Self-Injection Locking Pushes Integrated Radar Below Millimeter Resolution

Electronic sub-terahertz FMCW radars are attractive because the source, receiver, and control circuitry can be integrated, but their available sweep bandwidth is typically below 100 GHz. That constrains conventional range resolution to the millimeter scale, while widening a VCO sweep worsens phase noise and thus range accuracy. The authors present a fully integrated autodyne frequency-modulated (AFM) radar that uses self-injection locking (SIL) to address both constraints: feedback stabilizes the oscillator, and nonlinear receiver dynamics create a comb of useful intermediate-frequency lines.

Core Contribution

The central departure from conventional FMCW processing is that the authors do not treat the physical chirp bandwidth as the sole determinant of resolution. Their SIL AFM radar has a primary 191–258 GHz chirp, or 67 GHz bandwidth, but frequency hopping during the locked dynamics produces IF pulses and higher-order comb lines. The paper models these as an ensemble of synchronized, fictitious FMCW radars with expanded effective bandwidth. Selecting a comb line of order $N$ changes the nominal resolution relation from $\Delta R_{\min}=\alpha c/(2B)$ to $\Delta R_{\min}=\alpha c/(2NB)$.

That is a specific claim about signal representation rather than a literal terahertz-wide transmitted sweep: the supplement explicitly states that only the first line is an actual radiated frequency and that the remaining lines arise from nonlinear SIL behavior. The practical value is that range discrimination can be improved without building a photonic source or a physically wider electronic sweep.

Technical Approach

Figure 1 lays out the arrangement: a voltage-controlled oscillator drives the sub-THz transmit/receive path, while a sampled return is reinjected to establish self-injection locking. A quadratic receiver produces the IF current; its nonlinear dynamics convert transitions between stable locked branches into sharp IF pulses. The recorded pulse train is then analyzed as a frequency comb. The reported VCO is tunable at its 220 GHz second harmonic from 54 GHz at $V_B=3.6$ V to 68.6 GHz at $V_B=2.9$ V.

The implementation also depends on chirp linearity. The authors use a two-stage correction: particle-swarm optimization programs a deformed arbitrary-waveform-generator ramp to compensate the VCO tuning curve, then software resamples the IF signal at fixed phase rather than fixed time. They report that this second step improves linearity for frequency lines below $N<15$; higher-order operation therefore is not simply a matter of choosing a larger comb index.

SIL is also intended to suppress short-term frequency error. In a 1 m example, the paper gives a phase-noise-cancellation factor of 27.6 dB for a 1 MHz offset and a 6.66 ns round-trip delay. Its long-term stabilization measurement at 242.484 GHz over 1,000 s reports frequency standard deviations of 0.44 MHz with SIL and 2.32 MHz free-running, a 5.3-fold reduction.

Results and Analysis

The ranging experiment separates an aluminum target and a copper reference by 570 μm, placing five 100 μm paper sheets in front of the aluminum target. This supports the paper's sub-millimeter-resolution claim under a controlled two-target condition. The imaging experiments then reconstruct a printed M and ABC patterns from the phase of the first IF-comb line while scanning the target in 1 mm steps. The M remains readable after 20 paper pages, although the authors report degrading image SNR as layer-to-layer paper thickness and refractive-index variation accumulate.

Figure 5 provides the clearest accuracy evidence. At 185.5 mm, 30 repeated measurements at each 1 μm stage increment yield 3.4 μm accuracy at $k=2$, corresponding to 95.5% confidence, using the first comb-line phase. Across 18.5–81 cm, averaging 100 IF signals reduces reported accuracy from 16.4–62.6 μm without averaging to 3.4–13.2 μm. The result is less than 0.002% range error, but the low-end figure is contingent on substantial averaging and a translation-stage calibration experiment rather than an unconstrained scene.

The evidence is persuasive for a chip-based, short-range metrology and imaging instrument: it combines SIL dynamics, frequency-stability measurement, target-separation testing, and reconstructed images. It does not yet establish general-purpose radar performance. The range window is limited by comb-line spacing, so off-window scatterers can interfere with adjacent lines; more complex scenes will need higher chirp rates. The authors also identify advanced hardware and software chirp linearization as necessary for large $N$, precisely where the proposed resolution scaling is most appealing.