Edinburgh Napier University

  The dramatic increase in smart services makes adaptive management of communication networks more critical. Especially for Internet of Things (IoT) networks, adaptive management faces several challenges, like fluctuating network conditions, heterogeneity in data sources, and rapid response capabilities. These challenges lead to performance degradation and data losses in IoT applications if not handled. Even though traditional AI algorithms are applied in most network topologies, they fall short of handling these adaptive management challenges without requiring additional software developments. Therefore, we propose a Generative AI-powered Digital Twinning (GenTwin) framework to create digital twin models with generative AI algorithms. In this framework, we design two novel mechanisms: Priority Pooling and Twin Adapter. Priority Pooling is to extract the dynamic relations within the topology before performing model training. We theoretically formulate the priority levels and corresponding weights with a novel presence parameter to present a modular architecture to increase training efficiency. The Twin Adapter is to interact with the GAI architecture and fine-tune the model for the adaptive twin modelling task in IoT networks. After creating the adaptive twin models, we test the rapid response capabilities of GenTwin with what-if analysis. According to our simulation results, we note that the proposed pooling mechanism extracts the data relations 19% more by enhancing the training accuracy. In addition, we show that GenTwin surpasses the traditional twin performance in terms of rapid response capabilities by reducing the response time 53% when the dynamicity is maximum.