Resin Transfer Moulding is a technology that can produce high-performance composite structures with a low-cycle process. The primary issue with this process is the absence of information during the injection phase regarding the resin flow front. Typically, a grid of sensors is placed along the mold to better understand the phenomenon, but this comes with a high cost; the sensors could interfere with the composite structure, and the only information given is related to the number of sensors placed. The aim of this work is to develop a surrogate model for the RTM process that is less computationally demanding and can potentially replace the traditional finite elements model. This surrogate model could estimate a more realistic position of the resin flow front based on data collected from sensors by solving the inverse problem. This study simulates the resin injection using Abaqus to build a training dataset. We modelled a composite structure of alternating Kevlar® and carbon fibre layers, determining material properties experimentally. We derived key parameters such as density, permeability, and capillary pressure to simulate resin propagation in unsaturated conditions. The study highlights the critical influence of permeability on simulation accuracy and convergence, serving as a basis for future adjustments upon experimental validation. This work allowed us to define a surrogate model for the injection process based on Neural Networks, which is less demanding in computational terms compared with the Abaqus numerical model. This surrogate model serves as a digital twin to monitor in real time the resin flow inside the mould, leading to the concept of virtual sensors, allowing to lower the sensor cost of the process while minimising the effect in mechanical performance of the composite structure due to sensor interference.