Assessment of Physics-Informed Neural Networks for the mechanical char- acterization of viscoplastic materials

Resumo

Recent works have introduced the concept of Physics-Informed Neural Networks (PINN), which are
trained to solve supervised learning tasks while fulfilling laws of physics described by nonlinear partial differential
equations. In this work, this approach is compared to established constitutive models that address viscoplastic
material behaviour. The materials analysed are the Poly-lactic acid (PLA) and the Acrylonitrile Butadiene Styrene
(ABS). Two constitutive models are studied: Mulliken-Boyce, and an enhanced Zapas’ model for viscoplasticity.
The finite element method is employed for the discretization of solids under frictional contact and large strains.
The output of the numerical model is coupled to a PINN, which is trained to replicate the experimental data and
to be compared to the results provided by the constitutive models. The objective of this assessment is to verify
the benefits and liabilities of each method. The focus is given to the accurate representation of stress-strain and/or
force-displacement relationships. Other features such as computational time and memory bandwidth are also
assessed.