---
title: 'Pinductor: Cross-Disciplinary Inductive Devices'
url: https://www.emergentmind.com/topics/pinductor
type: topic
---

# Pinductor: Cross-Disciplinary Inductive Devices

Pinductor is a term that appears in arXiv literature with multiple domain-specific meanings rather than a single canonical definition. In electromagnetics and microsystems it denotes several classes of inductive or inductive-adjacent devices, including Hall-steered nanoscale inductors on topological insulators, PCB-based inductive position sensors, printable nanocomposite-core microinductors, digitally programmable MEMS microinductors, and pinned superconducting superinductors. In machine learning, it denotes the “POMDP-inductor,” a method for inducing executable POMDP world models from observation–action trajectories alone. This suggests that the term functions as a cross-disciplinary label for systems in which inductive behavior, inductive sensing, or model induction is organized around a “pinned,” structured, or programmatically induced mechanism rather than a single standardized device class [1706.07381], [2503.14627], [2507.23522], [2212.08899], [1901.01515], [2605.13740].

## 1. Nomenclature and cross-disciplinary usage

The term has been used in at least six distinct technical senses in the cited literature.

| Domain | Meaning of “Pinductor” | Representative arXiv id |
|---|---|---|
| Topological nanoelectronics | Hall-steered topological inductor interpretation | [1706.07381] |
| Automotive sensing | PCB-based inductive position sensor | [2503.14627] |
| Printed magnetics | Printable nanocomposite-enabled microinductor core | [2507.23522] |
| MEMS passives | Digitally controlled variable microinductor | [2212.08899] |
| Superconducting circuits | Weakly pinned superinductor | [1901.01515] |
| Machine learning | POMDP-inductor for world-model learning | [2605.13740] |

The resulting ambiguity is substantive rather than merely terminological. In some cases the word designates an electromagnetic component with stored magnetic energy; in others it denotes an inductive sensor; and in the machine-learning usage it names an induction procedure over latent dynamical models. The literature therefore does not support a unique field-independent definition. A common misconception is that “Pinductor” identifies one established hardware platform. The published record instead supports a heterogeneous usage pattern, with each field grounding the term in its own operative mechanism.

## 2. Hall-steered topological inductors on topological insulator surfaces

In nanoscale RF and power microelectronics, a closely related “Pinductor” concept is the topological inductor proposed on the surface of a 3D time-reversal-invariant topological insulator. The structure consists of a TI substrate whose surface hosts Dirac electrons with spin–momentum locking, together with square ferromagnetic islands having alternating out-of-plane magnetizations \(+z\) and \(-z\). A bias applied along \(x\) drives surface current, while the chemical potential \(\mu\) is tuned either into the magnetically induced surface gap, producing the quantum anomalous Hall regime, or into the surface bands, producing the anomalous Hall regime [1706.07381].

The underlying surface dispersion is
\[
E_{\mathrm{surf}}=\sqrt{\hbar^2 v_F^2 |k|^2 + M^2},
\]
with exchange \(M\) opening a Dirac mass gap. In the QAH regime, the Hall conductivity is approximately
\[
\sigma_{xy}\approx \nu_{\mathrm{occ}}\frac{e^2}{h},
\]
with \(\nu_{\mathrm{occ}}=\pm 1\) depending on FI magnetization. In both QAH and AHE regimes, the transverse Hall response deflects the longitudinal current around each ferromagnetic island, creating counter-rotating current loops. For alternating FI magnetization, the current circulates counter-clockwise around a \(+z\) island and clockwise around a \(-z\) island. The magnetic fields from neighboring loops then link, increasing stored magnetic energy and magnetic flux density \(B_z\) over and between the islands.

The device was modeled by an AC non-equilibrium Green’s function formalism self-consistently coupled to a full 3D electrodynamics solver of Maxwell’s equations in the frequency domain. In Lorenz gauge, the electrodynamic subsystem solves
\[
(\nabla^2+\omega^2/c^2)V=-\rho/\epsilon,\qquad
(\nabla^2+\omega^2/c^2)A=-\mu J,
\]
with \(A\) entering the tight-binding Hamiltonian through a Peierls phase. Inductance is extracted from magnetic energy,
\[
E_B=\int dV\, \frac{1}{\mu_0}|B|^2,\qquad
L=\frac{2E_B}{I^2},
\]
consistent with \(L=\Phi/I\).

The modeled TI has a \(1\ \mathrm{eV}\) bulk gap and ferromagnetic islands characterized by \(M=\pm 0.2\ \mathrm{eV}\). The NEGF domain is \((12a_0,10a_0,5a_0)\) with \(a_0=1\ \text{\AA}\); FI side length is \(0.3\ \mathrm{nm}\); spacing is varied from \(0.2\) to \(1.7\ \mathrm{nm}\); the bias is AC \(10\ \mathrm{mV}\); and frequency spans \(10\ \mathrm{MHz}\) to \(1\ \mathrm{THz}\). In the simplest two-island geometry, the reported inductance density is \(225\ \mathrm{nH/mm^2}\) from \(10\ \mathrm{MHz}\) to \(1\ \mathrm{THz}\). In the comparison benchmark, the reported value is \(930\ \mathrm{nH/mm^2}\) with a cut-off frequency of \(1000\ \mathrm{GHz}\). By contrast, a bare TI surface without FIs yields \(<1\ \mathrm{pH/mm^2}\).

Inter-island spacing is critical. Inductance is maximized when the gap captures roughly a half period of the sinusoidal inter-island Hall current \(I_y\), and for the modeled system the numerical optimum is approximately \(1.2\ \mathrm{nm}\). As \(\mu\) moves from the QAH gap into the AHE regime and then toward bulk bands, circulating current and \(B_z\) increase, but terminal current increases faster, reducing net inductance; nevertheless, the reported inductance drop is only about \(4\%\) at \(\mu=0.7\ \mathrm{eV}\). Simulations with on-site disorder in \([ -D/2, D/2]\) show that \(L\) can increase up to \(1.5\ \mu\mathrm{H/mm^2}\) at \(D=0.6\ \mathrm{eV}\), albeit with larger variance, through skew-scattering-induced “flux pockets.” This motivates the interpretation of the device as a Pinductor whose current vortices are effectively pinned by FI magnetization patterns.

## 3. PCB-based inductive position sensors in automotive applications

In automotive and industrial control, Pinductor denotes a PCB-based inductive position sensor using coupling between transmitter and receiver coils and a passive conductive target. A PCB stator contains a transmitter coil, several receiver coils, and the signal-processing ASIC, while the rotor carries a thin conductive coupling element with a winged pattern matched to the sensor periodicity \(p\). The transmitter operates typically at \(3\text{–}5\ \mathrm{MHz}\), generating a time-varying field that induces eddy currents in a non-ferromagnetic target such as stainless steel, aluminum, or copper. Those eddy currents modulate the effective mutual coupling between TX and RX as a function of rotor angle [2503.14627].

The key signal relation is
\[
v_r(\theta)=j\omega M(\theta) i_t.
\]
Two receiver coils are spatially phase-shifted by \(90^\circ/p\), producing approximate quadrature outputs after demodulation,
\[
V_{\cos}(\theta)\approx A\cos\theta,\qquad
V_{\sin}(\theta)\approx A\sin\theta,
\]
with angle recovered by
\[
\theta=\operatorname{atan2}(V_{\sin},V_{\cos}).
\]
The sensor’s electrical angle repeats \(p\) times per mechanical revolution, so the uniqueness range is \(2\pi/p\). In motor-control applications, \(p\) can be matched to motor pole-pair count.

A defining feature is the differential RX geometry. Alternating surface normals over coil segments suppress direct TX-to-RX coupling, so the dominant received signal arises from target-induced eddy-current modulation. PCB thickness is typically \(0.8\text{–}1.6\ \mathrm{mm}\), four-layer boards are common, and coil layers are placed on the side facing the target to minimize air gap. Designs can operate at about \(1\ \mathrm{mm}\) air gap and larger, although increasing gap reduces signal amplitude and SNR.

The frequency-domain design is tied to skin depth,
\[
\delta=\sqrt{\frac{2}{\omega\mu\sigma}}.
\]
At \(3\text{–}5\ \mathrm{MHz}\), the reported values are approximately \(30\text{–}38\ \mu\mathrm{m}\) for copper, \(38\text{–}49\ \mu\mathrm{m}\) for aluminum, \(190\text{–}246\ \mu\mathrm{m}\) for 1.4301 stainless steel with \(\mu_r\approx 1\), and \(\lesssim 20\ \mu\mathrm{m}\) for ferromagnetic steel. Non-ferromagnetic targets are preferred because they provide larger \(\delta\) and stronger, better-distributed eddy currents around the target wings.

The paper reports robustness against magnetic stray fields exceeding \(4000\ \mathrm{A/m}\), attributed to differential RX geometry and narrow-band synchronous demodulation. RX signals are typically \(5\text{–}100\ \mathrm{mV}\), demodulation bandwidth can be as narrow as \(\pm 5\ \mathrm{kHz}\) around the TX excitation, and high-speed operation up to approximately \(300{,}000\) electrical rpm is supported. A sample design with \(p=5\), \(n_w=3\), and sinusoidal RX centerlines exhibits mechanical angle error less than \(0.02^\circ\) (\(0.1^\circ\) electrical) in both finite-element simulation and test-bench measurement at \(1\ \mathrm{mm}\) air gap. The transfer ratio is roughly \(1\ \mathrm{V}\) demodulated RX per \(20\ \mathrm{V}\) TX drive at that air gap, while typical ASICs can operate with transfer ratios down to about \(1\ \mathrm{V}/250\ \mathrm{V}\).

## 4. Printable nanocomposite Pinductors and microinductor cores

In printed magnetics, Pinductor refers to microinductors enabled by printable or castable magnetic nanocomposites. The material system reported for this application consists of superparamagnetic \(\gamma\)-Fe\(_2\)O\(_3\) nanoparticles in an insulating poly-vinyl alcohol matrix with UV curing. The particle diameter is \(11\pm 3\ \mathrm{nm}\) by TEM, and volume fractions between \(10\%\) and \(45\%\) are achieved with excellent dispersion; small-angle neutron scattering shows that for \(18\text{–}45\ \mathrm{vol}\%\) loadings, at least \(95\%\) of nanoparticles are isolated rather than aggregated [2507.23522].

The magnetic volume susceptibility rises with loading and reaches \(\chi_v\approx 17\) for the \(44.7\text{–}45\ \mathrm{vol}\%\) sample, corresponding to
\[
\chi_v=\mu_r-1,
\]
so \(\mu_r\approx 18\). The AC susceptibility is reported as flat in-phase up to approximately \(10\ \mathrm{kHz}\); above that, the in-phase component decreases and the out-of-phase component increases due to blocking of the largest particles in the \(11\pm 3\ \mathrm{nm}\) distribution. Debye-model fits give \(K\approx 12.7\ \mathrm{kJ/m^3}\) and \(\tau_0\approx 2\ \mathrm{ns}\), with Néel relaxation
\[
\tau_N=\tau_0 \exp(KV/k_B T),
\qquad
f_{\mathrm{crit}}\approx \frac{1}{2\pi \tau_N}.
\]

The paper emphasizes that the composite is eddy current-free. The PVA matrix is insulating, particles are physically isolated, and measured loss scaling is inconsistent with an eddy-current-dominated \(f^2\) law. Above roughly \(100\ \mathrm{kHz}\), the significant loss mechanism is dynamic hysteresis associated with the largest nanoparticles transitioning from the superparamagnetic to the blocked regime. The reported scaling is
\[
P_{\mathrm{loss}}(f,B)\propto B^2 f^\alpha,\qquad \alpha\in[1,1.3].
\]
Measured loss magnitudes span \(10^2\text{–}10^5\ \mathrm{kW/m^3}\) across \(160.6\text{–}922.7\ \mathrm{kHz}\) and induced \(B\approx 7\text{–}107\ \mathrm{mT}\); at about \(1\ \mathrm{MHz}\) and \(30\ \mathrm{mT}\), the reported range is \(10^3\text{–}10^4\ \mathrm{kW/m^3}\).

The composite was integrated into a PCB-based 3-turn inductor with a closed, 3-leg core formed by filling magnetic vias and printing top and backside magnetic paths. The device was successfully tested in a power converter up to \(3.5\ \mathrm{MHz}\) switching, and small-signal inductance was measured from \(1\ \mathrm{kHz}\) to \(100\ \mathrm{MHz}\). The design relation used for extraction is
\[
L=\mu_0 \mu_r \frac{N^2 A}{\ell_{\mathrm{eff}}}.
\]
Because \(\chi_v\approx 17\) at \(45\ \mathrm{vol}\%\), the closed printed core provides a substantial inductance increase over the corresponding air-core geometry. The principal materials recommendation is to narrow the particle size distribution in order to suppress early blocking, flatten \(\chi'(f)\), and reduce \(\chi''\) in the target operating band.

## 5. Digitally controlled MEMS variable inductors

Another hardware usage of Pinductor is the digitally programmable microinductor formed by monolithically integrating five identical 3D microcoils with five in-plane electrostatic MEMS switches. Each coil is wound in two stacked layers around a vertical magnetic core and anchored on chip; one terminal is tied to the input and the other is routed through its associated switch network. By actuating the MEMS cantilevers, the coils can be connected in series, in parallel, or in mixed series/parallel configurations, yielding a discrete inductance range from one-fifth of a single-coil inductance to five times that inductance [2212.08899].

For \(n\) identical coils, the design yields
\[
\text{Number of steps} = \frac{n(n+1)}{2}.
\]
With \(n=5\), fifteen discrete values are reported:
\(0.20L\), \(0.25L\), \(0.33L\), \(0.50L\), \(1.00L\), \(1.25L\), \(1.33L\), \(1.50L\), \(2.00L\), \(2.33L\), \(2.50L\), \(3.00L\), \(3.50L\), \(4.00L\), and \(5.00L\), where \(L\) denotes the single-coil inductance. The baseline inductance formula is
\[
L=\mu_0 \mu_r \frac{N^2 A}{l},
\]
and finite-element extraction uses
\[
L=\frac{2W_m}{I^2}.
\]

The process targets coil lengths from approximately \(100\ \mu\mathrm{m}\) to \(1000\ \mu\mathrm{m}\) and approximately \(10\) to \(100\) turns. Performance is constrained by standard optical lithography limits on line width, spacing, and coil diameter, as well as by the maximum core height available in the five-layer PolyMUMPs process. The Ni-based magnetic core is reported to improve the response by more than five times relative to a Si core. Representative COMSOL values for a 10-turn segment and a \(10\ \mu\mathrm{m}\)-thick core are \(L\approx 4.624\ \mathrm{nH}\), \(R_s\approx 3.101\ \Omega\), and \(W_m\approx 2.312\ \mathrm{nJ}\) for Ni; for a \(20\ \mu\mathrm{m}\)-thick Ni core the corresponding values are \(L\approx 5.328\ \mathrm{nH}\), \(R_s\approx 3.932\ \Omega\), and \(W_m\approx 2.664\ \mathrm{nJ}\).

The MEMS actuation voltage is simulated over \(1\text{–}15\ \mathrm{V}\). The reported mechanical modes are \(32.772\ \mathrm{kHz}\) for the fundamental in-plane mode, \(81.848\ \mathrm{kHz}\) for the second out-of-plane mode, and \(205.33\ \mathrm{kHz}\) for the third in-plane mode. AC behavior is interpreted through the standard relations
\[
Q=\frac{\omega L}{R_s},
\qquad
f_s=\frac{1}{2\pi \sqrt{LC_p}}.
\]
This usage of Pinductor is therefore not a new physical inductive mechanism, but a digitally reconfigurable inductance architecture realized within a fully MEMS-compatible fabrication flow.

## 6. Weakly pinned superconducting superinductors

In superconducting circuit physics, Pinductor denotes a superinductor whose collective charge mode is weakly pinned by disorder and quantum phase slips. The physical system is a one-dimensional chain of \(N\) small Josephson junctions in series. Each island has junction capacitance \(C\) to its neighbors and ground capacitance \(C_g\), with typical parameters \(N\sim 10^2\), \(E_J/E_C\sim 20\), and \(E_C/E_g\sim 10^{-4}\), while longer chains of order \(10^4\) junctions make weak pinning experimentally relevant [1901.01515].

At long wavelength, the clean chain is described by a Luttinger liquid with parameter
\[
K=\pi \sqrt{\frac{E_J}{2E_g}}=\frac{\pi \hbar}{4e^2 Z},
\]
plasmon velocity
\[
v=a\sqrt{\frac{2E_JE_g}{\hbar^2}},
\]
and dispersion
\[
\omega(q)=\frac{v|q|}{\sqrt{1+(vq/\Omega)^2}},
\qquad
\Omega=\sqrt{\frac{2E_JE_C}{\hbar^2}}.
\]
Disorder enters through a phase-slip term
\[
H=H_0-\frac{\lambda}{a}\int dx\, \cos[2\theta(x)+\chi(x)],
\]
where random offset charges produce the phase field \(\chi(x)\). Under renormalization,
\[
\frac{dD}{dl}\simeq (3-2K)D,
\]
so the Bose-glass transition occurs at \(K_c=3/2\). In the localized phase, the charge-density wave is pinned.

The pinning scale is set by the Larkin length
\[
R_* = a \left(\frac{\hbar \Omega}{K\lambda}\right)^{\frac{2}{3-2K}}
\left(\frac{C}{C_g}\right)^{\frac{1-K}{3-2K}},
\]
with associated frequency
\[
\omega_*=\frac{v}{R_*}.
\]
For \(\omega\gg \omega_*\), the response is that of weakly disordered standing plasmons; for \(\omega\ll \omega_*\), excitations become localized and the low-frequency spectrum is that of a pinned charge density wave. The disorder-induced mean free path obeys
\[
\ell(\omega)=R_*\left(\frac{\omega}{\omega_*}\right)^2.
\]

Microwave reflection directly probes this transition. With a transmission line coupled to one end of the chain, the real part of the reflection amplitude satisfies
\[
r'(\omega)=-1+\frac{K}{K_0}\frac{\nu(0,\omega)}{\nu_0},
\]
where \(\nu(0,\omega)\) is the edge local density of states. In the pinned regime, the paper finds
\[
\langle \nu(0,\omega)\rangle = C \nu_0 \left(\frac{\omega}{\omega_*}\right)^4,
\qquad C\approx 0.032,
\]
so \(\langle r'(\omega)\rangle \to -1\) as \(\omega\to 0\). At high frequency, well-resolved half-integer-spaced standing-wave modes appear with nominal spacing \(\Delta=\pi v/d\), and the quality factor scales as
\[
Q(\omega)=\frac{\omega}{(4\ln 2)\,\delta(\omega)}\propto \omega^2.
\]
This “pinned” superinductor usage is therefore conceptually close to the ordinary linguistic reading of Pinductor: the inductive element remains a superinductor, but its collective mode is pinned by weak disorder.

## 7. POMDP-inductor in world-model learning

In machine learning, Pinductor is the name of a method rather than a physical component. The term is explicitly expanded as POMDP-inductor and denotes an LLM-assisted procedure for inducing executable POMDP world models from observation–action–reward trajectories alone, under strict partial observability: latent states are never revealed, either online or post hoc [2605.13740].

The target model is a POMDP
\[
(\mathcal{S},\mathcal{A},\mathcal{O},T,O,R,\rho_0,\gamma).
\]
Given a dataset of trajectories
\[
\tau=(o_0^\tau,a_0^\tau,r_1^\tau,d_1^\tau,o_1^\tau,\ldots,a_{H_\tau-1}^\tau,r_{H_\tau}^\tau,d_{H_\tau}^\tau,o_{H_\tau}^\tau),
\]
Pinductor asks an LLM to propose code for \(\rho_0^m\), \(T^m\), \(O^m_{\mathrm{LLM}}\), and \(R^m\). Candidate models are then evaluated by a belief-based likelihood computed from the model’s own filtered beliefs rather than privileged hidden states. The Bayes filter is
\[
b_{t+1}(s')=\eta\, O(o_{t+1}\mid s',a_t)\sum_s T(s'\mid s,a_t)b_t(s),
\]
and the softened observation model is defined by a distance kernel,
\[
O^m(o\mid s,a)\propto \exp\!\left(-\frac{d_{\mathrm{obs}}(o^m(s,a),o)}{\kappa}\right).
\]

Particle filtering with \(K=10\) particles is used for belief propagation, and the score is the belief-based expected log-likelihood
\[
\mathcal{L}(P^m;\mathcal{D})
=
\sum_{\tau\in\mathcal{D}}\sum_{t=0}^{H_\tau-1}
\mathbb{E}_{s_{t+1}\sim Q_{t+1}^{m,\tau}}
\left[\log O^m(o_{t+1}^\tau\mid s_{t+1},a_t^\tau)\right].
\]
Model search proceeds by generate–evaluate–refine. Each round proposes \(M=5\) candidate programs, evaluates them by particle filtering, extracts diagnostics, and uses disagreement information from a query-by-committee signal
\[
\operatorname{VE}(s,a)
=
-\frac{1}{\log |\mathcal{C}_j|}
\sum_y
\frac{V_{s,a}(y)}{|\mathcal{C}_j|}
\log
\frac{V_{s,a}(y)}{|\mathcal{C}_j|},
\]
together with UCB1 parent selection
\[
\mathrm{UCB}(p)=\bar S(p)+c_{\mathrm{ucb}}
\sqrt{\frac{\log n_{\mathrm{parent}(p)}}{n_p+1}},
\qquad c_{\mathrm{ucb}}=1.0.
\]
Final selection samples from a near-best set with softmax temperature \(T=0.1\). Default settings include \(\kappa=0.2\), a planner budget of \(5{,}000\) belief-state expansions, and an entropy coefficient of \(1.0\).

Experiments are reported on MiniGrid tasks including Empty, Corners, Lava, Four Rooms, and Unlock, with an offline buffer of \(N=10\) mixed success and failure trajectories per environment. Pinductor matches the performance and sample efficiency of POMDP Coder despite using less information, and substantially outperforms a tabular POMDP baseline that is granted hidden-state access. Belief entropy decreases smoothly over episodes and posterior mass on the true latent state increases with observations. Performance also scales with LLM capability: on Lava, Qwen3 14B achieves \(0.07\pm 0.14\) mean reward with \(10\%\) win rate, whereas Qwen 3.6 Plus achieves \(0.59\pm 0.10\) with \(83\%\) win rate and Claude Opus 4.7 achieves \(0.61\pm 0.07\) with \(87\%\) win rate; on Unlock, the corresponding values are \(0.00\pm 0.00\) and \(0\%\), \(0.36\pm 0.14\) and \(53\%\), and \(0.37\pm 0.22\) and \(50\%\). Performance degrades when semantic information is withheld, indicating that the method depends jointly on language-model priors and trajectory feedback.

This usage is the clearest case in which “Pinductor” is a formal method name. It is unrelated to electromagnetic inductors except by analogy: the system “induces” a latent world model, and the paper explicitly positions language-model priors as a practical tool for sample-efficient world-model learning under partial observability.

Source: https://www.emergentmind.com/topics/pinductor