Performance of Deep Material Networks for Multiscale Material Modeling (arxiv.org)

Recent research has evaluated the performance of Deep Material Networks (DMNs), innovative machine learning models designed for multiscale material modeling that combine micromechanical principles with strong extrapolation capabilities. These models stand out for their ability to be trained solely on linear elastic data before effectively applying this knowledge to nonlinear inelastic scenarios during real-time predictions. This systematic assessment focuses on various factors influencing their performance, such as initialization, batch size, and training data size, revealing that increasing training data significantly reduces both prediction error and variance. The study also introduces the rotation-free Interaction-based Material Network (IMN) formulation, which demonstrates a remarkable 3.4x to 4.7x speed-up in offline training while retaining accurate online predictions. These insights are crucial for the AI/ML community, as they underline the balance between model complexity and efficiency in material networks, offering essential considerations for deploying DMNs in practical applications. Overall, this research enhances the understanding of how to optimize DMN architectures for better performance, paving the way for advancements in material science and engineering through machine learning.