Data generation and deep neural network predictions for aged mechanical properties

The aim of this work is the data generation of aged mechanical properties following the Arrhenius equation and large deformation theory for a transversely isotropic bio-based polyurethane foam, and the application of this dataset
in the training process of different deep neural network architectures to evaluate their capacity to predict the full stress–strain behavior of this material after being exposed to different temperatures and long periods of time. To investigate the transversely isotropic behavior, the mechanical properties were
divided into two, longitudinal and transverse to the expansion direction, and a total of 4200 simulations with different temperature and age parameters were used in a UMAT subroutine calculating the stress by the Jaumann stress rate
for logarithmic strain levels of up to 1.2 mm/mm. The strain level obtained by the simulation, as well as the material direction, temperature, and time of exposure were used as input parameters in the DNN, with the output being stress level, yield strength, and stiffness. Tensorflow library was used to model
the DNN with two and three hidden layers of depth and width varying from 128 to 1024 neurons. We also investigated the differences between using the ReLU and ELU activation functions in this problem. Our findings highlight the impressive predictive capabilities of DNNs, with the ReLU-512-512-512 architecture demonstrating superior performance in terms of accuracy and
computational efficiency, reaching error values for the stress prediction of 5.5 kPa. Overall, this research underscores the potential of DNNs as a costeffective and efficient tool for predicting material behavior, offering valuable insights for researchers and engineers across various fields.