simulating material collisions, but it leads to increased research and development (R&D) costs and lead times. Using machine learning, this work establishes the mapping relationship among material composition, experimental dimensions, mechanical properties at low strains and mechanical properties at high strain rates, and by comparing four algorithms, artificial neural network (ANN) is considered as the best prediction model. Most importantly, a neural network with improved ANN structure is proposed to realize feature fusion by effectively embedding the mechanical properties under low strain into the network. The results show that the model has good prediction performance, and the prediction performance tends to be further enhanced with the increase of embedded low-strain data, and its seven-fold average R² of yield strength (YS), ultimate tensile strength (UTS) and fracture elongation (FE) at 500/s can be predicted up to 0.92, 0.94 and 0.82. Especially, the model can be generalized to other working conditions and materials, helping to reduce physical testing requirements and costs, and laying the groundwork for data and technical support to reduce the development cycle of crash simulation.