AUTOMATIX · AUTOmating MATerial modeling for composable and learnable behaviors

AUTOMATIX addresses challenges in constitutive material modeling by integrating machine learning with existing material knowledge in solid mechanics. Constitutive models are crucial for predicting material behavior under various loading and environmental conditions, yet traditional approaches often struggle to represent complex, non-linear, and time-dependent behaviors, limiting their accuracy across engineering applications. This project aims to bridge this gap by developing Material-Informed Neural Networks (MINNs), which combine empirical data with established mathematical structures to enhance interpretability, data efficiency, and predictive accuracy. By creating a modular, high-performance open-source library, the project will enable flexible modeling of complex material behaviors like plasticity, viscoelasticity, and damage mechanics. To improve generalizability and data efficiency compared to black-box ML models, the AUTOMATIX framework incorporates mathematical structures and partial material knowledge directly into a modern machine learning architecture. This gray-box approach allows MINNs to require less data while providing interpretable predictions aligned with known physical principles like thermodynamics. The framework is expected to enhance modeling accuracy in civil and mechanical engineering while advancing data-driven material modeling in multiphysics and multiscale systems. The project will test its approach on real-world applications of 3D printed fiber-reinforced concrete, which presents distinct challenges like layer-wise anisotropy, complex curing conditions, and various damage mechanisms, necessitating models that capture both microstructural and macroscale responses. By merging innovations from machine learning with material knowledge, the project aims to provide a reliable framework for improved material behavior prediction, offering valuable tools for researchers and engineers across multiple fields.