Recent literature reveals an escalating trend of incorporating data-driven methodologies in the modeling of viscoelastoplastic materials, especially of deep neural networks (DNNs). These networks are broadly classified into three categories: black box NNs, NNs that incorporate physics in a weak manner, and NNs that enforce physics strongly. In our study, we employ the Neural Ordinary Differential Equations (NODEs), a specialized variant of DNNs. In 2018, Chen et al. introduced NODEs, which utilizes a general multi-layer perceptron (MLP) as the driving component on the right-hand side of a system of ordinary differential equations (ODEs) [1]. This method notably integrates the time-step scaling of dynamics, an element not present in recurrent neural networks (RNNs). NODEs were integrated into the flow of internal states, following the Coleman-Gurtin internal state variable theory [2]. In addition, the development of automatically convex data-driven creep potential functions, utilizing NODEs and focusing strongly on the enforcement of physics, was successfully carried out [3]. It has been demonstrated that the creep potential proposed by Reese and Gonvindjee [4], represents a specific instance within this methodology. Nevertheless, it remains to be established whether it effectively captures the deformation-enhanced shear thinning of the creep potential, a key aspect for accurately describing the Payne effect observed in filled rubber [5]. Our objective is to assess the capability of NODEs in characterizing the behavior of filled rubber under varying frequencies and strain amplitudes. NODEs demonstrate superior performance over phenomenological models, particularly in capturing the Payne Effect. It is evident that NODEs represent a broader family of viscous potentials than those discussed in existing literature. They intrinsically incorporate a viscosity that varies with the invariants of stress, and hence, changes with the rate of deformation.