Predictive multiscale modelling is revolutionizing composite design and analysis by linking manufacturing-induced microstructures to macroscopic mechanical behaviours. Although traditional FEM and FFT-based homogenization offer high fidelity, their prohibitive computational costs often preclude industrial-scale application. To bridge this gap, we present the Deep Material Network (DMN), a physics-informed AI method designed to accelerate multiscale simulations. Unlike "black-box" neural networks, DMN utilizes mechanistic building blocks that strictly satisfy the essential micromechanics and thermodynamic consistency. A key advantage of DMN is its data efficiency: by offline training on linear elastic microstructural data, DMN encodes the complex physics of Representative Volume Elements (RVEs), enabling rapid, high-fidelity predictions of nonlinear multiscale material behaviors under arbitrary loading paths. This presentation details recent advancements in DMN, including a transfer learning approach that captures the anisotropic effects arising from spatially varying fiber orientations in short fiber-reinforced polymers, an interface-enrichment network architecture that predicts fiber-matrix interface debonding in continuous carbon fiber-reinforced composites, and the integration of DMN module into Ansys LS-Dyna under an Integrated Computational Materials Engineering (ICME) framework. Validated industrial applications demonstrate that DMN achieves high accuracy comparable to high-fidelity FEA while reducing computational time by several orders of magnitude, enabling AI-empowered multiscale CAE for advanced composites.