GALCIT Colloquium
Accurate prediction of stress distribution within the microstructure of composites is an
important step in optimizing materials design. Stress distribution drives damage initiation
and evolution, which directly affect the durability and lifespan of composite materials.
Traditional methods, such as finite element analysis (FEA), provide detailed insights into
composite behavior but are often computationally expensive, especially when dealing with
complex geometries and nonlinear material behaviors.
In this talk, I will introduce CompINet, a novel deep learning framework that accelerates
stress analysis in fiber-reinforced composites by serving as a fast surrogate for FEA. Unlike
conventional data-driven approaches that rely on large, computationally intensive datasets,
CompINet requires 20 times less training data. This improvement is achieved through the
incorporation of geometric features of composite microstructures during training. By
combining graph and convolutional neural networks, CompINet effectively captures fiber
interactions and predicts mechanical fields with high accuracy, offering a scalable solution
to complex stress analysis problems.