[Science Advances 2026] Disaggregated machine learning via in-physics computing at radio frequency

[paper] [news]

Concept Introduction

The WISE architecture enables disaggregated model access and energy-efficient machine learning (ML) for multiple clients in wireless edge networks.

a. A central radio broadcasts frequency-encoded model weights, W, onto a radio-frequency (RF) signal at the carrier frequency $F_w$, which is precoded to V to mitigate distortion introduced during propagation over the wireless channel, H.

b. Each client equipped with WISE encodes the inference request x at the carrier frequency $F_x$, and performs local ML inference for y at the carrier frequency $F_y$. Each fully connected (FC) layer in the ML model, corresponding to a matrix-vector multiplication (MVM), is realized using a passive computing mixer.

c. Illustration of the in-physics MVM computation during frequency down-conversion with frequency-encoded W, x and y.

Code Usage

In this repository, you can find two components of code in WISE:

a) the complex-valued machine learning model training and testing in digital here.

b) the analog computing simulations with randomized inputs and weights here.

(The experimental codes are also included in this repo, but the instructions are upon request)

Reference

If you find our work useful in your research, please consider citing our paper:

@article{gao2026disaggregated,
  title = {Disaggregated deep learning via in-physics computing at radio frequency},
  author = {Gao, Zhihui and Vadlamani, Sri Krishna and Sulimany, Kfir and Englund, Dirk and Chen, Tingjun},
  journal = {Science Advances},
  year = {2026},
  publisher = {American Association for the Advancement of Science},
}

If you have any further questions, please feel free to contact us at :D

zhihui.gao@duke.edu