Physics-augmented neural networks (PANNs) have gained increasing attention for the modeling of hyperelastic materials. Incorporating fundamental physical constraints into the network architecture combines flexibility with numerical robustness and thermodynamic consistency. However, existing approaches typically focus either on compressible materials [1] or on fully incompressible materials [2], employing different network architectures to represent their respective volumetric behaviors. PANNs for compressible hyperelastic materials are most commonly formulated using invariant-based representations of the right Cauchy--Green deformation tensor, such as the classical invariants I1, I2 and I3, while incompressible formulations often rely on the isochoric invariants Ī1 and Ī2. Due to the incompressibility constraint (J=1), mixed finite element formulations, such as u-p elements [3] are required. As an alternative, a volumetric penalty term may be added to the strain energy density function of incompressible PANNs, thereby relaxing the incompressibility constraint. In this work, we propose a new PANN architecture that consistently employs a volumetric-isochoric split and models both contributions using separate neural networks. The resulting framework provides a unified description for a wide range of material behaviors, covering both compressible and (nearly) incompressible materials. The data-driven representation of the volumetric part of the strain energy density function offers increased flexibility compared to commonly used analytical volumetric terms, such as quadratic or logarithmic formulations. Within this approach, the initial bulk modulus is learned implicitly as part of the volumetric PANN. The formulation naturally recovers the incompressible limit, while remaining applicable to fully compressible materials. In addition, explicit volumetric-isochoric separation aligns well with phase-field and fracture modeling approaches, where volumetric effects play a critical role. An additional advantage is that, for sufficiently large bulk moduli, indicating nearly incompressible behavior, the isochoric network can again be employed independently within a u-p formulation [3]. References: [1]: Linden L. et al.: https://doi.org/10.1016/j.jmps.2023.105363 [2]: Dammaß F. et al.: https://doi.org/10.1016/j.cma.2025.117937 [3]: Boffi D. et al.: https://doi.org/10.1007/978-3-642-36519-5