Session: 03-03-01: Applications of AI

Paper Number: 151384

151384 - Direct Prediction of Composite Material Stress-Strain Curves Using Cgan 

The talk introduces a deep learning model based on a conditional generative adversarial network (cGAN) for directly predicting stress-strain curves of composite materials. The model addresses a scenario involving 16 by 16 soft-hard binary component materials with an initial crack. A tensile test is conducted to generate the stress-strain response until the crack propagates across the sample. While the phase-field method is capable of simulating the failure process in composites, it is computationally expensive and impractical for predicting stress-strain responses for all possible combinations of binary components in the design space.

To overcome this limitation, machine learning, particularly the cGAN technique, is employed. The model uses a set of fully connected layers followed by a U-net generator to predict stress-strain curves directly, bypassing the need for extensive simulations and post-analysis. A PatchGAN discriminator is integrated to assess the realism of the generated curves, with the interplay between the generator and discriminator refining the results continuously. The dataset comprises random and bio-mimetic structures, and despite its limited size, the model achieves highly accurate predictions of stress-strain curves. It demonstrates impressive performance in acquiring mechanical properties, with an R² > 0.946 and a MAPE < 5.53% for the testing data. This work validates the feasibility of using cGAN for image-to-vector prediction and highlights its potential in solving inverse design problems.

Presenting Author: Kuan-Ying Chen Institute of Applied Mechanics

Presenting Author Biography: Kuan-Ying Chen is a Ph.D. student at the Institute of Applied Mechanics, National Taiwan University. His research interests focus on fracture mechanics and machine learning. Recently, he contributed a paper on machine learning-aided composite materials design, published in the Journal of the Mechanics and Physics of Solids (JMPS).