Iterative Bidirectional Recurrent Network for Predicting Failure of Multilayer Composites

Abstract

This work addresses the problem of predicting the ultimate strength state of multilayer composite materials based on their structural description. The composite is represented as an ordered sequence of layers with known fiber orientations and a set of physico-mechanical parameters. To solve this problem, an iterative bidirectional recurrent architecture is proposed, in which the hidden states of each layer are updated considering neighboring layers above and below over multiple iterations. This mechanism enables repeated interlayer state alignment, analogous to stress redistribution in the composite, and outperforms classical and bidirectional LSTM models, where information is propagated only once. The resulting layer representations are aggregated and used to predict an integrated failure index. A comparison with LSTM demonstrates that the proposed architecture more accurately captures nonlinear interlayer interactions and improves prediction accuracy.