Understanding the thermo-mechanical behavior of thermoplastic vulcanizates (TPVs) is crucial to improve their design and development workflow for industrial applications. TPVs ---dynamically vulcanized blends of fully cured EPDM rubber particles within a polypropylene matrix--- are aimed to replace traditional elastomers due to their processing ease, recyclability, and low density, but they face challenges in the competitive transport sector due to the time and resource-intensive classic testing methods and simulations. Machine learning (ML) is increasingly being integrated into various scientific and engineering fields, showing its greatest utility where conventional methods are inefficient. We expect that ML, particularly neural networks (NNs), can circumvent the limitations of physics-based thermo-mechanical modeling of TPVs. In this work, we assessed the capacity of physics-constrained NN-based constitutive models in the literature [1-3] to fit hyperelastic data, laying the groundwork to model more complex behaviors. To train the models, we conducted a tensile test campaign for a TPV-based test specimen, obtaining the necessary uniaxial tension, pure shear and equibiaxial tension strain-stress data. Then, we benchmarked the models in the style of Tac et al. [4], with the future aim of integrating them into a FEM package.