In this work, two ML based approaches to multi-scale modelling will be presented. The first approach utilises Deep Recurrent Convolutional Neural Networks (DRCNN) to build a surrogate for the material response in a structural scale simulation. A DRCNN has the ability to learn the material strain history dependency and consequently is capable of capturing the impact of damage initiation and progression. Additionally, this category of neural networks can learn topological features and their impact on the material performance, making them ideal for capturing the effect of the internal material architecture, including defects. In this approach, the ML model accelerates the multiscale simulation by eliminating the need for online high-fidelity calculations on the finer length scales. The second complementary ML approach accelerates the multiscale simulation by reducing the multiscale model dimensionality. Here, a data clustering approach, using 3D image registration, is employed to detect repeating material clusters in a high fidelity simulation. A tree database of material clusters and their historical response, under various loading conditions, is populated through the analysis of numerous high-fidelity simulations. In a multiscale simulation, the cluster database is queried to identify material patterns that has been seen previously and their historical responses. The pre-computed responses eliminate the need to perform the online multiscale simulations on these clusters in real-time. Examples of the application of both approaches on laminated and woven composites will be presented.