Learnable Wireless Digital Twins: Reconstructing Electromagnetic Field with Neural Representations

TL;DR

This work addresses the overhead of acquiring accurate MIMO channels by proposing a learnable digital twin that reconstructs the 3D EM field from crowd-sourced wireless samples. It builds an end-to-end framework combining a neural object ensemble (per object EM properties) and neural interaction modules (per interaction type) with geometric ray tracing on a shared 3D geometry map, enabling proactive channel prediction without pilot signals. The framework learns EM properties and interaction behaviors from data, decouples ray-independent and ray-dependent components, and demonstrates strong channel prediction accuracy and robustness to environmental changes (e.g., Office to Office-mod) with NMSEs below −26 dB for most cases. These results highlight the potential of digital twin aided communications to enable pilot-free, environment-aware wireless systems that can adapt to dynamic real-world settings.

Abstract

Fully harvesting the gain of multiple-input and multiple-output (MIMO) requires accurate channel information. However, conventional channel acquisition methods mainly rely on pilot training signals, resulting in significant training overheads (time, energy, spectrum). Digital twin-aided communications have been proposed in [1] to reduce or eliminate this overhead by approximating the real world with a digital replica. However, how to implement a digital twin-aided communication system brings new challenges. In particular, how to model the 3D environment and the associated EM properties, as well as how to update the environment dynamics in a coherent manner. To address these challenges, motivated by the latest advancements in computer vision, 3D reconstruction and neural radiance field, we propose an end-to-end deep learning framework for future generation wireless systems that can reconstruct the 3D EM field covered by a wireless access point, based on widely available crowd-sourced world-locked wireless samples between the access point and the devices. This visionary framework is grounded in classical EM theory and employs deep learning models to learn the EM properties and interaction behaviors of the objects in the environment. Simulation results demonstrate that the proposed learnable digital twin can implicitly learn the EM properties of the objects, accurately predict wireless channels, and generalize to changes in the environment, highlighting the prospect of this novel direction for future generation wireless platforms.

Figures (12)

• Figure 1: This figure captures the components of the digital replica, key challenges in digital twin construction, and the applications of digital-twin aided communications. In particular, the key challenges in digital twin construction includes: (i) how to obtain geometry information, (ii) how to obtain EM property information, and (iii) how to model EM distortion effect.
• Figure 2: This figure illustrates the $l$-th propagation path. The electric field emitted by the transmit antenna goes through two interactions and then reaches the receive antenna.
• Figure 3: This figure shows the $i$-th interaction at position $\mathbf{p}_{l, i}$ along the $l$-th propagation path. The $\mathbf{d}_{l, i}^{\mathrm{AoA}}, \mathbf{d}_{l, i}^{\mathrm{AoD}}$ are unit vectors representing the direction of the incoming and outgoing rays, respectively. $\mathcal{I}_{l, i}$ contains the interaction geometry information.
• Figure 4: This figure illustrates the process of using the neural object and neural interaction models to predict the outgoing E-field of the $i$-th interaction in the $l$-th path. The E-field can be directly predicted by the transfer function (for scattering) or can be obtained by multiplying the incoming E-field by the predicted transfer matrix (for reflection and diffraction). The dimensions of the input, intermediate, and output vectors are annotated.
• Figure 5: This figure shows the end-to-end process of using the learnable digital twin to predict the wireless channel. In particular, it first employs the neural object and neural interaction models to predict the outgoing E-fields of all paths, and then synthesize the channel with these E-fields.