We present a variational physics-informed deep learning framework for the simulation of brittle fracture in anisotropic media. The approach builds on the phase-field method, where crack initiation and propagation are described implicitly through an auxiliary damage variable derived from energy minimisation principles. Particular attention is given to anisotropic surface energies, including strongly anisotropic and non-convex cases, which remain challenging for conventional numerical methods. The governing problem is formulated through a generalised crack density functional incorporating higher- order gradient terms. The resulting non-convex total energy functional is solved using the Deep Ritz Method, in which neural networks approximate the displacement and phase-field variables and are trained by direct energy minimisation. To ensure stable and accurate representation of higher-order derivatives, the neural trial space is enriched using higher-order B-spline basis functions, reducing reliance on high-order automatic differentiation. The proposed framework is assessed through representative two-dimensional benchmark problems, including isotropic, weakly anisotropic, and strongly anisotropic fracture scenarios. The results demonstrate direction-dependent crack growth, preferred propagation paths and crack “kicking”, without the need for explicit crack tracking or ad-hoc criteria. The study highlights the potential of variational deep learning as a robust and flexible alternative for modelling complex fracture processes in anisotropic ma- terials.