Integration of Peridynamics and Deep Learning for Efficient and Accurate Thermomechanical Modeling

Accurate and efficient modeling of thermomechanical failure in critical structures under extreme conditions remains a great challenge. Traditional local methods struggle with discontinuities, such as fractures, while peridynamics (PD) is computationally intensive. This study presents a rapid prediction framework that combines sequential PD thermomechanical coupling simulations with a deep learning (DL) surrogate model. The framework adopts bond-based PD to solve the deformation field, accounting for thermal expansion, whereas the temperature field is handled using the peridynamic differential operator to address boundary effects and enhance transient accuracy. The validation results show that the PD thermomechanical coupling model achieved high accuracy. For example, the cooling simulation results of a 2D plate using PD and FEM show that the results had a global error in temperature and displacement of less than 0.7%. In the Al2O3 ceramic quenching simulations, the crack propagation path is accurately reproduced using the PD model, which matches the experimental data well. To improve the computational efficiency, the DL surrogate model was trained on a large dataset generated by PD simulations. The inputs include the crack geometries and loads, and the outputs are the predicted crack openings, average radial displacement/strain, and circumference change rates. The optimized deep neural network (DNN) consisted of two hidden layers, each with nine neurons. The DNN model predicted complex multi-output responses in approximately 0.5 s, about 1200 times faster than direct PD simulation, maintaining high accuracy. The PD-DL framework offers a new direction for assessing the thermomechanical damage and structural integrity in engineering applications.