Spiking neural networks save energy because neurons stay silent unless they have something to say. A spike costs one addition rather than a multiplication. Most of the energy numbers in this literature, however, come from counting operations on ordinary silicon, not from measuring physical devices. One candidate physical substrate is a superconducting nanowire. It "fires" by briefly switching from superconducting to resistive and then cools back down. For such a device, two numbers set the real energy bill. The first is how much energy a switch dumps. The second is how fast the resulting hot electrons shed that heat to the crystal lattice, which sets how quickly the neuron can fire again. The second number is controlled by electron-phonon coupling: how strongly electrons exchange energy with lattice vibrations (phonons). There is also a third, overhead cost: the refrigerator that keeps everything cold, which becomes much cheaper as the operating temperature rises.
The observation behind this proposal is simple. In a conventional superconductor, one measured curve, the Eliashberg function $\alpha^2F(\omega)$, controls both the superconducting transition temperature $T_c$ and the rate at which electrons cool. This curve records how strongly electrons couple to lattice vibrations at each frequency $\omega$. Because one quantity drives both, a conventional material cannot choose its operating temperature and its reset speed independently. That constraint should trace out a frontier in energy-per-spike versus firing rate. Unconventional superconductors are materials where first-principles calculations show phonons are far too weak to explain $T_c$, and pairing is attributed to magnetic (spin) fluctuations. In these materials the two quantities come apart. We propose to compute the conventional frontier from first principles and place unconventional materials against it. The question is whether they sit beyond it or behind it, measured in wall-plug energy per spike at useful firing rates, on real spiking workloads.
Core Claim
The claim is specific. For phonon-mediated superconductors, $T_c$ and the electron-phonon cooling rate are both functionals of $\alpha^2F(\omega)$. Together with the refrigeration cost, which falls steeply with $T_c$, this defines a computable Pareto frontier for superconducting spiking neurons. Materials with high $T_c$ but tiny phonon coupling lie off that frontier by construction. The frontier exists because the same spectral function sets both axes, and these materials break that link. Examples are the pressurized bilayer nickelate La$_3$Ni$_2$O$_7$, with $\lambda = 0.14$ against an observed $T_c \sim 80$ K (Absence of electron-phonon coupling superconductivity in the bilayer phase of La3Ni2O7 under pressure), and iron pnictides whose carriers couple to a broad non-phonon boson spectrum (Optical Spectroscopy of Superconducting Ba0.55K0.45Fe2As2).
Which side of the frontier they land on is genuinely open. Weak phonon coupling means cheap cooling at high $T_c$ but slow electron cooling, and so a risk of the wire latching in its resistive state. Coupling to spin fluctuations may give hot electrons an extra, fast escape route. Settling this sign is the contribution. No paper in the file connects pairing-mechanism physics to neuromorphic energy budgets. The closest device work, A Superconducting Nanowire-based Architecture for Neuromorphic Computing, treats the nanowire material as given and does not ask how its microscopic coupling constrains energy or speed.
How the pieces fit
Direction one supplies the mechanism and the data. Standard first-principles machinery produces $\alpha^2F(\omega)$ and the coupling $\lambda = 2\int \alpha^2F(\omega)/\omega\,d\omega$. Allen-Dynes-type formulas map these to $T_c$. Machine learning of superconducting critical temperature from Eliashberg theory used symbolic regression to fit correction factors to the Allen-Dynes formula. It cut held-out hydride error from 18.7% to 5.8% relative RMSE. We reuse its corrected formula as a method, since it is the most accurate cheap map from $\alpha^2F$ to $T_c$ in the file. We also reuse its 13 first-principles, 42 literature, and 2,819 synthetic spectra as the population that traces the conventional frontier. The same $\alpha^2F$ gives the phonon-limited electron scattering rate at temperature $T$:
$$\frac{\hbar}{\tau_{ep}(T)} = 4\pi\int_0^\infty d\omega\,\alpha^2F(\omega)\,\big[n_B(\omega,T) + f(\omega,T)\big],$$
Here $n_B$ and $f$ are the Bose and Fermi occupation functions. At high temperature this reduces to the familiar $2\pi\lambda k_BT$. We will also compute the full energy-relaxation integral, because energy relaxation, not scattering, sets how fast a hotspot cools. The scattering rate is shown as the readable form.
Direction two supplies the objective. The file's own list of open problems stresses that theoretical energy per operation (0.9 pJ per addition versus 4.6 pJ per multiply-add at 45 nm) diverges from physical energy. We take that criticism seriously by pricing every spike at the wall plug. Heat $Q$ dissipated at operating temperature $T$ costs
$$W_{\text{wall}} = Q\,\frac{T_0 - T}{\eta\,T},$$
where $T_0$ is room temperature and $\eta$ is the refrigerator's fraction of Carnot efficiency. This factor is what makes the $T_c$ axis matter so much. The energy $Q$ per switching event scales with the superconducting condensation energy of the wire volume. That energy grows with the gap, and therefore with $T_c$. So the three quantities are all tied to one spectrum in phonon superconductors: switching energy, cooling-overhead factor, and reset time $\sim\tau_{ep}$.
Research Program
The work has three stages.
Stage one: the conventional frontier. For every $\alpha^2F$ in the Eliashberg dataset, we compute $T_c$ (corrected Allen-Dynes), $\tau_{ep}$ at the operating temperature $T = T_c/2$, and a switching energy. We insert these into a standard electrothermal nanowire model to get energy per spike and maximum firing rate before latching. The synthetic Gaussian spectra fill the space densely. The first-principles and literature spectra anchor it to real materials. HRF neurons are relevant here: A Spiking Heterogeneous Harmonic Resonate-and-Fire State Space Model for Time Series models neurons as damped oscillators and needs sustained high firing rates on long sequences. That makes maximum firing rate a first-class constraint rather than a footnote.
Stage two: unconventional placements. For materials where computed phonon coupling cannot explain $T_c$, we pair the computed $\alpha^2F$ with the experimental $T_c$. These are La$_3$Ni$_2$O$_7$, LaBi (First-principles study of electron-phonon coupling and magnetoresistance of LaBi under pressure, computed $T_c$ 0.12 K versus 4–8 K measured), and the pnictide Ba$_{0.55}$K$_{0.45}$Fe$_2$As$_2$. The pairing is the point: their $T_c$ is set by something other than the spectrum that cools their electrons. We then add a second relaxation channel by inserting a non-phonon boson spectrum alongside $\alpha^2F$ in the rate integral. For the pnictide we use the optically extracted spectrum. For others we use RPA spin-fluctuation spectra built from multi-orbital models of the kind in Effective six-band model and unconventional spin-singlet pairing in Kagome superconductor CsV3Sb5. Energy dumped into spin fluctuations must still reach the lattice, so the model tracks the spin subsystem as a separate heat reservoir with its own coupling to phonons. This avoids simply adding the two rates.
Two honest caveats shape this stage. The nickelate and LaBi require 20–40 GPa and roughly 15 GPa respectively, so they are physics probes of the decoupled regime, not device candidates. The pnictide and kagome families are ambient-pressure and are the realistic device targets. Also, the LaBi superconductivity may come from impurity phases, as the source paper itself notes, so it enters only as a sensitivity case.
Stage three: system energy on real workloads. A per-spike cost means little without spike counts. We take spike statistics from trained models on three workloads. The first is a directly trained spiking ResNet with threshold-dependent batch normalization on DVS-Gesture (Going Deeper With Directly-Trained Larger Spiking Neural Networks, 96.87% accuracy). The second is the HRF state-space model on 18,000-step EigenWorms. The third is the spiking graph network GSG on the CS citation graph (Geometry-Aware Spiking Graph Neural Network). For each workload we recompute energy per inference with device-priced spikes, cooling overhead included. Any model whose required firing rate exceeds a material's latching limit is marked infeasible on that material rather than silently slowed.
Baselines
The naive reference point is the standard operation-counting estimate used across the spiking literature: each spike-driven addition at 0.9 pJ on 45 nm CMOS, with no routing, leakage, or cooling. Its reported values anchor the comparison. GSG is reported at 0.051 mJ per inference on CS versus 18.444 mJ for a dense GCN (Geometry-Aware Spiking Graph Neural Network). The HRF model is reported as about 69 times more efficient than non-spiking LinOSS on EigenWorms. We apply both accountings to identical spike counts from the same trained checkpoints, so any gap comes from the energy model alone.
The strongest competitor on the actual question is the best conventional superconductor on our computed frontier. This is whichever of the 13 first-principles, 42 literature, or HTSC-2025 materials (HTSC-2025: A Benchmark Dataset of Ambient-Pressure High-Temperature Superconductors) minimizes wall-plug energy per spike at the firing rate each workload needs. HTSC-2025 hydrides reach predicted ambient-pressure $T_c$ up to 160 K for Mg$_2$IrH$_6$ with strong phonon coupling, so they may dominate outright. For HTSC-2025 we need $\alpha^2F$, not just $T_c$. Where spectra are unavailable those materials are excluded rather than guessed. For device-level superconducting neurons there is no reported result found on Lacuna for energy per spike or firing rate in A Superconducting Nanowire-based Architecture for Neuromorphic Computing. We therefore build that baseline ourselves with the same electrothermal model and parameters for every material. This keeps the comparison fair: one device model, one refrigerator efficiency, one set of workloads, differing only in material spectra.
An unconventional material beats this baseline only if both of the following hold. It must achieve lower wall-plug energy per inference than the frontier's best point on at least one workload. And it must meet that workload's firing-rate requirement without latching. Both conditions must survive a factor-of-two variation in refrigerator efficiency and wire geometry.
The key ablation isolates the ingredient that makes these materials special, the decoupling of $T_c$ from $\alpha^2F$. We rerun each unconventional material with its $T_c$ replaced by the value its own $\alpha^2F$ predicts (near zero for the nickelate). A second ablation turns the spin-fluctuation relaxation channel on and off. Together these separate three effects: "warm operation is cheap," "phonon cooling is slow," and "spin fluctuations rescue cooling."
Why not just…?
Why not just use the highest-$T_c$ conventional superconductor, such as a stable hydride? This is the most serious threat, and the proposal is built to measure it rather than dismiss it. Strong phonon coupling gives both high $T_c$ and fast cooling, so conventional hydrides could dominate every unconventional candidate. Hydrides also carry uncertainty: HTSC-2025 reports an ALIGNN prediction error of 25 K on high-$T_c$ candidates, and their switching energy grows with the gap. If stage one shows the frontier's best point beats all unconventional materials on every workload, that negative result is itself the finding. It would mean phonon-free pairing buys nothing for neuromorphic hardware, and device work should chase phonon coupling, not exotic pairing.
Why not just count operations on CMOS, as the field already does? Counting additions treats energy as a property of the algorithm. The research file flags that silicon measurements routinely diverge from counts because of routing, peripheral circuits, and leakage. For cryogenic hardware the divergence is worse, because cooling overhead can exceed the device energy by orders of magnitude at low temperature. Stage three reports both accountings on the same spike counts, so the size of this gap is measured, not asserted.
Why not just use room-temperature spintronic neurons and skip refrigeration entirely? Spin Neurons: A Possible Path to Energy-Efficient Neuromorphic Computers and Proposal For Neuromorphic Hardware Using Spin Devices propose domain-wall and spin-current neurons with very low switching energy and no cooling cost. This is a real competitor at the system level. There is no reported result found on Lacuna for their per-spike energy. We will therefore include them only as a stated break-even line: the per-spike energy a room-temperature device would need to match our best superconducting point. We will not fabricate a comparison number.
Why not just fit relaxation times phenomenologically instead of computing them from first principles? Measured hotspot relaxation times exist for only a handful of device materials, and none are in the file. Screening across hundreds of spectra therefore requires computing $\tau_{ep}$ from $\alpha^2F$. The risk is that first-principles rates are off by a constant factor. We handle this by reporting results under a calibrated scale factor spanning an order of magnitude. A conclusion counts only if it holds across that range.
Decisive Checks
The first check is whether the frontier is real. Across the Eliashberg population, $T_c$ and $\tau_{ep}(T_c/2)$ should be tightly linked, with residual spread well below the spread in wall-plug energy per spike. If the corrected Allen-Dynes $T_c$ and the rate integral turn out nearly independent across realistic spectra, the premise fails and the project reduces to routine materials screening. We report this correlation first.
The second check is the placement and its sign. Each unconventional material's off-frontier distance is reported with and without the spin-fluctuation cooling channel. A robust finding would be either of two outcomes. In one, pnictide-class materials sit beyond the frontier only when spin-fluctuation relaxation is included. In the other, they sit behind it because weak phonon coupling causes latching at the rates HRF models need.
The third check is the workload test. For DVS-Gesture, EigenWorms, and CS, we compare wall-plug energy per inference for the best conventional point, the best unconventional point, and the CMOS operation-count estimate, all at matched accuracy from the same checkpoints. Either outcome is informative. A win for phonon-free materials would give device designers a pairing-mechanism criterion for choosing superconducting neuron materials. A loss would show that the electron-phonon coupling limit, which is a nuisance in superconductivity theory, is a hard floor in neuromorphic hardware.
References
- Absence of electron-phonon coupling superconductivity in the bilayer phase of La3Ni2O7 under pressure
- Optical Spectroscopy of Superconducting<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mi>Ba</mml:mi><mml:mn>0.55</mml:mn></mml:msub><mml:msub><mml:mi mathvariant="bold">K</mml:mi><mml:mn>0.45</mml:mn></mml:msub><mml:msub><mml:mi>Fe</mml:mi><mml:mn>2</mml:mn></mml:msub><mml:msub><mml:mi>As</mml:mi><mml:mn>2</mml:mn></mml:msub></mml:math>: Evidence for Strong Coupling to Low-Energy Bosons
- A superconducting nanowire-based architecture for neuromorphic computing
- Machine learning of superconducting critical temperature from Eliashberg theory
- A Spiking Heterogeneous Harmonic Resonate-and-Fire State Space Model for Time Series
- First-principles study of electron-phonon coupling and magnetoresistance of LaBi under pressure
- Effective six-band model and unconventional spin-singlet pairing in Kagome superconductor CsV3Sb5
- Going Deeper With Directly-Trained Larger Spiking Neural Networks
- Geometry-Aware Spiking Graph Neural Network
- HTSC-2025: A benchmark dataset of ambient-pressure high-temperature superconductors for AI-driven critical temperature prediction
- Spin Neurons: A Possible Path to Energy-Efficient Neuromorphic Computers
- Proposal For Neuromorphic Hardware Using Spin Devices