Why a microwave oven is not tuned to water

A widely repeated explanation of the microwave oven says that 2.45 GHz is “the resonant frequency of water.” The model used here for the dominant microwave response of bulk liquid water—Debye relaxation—contains no resonance at all. Its dielectric loss factor peaks at 19.2 GHz, not 2.45 GHz, but even that comparison needs care: a maximum of \(\varepsilon''\) is not automatically a maximum of absorbed power or wave attenuation.^1,2

The explanatory question for this note is: how does Debye relaxation turn an oscillating electric field into heat, and what does the model actually predict for dielectric loss and penetration depth? The route runs from a collective time constant, through the complex permittivity it implies, to the loss tangent and plane-wave penetration depth, and finally to ice, where the same idealized channel becomes extremely weak.

A dipole in a crowd

A water molecule carries a permanent electric dipole moment. In a static field, thermal motion keeps the dipoles mostly randomized, but a slight average alignment survives. This orientation polarization is what makes liquid water’s static relative permittivity so large: the value selected here is \(\varepsilon_s =\) 78.4 at 25 °C.^3,4

The dynamics live in one question: if the field is suddenly switched off, how fast does that collective alignment decay? Debye represented the relaxation as a single exponential,

\[ P(t) = P(0)e^{-t/\tau}. \]

The dominant liquid-water dielectric process is a collective relaxation of the orientational polarization, coupled to rearrangement of the hydrogen-bond network; \(\tau\) should not be read as the literal turn time of one isolated molecule. The selected value at 25 °C is 8.3 ps.^3,5

An exponential decay is the signature of an overdamped process. A resonance—an underdamped response with a natural frequency—would ring while it decayed. For bulk liquid water at 2.45 GHz, the modeled response is instead a broad classical relaxation. A photon at that frequency carries 10 \(\mu\)eV, while \(k_BT\) at 298.15 K is 25.7 meV: the photon energy is lower by a factor of 2.5×10³. That energy comparison locates the response in the classical regime; the observed dielectric spectrum, rather than the comparison alone, is what distinguishes a broad relaxation from a narrow molecular line.²

From a decay time to a dispersion curve

Take fields proportional to \(e^{-i\omega t}\). With that convention, passive loss has positive imaginary permittivity, \(\hat\varepsilon=\varepsilon'+i\varepsilon''\), and the one-sided response of the exponential is

\[ \hat{\varepsilon}(\omega)=\varepsilon_\infty+ \frac{\varepsilon_s-\varepsilon_\infty}{1-i\omega\tau}. \]

Separating real and imaginary parts gives

\[ \varepsilon'(\omega)=\varepsilon_\infty+ \frac{\varepsilon_s-\varepsilon_\infty}{1+\omega^2\tau^2}, \qquad \varepsilon''(\omega)= \frac{(\varepsilon_s-\varepsilon_\infty)\omega\tau} {1+\omega^2\tau^2}. \]

\(\varepsilon'\) is the in-phase response associated with refraction and energy storage. \(\varepsilon''\) is the quadrature component responsible for net work, with time-averaged volumetric dissipation

\[ \langle p\rangle=\tfrac12\varepsilon_0\omega\varepsilon''|E|^2. \]

The two parts are one causal response viewed in quadrature, the same structure explored at optical frequencies in refraction is absorption you cannot see.⁶

The one-pole model uses \(\varepsilon_\infty=\) 5.2 as an effective background for processes faster than the dominant relaxation. It is not the literal infinite-frequency or optical permittivity: reference models resolve additional faster relaxations and resonances.²

The dielectric loss factor \(\varepsilon''\) has a single maximum at \(\omega\tau=1\),

\[ f_{\varepsilon'',\mathrm{peak}}=\frac{1}{2\pi\tau}. \]

For the selected relaxation time, that is 19.2 GHz. This is specifically the Debye loss-factor peak. It is not an absorption or attenuation maximum: \(\langle p\rangle\) contains the additional factor \(\omega\), while propagation also depends on \(\varepsilon'\) and the frequency-dependent wave number. The 60 GHz single-Debye extrapolation below has a smaller \(\varepsilon''\) than the peak but a still shorter penetration depth.

Nor does the loss-factor peak explain the oven frequency. The 2.4–2.5 GHz band, centered at 2.45 GHz, is internationally designated for industrial, scientific, and medical applications. In ITU Region 2, 902–928 MHz, centered at 915 MHz, is designated as well. Those allocations establish where heating equipment may operate; they do not identify a molecular resonance.⁷

Putting numbers in

For a homogeneous, nonmagnetic medium and a plane wave, write

\[ k=\frac{\omega}{c}\sqrt{\varepsilon'+i\varepsilon''},\qquad \alpha=\operatorname{Im}k,\qquad D_p=\frac{1}{2\alpha}. \]

The field amplitude decays as \(e^{-\alpha z}\), so \(D_p\) is the distance over which transmitted power falls to \(1/e\) relative to its value just inside the material. The complete standard-library calculation validates the declared inputs and writes canonical JSON. Code 1 shows only its numerical core.

omega_tau = 2 * math.pi * frequency_hz * relaxation_time_s
epsilon = epsilon_infinity + (epsilon_static - epsilon_infinity) / (1 - 1j * omega_tau)
wave_index = cmath.sqrt(epsilon)
alpha = (2 * math.pi * frequency_hz / speed_of_light) * wave_index.imag
power_penetration_depth_m = 1 / (2 * alpha)

Code 1. Core of the single-Debye evaluation using the declared \(e^{-i\omega t}\) convention. The complete program adds input validation, canonical output, independent consistency checks, and non-writing --check mode.

Table 1 is the generated publication projection for liquid water. Its final row is explicitly a one-pole extrapolation, included to expose the difference between a loss-factor peak and attenuation rather than to claim high-frequency accuracy.

Table 1. Single-Debye relative permittivity, loss tangent, and \(1/e\) power penetration depth for liquid water at 25 °C.

At 2.45 GHz, the model gives \(\tan\delta=\) 0.119; the corresponding total dielectric loss angle, \(\arctan(\varepsilon''/\varepsilon')\), is 6.8°. Moving up to the \(\varepsilon''\) peak makes the modeled penetration depth only 0.5 mm, and moving to 60 GHz makes it shorter still. Maximizing \(\varepsilon''\) is therefore not the same engineering objective as depositing energy through a volume.

At 2.45 GHz, this homogeneous pure-water model gives a power depth of 1.9 cm; at 915 MHz it gives 13.2 cm. These are not penetration depths for a particular food. Salt, fat, bound water, temperature, geometry, interfaces, and the cavity field all change the result, and measured depths in foods can be much shorter. The direction of the comparison does survive in practice: 915 MHz generally penetrates foods more deeply than 2.45 GHz and is used in large-scale processing.⁸

The ice test

The model makes a sharp phase comparison, but it needs temperature-consistent inputs. For polycrystalline ice at -10 °C, a measured relaxation frequency of 2.80 kHz corresponds to \(\tau=\) 5.68×10⁻⁵ s, about 6.8 orders of magnitude slower than the selected liquid-water process.⁹

Table 2 runs the same calculation for both phases at 2.45 GHz. The approximate ice values \(\varepsilon_s=\) 97.0 and \(\varepsilon_\infty=\) 3.2 are representative inputs to this idealized comparison, not a complete microwave model for frozen food.³

Table 2. Liquid water at 25 °C and idealized ice at -10 °C evaluated at 2.45 GHz with the same single-Debye equations.

For ice, \(\omega\tau=\) 8.75×10⁵. The slow orientational increment therefore contributes negligibly at the oven frequency; faster polarization still supplies \(\varepsilon'\approx\) 3.20, while the modeled loss is tiny. Real frozen food is not pure ice, however. It contains solutes, interfaces, and unfrozen or newly thawed water, so energy deposition changes sharply across a thawing portion. Already-thawed regions can heat preferentially and produce thermal runaway while other regions remain frozen.¹⁰

Reduced-power defrost modes lower the time-averaged input. Conventional ovens commonly do that by cycling the magnetron, leaving off periods in which heat can redistribute; the strategy mitigates the contrast but does not guarantee uniform thawing.^10,11

Where the model stops

The single-relaxation-time model earns its place by being solvable, but its boundary should be drawn plainly. It does not explain microscopically why the selected liquid-water time is 8.3 ps; that value enters from measurement. Real water has several relaxation components, so \(\varepsilon_\infty=\) 5.2 is only an effective one-pole background and the 60 GHz row is an extrapolation.²

The parameters vary strongly with temperature. Ionic conduction adds a term \(\sigma/(\omega\varepsilon_0)\) in salty material. Food adds bound water, fats, pores, interfaces, and spatially varying composition. A kitchen oven adds reflection at the surface, standing-wave cavity modes, finite geometry, and thermal transport. The ice calculation omits impurities and liquid inclusions, which can dominate the loss of real frozen material.

Within that boundary, the model answers the opening mechanism question: microwave heating of liquid water comes from work done against a lagging, collective polarization. It is a broad relaxation, not a narrow resonance at the oven frequency. What the model does not do is predict the temperature field in dinner.

Reproducibility

The reviewed experiment bundle contains the declared inputs, complete calculation, canonical results, metrics generator, publication projection, environment record, and reviewed public-file manifest. Code 2 checks the complete chain.

python3 research/microwave-debye-relaxation/calculate.py --check
node research/microwave-debye-relaxation/generate-metrics.mjs --check
node scripts/verify-metrics.mjs
stack test
stack exec site rebuild
node scripts/verify-site.mjs

Code 2. Recalculate the canonical result in non-writing check mode, verify the typed projection and source fingerprints, test and rebuild the site, and inspect the generated links and failure markers.

The calculation uses only the Python standard library, with no random inputs, downloads, services, or credentials. In the documented environment it is end-to-end reproducible: declared inputs regenerate canonical results and the typed publication projection. The post’s generated values resolve from that projection rather than from copied output. This establishes their chain of custody, not the correctness of the physical model or its interpretation. If a reader sees a better parameter set, a missing loss channel, or an overextended inference, that is exactly where this account should be corrected.

References