Designing Non-periodic 3D Woven Composite Preforms Using LSTM Deep Learning Networks

Advanced 3D woven composites show significant promise for use in high-performance applications. Nonetheless, automating the design process for customized non-periodic woven architectures presents challenges due to the complexities arising from large-scale combinatorial design issues. This often leads to inefficient utilization of the tow reinforcement. In this work, by treating the woven design problem as a multi-agent system, we leverage an integrated framework combining the Hungarian Algorithm and Long Short-Term Memory (LSTM) networks. This allows for effective alignment of woven tows along loading paths, while enabling considerations of through-thickness reinforcement.