QED-Nano: Nano QED & Theorem-Proving ML

• QED-Nano is a heterogeneous term denoting both nanoscale quantum electrodynamics phenomena and a separate 4B machine-learning model for advanced theorem proving.
• In physics, QED-Nano covers perturbative methods, circuit-QED implementations, and nano-optical designs that refine light–matter interactions in semiconductors, plasmonics, and other nanostructures.
• The ML usage of QED-Nano offers a compact model trained via supervised and reinforcement learning to efficiently tackle Olympiad-level theorem proving.

QED-Nano is used in current literature in two distinct senses. In nanoscale physics, it denotes quantum-electrodynamic phenomena, devices, and computational frameworks in which strong confinement, material dispersion, dielectric screening, and open-system effects modify light–matter interaction at nanometer to micrometer scales. In a separate 2026 machine-learning paper, “QED-Nano” names a 4B open model for Olympiad-level theorem proving. This suggests that the expression is presently heterogeneous rather than terminologically fixed; its dominant physics usage nevertheless concerns how quantized fields, effective couplings, and Green-tensor-mediated interactions behave in semiconductor, nanophotonic, plasmonic, and dispersion-force settings (Kolomeisky, 2014, Moore et al., 12 Jun 2025, LM-Provers et al., 6 Apr 2026).

1. Asymptotic perturbation theory and effective couplings

A foundational use of QED-Nano in the physics sense is the analysis of how perturbative QED changes when transplanted from vacuum electrodynamics to condensed-matter systems with pseudo-relativistic quasiparticles. A quantitative refinement of Dyson’s instability argument maps the breakdown of the QED series to the Chandrasekhar–Landau collapse problem and yields the critical order

$N_c = 3.1\,\alpha^{-3/2},$

so that for vacuum QED, with $\alpha \approx 1/137$, the optimal truncation is

$$N_{\mathrm{opt}} \approx 3.1\,(137)^{3/2} \approx 5000.$$

The many-body estimate uses the semi-relativistic Hamiltonian

$$H=\sum_{i=1}^{N} c\sqrt{p_i^2+(mc)^2}-e^2\sum_{i<j}\frac{1}{|\mathbf r_i-\mathbf r_j|},$$

together with Pauli-constrained scaling, which changes Dyson’s original bosonic $N_c\sim \alpha^{-1}$ law to the fermionic Migdal–Krainov form $N_c\sim \alpha^{-3/2}$ (Kolomeisky, 2014).

The condensed-matter translation replaces $\alpha$ by the screened effective fine-structure constant

$\beta=\frac{e^2}{\hbar v_F\epsilon}.$

For narrow band-gap semiconductors and Weyl semimetals, the same scaling gives

$N_{\mathrm{opt}}\approx 3.1\,\beta^{-3/2},$

with representative values “around 80.” In InSb-type systems, $v_F\sim 4.3\times10^{-3}c$ and $\alpha \approx 1/137$0 give $\alpha \approx 1/137$1 and $\alpha \approx 1/137$2; in Weyl semimetals, $\alpha \approx 1/137$3 and $\alpha \approx 1/137$4 give $\alpha \approx 1/137$5 and $\alpha \approx 1/137$6 (Kolomeisky, 2014).

Graphene is qualitatively different because the relevant scaling is two-dimensional:

$\alpha \approx 1/137$7

With $\alpha \approx 1/137$8, suspended graphene with $\alpha \approx 1/137$9 gives $$N_{\mathrm{opt}} \approx 3.1\,(137)^{3/2} \approx 5000.$$0 and $$N_{\mathrm{opt}} \approx 3.1\,(137)^{3/2} \approx 5000.$$1, while graphene on quartz with $$N_{\mathrm{opt}} \approx 3.1\,(137)^{3/2} \approx 5000.$$2 gives $$N_{\mathrm{opt}} \approx 3.1\,(137)^{3/2} \approx 5000.$$3 and $$N_{\mathrm{opt}} \approx 3.1\,(137)^{3/2} \approx 5000.$$4. The derivation assumes $$N_{\mathrm{opt}} \approx 3.1\,(137)^{3/2} \approx 5000.$$5, so these small values are indicative rather than precise. Even so, the physical conclusion is clear: perturbation theory is practically robust in vacuum QED, narrower but still useful in narrow-gap and Weyl systems, and severely limited in graphene, where resummation, RG-improved expansions, lattice or Monte Carlo methods, and large-$$N_{\mathrm{opt}} \approx 3.1\,(137)^{3/2} \approx 5000.$$6 approaches become the natural alternatives (Kolomeisky, 2014).

2. Circuit-QED implementations in semiconductor nanostructures

In circuit and semiconductor nanostructures, QED-Nano denotes the embedding of charge or spin degrees of freedom into microwave resonators with sufficiently large zero-point fields to produce measurable dispersive shifts, coherent exchange, or strong coupling. A representative realization is the GaAs/AlGaAs gate-defined double quantum dot coupled to a flux-tunable high-impedance resonator made from a 32-SQUID array. The coupled system is modeled by the Jaynes–Cummings Hamiltonian

$$N_{\mathrm{opt}} \approx 3.1\,(137)^{3/2} \approx 5000.$$7

with $$N_{\mathrm{opt}} \approx 3.1\,(137)^{3/2} \approx 5000.$$8, $$N_{\mathrm{opt}} \approx 3.1\,(137)^{3/2} \approx 5000.$$9, and $$H=\sum_{i=1}^{N} c\sqrt{p_i^2+(mc)^2}-e^2\sum_{i<j}\frac{1}{|\mathbf r_i-\mathbf r_j|},$$0. The device achieved $$H=\sum_{i=1}^{N} c\sqrt{p_i^2+(mc)^2}-e^2\sum_{i<j}\frac{1}{|\mathbf r_i-\mathbf r_j|},$$1 MHz, $$H=\sum_{i=1}^{N} c\sqrt{p_i^2+(mc)^2}-e^2\sum_{i<j}\frac{1}{|\mathbf r_i-\mathbf r_j|},$$2 MHz, and $$H=\sum_{i=1}^{N} c\sqrt{p_i^2+(mc)^2}-e^2\sum_{i<j}\frac{1}{|\mathbf r_i-\mathbf r_j|},$$3 MHz, satisfying $$H=\sum_{i=1}^{N} c\sqrt{p_i^2+(mc)^2}-e^2\sum_{i<j}\frac{1}{|\mathbf r_i-\mathbf r_j|},$$4 and yielding $$H=\sum_{i=1}^{N} c\sqrt{p_i^2+(mc)^2}-e^2\sum_{i<j}\frac{1}{|\mathbf r_i-\mathbf r_j|},$$5. The high impedance $$H=\sum_{i=1}^{N} c\sqrt{p_i^2+(mc)^2}-e^2\sum_{i<j}\frac{1}{|\mathbf r_i-\mathbf r_j|},$$6–$$H=\sum_{i=1}^{N} c\sqrt{p_i^2+(mc)^2}-e^2\sum_{i<j}\frac{1}{|\mathbf r_i-\mathbf r_j|},$$7 k$$H=\sum_{i=1}^{N} c\sqrt{p_i^2+(mc)^2}-e^2\sum_{i<j}\frac{1}{|\mathbf r_i-\mathbf r_j|},$$8 enhanced the zero-point voltage and produced an approximately sixfold coupling enhancement relative to standard $$H=\sum_{i=1}^{N} c\sqrt{p_i^2+(mc)^2}-e^2\sum_{i<j}\frac{1}{|\mathbf r_i-\mathbf r_j|},$$9 resonators (Stockklauser et al., 2017).

A different implementation places a single electron on solid neon inside a magnetic-field-compatible NbTiN nanowire resonator. After neon deposition and electron loading, the resonator frequency was $N_c\sim \alpha^{-1}$0 GHz with $N_c\sim \alpha^{-1}$1 MHz, while the avoided crossing gave $N_c\sim \alpha^{-1}$2 MHz and a quasi-static qubit linewidth $N_c\sim \alpha^{-1}$3 MHz. The charge qubit operated at a sweet-spot frequency $N_c\sim \alpha^{-1}$4 GHz, showed Rabi frequencies up to $N_c\sim \alpha^{-1}$5 MHz, and exhibited $N_c\sim \alpha^{-1}$6 ns, Hahn-echo $N_c\sim \alpha^{-1}$7 ns, and $N_c\sim \alpha^{-1}$8s, increasing to $N_c\sim \alpha^{-1}$9s after annealing and reloading. The same platform was analyzed for spin-qubit compatibility via micromagnet-enabled spin–charge hybridization, with estimated spin Rabi frequency $N_c\sim \alpha^{-3/2}$0 MHz and average single-qubit fidelity estimates ranging from $N_c\sim \alpha^{-3/2}$1 to $N_c\sim \alpha^{-3/2}$2 depending on assumptions about $N_c\sim \alpha^{-3/2}$3 and magnetic noise (Wang et al., 29 May 2026).

Not all nano-cQED architectures target vacuum-Rabi splitting. A modular NbTiN-on-sapphire microwave assembly interfaced to a silicon-on-insulator nanowire transistor reached $N_c\sim \alpha^{-3/2}$4 GHz, $N_c\sim \alpha^{-3/2}$5, and $N_c\sim \alpha^{-3/2}$6, with minimum integration times of $N_c\sim \alpha^{-3/2}$7 ns for a single-dot dot–reservoir transition and $N_c\sim \alpha^{-3/2}$8 ns for a double-dot interdot transition. The same platform measured charge noise over nine decades of frequency up to $N_c\sim \alpha^{-3/2}$9 kHz and found a $\alpha$0 spectrum with amplitude $\alpha$1 at $\alpha$2 Hz (Horstig et al., 2023). This establishes a second circuit-QED function within QED-Nano: dispersive metrology and noise spectroscopy rather than coherent excitation exchange.

At still larger coupling ratios, circuit-QED spectra cease to be reducible to cavity-QED intuition. For a flux-qubit–LC-resonator system in the ultrastrong-coupling regime, the closed system is described by a quantum Rabi Hamiltonian, but the measured spectra depend on whether the output line couples inductively or capacitively. Capacitive readout samples an operator proportional to $\alpha$3 and reproduces cavity-QED-like spectra at $\alpha$4, whereas inductive readout samples a flux-derived operator with an additional $\alpha$5 contribution. In the ultrastrong regime, this changes transition visibility and can quench lines that remain bright in the capacitive case. A common simplification—using $\alpha$6 as the measured operator independently of port topology—therefore fails in precisely the parameter region where circuit QED most strongly departs from optical cavity QED (Napoli et al., 2024).

3. Nano-optical, plasmonic, and waveguide realizations

In nanophotonics, QED-Nano emphasizes the modification of spontaneous emission and collective optical response by deeply subwavelength confinement. A strong claim in this direction is that classical electromagnetism becomes inadequate for metallic nanostructures at localized surface plasmon resonance. The argument proceeds from the induced dipole $\alpha$7, with $\alpha$8 on the order of $\alpha$9 for a $\beta=\frac{e^2}{\hbar v_F\epsilon}.$0 nm-radius sphere, which produces dipoles in the range of $\beta=\frac{e^2}{\hbar v_F\epsilon}.$1–$\beta=\frac{e^2}{\hbar v_F\epsilon}.$2 Debye. Coupling this dipole to the vacuum field yields interaction energies on the order of $\beta=\frac{e^2}{\hbar v_F\epsilon}.$3–$\beta=\frac{e^2}{\hbar v_F\epsilon}.$4 eV, comparable to plasmon linewidths. The resonance is then written as

$\beta=\frac{e^2}{\hbar v_F\epsilon}.$5

where $\beta=\frac{e^2}{\hbar v_F\epsilon}.$6 is the classical eigenfrequency and $\beta=\frac{e^2}{\hbar v_F\epsilon}.$7 is a negative QED shift. In this framework, the QED correction reverses the incorrect size trend predicted by purely classical models and reproduces experimental data for Ag nanospheres, Au nanorods, and Au, Al, and Pt nanoplates (Zhang et al., 2015).

A fully layered waveguide realization was demonstrated in an all-van der Waals heterostructure consisting of a photoluminescent monolayer MoTe$\beta=\frac{e^2}{\hbar v_F\epsilon}.$8 embedded between two WSe$\beta=\frac{e^2}{\hbar v_F\epsilon}.$9 slabs on Au. The electromagnetic environment was analyzed through the dyadic Green tensor, with spontaneous emission rate

$N_{\mathrm{opt}}\approx 3.1\,\beta^{-3/2},$0

and Purcell factor $N_{\mathrm{opt}}\approx 3.1\,\beta^{-3/2},$1. Interferometric nano-photoluminescence directly imaged spatial oscillations arising from emission into guided slab modes. For the TE$N_{\mathrm{opt}}\approx 3.1\,\beta^{-3/2},$2 branch near $N_{\mathrm{opt}}\approx 3.1\,\beta^{-3/2},$3 eV in a $N_{\mathrm{opt}}\approx 3.1\,\beta^{-3/2},$4 nm slab, the group velocity was $N_{\mathrm{opt}}\approx 3.1\,\beta^{-3/2},$5 with $N_{\mathrm{opt}}\approx 3.1\,\beta^{-3/2},$6, the propagation length was $N_{\mathrm{opt}}\approx 3.1\,\beta^{-3/2},$7m, the mode-aggregated Purcell enhancement reached $N_{\mathrm{opt}}\approx 3.1\,\beta^{-3/2},$8, and the mode-aggregated radiative yield $N_{\mathrm{opt}}\approx 3.1\,\beta^{-3/2},$9 was $v_F\sim 4.3\times10^{-3}c$0–$v_F\sim 4.3\times10^{-3}c$1 at room temperature, with modeling indicating $v_F\sim 4.3\times10^{-3}c$2 if nonradiative loss were suppressed. Because the relevant modes are traveling-wave guided states rather than discrete cavity resonances, the language of $v_F\sim 4.3\times10^{-3}c$3, $v_F\sim 4.3\times10^{-3}c$4, and $v_F\sim 4.3\times10^{-3}c$5 is replaced here by LDOS, dispersion, and radiative versus nonradiative channel decomposition (Moore et al., 12 Jun 2025).

Nanophotonic cavity QED with thermal atoms provides another variant. A suspended SiN nanobeam photonic-crystal cavity designed near the Rb D$v_F\sim 4.3\times10^{-3}c$6 line combined $v_F\sim 4.3\times10^{-3}c$7 and $v_F\sim 4.3\times10^{-3}c$8, giving a peak coupling $v_F\sim 4.3\times10^{-3}c$9 GHz. Thermal Rb atoms cross the cavity mode in $\alpha \approx 1/137$00–$\alpha \approx 1/137$01 ns, so transit-time broadening is on the order of $\alpha \approx 1/137$02 GHz, yet the analysis still found transient strong coupling and Rabi flopping, with reported cooperativity $\alpha \approx 1/137$03 after including motion and Casimir–Polder effects (Alaeian et al., 2019). This broadens the scope of QED-Nano beyond solid-state emitters: subwavelength cavity volumes can compensate, at least probabilistically, for short interaction times and motional disorder.

4. Macroscopic QED, Green tensors, and dispersion forces

A major theoretical strand within QED-Nano is macroscopic QED, in which geometry, dispersion, and absorption enter exclusively through the classical dyadic Green tensor. The open-source package MQED-QD formalizes this as a workflow: construct $\alpha \approx 1/137$04 using analytic planar solvers or boundary-element simulations, map $\alpha \approx 1/137$05 and $\alpha \approx 1/137$06 into coherent and dissipative couplings, and propagate the resulting master equation with QuTiP. The central relations are

$\alpha \approx 1/137$07

and

$\alpha \approx 1/137$08

Case studies for molecular chains near silver showed that a planar surface mainly alters nearest-neighbor transport, whereas a silver nanorod supports surface-plasmon-polariton-mediated long-range dipole–dipole interactions that accelerate exciton delocalization and increase the participation ratio relative to planar geometries (Liu et al., 5 Mar 2026).

The same Green-tensor logic governs dispersion forces. For a one-dimensional nano-grating, the scattering Green tensor can be derived by a Rayleigh expansion over diffraction orders and inserted into the zero-temperature Casimir–Polder potential

$\alpha \approx 1/137$09

For a rectangular Au grating with period $\alpha \approx 1/137$10m, bar width $\alpha \approx 1/137$11m, and depth $\alpha \approx 1/137$12 nm, the $\alpha \approx 1/137$13-polarized component of the interaction becomes repulsive in the normal direction for separations $\alpha \approx 1/137$14 nm at $\alpha \approx 1/137$15. The result directly contradicts the common expectation that Casimir–Polder forces near passive nanostructures are necessarily attractive; the repulsion is not generic, but it is allowed when strong anisotropy selects diffraction and polarization channels with the appropriate sign structure (Buhmann et al., 2015).

These macroscopic-QED formulations also delimit their own regime of validity. MQED-QD assumes weak light–matter coupling, a Born–Markov description, and typically a single-excitation manifold; the nano-grating analysis assumes linear, causal, passive media and relies on convergent Rayleigh expansions. This suggests that QED-Nano, in its formal sense, is not a single theory but a hierarchy of geometry-aware reductions from Maxwell theory to open quantum dynamics and dispersion forces (Liu et al., 5 Mar 2026, Buhmann et al., 2015).

5. Fabrication, integration, and control infrastructures

Experimental QED-Nano depends on fabrication strategies that preserve material cleanliness while remaining compatible with microwave circuitry. A notable example is the deterministic transfer of clean, suspended carbon nanotubes into hybrid cQED devices. The method uses a separate growth chip with $\alpha \approx 1/137$16 cantilevers, low-voltage SEM pre-selection, storage and transfer at $\alpha \approx 1/137$17 mbar, deep trenches of $\alpha \approx 1/137$18–$\alpha \approx 1/137$19m on the circuit chip, and a vacuum stamping procedure in which outer CNT segments are cut by currents of $\alpha \approx 1/137$20–$\alpha \approx 1/137$21A. The cutting step locally anneals the CNT–metal interface and reduces contact resistance to below $\alpha \approx 1/137$22 M$\alpha \approx 1/137$23. Across eight devices, the approach supported Au, PdNi/Pd, and Nb/Pd contacts, yielded resonator quality factors of $\alpha \approx 1/137$24–$\alpha \approx 1/137$25 where reported, and preserved clean single- and double-dot transport, including a superconducting gap $\alpha \approx 1/137$26–$\alpha \approx 1/137$27 meV in Nb/Pd-contacted devices (Cubaynes et al., 2020).

This work is also a useful corrective to a frequent misconception. Clean transport and cQED compatibility do not by themselves imply that the full cavity-QED parameter set has been measured. The CNT study reports neither $\alpha \approx 1/137$28, $\alpha \approx 1/137$29, $\alpha \approx 1/137$30, nor $\alpha \approx 1/137$31, and therefore does not claim a Jaynes–Cummings-level characterization. Its significance lies instead in solving a bottleneck of QED-Nano fabrication: integrating suspended, low-disorder nanostructures into pre-fabricated microwave circuits without sacrificing compatibility with superconducting or ferromagnetic contacts (Cubaynes et al., 2020).

Control electronics are becoming comparably important. An RFSoC-based extension of the open-source QICK framework achieved an effective timing resolution of $\alpha \approx 1/137$32 ps by exploiting the native DAC grid at $\alpha \approx 1/137$33 GHz while the FPGA sequencer operated at $\alpha \approx 1/137$34 MHz. The method precomputes waveform replicas shifted by integer DAC samples and decomposes each delay into coarse sequencer cycles plus a fine DAC-sample offset, while preserving NCO phase continuity. Applied to NV-center dynamical-decoupling spectroscopy, this resolved hyperfine couplings with sub-kHz precision and synchronized microwave control, laser gating, and photon counting within a single timing domain (Marcenac et al., 13 Apr 2026). Although this is not itself a cavity or waveguide QED platform, it shows that QED-Nano increasingly includes the control stack needed to interrogate narrow spectral features in nanoscale quantum hardware.

6. Distinct machine-learning usage

A separate and unrelated meaning of QED-Nano appears in the paper “QED-Nano: Teaching a Tiny Model to Prove Hard Theorems,” where the term denotes a 4B open model for Olympiad-level proof generation rather than a quantum-electrodynamic platform. The model is built from Qwen3-4B (Thinking) and trained in three stages: supervised fine-tuning distilled from DeepSeek-Math-V2, reinforcement learning with rubric-based rewards, and reinforcement learning with a Reasoning Cache that decomposes long proofs into summarize-and-refine cycles (LM-Provers et al., 6 Apr 2026).

Its reported evaluation underscores that this usage is semantically separate from physics despite sharing the same label. On IMO-ProofBench, Qwen3-4B-Thinking-2507 scored $\alpha \approx 1/137$35, QED-Nano-SFT $\alpha \approx 1/137$36, QED-Nano-RL $\alpha \approx 1/137$37, and QED-Nano with the RSA scaffold $\alpha \approx 1/137$38 avg@3 grade percent; corresponding ProofBench scores were $\alpha \approx 1/137$39, $\alpha \approx 1/137$40, $\alpha \approx 1/137$41, and $\alpha \approx 1/137$42, while IMO-AnswerBench scores were $\alpha \approx 1/137$43, $\alpha \approx 1/137$44, $\alpha \approx 1/137$45, and $\alpha \approx 1/137$46 (LM-Provers et al., 6 Apr 2026). The full pipeline, datasets, and code were released publicly.

A common misconception is therefore easy to state and easy to correct: “QED-Nano” does not uniquely denote nanoscale quantum electrodynamics. In current arXiv usage, it also names a theorem-proving LLM. For literature searches, database indexing, and citation practice, the distinction matters. In physics, the term points toward nanoscale QED platforms, Green-tensor methods, and geometry-dependent light–matter interaction; in machine learning, it denotes an open small-model reasoning system trained for long-form proofs (LM-Provers et al., 6 Apr 2026).