Machine learning (ML)-driven reduced-order modelling is applied to accelerate steady-state convergence in three-dimensional, nonlinear, flux-driven two-fluid simulations of boundary plasma turbulence. A parametric scan of plasma resistivity, heating and density sources is performed to generate comprehensive datasets across various turbulent regimes for model training and validation. To efficiently manage and interpret these datasets, we apply the proper orthogonal decomposition technique to reduce the dimensionality of key plasma quantities such as plasma density, temperature, electric potential and vorticity. Data-driven models are trained to map physical parameters to low-dimensional representation, enabling the rapid generation of quasi-steady-state plasma profiles. The results demonstrate that density, temperature and electric potential are qualitatively well captured with a relatively low number of bases, whereas vorticity requires a larger number of bases due to its fine spatial structures. A comparison between ML-generated restarts and simulations from scratch demonstrates a significant computational advantage of the ML approach, reducing simulation time by up to a factor of three. This hybrid framework, combining data-driven reduced-order modelling with first-principles simulations, highlights the potential of ML to accelerate plasma turbulence modelling, making high-fidelity simulations more computationally feasible for large-scale fusion devices, such as ITER and DEMO.