This study uses Universal Differential Equations and differentiable physics to recover the unknown terms of nonlinear viscoelastic-fluid constitutive models (UCM, Johnson-Segalman, Giesekus, ePTT) from synthetic shear and normal stress data in oscillatory and startup flows. A neural network replaces the unknown constitutive terms while known physics is retained, and training differentiates through an ODE solver. Learners study how partial physical knowledge plus a universal approximator enables data-driven constitutive discovery, where the method succeeds and fails (ePTT), and how model distillation simplifies the recovered models.
This scientific-machine-learning study reconstructs the unknown terms of viscoelastic-fluid constitutive equations using Universal Differential Equations (UDEs) and differentiable physics. A UDE keep…