In the context of Finite Volume schemes, for tasks that require exact conservation and involve multiple materials, the Volume Of Fluid method (VOF) is a popular technique used to advect sharp interfaces. Given all the volume fractions in all the cells of a computational mesh, the goal of the VOF method is to predict the flux across the cell interfaces for each material. This flux is locally computed using the volume fractions in a stencil, i.e. a small set of contiguous cells, and is used in a standard advection scheme. There exist numerous VOF schemes, some of the most famous are PLIC, YOUNGS, LVIRA, ELVIRA and GRAD. In [1], the VOF-ML method is proposed for two-dimensional simulations. It is a VOF scheme that leverages the non-linearity and representation capability of deep neural networks to compute more accurate fluxes in situations where standard schemes introduce a significant artificial diffusion. This work is extended in [2] to problems with more than 2 materials. In this talk, we present the extension of the VOF-ML method to three-dimensional problems, with particular care on the final accuracy of the method. We discuss how to efficiently construct large datasets to train the involved neural network on accurate synthetic data, how to enforce physical and geometrical constraints, and how to efficiently incorporate the flux computation into a Finite Volume scheme. We numerically compare the obtained VOF-ML scheme with other schemes available in literature and show improved performances and reduced diffusion.