MAGneT: Multifaceted Magnetic Innovations

• MAGneT is a multifaceted term that denotes distinct research domains including superconducting magnet R&D, deep-learning-enhanced magnetic reconstruction, spectral graph neural networks, and magnetostatic design software.
• Each usage applies domain-specific methodologies—from precision field control in accelerator magnets to complex Hermitian operators in directed graphs—highlighting unique engineering and computational challenges.
• These diverse applications advance practical solutions in accelerator experiments, electron tomography, graph-based data analysis, and neutron polarization analysis by leveraging core magnetic principles.

to=arxiv_search.search 天天中彩票为什么 手机上天天中彩票json {"query":"MAGneT OR MagNet superconducting magnet systems magnetic reconstruction directed graphs arXiv", "max_results": 10} to=arxiv_search.search 天天大奖彩票站? to=arxiv_search.search 《凤凰大参考json {"query":"MAGneT", "max_results": 10, "sort_by":"relevance"} MAGneT, or MagNet, is not a single technical object in the arXiv literature but a shared name used in several distinct research contexts centered on magnetism, magnetic-field computation, or mathematically “magnetic” operators. It denotes an integrated superconducting-magnet R&D program at KEK for precision, radiation-hard, and high-field accelerator applications; a deep-learning-enhanced framework for three-dimensional magnetic reconstruction in vector field electron tomography; a spectral graph neural network for directed graphs based on the magnetic Laplacian; and, in an earlier neutron-instrumentation study, the MagNet finite-element software used for magnetostatic field design. The term also sits against a broader background of large superconducting magnet systems in accelerators, detectors, and fusion devices (Ogitsu et al., 2022, Lyu et al., 2022, Zhang et al., 2021, Salhi et al., 2012, Védrine, 2015).

1. Terminology and scope

In the cited literature, the same label refers to technically unrelated entities. The commonality is not methodology but the centrality of magnetic fields, magnetic materials, or magnetic analogies in the underlying formalism.

Usage	Domain	Core content
MAGneT	Superconducting magnet R&D	KEK program on precision, rad-hard, and high-field magnets
MagNet	Electron tomography	3D U-Net corrector for VFET missing-wedge artefacts
MagNet	Graph machine learning	Spectral GNN using the magnetic Laplacian
MagNet software	Neutron instrumentation	Magnetostatic FEM for XYZ polarization analysis

A recurring source of confusion is the assumption that these usages are variants of the same framework. They are not. The KEK usage concerns superconducting hardware and cryogenic engineering; the tomography usage concerns learned post-reconstruction of magnetic induction fields; the graph-learning usage concerns complex Hermitian operators on directed graphs; and the TOPAS usage refers to a finite-element package employed in magnetostatic design (Ogitsu et al., 2022, Lyu et al., 2022, Zhang et al., 2021, Salhi et al., 2012).

2. Superconducting-magnet engineering usage

Within accelerator and fusion engineering, MAGneT appears as an integrated superconducting magnet R&D program at KEK’s Cryogenics Science Center. The program is organized around three goals: ppm-level 3D field control for precision muon experiments, radiation-hard magnets for high-intensity proton and muon facilities, and high-field accelerator magnets for future colliders. Its immediate application space includes the J-PARC muon g-2/EDM experiment, COMET, contributions to the LHC and HL-LHC, and longer-term development for FCC-hh (Ogitsu et al., 2022).

The precision-field branch targets a 3 T superconducting storage solenoid for the J-PARC muon g-2/EDM experiment with uniformity better than ±0.1 ppm across the cylindrical storage volume. The program uses a truncated singular value decomposition code for coil-position and coil-size optimization, commercial FEM tools for non-linear magnetic effects, passive shimming strategies derived from MRI technology, and ultra-high-precision CW-NMR magnetometry developed in US–Japan collaboration. In accelerator notation, transverse field quality is written as

$$B_x + i B_y = \sum_{n=1}^{\infty} (b_n + i a_n)\left[\frac{x+i y}{r_{\mathrm{ref}}}\right]^{n-1},$$

with normal and skew multipoles quantified around a reference radius, while uniformity is expressed by $\Delta B/B \lesssim 1\times 10^{-7}$ for the 3 T solenoid (Ogitsu et al., 2022).

The radiation-hard branch is driven by COMET. Its capture solenoid houses a 50 kW proton target, and the downstream transport system uses curved solenoids with superimposed dipole fields for charge and momentum selection. The materials program includes cyanate/bismaleimide-triazin-based GFRP qualified to about 100 MGy, neutron-irradiation studies of RRR degradation in high-purity Al and Cu stabilizers, and comparative irradiation tests on REBCO tapes with mineral insulation for a proposed conduction-cooled muon production solenoid operating at about 20 K (Ogitsu et al., 2022).

The high-field branch is tied to KEK’s LHC and HL-LHC magnet lineage. MQXA reached the highest conductor field of 8.6 T, while the HL-LHC D1 beam-separation dipole has a peak field of 6.6 T and coil stresses of about 100 MPa. KEK reports Nb3Sn strands at 1100 A/mm$^2$ at 16 T, with 5 km-scale test production, and targets a 12 T, 100 mm aperture D1-like magnet for FCC-hh, followed by hybrid 16–20 T accelerator magnets combining about 12 T Nb3Sn coils with about 4–8 T HTS inserts (Ogitsu et al., 2022).

This program sits within the broader regime of large superconducting magnet systems. Nb-Ti remains the workhorse low-temperature superconductor, with $T_c \approx 9.2$ K and usable accelerator dipole fields in the $\sim 8$–10 T class when operated at 1.9–4.5 K; the LHC reached 8.33 T at 1.9 K. Nb$_3$Sn, with $T_c \approx 18$ K and $B_{c2} \approx 25$–30 T at 4.2 K, opens the 13–20 T regime. HTS conductors such as REBCO and Bi-2212 are positioned for $>20$ T systems, but with engineering emphasis on strain tolerance, cabling, joining, and quench detection. In this broader literature, canonical design relations include $B\rho = p/q$ for beam bending, $\Delta B/B \lesssim 1\times 10^{-7}$0 for stored energy, $\Delta B/B \lesssim 1\times 10^{-7}$1 for Lorentz-force density, and $\Delta B/B \lesssim 1\times 10^{-7}$2 for electromagnetic pressure (Védrine, 2015).

The representative systems defining this background include the Tevatron, HERA, RHIC, and LHC in accelerators; the CMS 4 T solenoid with stored energy of about 2.7 GJ and the ATLAS solenoid-toroid system in detectors; and fusion systems such as Tore Supra, EAST, KSTAR, JT-60SA, W7-X, and ITER, where central plasma fields span about 2.68–5.3 T and peak coil fields extend to about 12.8 T in the ITER central solenoid (Védrine, 2015). A plausible implication is that KEK’s MAGneT program should be read not as a standalone hardware platform but as a coordinated R&D layer built on the mature design principles of large superconducting magnet systems.

3. Deep-learning-enhanced magnetic reconstruction

In electron tomography, MagNet denotes a framework for three-dimensional magnetic reconstruction under limited-angle acquisition. It addresses the missing-wedge artefact in vector field electron tomography, where restricted specimen tilt leaves a wedge-shaped region of unmeasured spatial frequencies in Fourier space and produces anisotropic resolution, elongation along the beam axis, ringing, and nonuniqueness in the inverse problem. The method couples a conventional VFET pipeline to a learned 3D post-reconstruction corrector based on a U-shaped convolutional neural network (Lyu et al., 2022).

The physical forward model begins with Lorentz TEM phase retrieval through electron holography or transport-of-intensity. In the line-integral formulation,

$\Delta B/B \lesssim 1\times 10^{-7}$3

and, in the specific VFET module used in the study,

$\Delta B/B \lesssim 1\times 10^{-7}$4

Two orthogonal tilt series reconstruct $\Delta B/B \lesssim 1\times 10^{-7}$5 and $\Delta B/B \lesssim 1\times 10^{-7}$6 by scalar tomography with k-space bilinear interpolation, after which $\Delta B/B \lesssim 1\times 10^{-7}$7 is obtained by enforcing $\Delta B/B \lesssim 1\times 10^{-7}$8 (Lyu et al., 2022).

MagNet inserts a learned correction stage after this defective reconstruction. The pipeline is: acquisition of limited-angle phase shifts for orthogonal tilt series; VFET reconstruction of an artefact-contaminated $\Delta B/B \lesssim 1\times 10^{-7}$9; and passage of this three-channel volume through a 3D U-Net that outputs an enhanced $^2$0. The network inherits the 3D-Unet skeleton, uses skip connections, and begins with a convolutional layer producing 64 output feature maps with ReLU activation. It takes a three-channel $^2$1 volume and returns a three-channel volume of the same shape. Physics is not imposed through explicit Maxwell-loss terms; instead, it enters implicitly through the VFET pre-reconstruction, cross-component filtering, and priors learned from micromagnetic simulations (Lyu et al., 2022).

Training data were generated with JuMag.jl by solving Landau–Lifshitz–Gilbert dynamics for cylindrical samples of radius 40 pixels and thickness between 10 and 80 pixels. The texture library includes vortices, cylindrical domains, single and lattice skyrmions, spin helices, conical phases, Néel domain structures, and single-domain states. The total library size is 210 volumes, split into 150 training and 60 test samples. The field of view is $^2$2 pixels per dimension, and losses are computed only inside the material mask. Training uses voxel-wise MSE, Adam with learning rate $^2$3, Keras, a single NVIDIA Tesla V100-SXM2-16GB GPU, about 329 seconds per epoch, and 100 epochs (Lyu et al., 2022).

Across the 60-volume test set and five tilt limits, the reported median NRMSE values are as follows.

Tilt limit	VFET median NRMSE	MagNet median NRMSE
$^2$4	49.6%	29.0%
$^2$5	36.3%	19.2%
$^2$6	27.6%	14.2%
$^2$7	15.0%	10.7%
$^2$8	5.2%	9.5%

These results show that MagNet consistently outperforms conventional VFET for incomplete tilt series, while VFET is superior for complete tilt at $^2$9, which the study interprets as expected behavior for a corrector trained specifically on missing-wedge degradation (Lyu et al., 2022).

The paper’s detailed skyrmion example makes the improvement concrete. For a Bloch-type skyrmion in a 400 nm diameter, 130 nm thick disk under $T_c \approx 9.2$0 tilt, MagNet reduces slice NRMSE on $T_c \approx 9.2$1 from 17.51% to 11.34% for $T_c \approx 9.2$2, 34.75% to 12.85% for $T_c \approx 9.2$3, and 15.80% to 6.80% for $T_c \approx 9.2$4. On the $T_c \approx 9.2$5 nm layer, it reconstructs the surface Néel cap that VFET$T_c \approx 9.2$6 misses, cutting NRMSE from 39.77% to 15.13% for $T_c \approx 9.2$7 and from 39.73% to 12.50% for $T_c \approx 9.2$8. Under 2% Gaussian noise in the phase shifts, VFET$T_c \approx 9.2$9 yields 16.84%, 16.84%, and 26.57% for $\sim 8$0, $\sim 8$1, and $\sim 8$2, whereas MagNet$\sim 8$3 improves these to 8.97%, 6.47%, and 9.15% (Lyu et al., 2022).

A common misconception is that MagNet is a physics-informed network in the strict sense. The paper explicitly states that it does not impose Maxwell constraints such as $\sim 8$4 as training losses. Its mechanism is instead a data-driven, component-coupled k-space filtering learned across $\sim 8$5, $\sim 8$6, and $\sim 8$7 (Lyu et al., 2022).

4. Spectral graph learning on directed graphs

In graph machine learning, MagNet is a spectral GNN for directed graphs built on the magnetic Laplacian. Its starting point is the observation that most GNNs were developed for undirected graphs, whereas directed data such as citation, website, traffic, and influence networks lose information under symmetrization and become difficult to treat spectrally because ordinary real symmetric Laplacians do not exist for general directed graphs (Zhang et al., 2021).

The construction begins from a directed graph $\sim 8$8 with asymmetric adjacency matrix $\sim 8$9. One defines the symmetrized adjacency

$_3$0

the diagonal degree matrix $_3$1, and a charge-dependent phase

$_3$2

The complex Hermitian adjacency is then

$_3$3

and the normalized magnetic Laplacian is

$_3$4

Because $_3$5 is skew-symmetric, $_3$6 and therefore $_3$7 are Hermitian. The paper proves that both normalized and unnormalized magnetic Laplacians are positive semidefinite and that the eigenvalues of $_3$8 lie in $_3$9 (Zhang et al., 2021).

These properties restore a Fourier-like spectral theory on directed graphs. If $T_c \approx 18$0, then the directed Fourier transform is $T_c \approx 18$1 and inversion is $T_c \approx 18$2. Spectral convolution is diagonal multiplication in this basis, and MagNet adopts a Chebyshev polynomial parameterization to avoid explicit eigendecomposition:

$T_c \approx 18$3

In the experiments, the model uses $T_c \approx 18$4, so the architecture is ChebNet-like; with the usual GCN renormalization and $T_c \approx 18$5, it reduces to the standard undirected case (Zhang et al., 2021).

Unlike real-valued GNNs, MagNet uses complex-valued filters and feature maps, together with a complex ReLU,

$T_c \approx 18$6

After the final convolutional layer, the complex embedding is “unwound” into concatenated real and imaginary parts before a final real linear layer and softmax. For link prediction, ordered-pair features are built by concatenating the corresponding node embeddings (Zhang et al., 2021).

The charge parameter $T_c \approx 18$7, typically cross-validated in $T_c \approx 18$8, tunes the extent to which directionality affects the spectrum, especially through directed cycles. The paper’s examples make this explicit: for directed cycles, the eigenvectors remain classical Fourier modes $T_c \approx 18$9, while the eigenvalues shift as

$B_{c2} \approx 25$0

Thus, $B_{c2} \approx 25$1 is not a superficial hyperparameter but part of the spectral encoding of directed motifs (Zhang et al., 2021).

Empirically, the paper reports strong performance on both node classification and link prediction. On node classification, MagNet attains 84.3±7.0 on Cornell, 83.3±6.1 on Texas, 85.7±3.2 on Wisconsin, 79.8±2.5 on Cora-ML, 67.5±1.8 on CiteSeer, and 87.6±2.9 on Telegram; on synthetic DSBM graphs it reports 99.8±0.1, 98.4±0.5, and 93.8±1.5 on the three main settings. On link prediction, it is often the top performer, with examples including 80.7±2.7 on Cornell, 79.5±3.7 on Texas, 83.6±2.8 on Wisconsin, 86.1±0.9 on Cora-ML, and 85.1±0.8 on CiteSeer for direction prediction. The study notes that $B_{c2} \approx 25$2 is optimal for citation-graph node classification, where directionality acts as noise, but nonzero $B_{c2} \approx 25$3 is optimal for link prediction on those same graphs, where edge direction is structurally informative (Zhang et al., 2021).

A common misunderstanding is that MagNet merely aggregates over out-neighbors with complex weights. The paper’s claim is stronger: polynomial filters based on the magnetic Laplacian aggregate over both nodes reachable from a node and nodes that can reach it within $B_{c2} \approx 25$4 hops, with the phase factors allowing those two sets to be treated differently (Zhang et al., 2021).

5. Magnetostatic design software for neutron polarization analysis

In neutron instrumentation, MagNet refers to magnetostatic finite-element software used to design the magnetic environment for XYZ polarization analysis on the TOPAS thermal time-of-flight spectrometer. The study models a PASTIS-based coil insert with a wide-angle banana-shaped $B_{c2} \approx 25$5He neutron spin filter cell, whose off-center geometry makes full three-dimensional field design necessary (Salhi et al., 2012).

The modeled configuration comprises three orthogonal Helmholtz coil pairs aligned with the instrument’s $B_{c2} \approx 25$6, $B_{c2} \approx 25$7, and $B_{c2} \approx 25$8 axes, centered on the sample, and augmented by mu-metal sheets introduced to enlarge the homogeneous region and reduce blind areas from coil supports. The $B_{c2} \approx 25$9He cell is positioned significantly off-center but must still remain inside a field region homogeneous enough to preserve both neutron-spin transport and $>20$0He nuclear polarization (Salhi et al., 2012).

The magnetostatic framework is standard:

$>20$1

The design target is a holding field in the millitesla range. The MagNet maps in the study show $>20$2–1.25 mT for the $>20$3 case and about 1.067–1.078 mT for the $>20$4 case. Homogeneity is expressed by $>20$5 and by gradient constraints tied to the $>20$6He relaxation time,

$>20$7

For $>20$8 bar, the criterion $>20$9 yields $B\rho = p/q$0 h, which the paper adopts as a practical requirement (Salhi et al., 2012).

The reported result is that the inner part of the field is highly homogeneous, including the position of the $B\rho = p/q$1He cell, and that the geometry supports a nearly $B\rho = p/q$2 vertical scattering window matching the detector-bank height. The use of mu-metal sheets is central: they act as passive shims and flux guides, expanding the homogeneous region while suppressing fringe gradients that could otherwise depolarize neutron spins or shorten $B\rho = p/q$3He $B\rho = p/q$4 (Salhi et al., 2012).

Here, MagNet is neither a neural network nor a hardware program. It is a field-solving software environment used in 3D magnetostatic analysis. This older usage is therefore conceptually separate from both the VFET corrector and the directed-graph GNN, even though all three share the same name (Salhi et al., 2012).

6. Comparative interpretation and recurrent themes

Across these usages, MAGneT/MagNet functions as a homonym spanning hardware R&D, inverse problems, graph signal processing, and FEM-based instrument design. The shared lexical choice reflects a common connection to “magnetic” structure, but the operative objects are different in each case: superconducting conductors and cryogenic hardware in the KEK program; reconstructed induction fields $B\rho = p/q$5 in electron tomography; complex Hermitian Laplacians in directed-graph learning; and magnetostatic field maps in neutron instrumentation (Ogitsu et al., 2022, Lyu et al., 2022, Zhang et al., 2021, Salhi et al., 2012).

The mathematical commonalities are limited but nontrivial. The tomography and neutron-instrumentation usages are both anchored in Maxwell-type relations, including $B\rho = p/q$6, though one appears inside a reconstruction pipeline and the other inside FEM design. The superconducting-magnet literature emphasizes energy, force, and protection relations such as $B\rho = p/q$7, $B\rho = p/q$8, and $B\rho = p/q$9, whereas the graph-learning usage imports magnetic intuition only through phase factors and Hermitian structure, not through physical fields (Védrine, 2015, Lyu et al., 2022, Zhang et al., 2021, Salhi et al., 2012).

Their engineering constraints also differ sharply. KEK’s MAGneT program is constrained by ppm field uniformity, MGy radiation hardness, 100 MPa-class coil stress, MIITs-based protection, and the transition from Nb-Ti to Nb$\Delta B/B \lesssim 1\times 10^{-7}$00Sn and HTS conductors (Ogitsu et al., 2022). MagNet for VFET is constrained by missing wedges, limited tilt ranges of $\Delta B/B \lesssim 1\times 10^{-7}$01 to $\Delta B/B \lesssim 1\times 10^{-7}$02, finite simulated texture diversity, and the absence of explicit physics losses (Lyu et al., 2022). MagNet for directed graphs is constrained by spectral scalability, the choice of charge parameter $\Delta B/B \lesssim 1\times 10^{-7}$03, and the fact that it is not designed for very large graphs without additional approximation strategies (Zhang et al., 2021). The TOPAS MagNet study is constrained by wide-angle geometric coverage, $\Delta B/B \lesssim 1\times 10^{-7}$04He depolarization limits, and field homogeneity over a strongly off-center volume (Salhi et al., 2012).

The most defensible encyclopedia-level interpretation is therefore disambiguation rather than unification. In current arXiv usage, MAGneT is best understood as a reused research name whose meaning is entirely domain-dependent: a superconducting magnet R&D program in accelerator science, a 3D reconstruction framework in magnetic microscopy, a spectral GNN in directed-graph learning, or a magnetostatic FEM tool in neutron-scattering instrument design (Ogitsu et al., 2022, Lyu et al., 2022, Zhang et al., 2021, Salhi et al., 2012).