Flash boiling of superheated liquid jets is encountered in a variety of engineering applications, including safety-critical releases of pressure-liquefied fluids and advanced fuel injection systems. The underlying physics involves strong thermodynamic non-equilibrium, rapid phase change, and complex turbulence–interface interactions, which pose substantial challenges for predictive Computational Fluid Dynamics (CFD). High-fidelity simulations typically require sophisticated closure models and long computational times, limiting their applicability in design optimisation, uncertainty analysis, and real-time decision support. This work investigates scientific machine learning strategies for accelerating CFD simulations of flash-boiling jets, addressing key opportunities and challenges in AI for CFD. Two complementary approaches are considered. The first is an end-to-end emulation strategy, in which a machine learning model is trained to reproduce CFD solutions for flashing flows by coupling a thermodynamic non-equilibrium Homogeneous Relaxation Model (HRM) with convolutional neural networks [1]. This approach aims to deliver rapid predictions of flow and phase-change features at a fraction of the computational cost of conventional solvers. The second approach follows a hybrid solver paradigm, embedding machine learning models within a traditional Navier–Stokes framework to replace selected closure relations [2]. In particular, neural networks are employed to substitute turbulence-model equations within the pressure–velocity coupling, targeting computationally expensive and model-sensitive components while preserving the overall numerical structure and stability of the solver. The proposed methods are evaluated against unseen numerical simulations and available experimental data. Accuracy, generalisation across operating conditions, and numerical robustness are assessed, and implications for data efficiency and physical consistency are discussed. The results highlight the potential of combining end-to-end emulation and hybrid closures as a scalable pathway toward accelerated CFD for complex multiphase flows with phase change.