Composite materials are inherently multiscale since the micro and meso-mechanical features govern the macroscale properties. Therefore accurate modelling requires representation of this multiscale behaviour. However, the length-scale disparity between micro and macroscale results in significant computational cost especially in non-linear scenarios such as damage and fracture. To overcome this challenge in a robust manner, a non-intrusive reduced order modelling method is required. Accordingly, a neural network based modelling is combined with Proper Orthogonal Decomposition (POD) to represent the progressive damage behaviour observed in composite materials. The schematic representation is shown in Figure 1. Mesoscale Representative Volume Elements (RVE) are developed to model the mesoscale behaviour of composite materials. These RVEs are built with delamination, matrix cracking and fibre failure criteria such that they represent accurate behaviour of this length-scale. Homogenised stress-strain curves are obtained from these RVEs for representative orthogonal loading conditions resulting in a mesoscale database. The neural network is developed to enable thermodynamic consistency and path dependency essential for damage modelling. This is achieved by training the neural network to output the strain energy and automatic differentiation is used to calculate the stress. Furthermore, damage variables are used in the network explicitly to minimise the use of memory limiting Long Short-Term Memory (LSTM) units. This network is trained on reduced dimensional data obtained with POD to accelerate the computations further. The results will be demonstrated using open-hole tensile test of multidirectional composite laminate clearly showing the efficiency of multiscale progressive damage modelling developed with neural networks.