Transfer-learning-based strategy for enhancing prediction accuracy and computational efficiency of nonlinear mechanical properties in composite materials.

In modern composite design, superior macroscopic mechanical properties can be achieved by optimizing the microstructures of materials. However, the direct prediction of the microstructure–property relationship under nonlinear conditions through numerical methods can be inefficient or imprecise. In this study, a transfer learning strategy based on the reduced order model (ROM) was proposed, offering an enhanced and rapid prediction of the mechanical responses of composite materials under nonlinear conditions. Initially, an extensive training dataset was generated by applying the ROM, used to pre-train the neural network model. Following this, a few data calculated from full-order finite element models were leveraged to fine-tune the network parameters. This approach exploits the efficient data generation capability of the ROM, with its potential computational inaccuracies in nonlinear scenarios mitigated, leading to an improvement in the accuracy and efficiency of the surrogate model. Numerical examples of a nonlinear hyper-elastic material with inclusions were examined, revealing that the computational cost in the offline stage of the transfer learning method is only half that of traditional neural network models, and it enable near real-time predictions in the online stage. Notably, it was shown that the accuracy loss of the developed surrogate model in scenarios of strong non-linearity is significantly less than that of the ROM. This method presents an innovative pathway for the swift and accurate evaluation of the effective mechanical properties of composite structures, with the potential to offer valuable insights for related methodological research. [Display omitted] [ABSTRACT FROM AUTHOR]