This paper presents a magnetostatic analysis of induction machines based on the Mesh-Based Generated Reluctance Network (MBGRN) approach. While MBGRN provides a systematic discretization comparable to numerical methods, its application to induction machines remains scarce. This work addresses this gap through two complementary modeling strategies: a pure MBGRN model formulated in mesh fluxes, and a Hybrid Analytical Model (HAM) in which the airgap is described analytically. Rotor motion is handled by Lagrange interpolation across the sliding line in the MBGRN model, and by analytical continuity of Maxwell’s equations in the HAM. The study is intentionally restricted to a magnetostatic framework to enable a controlled assessment of the magnetic models under both single- and double-excitation modes, independent of electrical and mechanical coupling. The double-excitation configuration is introduced as a controlled magnetostatic benchmark with simultaneously imposed stator and rotor magnetomotive forces. It is physically compatible with wound-rotor or doubly-fed induction-machine architectures, in which the rotor winding is electrically accessible, and does not represent the normal operating condition of a standard squirrel-cage induction motor, where rotor currents are induced by slip. Validation against Finite Element Method (FEM) simulations—covering airgap flux density, winding fluxes, and electromagnetic torque—yields relative errors below 1.2 % in single-excitation mode and below 2.9 % in double-excitation mode. The originality of this work lies in the extension and assessment of a systematically generated mesh-flux MBGRN formulation for induction machines, its coupling with an analytical airgap model, and the controlled FEM-based assessment of both formulations under different magnetostatic excitation conditions.
This paper presents the design, optimization, and experimental validation of a magnetic device for investigating electromagnetic field effects on human cells. The device employs paired sets of concentric current-carrying loops whose radii are optimized, through a constrained formulation solved by sequential quadratic programming, to maximize field strength at the midplane while ensuring uniformity over the sample region. The optimized 32-loop-pair configuration produces a flat-top axial field of 0.434 mT with a radial RMS non-uniformity of 2.94% across the target region, and a sensitivity study shows that the flat-top profile is retained under few-millimetre axial and gap misalignments, with a radial offset introducing a larger localized deviation. Measurements on a fabricated prototype reproduce the predicted normalized flat-top profile at the sampled positions within the target region, confirming the design as a magnetostatic proof of concept. The configuration is intended for prospective in vitro exposure of dermal and neural cells.
This paper extends the existing loss-surface hysteresis model (LSHM) to predict iron loss in non-oriented electrical steel under superimposed DC bias. The conventional three-dimensional (3D) LSHM accurately reconstructs B–H loops for non-sinusoidal flux-density waveforms with harmonics, but DC bias shifts B and H offsets and drives the loss-surface trajectory outside the LSHM domain, forcing extrapolation and causing significant estimation errors. A Full-4D extension with Bdc as an explicit dimension removes this limitation but requires very large 4D datasets and costly interpolation, greatly increasing memory usage and computation. To avoid this challenge, a harmonic-correction lookup table (HCLUT)-based Quasi-4D LSHM is proposed, which decomposes DC-bias-induced distortion into offset and harmonic amplitude/phase variations and applies a frequency-domain correction to the 3D LSHM output, significantly reducing model complexity and computational burden. Ring-core measurements on 25PNX1250F and 30PNF1600 electrical steel show that the Quasi-4D LSHM remains within 0.9% of the Full-4D reference while reducing computational cost by ≈40%. The approach therefore provides a practical, computationally efficient method for accurate iron loss prediction in high-performance non-oriented electrical steel with strong DC bias.
The variational formulation of Physics-Informed Neural Networks (PINNs) has recently gained significant attention as a promising approach for solving partial differential equations; however, the approach performs best when applied to regular domains, still showing difficulties for complex geometries. To address this limitation, we present and compare two advanced Variational PINN (VPINN) strategies designed to enhance geometric flexibility and computational robustness. The first approach is based on a VPINN framework employing quadrilateral local elements. The second approach relies on a non-conforming domain decomposition strategy, in which rectangular local elements are employed without the requirement of fully covering the computational domain. Their comparison highlights a trade-off between geometric accuracy and implementation efficiency. Numerical results on electromagnetic problems demonstrate that both approaches achieve accurate reconstruction of the scalar potential with stable convergence behavior, although their performance depends on the interplay between domain complexity, element distribution, and approximation properties. The implementation of the proposed method is publicly available at https://github.com/ShayanDodge/INI-VPINN.
Identifying stroke type is crucial because treatment differs between ischemic and hemorrhagic events. Electrical impedance tomography infers brain impedance distribution from bioimpedance changes at the scalp, offering a portable, radiation-free, real-time alternative to conventional imaging, but its reconstructions are degraded by the ill-conditioned inverse problem. To test whether machine learning can enhance bedside stroke classification, we trained a deep neural network (DNN) on multi-frequency EIT recordings from 22 stroke patients in a 33-subject clinical dataset. Under subject-wise five-fold cross-validation the model achieved an average accuracy of 87.8% in distinguishing ischemic from hemorrhagic strokes. A cohort of this size cannot fix the accuracy attainable by the modality, so this indicates feasibility rather than settled performance, but it suggests that a DNN-based approach may enable rapid stroke type assessment in clinical practice.
