Physics-informed neural network for efficient prediction of the mechanical behaviour of short fiber-reinforced composites
Topic(s) : Material and Structural Behavior - Simulation & Testing
Co-authors :
Minjun KWAK (KOREA, REPUBLIC OF), Hyon-Woo PANG , Wonjun LEE , Woong-Ryeol* YUAbstract :
Modeling the complex viscoelastic-viscoplastic behavior of polymers is essential for accurately predicting their performance, both alone and as part of fiber composite materials. Numerous constitutive equations have been developed to simulate these viscoplastic properties using return mapping methods based on Newton-Raphson iterations. While these methods achieve satisfactory accuracy, their reliance on Newton-Raphson iterations often incurs significant computational costs and time, leading to limitations in practical applications, such as pseudograin modeling of short fiber-reinforced plastics. Each pseudograin represents a family of aligned short fibers embedded within a plastic matrix, and its mechanical behavior can be simulated using Newton-Raphson iterations for every pseudograin calculation. This substantial computational burden can be alleviated by employing a data-driven approach in viscoplasticity modeling. In this study, we propose a physics-informed artificial neural network that can replace Newton-Raphson iterations and significantly reduce computational time.
We chose a viscoplastic constitutive model featuring a 3D Drucker-Prager yield function and isotropic hardening. Artificial neural network was applied to calculate the corrected stress from the trial stress in the return mapping algorithm with the flow stress, which was also used as input to account for the hardening. The training data for the neural network model was generated numerically using the conventional return mapping scheme with precise control over the direction and density of the stresses. To optimize data collection and network training, we implemented dimensionality reduction and restoration techniques in the data space, including principal stress and symmetry transformations. During training, we enforced physics-based constraints using physics in the return mapping algorithm to prevent physically inconsistent predictions. To evaluate both computational efficiency and prediction accuracy, we integrated the trained neural network into an ABAQUS user material subroutine (UMAT) originally designed for simulating the mechanical behavior of short-fiber reinforced plastics using the pseudograin approach. Tests are conducted for a single element with arbitrary linear and non-linear strain paths, followed by part-level multi-element tests. Our findings demonstrate that the simulation results obtained through our neural network-based model closely match those predicted by the traditional viscoelastic constitutive model.
We chose a viscoplastic constitutive model featuring a 3D Drucker-Prager yield function and isotropic hardening. Artificial neural network was applied to calculate the corrected stress from the trial stress in the return mapping algorithm with the flow stress, which was also used as input to account for the hardening. The training data for the neural network model was generated numerically using the conventional return mapping scheme with precise control over the direction and density of the stresses. To optimize data collection and network training, we implemented dimensionality reduction and restoration techniques in the data space, including principal stress and symmetry transformations. During training, we enforced physics-based constraints using physics in the return mapping algorithm to prevent physically inconsistent predictions. To evaluate both computational efficiency and prediction accuracy, we integrated the trained neural network into an ABAQUS user material subroutine (UMAT) originally designed for simulating the mechanical behavior of short-fiber reinforced plastics using the pseudograin approach. Tests are conducted for a single element with arbitrary linear and non-linear strain paths, followed by part-level multi-element tests. Our findings demonstrate that the simulation results obtained through our neural network-based model closely match those predicted by the traditional viscoelastic constitutive model.