One of the main objectives of WP2 is to develop reliable digital twin (DT) platforms for training and monitoring AI-AI methods. These platforms utilize virtual twins to simulate physical twins, enabling continuous cycles of simulation, prediction, analysis, and optimization. To ensure the reliability of DT systems, a Bayesian framework is proposed to manage model uncertainty arising from data limitations. This framework supports ensembling-based methods for enhanced control and prediction. Moreover, a novel calibration scheme for ray tracing is introduced. This scheme employs a variational expectation maximization algorithm to correct phase errors, significantly improving prediction accuracy for tasks such as beamforming and user positioning. Additionally, a DT-aided semi-supervised learning approach is proposed. This method enhances AI model training by leveraging synthetic labels and mitigating biases through a tuned cross-prediction-powered inference scheme. These solutions enhance the management and optimization of AI models within DT platforms, ensuring their efficacy and reliability.
Simeone et al. (Fri,) studied this question.