Enabling Wireless Design Automation through a Differentiable RF Digital Twin

Every wireless system rests on a model of how signals propagate through a physical space. Prof. Xinyu Zhang's group builds that model as a living replica. Their RF Digital Twin (RFDT) (MobiCom'26) enables fully differentiable RF ray tracing for the first time, maintaining a two-way link between a real environment and its digital copy: measured signals continuously calibrate the twin, and the calibrated twin is then optimized to design better wireless systems. Unlike conventional ray tracers or black-box neural models, RFDT stays physically grounded while remaining fully learnable, running faster than NVIDIA's Sionna and producing accurate gradients that match full-wave ground truth.

RFDT enables simulation-in-the-loop design of wireless communication and sensing systems, and can serve as a digital twin toolkit that lets AI agents automate design iteration. An open-source implementation is available at witwin.ai. Two companion systems extend the framework: RFCanvas (SenSys'24) builds a digital twin from ordinary camera imagery, using visual priors to predict a site's RF channel from sparse measurements, and RF Genesis (SenSys'23) turns existing vision and graphics datasets into synthetic RF training data, letting sensing systems generalize to users, motions, and environments they never saw during training.

Digital Twins for Site-Specific Channel Modeling

Prof. Nuria González-Prelcic's WiSeCom Lab centers its digital twin research on a site-specific, physically grounded model of the radio environment that the network builds, learns from, and continuously refines. The lab has long used high-fidelity ray tracing environments, including the widely adopted Raymobtime methodology, to design and train the physical layer of AI-based wireless systems, and develops site-specific compressive channel estimation and online codebook learning that adapt to each deployment's geometry.

Most recently, the lab's physics-informed channel estimation framework treats digital-twin-generated signal-strength maps as environmental priors, fusing electromagnetic propagation knowledge with learned refinement to estimate and forecast channels under severe pilot constraints. This work feeds a broader vision, articulated in the group's Proceedings of the IEEE ISAC tutorial, in which 6G infrastructure, devices, aerial platforms, and satellites cooperate to build and update a shared environmental model through a joint loop of sensing, communication, and learning.

Resource and Computationally Efficient RF Modeling for Digital Twins

To stay useful in dynamic settings, a wireless digital twin must rapidly reconstruct and update a 3D picture of its environment from limited measurements. Today's dominant AI approaches, adapted from Neural Radiance Fields in computer graphics, are data-hungry, computationally expensive, and hard to interpret, a poor fit for the sparse-data conditions of real deployments. Prof. Piya Pal's group closes this gap by combining classical signal processing with modern machine learning, exploiting the harmonic and low-rank structure of wireless propagation to build physically informed, sample-efficient learning algorithms.

The group's NEAR framework (ICML'25) leverages the harmonic structure of wave propagation across non-uniform antenna arrays to reconstruct 3D far-field radar scenes from remarkably few measurements, even at low signal-to-noise ratios. A follow-on framework for neural ray tracing (ICML'26) introduces a physically grounded stochastic transport model that captures uncertainty while enforcing channel reciprocity, achieving state-of-the-art reconstruction accuracy with a fraction of the training data required by NeRF-based methods. These advances together chart a path to practical, real-time digital twins for 6G networks.