Pith. sign in

REVIEW 4 major objections 4 minor 21 references

Capacitive Touch Sensor Modeling With a Physics-informed Neural Network and Maxwell's Equations

T0 review · 4 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A physics-informed neural network predicts electric field and potential for capacitive touch sensors across 0–25 mm finger distances with a few percent normalized error and roughly 10x faster inference than FEM.

desk verdict A reasonable 3D PINN surrogate for capacitive sensors, but the generalization claim rests on test distances that were already used in training, so the proof of generalization is weaker than it looks. read the letter →

arxiv 2412.08650 v1 pith:VJKA3ZT7 submitted 2024-11-23 physics.comp-ph cs.LGeess.SP

Ganyong Mo , Krishna Kumar Narayanan , David Castells-Rufas , Jordi Carrabina This is my paper · ORCID
classification physics.comp-phcs.LGeess.SP
keywords physics-informedneuralnetworkcapacitivetouchsensorelectrostaticsimulationsurrogatemodelMaxwell'sequationselectricfieldpredictionfiniteelementmethod
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a physics-informed neural network can serve as a fast surrogate for finite-element electrostatic simulation of a capacitive touch sensor. The network takes spatial coordinates and finger distance as inputs and outputs electric field, electric potential, and charge density, trained with a loss that combines sparse simulation data at only three finger positions with residuals of the electroquasistatic Maxwell equations. If correct, designers could query electric field and potential for any finger distance between 0 and 25 mm in seconds instead of running multi-physics simulations, with normalized errors mostly between 0.003 and 0.062 and roughly tenfold faster inference. This matters because capacitive touch sensors in automotive and consumer devices are tuned through expensive simulations.

What carries the argument

The central object is the combinatorial loss function that couples data and physics. $L_{\text{bc}}$ enforces Dirichlet boundary conditions $V=0$ on finger and PCB and $V=3.3$ V on the sensor; $L_{\text{phy}}$ computes residuals of the electroquasistatic Maxwell equations using automatic differentiation with respect to coordinates, including $E=-\nabla V$, $\nabla\times E=0$, and the Laplace/Poisson equations; $L_{\text{data}}$ measures mean-square error against the sparse low-resolution simulation dataset $\Omega_S$ at three fixed finger distances. Collocation points come from $\Omega_S$ plus a non-uniform refined 3D grid $\Omega_G$ denser near the sensor-finger setup, and finger distances for the grid points are sampled uniformly over $[0,25]$ mm. An adaptive loss-weighting algorithm balances the three loss terms during training. This combination lets the physics residuals regularize the network so it can interpolate to distances and locations not present in the sparse training data.

What would settle it

Generate a finite-element field solution for a capacitive sensor with a different electrode layout or a non-cylindrical finger, feed it to the trained network at a finger distance inside 0–25 mm, and compare predicted $E$ and $V$ with the simulation; if the normalized error jumps well above the reported 0.003–0.062 range, the surrogate does not generalize to changed experimental setups.

Watch

Extended reading notes

Core claim

In the paper's own terms, the contribution is a trained surrogate sensor model: a fully connected network with five hidden layers of 512 neurons and tanh activations that solves the electroquasistatic form of Maxwell's equations in a 3D domain. The inputs are $(x,y,z,d)$; outputs are $\mathbf{E}=(E_x,E_y,E_z)$, $V$, and $\rho$. The model is trained by minimizing $L_{\text{total}}=\lambda_1 L_{\text{bc}}+\lambda_2 L_{\text{phy}}+\lambda_3 L_{\text{data}}$, where the physics losses enforce $\nabla\times\mathbf{E}=0$, $E=-\nabla V$, Laplace's equation in free space, and Poisson's equation at conductor surfaces, while the data loss anchors the solution to sparse simulation output at three finger distances. Evaluated on 26 unseen finger distances grouped into near, hover, and far zones, the network reaches NRMSE between 0.003 and 0.062 for $E$ and $V$, with the largest deviations in the $z$-component of the electric field near the finger boundary, and delivers roughly 10x faster inference than the finite-element simulation.

Load-bearing premise

The claim that the trained model serves as a surrogate for different experimental setups rests on the premise that changing only the finger distance spans the relevant design space; the sensor and PCB geometry, finger shape, and material permittivity are fixed during training.

Editorial extensions

If this is right

  • A design engineer can query electric field and potential for any finger distance in 0–25 mm in seconds, without regenerating a finite-element mesh or rerunning the simulation.
  • Sparse simulation data at only three finger distances, combined with PDE residuals, is enough to generalize across 26 tested distances, reducing the cost of building the training dataset.
  • Predictions remain accurate in near, hover, and far proximity zones, with normalized errors between 0.003 and 0.062 for $E$ and $V$; the main error concentration sits at the finger boundary.
  • Because the network is mesh-free and GPU-accelerable, the same trained model can be sampled at arbitrary resolutions and query points, unlike fixed-resolution finite-element output.
  • The framework is positioned as a foundation for modeling sensor arrays and long-term performance, since the per-query cost stays low when the geometry input set is extended.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The fixed-geometry setting means the surrogate cannot represent a new electrode layout, sensor size, finger shape, or different dielectric environment without retraining; extending the input space to geometry parameters would be a natural next step beyond what is demonstrated.
  • Charge density $\rho$ is used only inside the physics loss and is not reported as an output metric, so the paper's accuracy claims apply to $E$ and $V$; a reader should not infer that $\rho$ is validated independently.
  • The roughly 10x inference speedup counts only inference time, not the cost of generating the sparse FEM dataset or the 800k-iteration training run; the practical gain depends on how many design queries amortize that fixed cost.
  • Since the model was trained with $\varepsilon_r=1$ throughout the domain, wet-finger or moisture scenarios would need additional training or a permittivity input before the surrogate is useful in humid environments.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a physics-informed neural network (PINN) surrogate for a capacitive touch sensor. The model takes spatial coordinates and finger distance as inputs and outputs the electric field E, electric potential V, and charge density ρ. The loss is a weighted combination of boundary-condition, physics-residual, and sparse FEM data terms; the physics residuals come from the electroquasistatic form of Maxwell's equations (Faraday's law, Gauss's law, the E–V gradient relation, and Poisson's equation). The network is trained on FEM data at three finger distances together with PDE collocation points sampled over 26 finger distances, and evaluated on what the paper calls an unseen dataset. Reported overall NRMSE values range from about 0.003 to 0.062 for E and V, with peak 25% NRMSE values of 0.82–0.88 near the finger boundary. The paper claims a 10x inference speedup over FEM and concludes that the PINN can serve as a surrogate for fast design optimization.

Significance. If the central claim is supported, the paper would contribute to an industrially relevant application of PINNs: accelerating capacitive touch sensor design. The approach is reasonable in that the physics loss is derived from Maxwell's equations rather than from the fitted model, and the combination of sparse FEM data with PDE residuals is a sensible way to regularize the solution. The reported errors on the evaluated distances are encouraging, and the demonstrated ability to predict fields in a 3D geometry with a finger at various distances is a useful proof of concept. However, the significance is substantially limited by the lack of a genuine held-out test set by distance, the fixed sensor/PCB geometry, the absence of code and data, and the lack of comparison to simpler baselines. These gaps, if not addressed, reduce the paper to a demonstration of interpolation over an input parameter that was already present in training, which is far weaker than the claimed generalization to 'different experimental setups.'

major comments (4)
  1. [Experiments and Results / Conclusions] The claim of evaluation on 'unseen test cases' is not substantiated, and the paper appears to have a train/test circularity with respect to finger distance. The Experiments and Results section states that 'PDE-based loss minimization during training is conducted with 26 different finger distances to the sensor, ranging from dmin = 0mm to dmax = 25mm,' and the Conclusions state that 'we generated experimental data for 26 different finger positions to the sensor and organized them into three proximity zones.' These are the same 26 distances, with no description of any train/test split by distance. Because finger distance is the only non-spatial input that is varied, all evaluated distances may already have been used during physics-loss training, so the reported NRMSE values measure interpolation over already-seen distances rather than generalization to new distances. To support the central claim, the authors must re-evaluate on distances that are excluded from all training losses (including the PDE loss), or explicitly report a train/test split and show that the test distances were never used in any loss term.
  2. [Neural Network Setup] The input space is only (x, y, z, d), and the manuscript explicitly states 'Sensor and PCB have fixed dimensions; thus, BCs are randomly sampled from their corresponding meshes during training.' The abstract and conclusion claim that the model can perform inference 'for different experimental setups,' but the only setup parameter that can vary in the current model is the finger distance for one fixed sensor and PCB geometry. Changes in electrode layout, sensor size, finger shape, or nearby environment cannot be represented. The claim in the abstract and conclusions is therefore overbroad; it should be restricted to 'different finger distances for a fixed sensor/PCB geometry' unless the model is shown to generalize across geometry variations, which it currently cannot by construction.
  3. [Loss Functions, Eq. (12)] The physics loss in Eq. (12) contains redundant or trivial terms. The term ||Ê + ∇V̂||^2 enforces E = -∇V. The second term, ||-∇ × (ε0εr∇V̂)||^2, is the curl of a gradient times a scalar when εr is constant; since εr is set to 1 throughout the experiments, this term is identically zero. The third and fourth terms both contain ∇·(ε0εrÊ), and the fourth term is a residual between ρ̂ and the divergence. The authors should clarify on which subdomains each residual is applied (e.g., Laplace in free space, Poisson on conductors), remove the redundant curl term or justify its presence with a spatially varying εr, and explain the role of each term. As written, part of the 'physics' in the loss imposes no constraint.
  4. [Table 1] The peak 25% NRMSE values for E components in the near zone are 0.819–0.880, which means that in the region closest to the finger the predictions are essentially unreliable, and this is precisely the region that matters most for touch detection. The paper attributes this to sharp transitions around the finger boundary, but does not provide error maps, the fraction of points affected, or a distance-from-boundary analysis. Without such information, the practical value of the surrogate for design decisions in the near zone remains unclear. The authors should quantify how localized these high errors are and, if possible, report errors on derived quantities such as capacitance, which is the most relevant output for capacitive sensor functionality.
minor comments (4)
  1. [Methodology, Governing Equations] There is a typo: 'gradiant' should be 'gradient' in the sentence 'Additionally, the gradiant and the curl of E are calculated...'.
  2. [General] The paper does not provide code, data, training details such as exact hyperparameter values, or the procedure for generating the 'unseen' dataset. For reproducibility and for the reader to assess whether the reported results are robust, the authors should make the data and code available, or at least describe the exact test data generation procedure and the train/test split.
  3. [Experiments and Results] The paper does not compare the PINN against a simpler baseline, such as a fully data-driven neural network trained on the same FEM data, or against FEM interpolation. A baseline comparison would help assess whether the physics loss actually provides the claimed benefit in accuracy and generalization, rather than merely acting as a regularizer.
  4. [Conclusions] The statement that 'PINNs maintain their fast inference capabilities across different resolutions of input' is not demonstrated anywhere in the paper; no resolution study is reported. Please either add supporting experiments or remove the claim.

Circularity Check

1 steps flagged · score 6.0 of 10

The 'unseen test cases' are evaluated at the same 26 finger distances used for PDE-loss training, so the reported generalization is in-sample in the only varied setup parameter.

  1. fitted input called prediction [Experiments and Results (training distances); Conclusions (evaluation distances); Abstract ('unseen test cases')]
    "PDE-based loss minimization during training is conducted with 26 different finger distances to the sensor, ranging from a finger touching the sensor at dmin = 0mm to a relatively far position of dmax = 25mm. ... we generated experimental data for 26 different finger positions to the sensor and organized them into three proximity zones."

    The only varied setup input is finger distance d. Training minimizes the PDE/physics loss at 26 distances from 0 to 25 mm, and the evaluation set is described as 26 finger positions over the same range; no distance-based train/test split is reported. As written, the test distances are not held out: the NRMSE in the near, hover, and far zones measures performance at distances already used as PDE collocation inputs, not prediction for a genuinely new experimental setup. The abstract's 'unseen test cases' and the conclusion's 'generalization capability at all distances' therefore rest on in-sample distance inputs, making the generalization evidence partially circular by construction.

full rationale

The core PINN formulation is not circular: the physics loss is taken from Maxwell's equations in their electroquasistatic form, the BCs are independent physical constraints, and the data loss uses FEM outputs at only three finger distances. There is no self-citation chain, no imported uniqueness theorem, and no ansatz smuggled in via citation. The one significant circular step is in the validation of the central generalization claim. Because finger distance is the only continuous non-spatial input and the only setup variable changed, holding out distance is essential to demonstrate inference for 'different experimental setups.' The paper says PDE-loss training uses 26 finger distances from 0 to 25 mm and then says the unseen evaluation dataset was generated for 26 finger positions organized into the same three proximity zones, without describing any split. Thus the reported errors do not establish extrapolation to new finger distances; they are in-sample with respect to the distance input. This is a partial circularity in the evidence for the main claim, though the model itself is not derived from its evaluation targets.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The model introduces no new physical entities, but it relies on a simplified electrostatic model and several hand-chosen training choices. The main load-bearing assumptions are the electroquasistatic approximation, the ideal-conductor boundary conditions, and the homogeneous epsilon_r=1 domain; the free parameters are loss weights, training distances, and grid design. None are fitted physical constants, so the circularity burden is low.

free parameters (4)
  • Loss weights lambda_1, lambda_2, lambda_3 = not reported (adaptive weighting)
    Eq. 10 combines boundary, physics, and data losses; the adaptive algorithm adjusts weights during training, but no final values are given, and model performance depends on this trade-off.
  • Training finger distances d0, d1, d2 = dmin=0 mm, dmax=25 mm, one random middle
    Data supervision comes only from these three FEM distances; the generalization claim over 0 to 25 mm depends on this choice.
  • Non-uniform 3D collocation grid density = not specified
    The grid Omega_G is denser near the sensor and finger and coarser at edges; the exact construction is not given, so the physics loss coverage is underdetermined.
  • Network architecture hyperparameters = 5 hidden layers x 512 tanh, 800k Adam iterations
    Chosen by hand; the surrogate result depends on these choices, though they are standard.
assumptions (5)
  • domain assumption Electroquasistatic approximation: curl E = 0, magnetic fields neglected
    Invoked in Eq. 1 with reference to Baxter 1996 for low-frequency capacitive sensing.
  • domain assumption Finger, sensor, and PCB are ideal conductors with fixed potentials (0 V, 3.3 V, 0 V)
    Dirichlet BCs in Eqs. 8 and 9; real fingers have finite conductivity and dielectric properties.
  • domain assumption Whole domain is homogeneous with relative permittivity epsilon_r = 1
    Set in Experiments; ignores sensor dielectric layers, finger tissue, coatings, humidity, and temperature, which the introduction says matter.
  • domain assumption Conductors can be approximated as 2D surfaces with negligible thickness, giving surface charge relation n dot D = rho_s
    Eq. 3 is used to derive the Poisson boundary condition in Eq. 6.
  • domain assumption Neural network with combinatorial loss converges to the electrostatic solution at collocation points
    PINN methodology assumes the minimizer of the physics residual approximates the PDE solution; convergence is not proven, and the paper reports high errors near finger boundaries.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Capacitive Touch Sensor Modeling With a Physics-informed Neural Network and Maxwell's Equations." pith.science (2026). https://pith.science/paper/VJKA3ZT7

@misc{pith2026241208650,
  author       = {Pith},
  title        = {Pith review of: Capacitive Touch Sensor Modeling With a Physics-informed Neural Network and Maxwell's Equations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VJKA3ZT7}},
  note         = {Machine review of arXiv:2412.08650}
}
read the original abstract

Maxwell's equations are the fundamental equations for understanding electric and magnetic field interactions and play a crucial role in designing and optimizing sensor systems like capacitive touch sensors, which are widely prevalent in automotive switches and smartphones. Ensuring robust functionality and stability of the sensors in dynamic environments necessitates profound domain expertise and computationally intensive multi-physics simulations. This paper introduces a novel approach using a Physics-Informed Neural Network (PINN) based surrogate model to accelerate the design process. The PINN model solves the governing electrostatic equations describing the interaction between a finger and a capacitive sensor. Inputs include spatial coordinates from a 3D domain encompassing the finger, sensor, and PCB, along with finger distances. By incorporating the electrostatic equations directly into the neural network's loss function, the model captures the underlying physics. The learned model thus serves as a surrogate sensor model on which inference can be carried out in seconds for different experimental setups without the need to run simulations. Efficacy results evaluated on unseen test cases demonstrate the significant potential of PINNs in accelerating the development and design optimization of capacitive touch sensors.

Figures

Figures reproduced from arXiv: 2412.08650 by the authors.

Figure 1
Figure 1. System Architecture of the PINN Process ger. A coarse-resolution dataset at three specific distances is generated from a multi-physics simulation. Loss computed from this sparse data provides the initial estimate for the PINN which is then complemented with the physics residual losses arising from the governing partial differential equa￾tions (PDEs). This loss acts as a regularization agent to limit the admissible s… view at source ↗
Figure 2
Figure 2. Comparison of PINN Output and Simulation [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter doi edition editor eid howpublished institution isbn issn journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1...

    work page

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

    work page

  3. [3]

    Capacitive Sensors: Design and Applications

    Baxter L.K., 1996. Capacitive Sensors: Design and Applications. Wiley-IEEE Press

    work page 1996

  4. [4]

    Enforcing Dirichlet boundary conditions in physics-informed neural networks and variational physics-informed neural networks

    Berrone S.; Canuto C.; Pintore M.; and Sukumar N., 2023. Enforcing Dirichlet boundary conditions in physics-informed neural networks and variational physics-informed neural networks. Heliyon, 9, no. 8

    work page 2023

  5. [5]

    Temperature influence on capacitive sensor structures

    Gebhardt S.; Scheinert G.; and Uhlmann F.H., 2006. Temperature influence on capacitive sensor structures. In Information technology and electrical engineering-devices and systems, materials and technologies for the future. vol. 51

    work page 2006

  6. [6]

    Softadapt: Techniques for adaptive loss weighting of neural networks with multi-part loss functions

    Heydari A.A.; Thompson C.A.; and Mehmood A., 2019. Softadapt: Techniques for adaptive loss weighting of neural networks with multi-part loss functions. arXiv preprint arXiv:191212355

    work page 2019

  7. [7]

    and Psaltis D., 2022

    Lim J. and Psaltis D., 2022. MaxwellNet: Physics-driven deep neural network training based on Maxwell’s equations. Apl Photonics, 7, no. 1

    work page 2022

  8. [8]

    A combined data-driven and physics-driven method for steady heat conduction prediction using deep convolutional neural networks

    Ma H.; Hu X.; Zhang Y.; Thuerey N.; and Haidn O.J., 2020. A combined data-driven and physics-driven method for steady heat conduction prediction using deep convolutional neural networks. arXiv preprint arXiv:200508119

    work page 2020

Show all 21 references
  1. [9]

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

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

    2019

  2. [10]

    T., 2023

    Sharma P.; Evans L.; and et al. T., 2023. Stiff-PDEs and Physics-Informed Neural Networks. Arch Computat Methods Eng, 30, 2929--2958

    2023

  3. [11]

    A Maxwell's Equations Based Deep Learning Method for Time Domain Electromagnetic Simulations

    Zhang P.; Hu Y.; Jin Y.; Deng S.; Wu X.; and Chen J., 2020. A Maxwell's Equations Based Deep Learning Method for Time Domain Electromagnetic Simulations. In 2020 IEEE Texas Symposium on Wireless and Microwave Circuits and Systems (WMCS). 1--4

    2020

  4. [12]

    Banks, J., J. S. Carson, B. L. Nelson, and D. M. Nicol. 2000. Discrete-event system simulation\/ . 3d ed. Upper Saddle River, New Jersey: Prentice-Hall, Inc

    2000

  5. [13]

    Cheng, R. C. H., and J. P. C. Kleinen. 1999. Improved design of queueing simulation experiments with highly heteroscedastic responses. Operations Research\/ 47 (3): 762--777

    1999

  6. [14]

    Cheng, R. C. H. 1994. Selecting input models. In Proceedings of the Winter Simulation Conference , ed.\ J. D. Tew, S. Manivannan, D. A. Sadowski, and A. F. Seila, 184--191. Piscataway, New Jersey: Institute of Electrical and Electronics Engineers, Inc

    1994

  7. [15]

    Goldsman, D., and B. L. Nelson. 1994. Ranking, selection, and multiple comparisons in computer simulation. In Proceedings of the Winter Simulation Conference , ed.\ J. D. Tew, S. Manivannan, D. A. Sadowski, and A. F. Seila, 192--199. Piscataway, New Jersey: Institute of Electr...

    1994

  8. [16]

    S., and K

    Gupta, S. S., and K. J. Miescke. 1988. On the problem of finding the largest normal mean under heteroscedasticity. In Statistical Decision Theory and Related Topics IV , ed.\ S. S. Gupta and J. O. Berger, Volume 2. New York: Springer Verlag

    1988

  9. [17]

    Gupta, S. S., K. Nagel, and S. Panchapakesan. 1973. On the order statistics from equally correlated normal random variables. Biometrika\/ 60:403--413

    1973

  10. [18]

    M., and D

    Hammersley, J. M., and D. C. Handscomb. 1964. M onte C arlo methods . London: Methuen

    1964

  11. [19]

    Hsu, J. C. 1996. Multiple comparisons: Theory and methods . New York: Chapman & Hall

    1996

  12. [20]

    M., and W

    Law, A. M., and W. D. Kelton. 2000. Simulation modeling & analysis\/ . 3d ed. New York: McGraw-Hill, Inc

    2000

  13. [21]

    Rinott, Y. 1978. On two-stage selection procedures and related probability-inequalities. Communications in Statistics\/ A7:799--811

    1978

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.