Machine-learning-assisted material and geometry characterization from Casimir force measurement

Hideo Iizuka; Shanhui Fan

arxiv: 2604.15763 · v1 · submitted 2026-04-17 · 🪐 quant-ph · physics.optics

Machine-learning-assisted material and geometry characterization from Casimir force measurement

Hideo Iizuka , Shanhui Fan This is my paper

Pith reviewed 2026-05-10 09:30 UTC · model grok-4.3

classification 🪐 quant-ph physics.optics

keywords Casimir forcemachine learningthin filmpermittivitymaterial characterizationquantum vacuum fluctuationsinverse problem

0 comments

The pith

Machine learning recovers thin-film thickness and permittivity spectrum from Casimir force-spacing curves

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

The paper shows that the Casimir force between parallel plates, one coated with a nanoscopic film, varies with separation in a manner that encodes both the film's thickness and its permittivity as a function of frequency over a broad range. A machine learning model trained on simulated force data learns to invert this relationship and extract the two quantities from a measured curve. This works because quantum vacuum fluctuations provide a naturally broadband electromagnetic probe whose effective frequency content changes with plate spacing. A sympathetic reader would care because the result turns the Casimir effect into a contact-free metrology technique that requires no external light source or additional probes.

Core claim

Using a machine learning approach, we show that one can infer both the thickness of the film and its permittivity over a broad frequency range, starting from the dependency of the Casimir forces on the spacing between the two plates.

What carries the argument

A supervised machine learning model trained on computed Casimir force versus separation curves that learns the mapping from those curves to film thickness and the frequency-dependent permittivity function.

If this is right

Both geometric thickness and material permittivity spectrum can be extracted from a single type of force measurement.
The inverse problem of deducing material properties from Casimir forces becomes solvable by data-driven methods.
Quantum vacuum fluctuations can serve as the electromagnetic source for broadband material characterization.
The approach remains viable when realistic levels of noise are present in the force data.

Where Pith is reading between the lines

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

The same inversion strategy could be applied to multilayer films or non-planar geometries once sufficient training data are generated.
Experimental validation on real Casimir apparatus would be the direct next step to assess practical accuracy beyond simulations.
Related force or torque measurements in other quantum-vacuum setups might admit similar machine-learning treatment for property extraction.

Load-bearing premise

The Casimir force versus spacing curve contains sufficient independent information to uniquely determine both film thickness and the full permittivity spectrum over a broad frequency range, with the ML model able to generalize under realistic noise and model assumptions.

What would settle it

Measure the Casimir force-spacing curve for a thin film whose thickness and permittivity spectrum have been independently verified by other methods, then test whether the trained model recovers those known values within experimental uncertainty.

Figures

Figures reproduced from arXiv: 2604.15763 by Hideo Iizuka, Shanhui Fan.

**Figure 1.** Figure 1: Material and geometry characterization from Casimir force measurement. We consider a geometry consisting of two gold regions, with one of the gold regions coated with a dielectric thin film. The two regions are separated by a vacuum gap 𝑑. The permittivity 𝜖𝑑 (𝜔) and the thickness 𝑡 of the thin film are determined using a neural network. The input to the neural network is the Casimir forces as a function o… view at source ↗

**Figure 2.** Figure 2: Predicted film thickness 𝑡 and resonance frequency 𝜔02 in the two-pole model for 𝜖𝑑 (𝜔) in Eq. (1). 200 testing data points are plotted in each panel. The fact that these data points are along the linear line indicates the agreement between the predicted values and the true values. RMSE is given at the right bottom of each panel. (a) Film thickness 𝑡 ( 10𝑛𝑚 ≤ 𝑡 ≤ 500𝑛𝑚 ) and (b) resonance frequency 𝜔02 (3 … view at source ↗

**Figure 3.** Figure 3: (a),(b) Permittivity spectra for the two-pole model. Two examples are selected from each of the two testing data sets, corresponding to [PITH_FULL_IMAGE:figures/full_fig_p025_3.png] view at source ↗

**Figure 4.** Figure 4: Predicted film thickness 𝑡 and pole parameters in the four-pole model for 𝜖𝑑 (𝜔) in Eq. (1). 400 testing incidences are plotted in each panel. RMSE is given at the right bottom of each panel. (a) Film thickness 𝑡. (b)-(l) Pole parameters in Eq. (1) are shown in (b) for 𝜔𝑝1, (c) for 𝛾1, (d) for 𝜔02, (e) for 𝜔𝑝2, (f) for 𝛾2, (g) for 𝜔03, (h) for 𝜔𝑝3, (i) for 𝛾3, (j) for 𝜔04, (k) for 𝜔𝑝4, and (l) for 𝛾4. The … view at source ↗

**Figure 5.** Figure 5: (a)-(c) Predicted permittivity spectra for the four-pole model. These examples are selected from the testing data set presented in [PITH_FULL_IMAGE:figures/full_fig_p027_5.png] view at source ↗

**Figure 1.** Figure 1: The blue and pink solid lines represent the real and imaginary parts of predicted 𝜖𝑑 (𝜔). Blue and pink crosses in (d) and circles in (e) and (f) represent the real and imaginary parts of silicon permittivity in Ref. 52 and true 𝜖𝑑 (𝜔), respectively. The pole parameters in Eq. (1) for (d)-(f) are presented in Table S11 of Supplemental Material [40] [PITH_FULL_IMAGE:figures/full_fig_p030_1.png] view at source ↗

read the original abstract

A broadband electromagnetic source is important for scientific and technological applications. Quantum vacuum fluctuations, which manifest most prominently in the Casimir effect, provide a fundamentally broadband electromagnetic source. Here we explore a potential consequence of the broadband nature of quantum vacuum fluctuations, by showing that such fluctuations can enable measurement of material permittivity over a broad frequency range. Specifically, we consider the Casimir force in a parallel-plate geometry, with one plate covered by a nanoscopic thin film. Using a machine learning approach, we show that one can infer both the thickness of the film and its permittivity over a broad frequency range, starting from the dependency of the Casimir forces on the spacing between the two plates. Our work highlights the application potential of using vacuum fluctuations as a naturally-existing broadband electromagnetic source for material characterization, and shows that the inverse problem in Casimir force calculation can be solved with machine learning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper frames Casimir force-spacing data as input for ML to jointly recover film thickness and broadband permittivity, but the abstract supplies no validation or handling of the inverse problem's ill-posedness.

read the letter

The core claim is that machine learning can extract both a thin film's thickness and its permittivity spectrum over a broad frequency range from how the Casimir force varies with plate spacing. This turns the broadband vacuum fluctuations behind the Casimir effect into a characterization tool without external sources. The framing of the Casimir inverse problem as a joint ML task for geometry plus material response looks new compared to prior work on forward calculations or simpler inversions. The abstract does a clean job laying out the conceptual motivation: quantum vacuum fluctuations already span frequencies, so why not use the force curve they produce for material info. That part is straightforward and worth noting. The execution details are missing. No description of how the training data was generated, what network was used, what loss or metrics were tracked, or how noise and dispersion models were handled. The stress-test point lands: the Lifshitz formula computes the force as an integral over Matsubara frequencies of reflection coefficients that depend on permittivity at imaginary frequencies and on thickness. This is a smoothing transform, so many different permittivity functions can map to force curves that are indistinguishable within experimental noise. Without any analysis of the operator's null space, condition number, or sensitivity to high-frequency features, it is not obvious that unique recovery is possible beyond the training set. The paper may contain those checks in the full text, but nothing in the abstract or the visible rationale shows them. This work would interest people in nanophotonics or surface metrology who already measure Casimir forces and want to add material extraction. It is worth a reading-group discussion to see whether the full manuscript actually demonstrates reliable inversion or just shows results on synthetic data that match the training distribution. I would send it to peer review because the idea is concrete enough to test and the field could use new characterization routes, but the referees will need to see the validation and the handling of non-uniqueness before the central claim can be taken as supported.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a machine-learning method to solve the inverse Casimir problem: from the measured dependence of the Casimir force on plate separation in a parallel-plate geometry (one plate coated with a nanoscopic thin film), the approach infers both the film thickness and the film's permittivity spectrum over a broad frequency range, exploiting the broadband character of quantum vacuum fluctuations.

Significance. If the central claim holds, the work would establish a concrete route to using Casimir forces as a naturally broadband electromagnetic probe for material characterization, with potential impact on nanophotonics and thin-film metrology. The manuscript correctly identifies the forward map as the Lifshitz integral and frames the inverse task as amenable to data-driven methods; however, no machine-checked proofs, reproducible code, or parameter-free derivations are supplied.

major comments (2)

[Abstract and §3] Abstract and §3 (ML implementation): the assertion that the ML model successfully recovers both thickness and the full broadband permittivity is unsupported by any description of training-data generation, network architecture, loss function, validation metrics, error bars, or explicit treatment of experimental noise and dispersion-model mismatch. These omissions are load-bearing for the central claim.
[§2] §2 (theoretical setup): no analysis of the forward operator's null space, condition number, or sensitivity kernels is given. Because the Casimir force F(d) is obtained via the Lifshitz formula—an integral transform over imaginary Matsubara frequencies—the manuscript must demonstrate that distinct pairs (t, ε(ω)) produce distinguishable F(d) curves within realistic noise; without this, uniqueness of the inferred spectrum cannot be established beyond the training distribution.

minor comments (2)

[Figures] Figure captions and axis labels should explicitly state the frequency range over which permittivity is recovered and the noise level assumed in the synthetic data.
[Throughout] Notation for the permittivity on the imaginary axis (ε(iξ)) versus real-frequency ε(ω) should be introduced once and used consistently.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We address each major comment point by point below. Revisions have been made to strengthen the presentation of the machine-learning methodology and to provide additional theoretical context where feasible.

read point-by-point responses

Referee: [Abstract and §3] Abstract and §3 (ML implementation): the assertion that the ML model successfully recovers both thickness and the full broadband permittivity is unsupported by any description of training-data generation, network architecture, loss function, validation metrics, error bars, or explicit treatment of experimental noise and dispersion-model mismatch. These omissions are load-bearing for the central claim.

Authors: We agree that the original manuscript lacked sufficient detail on the machine-learning implementation. In the revised version we will expand §3 with a full description of the training-data generation procedure (including the sampling ranges for film thickness and the dispersion models used to generate permittivity spectra), the neural-network architecture (layer count, neuron numbers, activation functions, and regularization), the loss function, the validation and test metrics (including error bars on recovered parameters), and explicit handling of additive experimental noise together with possible mismatches between the training dispersion models and real materials. These additions will directly support the central claim. revision: yes
Referee: [§2] §2 (theoretical setup): no analysis of the forward operator's null space, condition number, or sensitivity kernels is given. Because the Casimir force F(d) is obtained via the Lifshitz formula—an integral transform over imaginary Matsubara frequencies—the manuscript must demonstrate that distinct pairs (t, ε(ω)) produce distinguishable F(d) curves within realistic noise; without this, uniqueness of the inferred spectrum cannot be established beyond the training distribution.

Authors: We acknowledge the value of a theoretical characterization of the forward map. While the empirical performance of the trained model on held-out and noisy test data already indicates practical distinguishability, we will add to the revised §2 a brief sensitivity analysis together with numerical examples that illustrate how distinct (t, ε(ω)) pairs produce measurably different F(d) curves under realistic noise levels. A complete analytical treatment of the null space and condition number of the infinite-dimensional Lifshitz operator lies outside the scope of the present work; we will therefore note this limitation and emphasize that the machine-learning results provide evidence of uniqueness within the physically relevant parameter domain explored during training. revision: partial

Circularity Check

0 steps flagged

No significant circularity; ML inverse mapping is independent of training inputs

full rationale

The paper demonstrates an ML-based solver for the inverse Casimir problem: a model is trained on forward-simulated pairs (F(d), {t, ε(ω)}) generated from the Lifshitz formula and then applied to new F(d) curves to recover t and the spectrum. This is a standard supervised-learning setup with no self-definitional equations, no fitted parameters renamed as predictions, and no load-bearing self-citations that collapse the central claim. The derivation chain consists of (1) forward simulation (independent), (2) supervised training (independent), and (3) inference on held-out data (independent). No step reduces by construction to its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on standard electromagnetic theory for computing Casimir forces from permittivity; no new physical entities or ad-hoc axioms are introduced beyond the usual assumptions of Lifshitz theory and the validity of the chosen ML architecture.

axioms (1)

standard math Casimir force between plates can be computed from frequency-dependent permittivity via Lifshitz formula or equivalent electromagnetic fluctuation theory
Invoked implicitly as the forward model that generates the training data for the ML inverse solver.

pith-pipeline@v0.9.0 · 5444 in / 1230 out tokens · 39851 ms · 2026-05-10T09:30:29.635756+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

1 extracted references · 1 canonical work pages

[1]

Learning representations by back-propagation errors,

[S1] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning representations by back-propagation errors,” Nature 323, 533-536 (1986). [S2] H. Iizuka and S. Fan, “Fictitious pole scheme for extending the spectrum of material permittivities in Casimir force calculations,” Phys. Rev. B 110, 235409 (2024). [S3] C. Henkel, K. Joulain, J.-Ph. Mulet, and J....

work page 1986

[1] [1]

Learning representations by back-propagation errors,

[S1] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning representations by back-propagation errors,” Nature 323, 533-536 (1986). [S2] H. Iizuka and S. Fan, “Fictitious pole scheme for extending the spectrum of material permittivities in Casimir force calculations,” Phys. Rev. B 110, 235409 (2024). [S3] C. Henkel, K. Joulain, J.-Ph. Mulet, and J....

work page 1986