Pith. sign in

REVIEW 5 major objections 4 minor 31 references

The magnetic behavior of a synchronous machine reduces to a single learned scalar energy field, whose gradients give current and torque exactly.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A gradient-network model trained on sparse flux-linkage/current data reproduces the saturable, angle-periodic magnetic maps of a 5.6-kW synchronous machine while enforcing reciprocity, convexity, and smoothness by construction.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A clean, modest adaptation of gradient networks to synchronous-machine magnetic modeling with genuinely good data efficiency, but the abstract overclaims extrapolation and an embedded demo that the paper does not actually contain. the 5 major comments →

arxiv 2602.14947 v3 pith:RCK55FZN submitted 2026-02-16 eess.SY cs.LGcs.SY

Gradient Networks for Universal Magnetic Modeling of Synchronous Machines

Junyi Li , Tim Foissner , Floran Martin , Antti Piippo , Marko Hinkkanen This is my paper
classification eess.SY cs.LGcs.SY
keywords gradient networksmagnetic saturationsynchronous machinesphysics-informed machine learningreciprocityenergy balancespatial harmonicselectric drives
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes that the magnetic subsystem of a synchronous machine—the nonlinear relationship among flux linkage, current, rotor angle, and torque—can be represented exactly as the gradient of a single scalar magnetic energy field. It introduces neural networks that output gradients directly, so current and torque are computed as partial derivatives of learned energy, guaranteeing energy balance and reciprocity by construction. Because the learned energy is convex, the current–flux map is monotone and invertible, giving smooth, physically plausible behavior even when trained on a small fraction of the available data. On measured and finite-element data from a 5.6-kW PM synchronous reluctance machine, the approach reports flux and torque errors around 0.01 per unit with 10% of training data, still works with 2–0.2% of the data, and runs in real time on an embedded controller. If correct, this gives a universal, data-efficient substitute for lookup tables and black-box networks in electric-drive modeling and control.

Core claim

Central claim: a saturable synchronous machine can be modeled as a lossless, conservative magnetic system in which stator current and electromagnetic torque are partial derivatives of one scalar field energy W(ψd, ψq, θm). The paper implements this with a gradient network g(x) = A0x + b0 + Aᵀσ(Ax+b) whose Jacobian is symmetric positive-semidefinite by construction, so g is the gradient of a convex energy. Fourier features in the rotor angle add spatial harmonics while preserving reciprocity and energy balance; convexity guarantees a unique invertible current–flux map. Validated on measured and FEM data from a 5.6-kW PM synchronous reluctance machine, the models give rms flux and torque error

What carries the argument

The gradient network g(x) = A0x + b0 + Aᵀσ(Ax+b), where A0 is symmetric positive semidefinite and σ is chosen so that its Jacobian is symmetric positive semidefinite (elementwise non-decreasing activations such as squareplus or algebraic sigmoid, or vector activations such as softmax or the proposed p-norm gradient). Its Jacobian A0 + AᵀJσ(z)A is automatically symmetric positive semidefinite, meaning g is the gradient of a convex scalar field energy W(x). The machine model takes x = [ψd, ψq] (plus Fourier features cos(kθm), sin(kθm) for spatial harmonics) and reads the current as g(x), with torque assembled from g and the angle derivatives. This architecture carries the argument: it converts

Load-bearing premise

The machine's magnetic field must be lossless and conservative—there must be a single scalar energy whose gradients give both current and torque—and the incremental inductance must be positive-definite; if iron losses, hysteresis, or eddy currents are significant, the architecture cannot represent them.

What would settle it

Measure the 2x2 incremental inductance matrix ∂(id,iq)/∂(ψd,ψq) on a real machine and test symmetry: if ∂id/∂ψq ≠ ∂iq/∂ψd at any operating point, no conservative energy function exists, and the gradient-network premise fails. A simpler proxy: drive the machine with a periodic flux-linkage cycle and check whether the closed-loop integral of i·dψ (the dissipated energy) is zero; a nonzero area shows non-conservative behavior the model cannot encode.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • The same architecture extends to any synchronous machine type and likely to other electromagnetic actuators, since it only assumes a lossless conservative magnetic model.
  • Control algorithms can invert the model reliably: convexity guarantees a unique flux-linkage for each current (and vice versa), so flux-vector control and optimal trajectories (MTPA, MTPV) can be computed without iterative solvers.
  • The smoothness of the learned surfaces removes the numerical noise that linear-interpolated lookup tables introduce, which matters for gradient-based optimization of drives.
  • Data requirements drop by orders of magnitude compared with black-box neural networks, since the physics is embedded rather than learned.
  • Real-time closed-loop operation on an embedded control platform indicates the models are cheap enough to run inside a drive.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The assumed losslessness means the model deliberately excludes iron losses, hysteresis, and eddy currents; if a drive model must capture these, the conservative-gradient structure would need extension (e.g., adding a dissipative term) rather than more layers.
  • The universal-approximation claim is relative to the class of monotone conservative fields; that class is exactly the set of physically feasible magnetostatic models, but not all data-driven magnetic maps with losses. So 'any physically feasible' should be read as 'any feasible within the lossless magnetostatic model.'
  • The p-norm gradient activation is a practical engineering contribution: it gives comparable accuracy to softmax at lower computational cost, because the fractional power is computed once per forward pass regardless of network width, which matters for embedded implementation.
  • A natural testable extension: apply the same architecture to a machine with significant iron loss or a multi-phase machine and check whether reciprocity still holds in measured data; if not, a dissipative extension would be needed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper proposes a physics-constrained neural-network architecture for modeling the magnetic characteristics of synchronous machines. The stator-current map and electromagnetic torque are represented as gradients of a learned scalar field energy, using "gradient networks" that enforce a symmetric positive-semidefinite Jacobian (reciprocity) and, when monotone, a convex energy leading to an invertible flux-current relation. Fourier features are added to capture rotor-angle spatial harmonics while preserving the gradient structure. The models are trained on measured and FEM data from a 5.6-kW PM synchronous reluctance machine, with reported RMS flux-linkage and torque errors on the order of 0.004–0.023 p.u. using 0.2–10% of the available grid data. The paper also describes application examples for simulation, flux-vector control, and MTPA/MTPV locus computation.

Significance. If the empirical claims are fully supported, the architecture is a useful contribution: it gives a smooth, invertible, physically consistent magnetic model with a small parameter count, and it reduces the data requirement relative to black-box approaches. The derivation from conservative magnetostatics is sound, the reciprocity and convexity-by-construction properties are mathematically clear, and the measured-dataset link is a practical asset. The universal-approximation claim, however, needs qualification, and several advertised advantages (extrapolation, data efficiency, superiority to black-box networks, real-time embedded demonstration) are not actually demonstrated in the present manuscript. The core idea is promising and likely publishable after strengthening the empirical validation.

major comments (5)
  1. [Abstract; Section IV] The abstract claims the method is “further demonstrated experimentally in real-time closed-loop operation on an embedded control platform,” but Section IV contains no such experiment. Figure 10 is a simulation trace, not a real-time embedded closed-loop result. Either add the experiment or remove this claim; as written it is unsupported.
  2. [Section I and Section IV (extrapolation claim)] The claim of “reliable extrapolation” is not tested. In all experiments, the training subsets (10%, 2%, 0.2%) are uniformly subsampled over the full measured/FEM grid, so every validation point lies inside the training envelope. Strong convexity guarantees monotonicity and invertibility but not accuracy outside the sampled region. The authors should train on a restricted envelope (e.g., current/flux magnitudes up to 1.2 p.u.) and test up to the full 2.4 p.u. range, comparing with a black-box MLP and a lookup table. This is load-bearing because extrapolation is a central advertised advantage.
  3. [Section IV, Tables II–IV] No comparison is made to black-box neural networks or lookup-table baselines. The abstract and introduction claim that the proposed model “requires less training data” and “offers advantages over black-box networks,” but Tables II–IV only compare activation variants of the proposed architecture. Identical training/validation splits should be used to report errors for a standard MLP and a linearly interpolated lookup table, otherwise the data-efficiency and generalization claims cannot be assessed.
  4. [Section IV, Tables II–IV] All numerical results are single-run point estimates with no variance. This is particularly concerning in Table II, where the 2% case uses only 12 training samples for a 41-parameter model; a different uniform subsample could change the reported errors substantially. The small differences between activations (e.g., softmax vs. p-norm at 0.2% in Table IV) may be within run-to-run variation. Report mean and standard deviation over multiple seeds or multiple subsampling replicates.
  5. [Abstract; Sections II and III-A] The phrase “universally approximate any physically feasible magnetic behavior” is too broad. The model assumes a lossless conservative system (Section II), and the authors themselves state in Section III-A2 that elementwise activations cannot universally approximate all monotone conservative fields. The universal-approximation property is inherited from [26] for the softmax vector activation, but it is not established for the p-norm gradient activation (14) used in the experiments. The claim should be qualified to “with the softmax vector activation, within lossless conservative magnetostatic models,” or a proof/reference for p-norm universality should be supplied.
minor comments (4)
  1. [Figure 9 caption] The caption says “the full measured dataset,” but the experiment in Section IV-B uses the FEM dataset. Replace “measured” with “FEM.”
  2. [Section IV-A1 / Table II] The relation between “2% of the full dataset” and “only 12 training samples per map” is not obvious because the displayed grid has 21×27 = 567 points while the unique measured dataset has 21×13 = 273 points. Clarify which grid the percentages refer to.
  3. [Section III-A3] For the p-norm activation (14), the claim that S(z) is convex and hence the activation is monotone is stated without proof. A one-line justification (norm convexity after the constant is absorbed) would help readers verify the property.
  4. [Section IV-B4 / Figure 11] The MTPA/MTPV loci are computed from a model without spatial harmonics while the simulation model includes them. State explicitly whether this inconsistency is intentional (e.g., because the control model deliberately neglects harmonics) to avoid confusion.

Circularity Check

0 steps flagged

No significant circularity: the physics constraints and universal-approximation theorem are external, and the reported accuracy is ordinary interpolation-quality regression.

full rationale

The derivation chain is not circular. The gradient/energy-balance identities (current and torque as gradients of field energy, Eqs. (2), (5), (10)) come from external textbooks [1], [2], and the universal-approximation capability of monotone gradient networks is imported from the external paper [26]. The conservation/convexity structure is an architectural constraint imposed on the learned map, not a quantity fitted to the validation data; therefore the held-out rms errors on the dense measured/FEM grids are genuine measures of interpolation accuracy. The author-group self-citations ([4], [20], [29]) are contextual and non-load-bearing: [20] is cited for identification routines and for p=6/8 exponent choices, and [29] for the flux-vector control example. None supports the central universal-approximation or energy-balance claims. Two non-circular caveats should be flagged: (i) the 'physically feasible' universality claim is explicitly conditional on a lossless conservative magnetic subsystem (Section II), so iron losses and hysteresis are outside the representation; (ii) the advertised 'reliable extrapolation' (Abstract, Section I) is not actually tested—the training subsets in Section IV are uniformly spread over the full operating grid, so validation points are interior interpolations, and convexity guarantees monotonicity/invertibility but not extrapolation accuracy. These are scope/correctness risks, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central model is a neural network fitted to data; the only physical content is the gradient/convexity constraint inherited from prior electromechanical theory and the universal-approximation theorem from [26]. No new physical entities are introduced; the implicit energy function is a mathematical latent. The hand-chosen hyperparameters (N, p, training fractions, Fourier order k) are the main freedom not justified by data.

free parameters (5)
  • Gradient network weights A, b, A0, b0 = 3N+5 parameters with N=12 (no harmonics); 5N+7 with N=48 (with harmonics)
    Trained on measured or FEM data via MSE loss; they define the learned map and are the model itself.
  • Activation shape parameter beta = learned; values not reported
    Appears in Eqs. (11)-(14); controls softness of saturation and is trained alongside the weights.
  • p-norm exponent p = 6 or 8
    Hand-chosen positive even integer matching exponents used in prior iron-saturation models [3],[4],[20]; not learned from data.
  • Hidden unit count N = 12 (without harmonics), 48 (with harmonics)
    Chosen by the authors without a reported selection criterion or sensitivity study.
  • Strong convexity coefficients mu_d, mu_q = not reported
    Diagonal entries of A0 enforcing strong monotonicity; not stated whether learned or fixed.
axioms (5)
  • domain assumption Lossless conservative magnetic field: currents and torque are gradients of a scalar field energy W(ψs, θm) (Eq. (2)).
    Standard magnetostatic assumption for synchronous-machine models, cited to [1],[2]; ignores iron losses, hysteresis, and eddy currents.
  • domain assumption Strict convexity/monotonicity of the current map in flux linkage (positive-definite incremental inductance).
    Section II-B1; needed for existence of unique inverse maps and robust model inversion.
  • standard math Universal approximation theorem for monotone gradient networks with softmax vector activation.
    Inherited from [26], not proved in this paper; supports the universality claim.
  • domain assumption Known electrical periodicity of spatial harmonics (60-degree period, k=6) for the example machine.
    Section IV-B1; the Fourier feature choice relies on this machine-specific symmetry.
  • standard math Coordinate transformation and chain rule used in Appendix A.
    Standard matrix-exponential rotation and calculus; no independent evidence needed.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Gradient Networks for Universal Magnetic Modeling of Synchronous Machines." pith.science (2026). https://pith.science/paper/RCK55FZN

@misc{pith2026260214947,
  author       = {Pith},
  title        = {Pith review of: Gradient Networks for Universal Magnetic Modeling of Synchronous Machines},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RCK55FZN}},
  note         = {Machine review of arXiv:2602.14947}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This paper presents a physics-constrained neural network framework for dynamic modeling of saturable synchronous machines, including spatial harmonics. The proposed architecture embeds gradient networks directly into the fundamental machine equations to model nonlinear, coupled electromagnetic behavior. By learning the gradient of magnetic field energy, the model satisfies reciprocity and energy-balance constraints by construction. The approach can universally approximate any physically feasible magnetic characteristics while offering key advantages over lookup tables and conventional black-box networks: monotonicity, smooth outputs, and improved generalization from limited data. These properties also support robust model inversion and trajectory optimization for control. The method is validated using measured and finite-element-method (FEM) data from a 5.6-kW permanent-magnet (PM) synchronous reluctance machine and is further demonstrated experimentally in real-time closed-loop operation on an embedded control platform. The results show accurate and physically consistent modeling performance, even with limited training data.

Figures

Figures reproduced from arXiv: 2602.14947 by Antti Piippo, Floran Martin, Junyi Li, Marko Hinkkanen, Tim Foissner.

Figure 1
Figure 1. Figure 1: Electromechanical dynamics of a generic synchronous machine: (a) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Gradient network used for the magnetic models. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Proposed magnetic models in rotor coordinates: (a) without spatial [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Vector activation 𝜎1 (𝑧1, 𝑧2 ) visualized in two-dimensional case: (a) softmax (13) with 𝛽 = 1; (b) 𝑝-norm gradient (14) with 𝑝 = 4 and 𝛽 = 1. activation types are used for the current and flux-linkage maps. For simplicity, we employ the same activation function for all hidden units within each network. For modeling current maps, rectifier-type activations, such as softplus or algebraic squareplus, can be … view at source ↗
Figure 6
Figure 6. Figure 6: Test bench including a 5.6-kW PM synchronous reluctance machine [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Flux-linkage maps: (a) 𝜓d (𝑖d, 𝑖q ); (b) 𝜓q (𝑖d, 𝑖q ). The surfaces show the predicted maps from the 𝑝-norm gradient model (14) with 𝑝 = 8 and 𝑁 = 12 hidden units, trained on a 10% subset. Markers show the same measured dataset as in [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Electromagnetic torque from the softmax model with [PITH_FULL_IMAGE:figures/full_fig_p007_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: Optimal control loci computed from the proposed model without [PITH_FULL_IMAGE:figures/full_fig_p008_11.png] view at source ↗
Figure 10
Figure 10. Figure 10: Simulation example showing acceleration from standstill to 1 p.u. The [PITH_FULL_IMAGE:figures/full_fig_p008_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 4 linked inside Pith

  1. [1]

    H. H. Woodson and J. R. Melcher,Electromechanical Dynamics. John Wiley & Sons, 1968. [Online]. Available: https://ocw.mit.edu/ans7870/ resources/woodson/textbook/emd part1.pdf

    1968

  2. [2]

    A. E. Fitzgerald, C. Kingsley, Jr., and S. D. Umans,Electric Machinery, 6th ed. Boston, MA: McGraw-Hill, 2003

    2003

  3. [3]

    Die Nachbildung von Magnetisierungskurven durch einfache algebraische oder transzendente Funktionen,

    J. Fischer and U. Moser, “Die Nachbildung von Magnetisierungskurven durch einfache algebraische oder transzendente Funktionen,”Arch. f ¨ur Electrotechnik, vol. 42, no. 5, pp. 286–299, 1956

    1956

  4. [4]

    Inclusion of magnetic saturation in dynamic models of synchronous reluctance motors,

    Z. Qu, T. Tuovinen, and M. Hinkkanen, “Inclusion of magnetic saturation in dynamic models of synchronous reluctance motors,” inProc. Int. Conf. Electrical Machines, Marseille, France, Sep. 2012, pp. 994–1000

    2012

  5. [5]

    Analytical modeling and simulation of highly utilized electrical machines considering nonlinear effects,

    M. G ¨artner, F. Schuller, N. Parspour, and P. Seibold, “Analytical modeling and simulation of highly utilized electrical machines considering nonlinear effects,” inProc. Int. Conf. Elect. Machines, Marseille, France, 2012, pp. 2786–2791

    2012

  6. [6]

    Flux maps spatial harmonic modeling and measurement in synchronous reluctance motors,

    A. Varatharajan, P. Pescetto, S. Ferrari, and G. Pellegrino, “Flux maps spatial harmonic modeling and measurement in synchronous reluctance motors,” inProc. IEEE ECCE, Nashville, TN, Oct./Nov. 2023, pp. 4868– 4873

    2023

  7. [7]

    A saturation model based on a simplified equivalent magnetic circuit for permanent magnet machines,

    F. Lelli, M. Hinkkanen, and F. G. Capponi, “A saturation model based on a simplified equivalent magnetic circuit for permanent magnet machines,” inProc. Int. Conf. Elect. Machines, Torino, Italy, Sep. 2024

    2024

  8. [8]

    Flux- observer-based high-performance control of synchronous reluctance mo- tors by including cross saturation,

    A. Vagati, M. Pastorelli, G. Franceschini, and V. Drogoreanu, “Flux- observer-based high-performance control of synchronous reluctance mo- tors by including cross saturation,”IEEE Trans. Ind. Appl., vol. 35, no. 3, pp. 597–605, May/Jun. 1999

    1999

  9. [9]

    A high-fidelity and computationally efficient model for interior permanent-magnet machines considering the magnetic saturation, spatial harmonics, and iron loss effect,

    X. Chen, J. Wang, B. Sen, P. Lazari, and T. Sun, “A high-fidelity and computationally efficient model for interior permanent-magnet machines considering the magnetic saturation, spatial harmonics, and iron loss effect,”IEEE Trans. Ind. Electron., vol. 62, no. 7, pp. 4044–4055, 2015

    2015

  10. [10]

    Modeling of interior permanent magnet machine considering saturation, cross coupling, spatial harmonics, and temperature effects,

    S. Li, D. Han, and B. Sarlioglu, “Modeling of interior permanent magnet machine considering saturation, cross coupling, spatial harmonics, and temperature effects,”IEEE Trans. Transport. Electrific., vol. 3, no. 3, pp. 682–693, 2017

    2017

  11. [11]

    Identification of IPMSM flux-linkage map for high-accuracy simulation of IPMSM drives,

    J. Lee, Y.-C. Kwon, and S.-K. Sul, “Identification of IPMSM flux-linkage map for high-accuracy simulation of IPMSM drives,”IEEE Trans. Power Electron., vol. 36, no. 12, pp. 14 257–14 266, 2021

    2021

  12. [12]

    The dq-theta flux map model of synchronous machines,

    S. Ferrari, G. Dilevrano, P. Ragazzo, and G. Pellegrino, “The dq-theta flux map model of synchronous machines,” inProc. IEEE ECCE, Vancouver, Canada, Oct. 2021, pp. 3716–3723

    2021

  13. [13]

    Experimental identification of the dq𝜃flux maps of synchronous machines,

    A. Bojoi, P. Pescetto, S. Ferrari, and G. Pellegrino, “Experimental identification of the dq𝜃flux maps of synchronous machines,” inProc. Int. Conf. Elect. Machines, Turin, Italy, Sep. 2024

    2024

  14. [14]

    Sensorless speed control of synchronous reluctance motor drives based on extended kalman filter and neural magnetic model,

    D. Pasqualotto, S. Rigon, and M. Zigliotto, “Sensorless speed control of synchronous reluctance motor drives based on extended kalman filter and neural magnetic model,”IEEE Trans. Ind. Electron., vol. 70, no. 2, pp. 1321–1330, Feb. 2023

    2023

  15. [15]

    RNN-based high fidelity permanent magnet synchronous motor emulator considering driving inverter switching faults,

    H. Mohajerani, U. Deshpande, and N. C. Kar, “RNN-based high fidelity permanent magnet synchronous motor emulator considering driving inverter switching faults,”IEEE J. Emerg. Sel. Topics Ind. Electron., vol. 5, no. 4, pp. 1420–1434, Oct. 2024

    2024

  16. [16]

    A neural-network-based electric machine emulator using neuro- fuzzy controller for power-hardware-in-the-loop testing,

    ——, “A neural-network-based electric machine emulator using neuro- fuzzy controller for power-hardware-in-the-loop testing,”IEEE Trans. Energy Convers., vol. 40, no. 3, pp. 2242–2255, Sep. 2025

    2025

  17. [17]

    Estimation of flux saturation model for SynRMs using artificial neural network,

    J.-H. Lee, Y.-J. Lee, M.-S. Lee, D.-S. Jin, and Y.-D. Yoon, “Estimation of flux saturation model for SynRMs using artificial neural network,”IEEE Trans. Ind. Appl., vol. 61, no. 2, pp. 3143–3151, 2025

    2025

  18. [18]

    Fast flux mapping technique for synchronous reluctance machines: Method description and comparison with full FEA and measurements,

    M. Carbonieri, L. D. Leonardo, A. Mahmoudi, M. Tursini, N. Bianchi, and W. L. Soong, “Fast flux mapping technique for synchronous reluctance machines: Method description and comparison with full FEA and measurements,”IEEE Trans. Ind. Appl., vol. 59, no. 4, pp. 4056–4065, Jul./Aug. 2023

    2023

  19. [19]

    Experimental identification of the magnetic model of synchronous machines,

    E. Armando, R. Bojoi, P. Guglielmi, G. Pellegrino, and M. Pastorelli, “Experimental identification of the magnetic model of synchronous machines,”IEEE Trans. Ind. Appl., vol. 49, no. 5, pp. 2116–2125, Sep. 2013

    2013

  20. [20]

    Sensorless self-commissioning of synchronous reluctance motors at standstill without rotor locking,

    M. Hinkkanen, P. Pescetto, E. M ¨ols¨a, S. E. Saarakkala, G. Pellegrino, and R. Bojoi, “Sensorless self-commissioning of synchronous reluctance motors at standstill without rotor locking,”IEEE Trans. Ind. Appl., vol. 53, no. 3, pp. 2120–2129, May/Jun. 2017

    2017

  21. [21]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,

    M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,”Journal of Computational Physics, vol. 378, pp. 686–707, 2019

    2019

  22. [22]

    Lagrangian neural networks,

    M. Cranmer, S. Greydanus, S. Hoyer, P. Battaglia, D. Spergel, and S. Ho, “Lagrangian neural networks,” inICLR Workshop on Integration of Deep Neural Models and Differential Equations, 2020. [Online]. Available: https://arxiv.org/abs/2003.04630

    Pith/arXiv arXiv 2020

  23. [23]

    Hamiltonian neural networks,

    S. Greydanus, M. Dzamba, and J. Yosinski, “Hamiltonian neural networks,” inAdv. Neur. Inf. Process. Syst., vol. 32, Vancouver, Canada,

  24. [24]

    Port-Hamiltonian neural networks for learning explicit time-dependent dynamical systems,

    S. A. Desai, M. Mattheakis, D. Sondak, P. Protopapas, and S. J. Roberts, “Port-Hamiltonian neural networks for learning explicit time-dependent dynamical systems,”Phys. Rev. E, vol. 104, p. 034312, Sep. 2021

    2021

  25. [25]

    van der Schaft and D

    A. van der Schaft and D. Jeltsema,Port-Hamiltonian Systems Theory: An Introductory Overview. Delft, Netherlands: now Publishers Inc., 2014

    2014

  26. [26]

    Gradient networks,

    S. Chaudhari, S. Pranav, and J. M. F. Moura, “Gradient networks,”IEEE Trans. Signal Process., vol. 73, pp. 324–339, 2025

    2025

  27. [27]

    Fourier features let networks learn high frequency functions in low dimensional domains,

    M. Tancik, P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. Barron, and R. Ng, “Fourier features let networks learn high frequency functions in low dimensional domains,” inAdv. Neur. Inf. Process. Syst., vol. 33, 2020. [Online]. Available: https://arxiv.org/abs/2006.10739

    Pith/arXiv arXiv 2020

  28. [28]

    Squareplus: A softplus-like algebraic rectifier,

    J. T. Barron, “Squareplus: A softplus-like algebraic rectifier,” 2021. [Online]. Available: https://arxiv.org/abs/2112.11687

    Pith/arXiv arXiv 2021

  29. [29]

    Design framework for sensorless control of synchronous machine drives,

    L. Tiitinen, M. Hinkkanen, and L. Harnefors, “Design framework for sensorless control of synchronous machine drives,”IEEE Trans. Ind. Electron., vol. 72, no. 2, pp. 1379–1390, Feb. 2025

    2025

  30. [30]

    Direct flux vector control of synchronous motor drives: Accurate decoupled control with online adaptive maximum torque per ampere and maximum torque per volts evaluation,

    A. Varatharajan, G. Pellegrino, and E. Armando, “Direct flux vector control of synchronous motor drives: Accurate decoupled control with online adaptive maximum torque per ampere and maximum torque per volts evaluation,”IEEE Trans. Ind. Electron., vol. 69, no. 2, pp. 1235– 1243, Feb. 2022. 10 Junyi Lireceived the B.Sc. degree in communication engineering ...

    2022

  31. [2019]

    Available: https://arxiv.org/abs/1906.01563

    [Online]. Available: https://arxiv.org/abs/1906.01563

    Pith/arXiv arXiv 1906

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.