Power-Plane Cavity Resonance and Edge Radiation

Two planes are a waveguide

In schematic diagrams, power and ground connections are drawn as simple node symbols or infinite zero-ohm buses. Even in PCB layout tools, planes appear as quiet, solid floods of continuous copper that comfortably carry DC currents. But when fast transient currents strike a circuit board, a pair of conductive planes behaves nothing like a static wire.

A power plane separated from a ground plane by a thin sheet of dielectric forms a two-dimensional electromagnetic cavity resonator — behaving exactly like an acoustic drum skin or a microwave waveguide. Because the boundaries of the PCB are open circuits, high-frequency electromagnetic waves propagating across the board cannot simply escape. They bounce back and forth off the perimeter. At particular natural resonant frequencies set purely by the board's length, width, and dielectric constant, these reflections interfere constructively into violent standing waves. At these resonant modes, the impedance between power and ground explodes into tall peaks, and the perimeter of the board turns into an accidental slot antenna radiating RF energy into the room.

Think about what happens when a chip suddenly draws current from a plane pair. Charge leaves the plane near the chip's power pins, and the disturbance spreads outward — as a wave, at the speed of light in the dielectric, in all directions.

That wave reaches the edge of the board. The edge is an open circuit, so it reflects, exactly as a signal reflects off an unterminated line end. The reflection travels back, meets the far edge, reflects again.

At most frequencies those reflections arrive back out of step with each other and average away. But at the frequencies where the board's dimensions happen to fit a whole number of half-wavelengths, every reflection arrives in phase with the next, and they reinforce. That is a standing wave, and the plane pair has become a resonant cavity.

Where the modes land

Two planes with dielectric between them are a waveguide. Energy injected between them propagates outward, reflects off the open board edges, and forms standing waves at the frequencies where the board dimensions fit a whole number of half-wavelengths:

f_mn = ( c / 2√D_k ) · √( (m/a)² + (n/b)² ) In the ideal rectangular model, a 100 × 80 mm plane pair in D_k 4.2 has its first mode at about 730 MHz, with the next few following close behind. Lateral dimensions and dielectric constant set this estimate; plane spacing mainly changes cavity impedance.

At a resonance the impedance between the planes is not uniform. It is lower near voltage nodes and higher near anti-nodes, so impact depends on where current is injected and voltage is observed. Moving a die or connector changes its coupling to the mode even when the stackup and netlist stay the same.

Why the usual fix does not apply

At 730 MHz, many conventionally mounted board-capacitor branches are already inductive and cannot be treated as ideal shunts. A capacitor or reactive branch placed at an anti-node can load and shift the resonance; meaningful damping requires enough loss in the branch and a mounting path that remains effective at the mode frequency.

What genuinely works, in rough order of cost:

Stitch the plane edges — between planes of the SAME net. Ground vias around the perimeter, tying a ground plane to the other ground planes above and below it, shorten that cavity's open boundary and cut the radiation from the edges. Cheap, and it helps emissions at the same time.
Be precise about which nets: this does not mean connecting a power plane to ground. A power/ground cavity has a DC voltage across it and cannot be stitched shut with vias. If you want to damp that cavity you need a deliberate network — capacitors around the perimeter, or a lossy embedded-capacitance layer — and the capacitors stop helping above their own mounting resonance, which is usually below the cavity modes. Draw the actual nets before acting on any generic instruction to “add stitching vias”.
Thin the plane-pair dielectric. Bringing the planes closer raises the capacitance per unit area and lowers the cavity impedance across the board. It also improves every decoupling loop on that pair.
Use a lossy or embedded-capacitance layer. A thin, high-D_k, deliberately lossy laminate between the planes can lower mode Q by removing energy.
Break up the plane. A smaller cavity resonates higher; splitting a large plane into regions can push every mode above the band of interest. This trades against return-path continuity, so it is rarely free.

The EMI connection An excited cavity radiates from its open edges, and this is one of the more common root causes of an emissions peak whose source is not obvious from the signal list. If its frequency is close to a calculated plane mode, map the field or probe response versus position and vary candidate excitations before choosing between boundary, stackup, and routing changes.

Why this decides how your board behaves

Impedance becomes a function of position. At a resonance the plane pair is near a short at the nodes and near an open at the anti-nodes. The same board can be fine with the die in one place and marginal in another because the port couples to a different part of the mode shape.
It radiates. A resonating cavity with open edges is an antenna whose aperture is the board's perimeter. Plane resonance is one of the most common causes of a radiated emissions failure that has no obvious signal-path source.
The cavity creates transfer impedance between locations. Current injected at one port can produce voltage at another through a shared plane-pair mode, even when the functional circuits are on separate nets.
Ordinary decoupling may have little effect at the mode frequency. When the mounted branch is already inductive, adding nominal capacitance does not create the intended low-impedance shunt. Evaluate the mounted network and mode location.

What to do about it

Stitch the plane edges — between planes of the same net. Ground vias around the perimeter tie a ground plane to the other ground planes above and below it, shortening that cavity's open boundary and cutting the radiating aperture. This is cheap and it is the first thing to do.

Thin the plane-pair dielectric. A thinner gap lowers the cavity's characteristic impedance, which lowers the voltage a given current develops at an anti-node. It also adds distributed capacitance and improves every capacitor's mounting loop. One change, three benefits.

Place the die away from an anti-node if you have the choice. Floorplanning beats component selection here, and it is free at the right moment and impossible later.

Damp deliberately if you must. Terminating the cavity edge with series RC networks — a resistor near the cavity's characteristic impedance in series with a capacitor to block DC — absorbs energy at the modes rather than relocating them. It costs parts and board area; unlike a nearly reactive load, its resistor is intended to remove modal energy.

Using this to find a fault

An emissions peak with no obvious signal harmonic at that frequency. Compute the board's first few modes from its dimensions. A match makes cavity excitation a useful hypothesis; confirm it with spatial probing, source changes, or a field model.
Noise that depends on where you probe the plane. That position dependence is the signature. A lumped PDN problem looks the same everywhere; a cavity mode does not.
A problem that changes with board outline. Lateral dimensions set the ideal mode frequencies; dielectric constant, cutouts, vias, and boundary conditions perturb them.
Two rails interfering with no shared net. Look for a shared plane pair and check whether the interference frequency is a cavity mode.

Use the closed-form modes as a screening calculation The equation assumes solid rectangular planes, uniform dielectric, and idealised open edges. Splits, antipad fields, connectors, stitching, loss, and attached structures change both mode frequency and Q. Use the first modes to choose investigation frequencies, then evaluate the real geometry with ports at the actual excitation and observation locations.

Go deeper — the mode formula, and why a capacitor loads rather than damps

For a rectangular cavity of dimensions a × b with open edges, the resonant frequencies are:

f_mn = ( c / 2√D_k ) · √( (m/a)² + (n/b)² ) m, n are integers naming the mode; (1,0) is the lowest, with half a wavelength across the long dimension

Notice what is absent: the plane separation. The modes are set by the board's lateral dimensions and the dielectric constant, and the stackup barely moves them. What the separation does change is the cavity's characteristic impedance, and therefore how much voltage a given injected current produces — which is why thinning helps even though it does not move the resonance.

Why a capacitor loads rather than damps. Damping requires removing energy, which requires resistance. Above its self-resonant frequency a capacitor is an inductance with a small series resistance, and connecting a reactance across a resonant cavity shifts its resonant frequency without absorbing much. The mode moves; the Q does not fall. That is the precise sense in which "add more decoupling" fails here, and it is different from the way it fails for anti-resonance.

Where the mode shape matters. At the (1,0) mode the voltage is maximum at the two short edges and zero along the centreline. A die sitting on the centreline sees almost nothing of that mode; one near an edge sees all of it. Higher modes have more complicated patterns, so a placement that avoids one may sit on another — which is why this is checked with a field solve rather than by hand past the first few.

One honest limitation of the cavity picture: it assumes solid, rectangular planes. Real planes have via antipads, splits and cutouts, and each of those perturbs the mode shapes and can create local resonances that the closed form does not predict. The formula tells you roughly where to look; it does not substitute for solving the actual geometry.

In the real world

Plane resonance is the clearest example on this site of a problem whose fix is structural and whose window closes early. Stitching, dielectric thickness and die placement are all decided before layout is finished, and none of them can be retrofitted onto a fabricated board.

Which makes the first few modes worth computing at the start of a project rather than after an emissions failure. It is one line of arithmetic from the board outline, and it tells you which frequencies to keep an eye on for the rest of the design.