Fiber dispersion in the right ventricle: A comparison of constitutive neural network predictions with experimental data.

The mechanical behavior of right ventricular (RV) myocardium is governed by its anisotropic microstructure, yet constitutive models that account for fiber dispersion and enable reliable parameter identification remain limited. In this study, we propose a physics-embedded constitutive neural network framework for automated discovery of strain energy functions and microstructural parameters from experimental data. The model is formulated within an incompressible, orthotropic hyperelastic setting using invariant-based representations. Fiber, sheet, and normal directions are incorporated through a rotated structural basis, and dispersion effects are modeled using a generalized structure tensor approach. The framework is trained on multi-axial mechanical data from ovine RV myocardium, including uniaxial tension-compression and simple shear tests. We investigate two training scenarios: (i) full datasets containing both tensile and compressive regimes and (ii) datasets restricted to tensile loading. In both cases, the model accurately reproduces the measured stress-strain responses and identifies sparse, interpretable constitutive models which involve isotropic, anisotropic, and coupling invariants. However, the identifiability of microstructural parameters strongly depends on the available loading conditions. While tensile-only data yield higher predictive accuracy, they result in non-unique or biased estimates of fiber dispersion. In contrast, inclusion of compressive data enables consistent identification of dispersion parameters by separating fiber and matrix contributions. These results highlight the importance of multi-axial loading data for robust parameter identification and demonstrate the capability of constitutive neural network-based approaches for data-driven modeling of anisotropic soft tissues.