How to Bridge the Sim-to-Real Gap in Digital Twin-Aided Telecommunication Networks

TL;DR

The paper tackles the data scarcity challenge in telecom AI by leveraging digital twins to generate site-specific data and bridge the sim-to-real gap. It proposes two core avenues: calibrating digital twins with real measurements (including differentiable simulation and phase-error-aware calibration) and robust training approaches that explicitly account for residual gaps (Bayesian environment modeling and prediction-powered inference). Three PPI-based training variants—Semi-Supervised PPI, Cross-PPI, and Context-Aware PPI—offer data-efficient strategies to mitigate simulator bias using limited real data. The work highlights practical benefits for network automation while acknowledging challenges like real-time fidelity, cross-site transfer, and continual adaptation.

Abstract

Training effective artificial intelligence models for telecommunications is challenging due to the scarcity of deployment-specific data. Real data collection is expensive, and available datasets often fail to capture the unique operational conditions and contextual variability of the network environment. Digital twinning provides a potential solution to this problem, as simulators tailored to the current network deployment can generate site-specific data to augment the available training datasets. However, there is a need to develop solutions to bridge the inherent simulation-to-reality (sim-to-real) gap between synthetic and real-world data. This paper reviews recent advances on two complementary strategies: 1) the calibration of digital twins (DTs) through real-world measurements, and 2) the use of sim-to-real gap-aware training strategies to robustly handle residual discrepancies between digital twin-generated and real data. For the latter, we evaluate two conceptually distinct methods that model the sim-to-real gap either at the level of the environment via Bayesian learning or at the level of the training loss via prediction-powered inference.

Figures (6)

• Figure 1: (left) Illustration of the sim-to-real gap between (top) a real-world urban wireless deployment and (bottom) a ray tracing-based digital twin (DT). The DT model may yield incorrect channel impulse responses (CIRs) due to a mismatch between the real geometry and material parameters and the ones assumed by the simulator. (right) When using the DT-predicted CIRs to train a beamforming model, the sim-to-real gap may cause the selection of a sub-optimal beam.
• Figure 2: As detailed in this paper, the sim-to-real gap can be partially bridged via three main complementary strategies: (a) Digital twin calibration (Sec. \ref{['sec:calibrate_dt']}), which uses real-world measurements to directly improve the adherence of the DT to the real world; (b)-(c) Robust AI training, which accounts for the residual sim-to-real gap by modeling the uncertainty on the true environment via Bayesian inference (panel (b), Sec. \ref{['sec:bayesian_dt']}), or by correcting the bias caused by the sim-to-real gap on the training objective (panel (c), Sec. \ref{['sec:ppi_dt']}). (Dashed lines indicate DT-generated synthetic data, while solid lines represent real-world measurements.)
• Figure 3: Relative power prediction errors of ray tracing-based DTs calibrated using phase error-oblivious (orange), uniform phase error (green), and phase error-aware (blue) calibration methods, where calibration is conducted using channel measurements at different bandwidths (figure adapted from ruah2024calibrating).
• Figure 4: Training a multiple access protocol using synthetic data from a digital twin (DT) modeling the access channel: (left) a Bayesian formulation of DT learning assigns different configurations (interference levels) of the DT to a probability distribution that depends on prior knowledge and available data. In contrast, a frequentist approach would only select the highest-probability model. (center) While the frequentist DT trains the policy inside the single selected model, the Bayesian DT generates data from multiple models, weighting the importance of multiple plausible DT models during policy optimization. (right) When calibration data is in limited supply, explicitly modeling the sim-to-real gap via uncertainty-aware Bayesian methods can significantly improve the performance of models trained using synthetic data from the DT (adapted from ruah2023bayesian).
• Figure 5: (a) Illustration of Cross-PPI sifaou2024semi: The real data is divided into $K$ folds, and $K$ calibration procedures are conducted, each using all real data except one fold, held out for bias estimation. The calibration bias is estimated and removed from the synthetic data loss (as shown in Fig. \ref{['subfig:method_ppi']}), obtaining the Cross-PPI loss. (b) Channel capacity as a function of the proportion of real data for Cross-PPI, PPI angelopoulos2023prediction, empirical risk minimization (ERM), and pseudo-empirical risk minimization (P-ERM). As detailed in sifaou2024semi, given real data combined with synthetic samples, a classification model is trained to map UE-location to the optimal downlink beam index. The performance is measured in terms of achievable downlink channel capacity.