D2.5 Digital Twin Training Sandbox

One of the primary objectives of WP2 is the development of robust digital twin (DT) platforms for training and monitoring AI-driven methods. These platforms employ virtual twins to emulate their physical counterparts, facilitating iterative cycles of simulation, analysis, and optimisation to enhance system performance. To ensure the reliability and generalizability of DT systems, we propose a semi-supervised learning (SSL) framework incorporating two novel training algorithms. The first method improves AI model training by leveraging synthetic labels and reducing biases through a tuned cross-prediction-powered inference scheme. This approach is compatible with state-of-the-art SSL techniques, allowing seamless integration for enhanced performance. Additionally, we introduce context-aware doubly-robust (CDR) learning, an SSL strategy that dynamically adjusts its reliance on pseudo-labelled data based on the varying fidelity levels of the DT across different operational contexts. Evaluated on downlink beamforming tasks, CDR demonstrates superior performance over existing methods, achieving a 24% reduction in loss compared to conventional context-agnostic learning in low labelled-data regimes. These contributions improve the robustness and reliability of AI models training within DT platforms, ensuring their scalability and real-world applicability.