Enhancing Mechanical Metamodels With a Generative Model-Based Augmented Training Dataset.

Area of Science:

• Computational mechanics
• Biomaterials science
• Machine learning applications

Background:

• Modeling biological soft tissue mechanics is challenging due to material heterogeneity and complex microstructures.
• Machine learning (ML) offers potential for simulating heterogeneous materials but is limited by small datasets.
• Existing methods struggle to characterize and simulate the intricate spatial patterns crucial for tissue behavior.

Purpose of the Study:

• To investigate the effectiveness of ML-based generative models and procedural methods for augmenting limited input pattern datasets in tissue modeling.
• To address the challenge of limited data for training ML models in the context of biological soft tissue simulations.
• To enable more comprehensive exploration of material parameter spaces in heterogeneous tissue simulations.

Main Methods:

• Utilized a style-based generative adversarial network (GAN) with adaptive discriminator augmentation.
• Trained the GAN on a limited dataset of 1000 example patterns.
• Generated diverse patterns resembling real microstructures for use in finite element simulations.

Main Results:

• The style-based GAN successfully generated authentic patterns from a small dataset of 1000 examples.
• Generated patterns, when resembling real ones, effectively augmented finite element simulation training datasets.
• Developed and released an open-access finite element analysis simulation dataset based on Cahn-Hilliard patterns.

Conclusions:

• ML-based generative models, particularly style-based GANs, can effectively augment limited datasets for simulating heterogeneous biological tissues.
• The generated patterns provide a viable method to expand training data, improving the accuracy and scope of tissue mechanics simulations.
• The open-access dataset facilitates future research in computational biomechanics and ML-driven material modeling.