Casimir Geometry: Force, Modes, and Inverse Probing

• Casimir geometry is the configuration of bodies that shapes the vacuum fluctuation spectrum and determines force laws and material dispersion.
• In planar thin-film setups, the gap (d) and film thickness (t) act as spectroscopic controls for inverse reconstruction of material properties.
• Beyond planar systems, curvature and corrugation induce non-specular scattering, underscoring geometry’s role in tailoring force laws and thermal effects.

Casimir geometry denotes the shape, arrangement, and relative positioning of bodies that determine the spectrum of vacuum or fluctuation modes and therefore the resulting Casimir interaction. In the scattering formulation, geometry enters through the basis used to represent fields, the matrix structure of the reflection operators, and the propagation or translation operators between objects (Lambrecht et al., 2010). In a narrower and explicitly operational sense, recent inverse-design work uses “Casimir geometry” to mean a highly controlled planar multilayer configuration—two parallel gold plates separated by a vacuum gap, with one plate coated by a nanoscopic dielectric thin film—so that the geometric parameters $d$ and $t$ act as spectroscopic control variables for the unknown film permittivity $\varepsilon_d(\omega)$ (Iizuka et al., 17 Apr 2026).

1. Geometry as the organizing variable of Casimir interactions

In Casimir theory, geometry is not a secondary perturbation on top of material response; it determines which electromagnetic or statistical modes are allowed, how they propagate, and how they are mixed. For two parallel plates, translational symmetry implies specular scattering and conservation of transverse wavevector and polarization. For arbitrarily shaped stationary objects, the general free energy takes the operator form

$$F(T)=k_B T \sum_{m=0}^{\infty}{}' \mathrm{Tr}\,\ln\Bigl(1-R_1(i\xi_m)e^{-KL}R_2(i\xi_m)e^{-KL}\Bigr),$$

so geometry enters both through the reflection operators $R_1,R_2$ and through the propagation operator $e^{-KL}$ (Lambrecht et al., 2010).

This perspective makes the term “Casimir geometry” broader than any single configuration. Parallel plates, corrugated gratings, plane–sphere, sphere–sphere, and patterned multilayers are not merely different experimental layouts; they correspond to different mode structures and therefore to different force laws, thermal corrections, and sensitivities to material dispersion. A recurring consequence is that approximations derived from the parallel-plate limit, especially the Proximity Force Approximation (PFA), are geometry dependent and can fail qualitatively once curvature, corrugation, or mode mixing becomes important (Lambrecht et al., 2010).

2. The canonical planar thin-film configuration

In the specific geometry analyzed for machine-learning-assisted characterization, the system consists of two parallel gold plates, layers 1 and 3, separated by a vacuum gap of width $d$ (layer 0), with one of the gold plates coated by a nanoscopic dielectric thin film (layer 2) of thickness $t$ and permittivity $\varepsilon_d(\omega)$. The layer sequence is

$$\text{Au (3)} \;|\; \text{film (2)} \;|\; \text{vacuum (0)} \;|\; \text{Au (1)}.$$

Gold is modeled with a Drude permittivity, and the film with an $t$0-pole Lorentz–Drude model,

$t$1

where $t$2, $t$3, and $t$4 are material parameters and are treated as unknowns together with $t$5 in the inverse problem (Iizuka et al., 17 Apr 2026).

The geometry is idealized in the planar Lifshitz limit. The theoretical expressions assume infinite parallel plates, planar symmetry, no edge effects, and pressure rather than total force. Interfaces are taken to be perfectly planar, materials are isotropic and non-magnetic with $t$6, fields are decomposed into plane waves with lateral wavevector $t$7, and the system is at thermal equilibrium at room temperature, $t$8. For machine-learning input generation, the gap is sampled at 20 logarithmically spaced distances, with $t$9 in the two-pole and four-pole studies and $\varepsilon_d(\omega)$0 in the more realistic scenario. The film thickness is reconstructed over $\varepsilon_d(\omega)$1 in the simple and broad-range studies, and over $\varepsilon_d(\omega)$2 in the more realistic case (Iizuka et al., 17 Apr 2026).

This highly constrained setup is important because it isolates a single unknown geometric scalar, $\varepsilon_d(\omega)$3, against a fixed and well-characterized substrate. The geometry therefore becomes a calibrated probe of material dispersion rather than an uncontrolled source of experimental variability.

3. Lifshitz formulation and geometric control of spectral sensitivity

For the planar multilayer, the Casimir pressure is written in standard Lifshitz form after Wick rotation to imaginary frequencies,

$\varepsilon_d(\omega)$4

with

$\varepsilon_d(\omega)$5

Here $\varepsilon_d(\omega)$6 is the Fresnel coefficient of the vacuum–gold interface, while $\varepsilon_d(\omega)$7 is the effective reflection coefficient of the vacuum–(film + gold substrate) stack (Iizuka et al., 17 Apr 2026).

For a single film on a substrate, the effective coefficient is

$\varepsilon_d(\omega)$8

so the geometry enters explicitly through two exponential propagation factors: $\varepsilon_d(\omega)$9 across the vacuum gap and $$F(T)=k_B T \sum_{m=0}^{\infty}{}' \mathrm{Tr}\,\ln\Bigl(1-R_1(i\xi_m)e^{-KL}R_2(i\xi_m)e^{-KL}\Bigr),$$0 across the film. The first controls the penetration of modes through the cavity; the second controls how strongly the coating modifies the substrate reflection. This is the direct mathematical expression of the statement that Casimir geometry selects the fluctuation spectrum.

In the planar setting, the relevant modes are labelled by Matsubara index $$F(T)=k_B T \sum_{m=0}^{\infty}{}' \mathrm{Tr}\,\ln\Bigl(1-R_1(i\xi_m)e^{-KL}R_2(i\xi_m)e^{-KL}\Bigr),$$1, in-plane wavevector $$F(T)=k_B T \sum_{m=0}^{\infty}{}' \mathrm{Tr}\,\ln\Bigl(1-R_1(i\xi_m)e^{-KL}R_2(i\xi_m)e^{-KL}\Bigr),$$2, and polarization $$F(T)=k_B T \sum_{m=0}^{\infty}{}' \mathrm{Tr}\,\ln\Bigl(1-R_1(i\xi_m)e^{-KL}R_2(i\xi_m)e^{-KL}\Bigr),$$3. Varying $$F(T)=k_B T \sum_{m=0}^{\infty}{}' \mathrm{Tr}\,\ln\Bigl(1-R_1(i\xi_m)e^{-KL}R_2(i\xi_m)e^{-KL}\Bigr),$$4 changes the spectral window that contributes significantly to the pressure through the characteristic cutoff

$$F(T)=k_B T \sum_{m=0}^{\infty}{}' \mathrm{Tr}\,\ln\Bigl(1-R_1(i\xi_m)e^{-KL}R_2(i\xi_m)e^{-KL}\Bigr),$$5

Smaller $$F(T)=k_B T \sum_{m=0}^{\infty}{}' \mathrm{Tr}\,\ln\Bigl(1-R_1(i\xi_m)e^{-KL}R_2(i\xi_m)e^{-KL}\Bigr),$$6 gives access to higher frequencies, while larger $$F(T)=k_B T \sum_{m=0}^{\infty}{}' \mathrm{Tr}\,\ln\Bigl(1-R_1(i\xi_m)e^{-KL}R_2(i\xi_m)e^{-KL}\Bigr),$$7 probes lower frequencies. Varying $$F(T)=k_B T \sum_{m=0}^{\infty}{}' \mathrm{Tr}\,\ln\Bigl(1-R_1(i\xi_m)e^{-KL}R_2(i\xi_m)e^{-KL}\Bigr),$$8 changes the effective reflection coefficient through $$F(T)=k_B T \sum_{m=0}^{\infty}{}' \mathrm{Tr}\,\ln\Bigl(1-R_1(i\xi_m)e^{-KL}R_2(i\xi_m)e^{-KL}\Bigr),$$9: when $R_1,R_2$0, the film is effectively thin and only weakly perturbs the gold substrate, whereas for $R_1,R_2$1, pronounced thickness-dependent interference and decay effects appear. Resonances in $R_1,R_2$2 modulate $R_1,R_2$3 and the Fresnel coefficients, producing surface polariton features that are imprinted in the force (Iizuka et al., 17 Apr 2026).

A common misunderstanding is that geometry merely rescales a material-dependent force amplitude. In the planar thin-film problem, geometry instead acts as a frequency selector: because $R_1,R_2$4 is varied over orders of magnitude, the mode cutoff sweeps a broad spectral range, and the force becomes sensitive to the film permittivity over more than two decades of frequency, from infrared to ultraviolet (Iizuka et al., 17 Apr 2026).

4. Beyond planarity: curvature, corrugation, diffraction, and thermal structure

Outside the specular parallel-plate limit, Casimir geometry is governed by non-specular scattering. Corrugated plates, gratings, and spheres generate mode mixing, couple different wavevectors and polarizations, and invalidate any treatment that depends only on local separation. In the scattering approach, this is precisely the distinction between diagonal reflection operators in plane-wave basis and full matrices that mix diffraction orders or spherical multipoles (Lambrecht et al., 2010).

For corrugated surfaces, the geometry introduces new scales such as corrugation wavelength, grating period, groove depth, and lateral shift. In perturbative sinusoidal gratings, the lateral Casimir interaction is controlled by a spectral sensitivity function $R_1,R_2$5, and the ratio $R_1,R_2$6 satisfies $R_1,R_2$7 for finite $R_1,R_2$8, meaning that PFA overestimates the lateral force. For deep gratings, full non-specular calculations are required, and the force can differ strongly from PFA, with the sign of the deviation depending on material class and geometric regime (Lambrecht et al., 2010). Experiments on interpenetrated rectangular gratings push this geometric sensitivity much further: just before interpenetration, deviations from PFA reach a factor of $R_1,R_2$9, and after interpenetration the force becomes non-zero and independent of displacement over a plateau region (Wang et al., 2020).

Plane–sphere geometry provides the clearest example of curvature effects. Exact multipolar calculations show that the leading correction to PFA depends on material model: for perfect reflectors at $e^{-KL}$0, the slope of the force-gradient correction is $e^{-KL}$1, while for real metals in the plasma model it is $e^{-KL}$2 (Lambrecht et al., 2010). In the classical high-temperature limit, the same geometry becomes universal in a different sense: for Drude metals, the plane–sphere free energy is described by a universal function of the aspect ratio $e^{-KL}$3, independent of the conductivity, and its deviation from PFA is better fitted by polynomial expansions in $e^{-KL}$4 than by powers of $e^{-KL}$5 alone (Canaguier-Durand et al., 2012).

Thermal effects are also geometry specific. In plane–plane geometry, the familiar high-temperature Drude/plasma discrepancy is a factor of 2; in plane–sphere geometry, the corresponding ratio becomes $e^{-KL}$6 for large spheres and can be as low as $e^{-KL}$7 for smaller ones (Lambrecht et al., 2010). Likewise, a negative interaction entropy is not exclusive to dissipative models; it arises already for perfect reflectors in plane–sphere geometry and therefore reflects geometry-dependent redistribution of fluctuation modes rather than a thermodynamic inconsistency (Lambrecht et al., 2010). For a Dirichlet scalar in sphere–plate and cylinder–plate geometries, the thermal Casimir force becomes non-monotonic below a critical temperature $e^{-KL}$8, because temperature reweights large worldlines of size $e^{-KL}$9 and curvature determines which of those fluctuations intersect both bodies (Weber et al., 2010).

5. Geometry as an inverse probe and data representation

The planar thin-film geometry is unusual in that it is not only a forward Casimir problem but also an inverse one. The task is: given the Casimir force or pressure as a function of separation, reconstruct the film thickness $d$0 and the parameters of $d$1. Because $d$2 is infinite-dimensional, the problem is reduced to finite dimension by the $d$3-pole Lorentz–Drude parametrization (Iizuka et al., 17 Apr 2026).

The machine-learning representation encodes geometry directly in the input vector. Instead of feeding the pressure itself, the method uses the spatial derivative of the normalized pressure,

$d$4

sampled as $d$5 at 20 logarithmically spaced separations. Each entry in the input therefore corresponds to a distinct geometric scale $d$6, and hence to a distinct spectral cutoff $d$7. The output vector contains $d$8 and all Lorentz–Drude parameters, and the loss is a mean squared error in logarithmic space so that nanometer and frequency scales contribute comparably (Iizuka et al., 17 Apr 2026).

The quantitative reconstruction results make the role of geometry explicit. In a two-pole model over $d$9 and $t$0, the RMSE in $t$1 is $t$2 to $t$3, and the RMSE in $t$4 is $t$5. In the broader four-pole model over the same geometric ranges, the RMSE in $t$6 is $t$7, resonance frequencies are reconstructed with RMSE $t$8, plasma frequencies with RMSE $t$9, and damping parameters with RMSE $\varepsilon_d(\omega)$0. In the more realistic range $\varepsilon_d(\omega)$1, $\varepsilon_d(\omega)$2, the RMSE in $\varepsilon_d(\omega)$3 improves to $\varepsilon_d(\omega)$4, and a crystalline silicon test reproduces the main Palik spectral features, with larger discrepancies only at the lowest frequencies (Iizuka et al., 17 Apr 2026).

These results show that Casimir geometry can act as a broadband probe. High-frequency resonances and plasma frequencies are relatively well constrained because changing $\varepsilon_d(\omega)$5 scans the infrared-to-ultraviolet range, whereas damping rates and low-frequency permittivity remain weakly constrained because the force in this geometry is much less sensitive to losses at room temperature (Iizuka et al., 17 Apr 2026).

6. Practical limits, experimental realism, and broader meanings of the term

The inverse framework is built in the ideal planar limit, but its practical relevance depends on realistic geometric constraints. Representative experimental distance windows cited for related measurements are roughly $\varepsilon_d(\omega)$6 for parallel plates, $\varepsilon_d(\omega)$7 for sphere–plate, and $\varepsilon_d(\omega)$8 for near-field radiative heat transfer. This motivates the more realistic training range $\varepsilon_d(\omega)$9 in the thin-film study (Iizuka et al., 17 Apr 2026).

At the same time, idealized geometry remains a major limitation. Finite-size corrections, alignment errors, surface roughness, and imperfect planarity are omitted from the forward model. For actual deployment, such effects would need to be incorporated into the solver or folded into the uncertainty budget; otherwise the inversion would misinterpret geometric imperfections as changes in $$\text{Au (3)} \;|\; \text{film (2)} \;|\; \text{vacuum (0)} \;|\; \text{Au (1)}.$$0 or $$\text{Au (3)} \;|\; \text{film (2)} \;|\; \text{vacuum (0)} \;|\; \text{Au (1)}.$$1. Noise handling is similarly geometry sensitive: a denoising autoencoder can mitigate moderate Gaussian noise with standard deviation $$\text{Au (3)} \;|\; \text{film (2)} \;|\; \text{vacuum (0)} \;|\; \text{Au (1)}.$$2 in normalized force, but residual denoising error directly degrades permittivity reconstruction (Iizuka et al., 17 Apr 2026).

More broadly, “Casimir geometry” is not restricted to electromagnetic vacuum forces between conductors. In statistical-mechanical film geometries, the finite-size force is determined by one finite thickness $$\text{Au (3)} \;|\; \text{film (2)} \;|\; \text{vacuum (0)} \;|\; \text{Au (1)}.$$3 and the imposed boundary condition; for the relativistic Bose gas with periodic boundary conditions, the Casimir force scales as $$\text{Au (3)} \;|\; \text{film (2)} \;|\; \text{vacuum (0)} \;|\; \text{Au (1)}.$$4 and has exact universal amplitude $$\text{Au (3)} \;|\; \text{film (2)} \;|\; \text{vacuum (0)} \;|\; \text{Au (1)}.$$5 (Dantchev, 2019). In quantum antiferromagnets, two domain walls defining a slab of thickness $$\text{Au (3)} \;|\; \text{film (2)} \;|\; \text{vacuum (0)} \;|\; \text{Au (1)}.$$6 generate a Casimir-type interaction of magnon fluctuations that decays as $$\text{Au (3)} \;|\; \text{film (2)} \;|\; \text{vacuum (0)} \;|\; \text{Au (1)}.$$7, with strong anisotropy set by wall orientation (Jagannathan et al., 2012). In two-component Bose–Einstein condensates confined by parallel plates, the plate geometry discretizes Bogoliubov modes and yields a Casimir force that is additive over components in one-loop approximation and vanishes for large plate spacing, zero intraspecies interaction, or full strong segregation (Thu, 2016).

In that extended usage, the phrase designates any configuration in which geometry and boundary conditions determine a fluctuation spectrum whose finite-size modification becomes measurable as a force, free-energy shift, or inverse-sensing signal. The electromagnetic thin-film planar problem makes this logic especially explicit: geometry, through $$\text{Au (3)} \;|\; \text{film (2)} \;|\; \text{vacuum (0)} \;|\; \text{Au (1)}.$$8 and $$\text{Au (3)} \;|\; \text{film (2)} \;|\; \text{vacuum (0)} \;|\; \text{Au (1)}.$$9, does not merely host the Casimir effect but performs the spectral selection that makes material characterization possible (Iizuka et al., 17 Apr 2026).