Battery-powered sensor nodes mostly die from idle listening, not from sending data. Idle listening means keeping the radio in receive mode just in case a packet arrives, and it can consume most of a node's energy. The standard fix is a wake-up receiver: a tiny, always-on detector that turns on the main radio only when it senses a specific radio pulse. The KRATOS platform shows the cost of this fix. Its wake-up receiver draws only 1.83 µW, but it hears signals only down to −50 dBm, while the LoRa data radio it guards hears down to −140 dBm (KRATOS). The wake-up path therefore reaches far less distance than the data path. Tunnel field-effect transistors (TFETs) are often proposed as the way out. These are switches that pass current by quantum tunneling rather than by thermally exciting carriers over a barrier, so their current can rise with voltage faster than the thermal limit allows. A detector built on a steeper switch produces more output per unit of input power. This proposal asks whether that promised gain survives realistic physics.
Our central claim is that the answer can be predicted, and that it is probably "mostly not." Recent first-principles simulations found that electron-electron collisions degrade a carbon-nanotube TFET's subthreshold swing from about 10 mV/dec in a collision-free (ballistic) model to about 50 mV/dec, and raise its off-state leakage by more than two orders of magnitude (Xia et al.). Subthreshold swing is the gate voltage needed to change the current tenfold. No one has turned that device result into the quantity a wake-up receiver designer cares about: detection curvature and sensitivity at a fixed microwatt budget. We propose to compute it directly with many-body quantum transport simulation, carry it through a detector model to sensitivity and wake-up range, and settle whether ballistic TFET projections overstate the wake-up range by several decibels.
Core Claim
A detector that works in the exponential part of a transistor's current-voltage curve turns small radio signals into a DC output in proportion to its curvature coefficient,
$$\mathcal{Q} = \frac{d^2 I/dV^2}{dI/dV},$$
where $I$ is the device current and $V$ the input (gate) voltage. When current grows as $I \propto 10^{V/S}$ with swing $S$, this ratio equals $\ln 10 / S$. The same identity turns the thermal limit of 60 mV/dec into about 38 V⁻¹. Applying it to the numbers of Xia et al. gives a back-of-envelope forecast. The ballistic swing of 10 mV/dec implies a curvature near 230 V⁻¹, about six times the thermal limit. The interacting swing of 50 mV/dec implies about 46 V⁻¹, only about 1.2 times the limit.
In a square-law detector, output voltage scales with curvature times input power. If the noise stayed fixed, the sensitivity gain over a thermally limited detector would therefore be roughly 7–8 dB under the ballistic picture but under 1 dB under the interacting picture. The extra leakage raises shot noise, which could push the realistic gain lower still. Free-space path loss grows 20 dB per decade of distance. A 7.7 dB gain would therefore mean about 2.4 times more wake-up range, while a sub-1 dB gain means almost none.
This forecast uses point swings from a logic-oriented study. The real curvature depends on where in the curve the detector is biased, on the bias current the microwatt budget allows, and on how many-body broadening reshapes the band edge. Settling these three things is the research.
Research Program
The work has three layers.
The device layer reuses the simulation method of Xia et al. and Deuschle et al.. It builds a first-principles Hamiltonian in a Wannier basis and solves it with non-equilibrium Green's functions (NEGF), the standard way to compute current through a nanoscale channel attached to contact reservoirs. It adds electron-electron scattering through the self-consistent GW approximation, a many-body correction that lets collisions and Auger processes broaden and shift energy levels. An Auger process is one where one carrier's recombination hands energy to another. Prior work reported swing and off-current for logic switching. We instead extract curvature, conductance, and noise current as functions of gate bias. We do this in the microwatt bias window a wake-up receiver can afford, on the same (8,0) nanotube device and on the monolayer GeSe TFET that Brahma et al. studied only ballistically.
Curvature is a second derivative, so it amplifies numerical error. A relative current error $\epsilon$ at a finite-difference step $h$ produces a curvature error of order $\epsilon/h^2$. With millivolt steps, residuals that are harmless for logic metrics can swamp the answer. Here finite-time convergence enters, as a numerical constraint rather than the headline. We require every bias point to reach a certified residual within a fixed iteration budget, so that a dense bias sweep costs a predictable amount of compute. Two options will be compared under that budget. The first is standard Anderson mixing, a scheme that combines recent iterates to speed convergence. The second is a continuous fractional-power damping of the potential residual, modeled on finite-time and fixed-time control laws (Cortés; Kuang et al.). These laws drive an error to exactly zero in bounded time without the rapid on-off switching of discontinuous controllers. The continuous-time settling guarantee does not carry over automatically to a discrete iteration. We therefore treat this as an ablation that we will drop if it does not beat Anderson.
The detector layer fits a compact current-voltage and noise model to the simulated data and places it in a standard envelope-detector circuit held to the KRATOS power envelope. It reports minimum detectable input power. The network layer converts sensitivity gains into wake-up range and into the guard-time and polling figures of the on-demand TDMA testbed (Piyare et al.). TDMA, time-division multiple access, gives each node its own transmission slot. This conversion shows whether the device physics changes node lifetime.
Baselines
The naive reference point is a thermally limited detector whose curvature is fixed at the thermal value, about 38 V⁻¹ at room temperature. Its absolute sensitivity is anchored to the reported KRATOS figure of −50 dBm at 1.83 µW in deep sleep with the wake-up receiver active (KRATOS). Every simulated device is scored as a decibel change relative to this point, under the same circuit topology, bias power, and bandwidth.
The ballistic NEGF TFET is the projection that circuit designers implicitly rely on. It is the strongest competitor in the sense that matters: if it predicts the interacting device's detector sensitivity to within about 1 dB, the expensive many-body machinery is unnecessary for this application. It runs on identical Hamiltonians, bias grids, and residual tolerances, with the collision self-energy switched off. The reported reference number is a swing of about 10 mV/dec ballistic versus about 50 mV/dec with GW, plus more than two orders of magnitude higher off-current (Xia et al.). For the GeSe device we found no reported ON/OFF ratio on Lacuna, and no reported result on Lacuna for curvature or detector sensitivity of any TFET.
The key ablation is a "broadening-only" model. It keeps the ballistic solver but adds a constant imaginary self-energy fitted to the band-edge broadening that the GW calculation produces. This separates the effect of smeared band edges from the separate Auger leakage channel. We count the many-body result as a real finding only if full GW differs from both ballistic and broadening-only by more than the solver's certified error bar on curvature.
For the solver, the baseline is Anderson mixing run with the same tolerance and a hand-tuned damping factor. Fractional-power damping must reach the target curvature accuracy in fewer total Green's-function evaluations across the whole sweep to count as a win.
Why not just…?
Why not just add a phenomenological broadening term to cheap ballistic simulations? This is a real threat, and it is exactly the ablation above. If broadening-only reproduces GW curvature to within the error bar across both devices, the practical recommendation becomes the cheap model. We would report that outcome as the result. The reason to expect it to fail is that Xia et al. attribute part of the leakage to Auger energy transfer, which a static broadening term cannot represent.
Why not just boost sensitivity with spreading codes and error correction? Shih et al. improve wake-up sensitivity through coding. That approach works at the signal-processing level and stacks with any front-end gain; it does not replace a better front end. Its correlation logic also draws power that the file does not quantify against the 1.83 µW budget. We treat coding as orthogonal. The question here is how many decibels the front end itself can contribute.
Why not just run Anderson mixing to a tighter tolerance instead of borrowing finite-time control? This may well be enough, and Anderson is our default. The fractional-power damping stays only if it lowers total compute at matched curvature accuracy on the GW sweeps, where the file's account suggests self-consistency is hardest. The main scientific claim does not depend on it.
Why not rely on measured TFET data? The file contains no measured detector-relevant TFET characteristics. Simulation is also the only way to separate the collision mechanisms that determine whether better materials or bias choices could recover the lost gain.
Proposed Experiments
The first experiment reproduces the reported swings on the (8,0) nanotube TFET, about 10 and about 50 mV/dec. This validates the pipeline before any new quantity is computed. We then compute curvature, conductance, and shot-noise current over gate bias for ballistic, broadening-only, and full GW models, restricted to bias currents consistent with a sub-2 µW detector, and repeat on monolayer GeSe.
The decisive check is whether full-GW curvature, at an affordable bias current and with leakage-induced noise included, beats the thermal limit by more than 3 dB of detector sensitivity. Below that threshold, TFET front ends are not worth their cost for wake-up receivers. A second decisive check is whether ballistic projections overstate sensitivity by more than 5 dB, the gap our back-of-envelope forecast predicts.
Finally, we convert the sensitivity figures into wake-up range and into projected lifetime. The lifetime model is anchored to the reported more-than-three-year, 1.4-times-over-listen-before-talk result of the on-demand TDMA testbed (Piyare et al.). This conversion states plainly how much of the gap between −50 and −140 dBm device physics could close.
References
- KRATOS: An Open Source Hardware-Software Platform for Rapid Research in LPWANs
- Influence of Carrier–Carrier Interactions on the Sub-Threshold Swing of Band-to-Band Tunnelling Transistors
- Electron-electron interactions in device simulation via nonequilibrium Green's functions and the GW approximation
- Scalability assessment of Group-IV mono-chalcogenide based tunnel FET
- Discontinuous dynamical systems
- Finite-time stabilization control of quantum systems
- On-Demand TDMA for Energy Efficient Data Collection with LoRa and Wake-up Receiver
- High sensitivity wake-up radio using spreading codes: design, evaluation, and applications