Deep long short-term memory neural network for accelerated elastoplastic analysis of heterogeneous materials: An integrated data-driven surrogate approach

In this work, an integrated data-driven surrogate approach based on the finite-volume direct averaging micromechanics (FVDAM) and the long short-term memory (LSTM) neural network is explored to predict the elastoplastic response of composite materials. In particular, the FVDAM is first applied to generate the uniaxial and cyclic response of unidirectional composites with various off-axis orientations. Next, a two-layered neural network is trained to associate the applied strains to the corresponding stresses, which is subsequently evaluated using the separate, hold-out testing dataset. The LSTM-estimated stress?strain responses coincide with the FVDAM reference results for all the loading cases. The advantage of the LSTM to naturally capture the history-dependent stress?strain behavior over the fully connected neural network is presented, with the percentage prediction errors of the former approach an order of magnitude lower than the latter. Moreover, the robustness of the LSTM surrogate model is examined by analyzing the training data with white noise. The proposed framework offers a viable alternative for the determination of the history-dependent response of composites directly from data analysis without the need to understand the underlying deformation mechanism in the techniques of homogenization, as well as provides a foundation for efficient multiscale analysis of composite materials and structures.