For many industrial applications, data-driven and ML-based modelling is crucial to provide fast simulations and real-time-capable models. Machine learning based reduced-order models have proven to be a worthy replacement for high-fidelity full-order models. However, in applications such as contact problems or structures with local plastic or creep deformation, model order reduction schemes should include additional substructuring techniques. Notable use cases include wheel-rail contact, where a more accurate representation of the contact behaviour is crucial. Similarly, in microelectronics, specific components such as solder joints between printed circuit boards (PCBs) and chips can experience creep deformation due to elevated temperatures during product test cycles. In these cases, access to system operators can be challenging and calculations are often performed redundantly for all combined components. Therefore, the combination of non-intrusive Craig-Bampton and ML-based technologies such as operator inference allows us to create more accurate and efficient digital twins, bringing us closer to realising the industrial metaverse.