Rank Reduction Autoencoders: A Novel Framework for Generative Design and Solutions in Composites
Please login to view abstract download link
The use of composite materials has become indispensable in a wide range of industrial applications due to their potential to provide customized mechanical and physical properties to meet specific industrial needs. This flexibility stems from their ability to combine different materials in precise arrangements and distributions, using specific manufacturing processes. However, to remain competitive in today's fast- evolving market, the composites industry must seek innovative and optimized solutions in the design of composite materials. Despite significant advances, significant challenges remain, particularly in the accurate and efficient simulation and optimization of composite materials. These challenges are largely attributed to the multi-scale effects intrinsic to these materials, where macroscopic performance is strongly influenced by microstructural factors. To address these challenges, this study introduces an innovative methodology using Rank Reduction Autoencoders (RRAEs), for rapid predictions in both forward and reverse design processes, thereby improving the efficiency of composite material optimization. The approach consists of creating two RRAE models: one for generating microstructures (Geometry) and the other for predicting the corresponding mechanical response (Solutions). These models are linked in latent space through regression techniques, facilitating quick transitions between Geometry and Solutions. This framework enables us not only to rapidly predict composite behavior as microstructures evolve, but also to efficiently generate microstructures tailored to specific properties and performance criteria. Furthermore, the methodology allows us to discover and explore new designs and geometries for composite microstructures by easily performing linear interpolation in latent space.
