Skip to main content

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

  • Article
  • Published:

An integrated microwave neural network for broadband computation and communication

Nature Electronics volume 8, pages 738–750 (2025)Cite this article

Subjects

This article has been updated

Abstract

The development of high-bandwidth applications, including multi-gigabit communication and radar imaging, demands faster processing. However, in the microwave regime, where frequencies exceed clock rates, sampling and computation become challenging. Here we report an integrated microwave neural network for broadband computation and communication. Our microwave neural network operates across tens of gigahertz but is reprogrammed with slow megabits per second control bitstreams. By exploiting strong nonlinearity in coupled microwave oscillations, it expresses its computation in a narrower spectrum, enabling easy read-out. The system searches bit sequences in multi-gigabits per second data and emulates digital functions without custom circuits. It accelerates radio-frequency machine learning by classifying encoding schemes and detecting frequency shifts to track flight trajectories from radar. The microwave neural network is fabricated with standard complementary metal–oxide–semiconductor technology. It occupies a sub-wavelength footprint of 0.088 mm2 on chip and has a sub-200-mW power consumption, supporting integration in a general-purpose analogue processor.

Access through your institution
Buy or subscribe

This is a preview of subscription content, access via your institution

Access options

Access through your institution

Buy this article

  • Purchase on SpringerLink
  • Instant access to full article PDF
Buy now

Prices may be subject to local taxes which are calculated during checkout

Fig. 1: Solution to ultrawideband computation and communication signal processing problems.
Fig. 2: Distributed microwave oscillatory modes form an ultrabroadband neural network.
Fig. 3: Evolution of the integrated MNN’s frequency response to broadband microwave signals and parametric drives.
Fig. 4: CMOS-integrated MNN mimics an ultrahigh-speed digital computer without the need for a clock.
Fig. 5: Simplified radar signal processing and communication decoding through the integrated MNN.

Similar content being viewed by others

Data availability

The measured spectral data used for backend training in digital emulation and radar tracking tasks are available via Zenodo at https://zenodo.org/records/14188849 (ref. 51).

Code availability

Code for backend training, radar flight scenario simulation and coupled-mode-theory-based numerical simulations that support the conclusions in this article are available via Zenodo at https://doi.org/10.5281/zenodo.14188849 (ref. 51).

Change history

  • 19 August 2025

    In the version of this article initially published online, the equal contributors footnote for Bala Govind and Maxwell G. Anderson was missing and now appears in the HTML and PDF versions of the article.

References

  1. Wang, P. et al. Multi-gigabit millimeter wave wireless communications for 5G: from fixed access to cellular networks. IEEE Commun. Mag. 53, 168–178 (2015).

    Article  Google Scholar 

  2. Li, K. et al. An integrated CMOS–silicon photonics transmitter with a 112 gigabaud transmission and picojoule per bit energy efficiency. Nat. Electron. 6, 910–921 (2023).

    Article  Google Scholar 

  3. Zhou, Z. et al. Communications with guaranteed bandwidth and low latency using frequency-referenced multiplexing. Nat. Electron 6, 694–702 (2023).

    Article  Google Scholar 

  4. Li, A. et al. Evaluating modern GPU interconnect: PCIe, NVLink, NV-SLI, NVSwitch and GPUDirect. IEEE Trans. Parallel Distrib. Syst. 31, 94–110 (2020).

    Article  Google Scholar 

  5. Gomes, W. et al. Ponte Vecchio: a multi-tile 3D stacked processor for exascale computing. In 2022 IEEE International Solid-State Circuits Conference (ISSCC) 42–44 (IEEE, 2022).

  6. Esmaeilzadeh, H. et al. Dark silicon and the end of multicore scaling. In 2011 38th Annual International Symposium on Computer Architecture (ISCA) 365–376 (IEEE, 2011).

  7. Gupta, U. et al. Chasing carbon: the elusive environmental footprint of computing. In 2021 IEEE International Symposium on High-Performance Computer Architecture (HPCA) 854–867 (IEEE, 2021).

  8. Rupp, K. Microprocessor trend data. GitHub https://github.com/karlrupp/microprocessor-trend-data/tree/master/50yrs (2024).

  9. Desislavov, R. et al. Trends in AI inference energy consumption: beyond the performance-vs-parameter laws of deep learning. Sustain. Comput. Inform. Syst. 38, 100857 (2023).

    Google Scholar 

  10. Clark, K. A. et al. Synchronous subnanosecond clock and data recovery for optically switched data centres using clock phase caching. Nat. Electron. 3, 426–433 (2020).

    Article  Google Scholar 

  11. Zhang, W. et al. Edge learning using a fully integrated neuro-inspired memristor chip. Science 381, 1205–1211 (2023).

    Article  Google Scholar 

  12. Zhou, Y. et al. Computational event-driven vision sensors for in-sensor spiking neural networks. Nat. Electron. 6, 870–878 (2023).

    Article  Google Scholar 

  13. Dong, B. et al. Higher-dimensional processing using a photonic tensor core with continuous-time data. Nat. Photon. 17, 1080–1088 (2023).

    Article  Google Scholar 

  14. Cai, H. et al. Brain organoid reservoir computing for artificial intelligence. Nat. Electron. 6, 1032–1039 (2023).

    Article  Google Scholar 

  15. Sludds, A. et al. Delocalized photonic deep learning on the internet’s edge. Science 378, 270–276 (2022).

    Article  Google Scholar 

  16. Wright, L. G. et al. Deep physical neural networks trained with backpropagation. Nature 601, 549–555 (2022).

    Article  Google Scholar 

  17. Senanian, A. et al. Microwave signal processing using an analogue quantum reservoir computer. Nat. Commun. 15, 7490 (2024).

    Article  Google Scholar 

  18. Gao, X. et al. Programmable surface plasmonic neural networks for microwave detection and processing. Nat. Electron. 6, 319–328 (2023).

    Article  Google Scholar 

  19. Ross, A. et al. Multilayer spintronic neural networks with radiofrequency connections. Nat. Nanotechnol. 18, 1273–1280 (2023).

    Article  Google Scholar 

  20. Zhu, M. et al. A reconfigurable linear RF analog processor for realizing microwave artificial neural network. IEEE Trans. Microwave Theory Techn. 72, 1290–1301 (2024).

    Article  Google Scholar 

  21. Feng, H. et al. Integrated lithium niobate microwave photonic processing engine. Nature 627, 80–87 (2024).

    Article  Google Scholar 

  22. Xu, Y. et al. Bidirectional interconversion of microwave and light with thin-film lithium niobate. Nat. Commun. 12, 4453 (2021).

    Article  Google Scholar 

  23. Moy, W. et al. A 1,968-node coupled ring oscillator circuit for combinatorial optimization problem solving. Nat. Electron. 5, 310–317 (2022).

    Article  Google Scholar 

  24. Dutta, S. et al. An Ising Hamiltonian solver based on coupled stochastic phase-transition nano-oscillators. Nat. Electron. 4, 502–512 (2021).

    Article  Google Scholar 

  25. Jouppi, N. et al. In-datacenter performance analysis of a tensor processing unit. In Proc. 44th Annual International Symposium on Computer Architecture (ISCA) 1–12 (ACM, 2017).

  26. Sounas, D. L. et al. Broadband passive isolators based on coupled nonlinear resonances. Nat. Electron. 1, 113–119 (2018).

    Article  Google Scholar 

  27. Govind, B. et al. Ultra-compact quasi-true time delay for boosting wireless channel capacity. Nature 627, 88–94 (2024).

    Article  Google Scholar 

  28. Razavi, B. The cross-coupled pair—part I [a circuit for all seasons]. IEEE Solid-State Circuits Mag. 6, 7–10 (2014).

    Article  Google Scholar 

  29. Babaie, M. et al. A class-F CMOS oscillator. IEEE J. Solid-State Circuits 48, 3120–3133 (2013).

    Article  Google Scholar 

  30. Jamali, B. et al. A fully integrated 50-280-GHz frequency comb detector for coherent broadband sensing. IEEE Trans. Terahertz Sci. Technol. 9, 613–623 (2019).

    Article  Google Scholar 

  31. Hussein, H. M. E. et al. Passive frequency comb generation at radiofrequency for ranging applications. Nat. Commun. 15, 2844 (2024).

    Article  Google Scholar 

  32. Shu, H. et al. Microcomb-driven silicon photonic systems. Nature 605, 457–463 (2022).

    Article  Google Scholar 

  33. Zhang, M. et al. Broadband electro-optic frequency comb generation in a lithium niobate microring resonator. Nature 568, 373–377 (2019).

    Article  Google Scholar 

  34. Haus, H. A. et al. Coupled-mode theory. Proc. IEEE 79, 1505–1518 (1991).

    Article  Google Scholar 

  35. Pereira, U. et al. Attractor dynamics in networks with learning rules inferred from in vivo data. Neuron 99, 227–238.e4 (2018).

    Article  MathSciNet  Google Scholar 

  36. Parker, T. S. et al. Practical Numerical Algorithms for Chaotic Systems (Springer, 1989).

  37. Sarkar, T. et al. An organic artificial spiking neuron for in situ neuromorphic sensing and biointerfacing. Nat. Electron. 5, 774–783 (2022).

    Article  Google Scholar 

  38. Du, Y. & Mordatch, I. Implicit generation and modeling with energy‑based models. Adv. Neural Inf. Process. Syst. 32, 8738–8748 (2019).

    Google Scholar 

  39. Dorkenwald, S. et al. Neuronal wiring diagram of an adult brain. Nature 634, 124–138 (2024).

    Article  Google Scholar 

  40. Frans, Y. et al. A 56-Gb/s PAM4 wireline transceiver using a 32-way time-interleaved SAR ADC in 16-nm FinFET. IEEE J. Solid-State Circuits 52, 1101–1110 (2017).

    Article  Google Scholar 

  41. Agazzi, O. E. et al. A 90 nm CMOS DSP MLSD transceiver with integrated AFE for electronic dispersion compensation of multimode optical fibers at 10 Gb/s. IEEE J. Solid-State Circuits 43, 2939–2957 (2008).

    Article  Google Scholar 

  42. Al-Eryani, J. et al. Fully integrated single-chip 305–375-GHz transceiver with on-chip antennas in SiGe BiCMOS. IEEE Trans. Terahertz Sci. Technol. 8, 329–339 (2018).

    Article  Google Scholar 

  43. Lin, Y. et al. An improved neural network pruning technology for automatic modulation classification in edge devices. IEEE Trans. Veh. Technol. 69, 5703–5706 (2020).

    Article  Google Scholar 

  44. Wang, Y. et al. MsmcNet: a modular few-shot learning framework for signal modulation classification. IEEE Trans. Signal Process. 70, 3789–3801 (2022).

    Article  Google Scholar 

  45. Zhang, H. et al. A data preprocessing method for automatic modulation classification based on CNN. IEEE Commun. Lett. 25, 1206–1210 (2021).

    Article  Google Scholar 

  46. Davis III, R. et al. RF-photonic deep learning processor with Shannon-limited data movement. Sci. Adv. 11, eadt3558 (2025).

    Article  Google Scholar 

  47. He, K. et al. Deep residual learning for image recognition. In Proc. IEEE Conference on Computer Vision and Pattern Recognition 770–778 (IEEE, 2016).

  48. Pedregosa, F. et al. Scikit-learn: machine learning in Python. J. Mach. Learn. Res. 12, 2825–2830 (2011).

    MathSciNet  Google Scholar 

  49. Paszke, A. et al. PyTorch: an imperative style, high-performance deep learning library. NeurIPS 32, 8024–8035 (2019).

    Google Scholar 

  50. Loshchilov, I. et al. Decoupled weight decay regularization. In International Conference on Learning Representations 1–10 (ICLR, 2019).

  51. Govind, B., Anderson, M., Wu, F., McMahon, P. & Apsel, A. Code and data for ‘An integrated microwave neural network for broadband computation and communication’ (data set). Zenodo https://doi.org/10.5281/zenodo.14188849 (2025).

Download references

Acknowledgements

B.G. and A.A. acknowledge GlobalFoundries and the Defense Advanced Research Projects Agency (DARPA), through its Wideband Adaptive RF Protection (WARP) program, for providing the chip fabrication facility. B.G. and A.A. also thank the Cornell NanoScale Facility, a member of the National Nanotechnology Coordinated Infrastructure (NNCI), which is supported by the National Science Foundation (grant no. NNCI-2025233) and where the work was done in part. M.G.A. and P.L.M. additionally acknowledge funding from the National Science Foundation (award no. CBET-2123862). F.O.W. acknowledges support from the Eric and Wendy Schmidt AI in Science Postdoctoral Fellowship, a program of Schmidt Sciences, LLC. We thank A. Aksaray, S. Huang, T. Tapen, F. Monticone, A. Senanian, S. Prabhu and V. Krementski for useful discussions.

Author information

Author notes

  1. These authors contributed equally: Bala Govind, Maxwell G. Anderson.

Authors and Affiliations

  1. School of Electrical and Computer Engineering, Cornell University, Ithaca, NY, USA

    Bala Govind & Alyssa Apsel

  2. Kavli Institute at Cornell for Nanoscale Science, Cornell University, Ithaca, NY, USA

    Bala Govind & Peter L. McMahon

  3. School of Applied and Engineering Physics, Cornell University, Ithaca, NY, USA

    Maxwell G. Anderson, Fan O. Wu & Peter L. McMahon

Authors
  1. Bala Govind
    View author publications

    Search author on:PubMed Google Scholar

  2. Maxwell G. Anderson
    View author publications

    Search author on:PubMed Google Scholar

  3. Fan O. Wu
    View author publications

    Search author on:PubMed Google Scholar

  4. Peter L. McMahon
    View author publications

    Search author on:PubMed Google Scholar

  5. Alyssa Apsel
    View author publications

    Search author on:PubMed Google Scholar

Contributions

B.G. conceived the structure of the integrated microwave neural network, designed the CMOS chips and set up the experiment. M.G.A. automated the experiment and performed the machine learning. B.G. and F.O.W. formulated the coupled model of the parametrically driven oscillators, with F.O.W. producing the numerical results. A.A. supervised the design of the chips and the experiment. P.L.M. supervised the machine learning. B.G. wrote the paper and produced the figures, with input from all authors.

Corresponding author

Correspondence to Bala Govind.

Ethics declarations

Competing interests

The authors have filed a provisional US patent application (no. 63/742,208) based on the neural processor presented in this article.

Peer review

Peer review information

Nature Electronics thanks Bowei Dong, Xingjun Wang and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data

Extended Data Fig. 1 Structure of the distributed nonlinear waveguide resonator.

a, The nonlinear transmission line comprises cascaded π sections of fixed inductive waveguide segments and drive-sensitive nonlinear capacitors. Each resonance absorbs part of the incoming microwave signal and transmits the rest to subsequent segments. The overall nonlinear response results from cumulative reflections back to the RFin port. b, Top view of the layout of the nonlinear waveguide. Nonlinear capacitors are inserted periodically along the pretzel-shaped trace. Their nominal bias points are set through analog voltages. c, In the Silicon-on-Insulator process used here, a single polynomial nonlinear capacitor consists of two pairs of antiparallel diodes. c.i, Compact layout of this component. c.ii, The schematic, wherein the diode pairs are biased at their mid-point. c.iii, SPICE / Spectre-model simulated characterization of the nonlinear capacitor across variation in bias voltages and input RF power shows that by injecting a constant RF input power of -10 dBm, the effective capacitance decreases with increasing bias voltage. c.iv, The effective capacitance is highly sensitive to the input RF drive’s power. The capacitance reduces with increasing field strength. c.v, The same nonlinearity, when considering its variance for different frequencies. Here, the capacitance varies with RF voltage as \({{\rm{C}}}_{{\rm{eff}}}({{\rm{V}}}_{{\rm{NL}}},{{\rm{V}}}_{{\rm{in}}})={\rm{a}}-{\rm{b}}{{\rm{V}}}_{{\rm{in}}}+{\rm{c}}{{{\rm{V}}}_{{\rm{in}}}}^{2}-{\rm{d}}{{{\rm{V}}}_{{\rm{in}}}}^{3}+\ldots {\rm{fF}}\) which is well-suited to generating expansive functions for a neural pre-processor.

Extended Data Fig. 2 Structure of the distributed linear waveguide.

a, Top view of the each tunable linear waveguide (B, C and D) that produces linear modes. Switches S1,2….6 are inserted periodically along the length of the transmission line and make it tunable in length, with subsegments TLine1,2…6. Shorting the line through these switches alters the effective length of the waveguide to support different fundamental frequencies. b, Cross-section of RF-optimized and digital metal layers in the 45 nm Silicon-on-Insulator CMOS metal stack. Here, the top three layers are via’ed together for low-loss transmission. The five metal layers below are used for routing control signals from a Serial-to-Parallel Interface, to the switches, and also via’ed together to form a low-loss return path to the power supply. c, A schematic of the linear waveguide resonators, with options to lengthen or shorten the return path of the microwave signal. In the experiments, however, for simplicity, only the shortest path (configuration with all switches turned on) was used for training in the machine learning tasks.

Extended Data Fig. 3 Reduction of the CMOS Microwave Neural Network circuit to a generalized coupled mode model.

a, The integrated microwave neural network consists of interconnected linear and nonlinear resonators. The linear resonator is designed as a single waveguide with an adjustable length, implemented through a cascade of sub-segments, each referred to as Lin. These sub-segments can be grounded via switches (S1, S2, S3, S4 and S5), which immediately terminate the microwave signal’s return path at the first switch that is shorted to ground. In contrast, the nonlinear resonator features a transmission line loaded with polynomially nonlinear capacitors. These capacitors form C-L-C π sections that are coupled by delays. Microwave power from input pads is distributed to these resonators through symmetrically arranged couplers (whose equivalent circuits are marked in purple). The left coupler divides power into two linear waveguides, while the right coupler feeds into a linear waveguide and into a nonlinear waveguide. Saturable gain elements, implemented as cross-coupled transistor pairs, connect the waveguides on opposite sides of the circuit, compensating for losses within the electromagnetic structures. Additionally, a pair of capacitor banks provides a small degree of tunability to the modes supported by the waveguides. Critically, there is parametric coupling between the circuit’s upper and lower halves through a pair of slow bitstream-driven switches. b, To simplify the complex circuit, we recognize that since the linear resonators support only a single natural frequency, they can be represented as tank circuits composed of Lj, and Cj. The symmetry in the bottom half of the circuit allows us to approximate the capacitor banks as two evenly split capacitors, contributing to the overall capacitance of the tank circuits. However, the asymmetry in the configuration of the resonators in the upper half does not permit such a simplification. c, To focus on the primary mechanism by which the system’s sensitivity to incoming signals is enhanced, we can largely ignore the left half of the circuit and concentrate on the interaction between the nonlinear distributed resonances and the linear resonator on the right half. These components interact only through the inductive path via a coupler and a coupling capacitor between the turns of the coupler. The source of regenerative gain through the cross-coupled pair is retained. For ease of analysis, we represent the parametrically driven switch as a tunable capacitor, which can be toggled between a very small value (open circuit) and a very large value (short circuit). d, The reduced circuit can be represented as an ensemble of coupled modes–a cascade of nonlinear resonators connected to a linear resonator via a parametrically varied switched coupling and a fixed phase delay (through the coupler). These modes interact with the incoming drive (fast Gigabit/sec microwave-speed data), with internal losses being compensated by saturable gain.

Extended Data Fig. 4 Effects of initial conditions and drive-detuning on spectral response and time evolution in the Microwave Neural Network, without parametric switch coupling.

a, The integrated circuit in the main Article is reduced to a coupled-mode model. This comprises a cascade of nonlinear resonances (waveguide A) coupled both passively and through parametric switching to a linear resonance. The power spectrum and Poincaré map of modes represented by the first nonlinear resonator (vNL,1) are observed. b, The MNN is driven with an external signal at 0.5ωNL,1. For a first set of initial conditions imposed on individual resonators’ responses, there exists a regime in which (b.i), pure harmonic oscillation is seen and, (b.ii), for another set of initial conditions, comb-like behavior is produced. These correspond to situations where the Poincaré map shows sparse and organized points. This suggests a quasi-stable dynamic state (b.iii), and another where the two islands reflect more unstable, dynamic responses (b.iv), indicating locally chaotic solutions. c, If, instead, the drive was fed to the MNN at 0.65ωNL,1, different working regimes are triggered by different initial conditions. True comb-like behavior (c.i), can be produced, as evidenced by isolated, longer-memory, coherent solutions on the Poincaré map (c.ii). In another instance, divergent chaotic behavior shown by dense, scattered points (indicative of chaotic dynamics) (c.iv) manifests itself by a less structured spectrum emitted by the nonlinear resonator(s) (c.iii). For all simulations, nonlinear coupling coefficients, gain and decay rates are normalized with respect to ωNL,1. Here, βi,i+1 = 0.02 and \({{{\boldsymbol{\gamma }}}_{{\bf{i}}}}^{{\bf{int}}}=0.03\), with i = 1, 2…7 and G equals 0.2. Since parametric coupling is absent, βpar = 0.

Extended Data Fig. 5 Manipulation of the comb-like spectra emitted by the Microwave Neural Network, with microwave-speed signals and slow parametric bitstreams.

In experiments involving Gigabit/sec data and radar returns, the accessible phase space of MNN is constrained by the configurations of incoming drive signals and the parametric bitstreams applied on the switch between the first non-linear mode and a linear mode (a). As a result, the output spectrum is shaped by the constrained, unstable nonlinear dynamics of the system, indicating extreme sensitivity to the modulation of GHz-speed microwave data (1-20 GHz) by slow MHz-speed parameters. For Drive 1, the Poincaré map for slow parameter bitstreams 1 and 2 shows clustered points with a few points that are more spread out (b.iv and b.v), suggesting quasi-periodic dynamics while Bitstream 3 gives tightly clustered points (b.vi), indicating stable, periodic behavior. For Drive 2, under action of all three parametric bitstreams, the formation of ring-like structures indicates quasi-periodic behavior with non-linear dynamics (c.iv, c.v and c.vi). Finally, under Drive 3, bitstreams 1 and 2 induce a combination of ring-like patterns in the Poincaré maps (d.iv and d.v), revealing a tendency towards chaotic behavior, while Bitstream 3 exhibits multiple, structured rings (d.vi), indicating quasi-periodic behavior with complex and unstable dynamics. For all simulations, nonlinear coupling coefficients, gain and decay rates are normalized with respect to ωNL,1. Here, βi,i+1 = 0.02 and \({{{\boldsymbol{\gamma }}}_{{\bf{i}}}}^{{\bf{int}}}=0.03\), with i = 1, 2…7 and G equals 0.2.

Extended Data Fig. 6 Experimental setup to record the microwave neural network’s response to microwave-speed digital bitstreams.

The CMOS microwave neuron chip is wire-bonded to a Printed Circuit Board, which connects it to external power supplies and bias voltages for the oscillators’ core, logic and drivers and Kerr-capacitors’ bias. a, A low-speed bit pattern (clocked under 200 MHz) drives switches that establish parametric coupling between oscillators, imparting a shape to the microwave comb neuron. A first pair of probes, forming a Ground-Signal-Ground-Signal-Ground (GSGSG) configuration cyclically transfer high bandwidth 0-12 gigabits per second bitstreams into the chip to interact with the comb. The resulting microwave computations are manifested as steady-state responses in the frequency domain and produce new combs, each unique to the combination of inputs, that is, the fast input bitstreams and slow parameter bitstreams. Output spectra from two ports are read off spectrum analyzers in a small band between 10 and 14 gigahertz, through a second pair of probes and Ground-Signal-Ground-Signal-Ground waveguides. b, The probe-station assembly consists of the CMOS die attached to a breakout PCB, interfaced with millimeter wave probes for input and output data. It also includes a low-speed BNC cable interface for feeding in the parametric bitstream drive.

Extended Data Fig. 7 Measured frequency-comb-like spectra generated by the coupled nonlinear and linear resonators.

a, The supply voltage on the saturable gain element alters the transconductance of the cross-coupled pair’s transistors and, thereby, the stability of the comb. It has a nominal value of 0.5 V, below which it collapses. b, The polynomial nonlinearity is tuned using bias voltages that set the sensitivity of the transmission line’s capacitances to incoming microwaves. c, The initial profile of the 12.5 GHz comb can be altered by feeding parameters through a 32-bit sequence, run cyclically at slow speeds under 150 MHz. This parameterized comb is exposed to incoming drive signals and performs computation on them.

Extended Data Fig. 8 Characterization of the comb generated by coupling nonlinear modes on chip.

a, Under nominal biasing conditions (power supply of 0.6 V and VNL = 0.6 V) and without the influence of parametric bits and incoming microwave signals, the measured undisturbed frequency comb is centered at 12.47 GHz and has a constant line spacing of 80 MHz. b, The phase noise is measured for various offsets from the central component. While this version of the chip has slightly higher close-in phase noise than conventional CMOS oscillators, microwave neurons built from this modality are stable and highly input sensitive to drive signals.

Supplementary information

Supplementary Information

Supplementary Figs. 1–7, Tables 1–4 and Notes 1–9.

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Govind, B., Anderson, M.G., Wu, F.O. et al. An integrated microwave neural network for broadband computation and communication. Nat Electron 8, 738–750 (2025). https://doi.org/10.1038/s41928-025-01422-1

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue date:

  • DOI: https://doi.org/10.1038/s41928-025-01422-1

Search

Advanced search

Quick links

Nature Briefing AI and Robotics

Sign up for the Nature Briefing: AI and Robotics newsletter — what matters in AI and robotics research, free to your inbox weekly.

Get the most important science stories of the day, free in your inbox. Sign up for Nature Briefing: AI and Robotics