Rethinking Simulation for Composites: A Framework for the Development of Novel AI/ML-FEM Methods
The evolution of simulation methods is driven by the development of novel approaches and, crucially, by the continuous expansion of computational capacity. While high-performance computing systems have witnessed substantial growth, the majority of computational workloads within the aerospace domain still rely on established numerical methods, primarily the finite element method and finite volume method. The recent surge in AI/ML research, coupled with the vast training data enabled by HPC systems, has spurred significant interest in machine-learned constitutive and surrogate models. Although these models offer considerable value, the fundamental numerical methods utilized for physics-based simulations have not kept pace with the rapid advancements in AI/ML. Existing efforts often involve the integration of PyTorch calls into legacy simulation tools, a practice that introduces significant overhead due to the necessity of data migration between disparate computational environments. To address this disparity and establish a robust framework that facilitates the seamless integration of AI/ML capabilities across all aspects of traditional numerical methods, we embarked on a project to translate the finite element method into the JAX Python library. This translation enables seamless calls to AI/ML models, native multi-GPU support, automatic differentiation, and other valuable features. Crucially, this project prioritizes modularity and anticipates a community-driven development model. The aim is to rethink how simulations are performed for composite materials, enabling a single simulation to leverage an ecosystem of machine-learned models for many aspects such as constitutive relations, crack growth, and progressive damage. As composite structures continue to become more complex with innovations in manufacturing and material design, we anticipate that a hybrid of data-driven and traditional numerical methods will be required. This paper and accompanying presentation will detail the architecture of the developed framework, present preliminary results for fiber-reinforced composite, and outline a forward-looking vision for the future of AI-enhanced numerical methods.
IMPACT
Accomplishment: Translated the finite element method to the JAX Python library; Result: Created a unified framework for the development of future hybrid AI/ML-FE methods that may greatly accelerate material and structural design and analysis.
PRESENTER
Ballard, Michael
michael.ballard.28@us.af.mil
817-733-5533
Air Force Research Laboratory
CO-AUTHOR(S)
Calvin Anderson
Calvin.Anderson@gdit.com
James Lill
James.Lill@gdit.com
Alberto Cattaneo
alberto.cattaneo.ctr@us.af.mil
CATEGORY
AI/ML for HPC
SECONDARY CATEGORY
Comp Structural Mechanics
SYSTEM(S) USED
Raider