A ferrite bead is not an inductor

Z = R + jX: how a bead changes with frequency

A ferrite bead is a conductor passed once through a ferrite core, not a coil. Its equivalent circuit combines the DC resistance Rdc, a resistance Rcore for the core's magnetic loss, the inductance L, and a stray capacitance Cp across the terminals. L and Rcore are in parallel, Rdc is in series with them, and Cp sits across the whole thing.

The impedance is Z = R + jX, and R (real part) and X (imaginary part) behave differently with frequency. At low frequency ωL ≪ Rcore, so current flows almost entirely through L and the inductive reactance X dominates. As in any inductor, the energy does not become heat; it moves back and forth between the supply and the capacitor.

As frequency rises and ωL approaches Rcore, part of the current flows into Rcore and R rises sharply. The datasheet's "600Ω at 100MHz" is |Z| in this band where R dominates. Modeled with typical values back-calculated to give 600Ω at 100MHz (L = 1.1µH, Rcore = 620Ω, Rdc = 0.08Ω, Cp = 2.1pF), R and X cross at about 60MHz, and at 100MHz R ≈ 616Ω, X ≈ 48Ω and |Z| ≈ 618Ω, almost a pure resistance.

Higher up, Cp dominates and X turns negative (capacitive). In this model the self-resonance (X = 0) is at about 105MHz, and at 200MHz R ≈ 257Ω, X ≈ −306Ω and |Z| ≈ 399Ω. From low frequency upward the bead has three faces, inductor, resistor, capacitor, and "600Ω at 100MHz" captures only the band where it acts as a resistor.

"600Ω at 100MHz" means a part that turns noise into heat

At the same impedance, a pure inductor can only produce X at high frequency, so the current it blocks is simply reflected back. The noise energy stays in the circuit and becomes ringing if there is something to resonate with.

A bead's Rcore is a real loss (a resistance), so when noise current passes at around 100MHz most of it is dissipated in the R = 616Ω part and disappears as the core's magnetic loss, as heat. Because it absorbs rather than reflects, it is less likely to add a new peak to unintended resonances in the surrounding wiring.

That is why choosing a part by "600Ω at 100MHz" alone is not enough. |Z| is the vector sum of R and X, so at the same 600Ω some parts are almost all R (suited to EMI suppression) and others still carry X (closer to low-frequency inductor use). How much R there is at the frequency you want to suppress can be read from an R-X plot, or from |Z| with the phase angle, or a table listing R and X separately.

DC current saturates the core: Z drops a lot even at half the rated current

When DC current (a bias current on top of the signal) flows through the ferrite core, the core approaches magnetic saturation and permeability falls. L and Rcore both fall, so the whole Z-f curve sags. This is the DC bias characteristic.

Saturation is not "fine up to the rated current"; the impedance falls continuously, well below the rated current. Applying the permeability drop to L and Rcore with a single factor (saturation factor s) in the model above, s = 0.55 (about 50% of rated current) takes Z at 100MHz from 618Ω to 311Ω (about 50%), and s = 0.3 (rated current) to 156Ω (about 25%). "Half the rated current, so it is safe" is not true: for EMI suppression the effect is already halved.

A datasheet's DC bias plot has DC current on the horizontal axis and |Z| at that current on the vertical (usually at one representative frequency such as 100MHz). Read Z at your operating current as an absolute value, not as a percentage of rating. A part whose Z halves at 50% of Irated is not unusual, and the smaller and lower-profile the part, the steeper the drop.

The rated current is also not set by saturation. Many manufacturers define it as the current at which self-heating under continuous load does not exceed a specified temperature rise (ΔT of about 20-40℃, depending on the part and maker). Saturation starts at a much lower current, so for EMI work check Z at the actual current on the DC bias plot.

Resonance with the downstream capacitor makes low-frequency noise worse

If you assume "a bead in series with the supply line is always safe," you overlook that the bead's L and the downstream decoupling capacitor form an LC low-pass filter. At low frequency a bead has almost no resistance (Rdc = 0.08Ω in the model above) and is nearly a pure L, so this filter peaks at resonance unless it is damped.

A 1.1µH bead (its low-frequency inductance) with a 10µF decoupling capacitor resonates at f0 = 1 / (2π√(LC)) ≈ 48kHz. With Q = (1/R)√(L/C) and only the bead's Rdc = 0.08Ω as R (the worst case, ignoring electrolytic ESR and wiring resistance), Q ≈ 4.15, and the capacitor-side voltage relative to the noise voltage ahead of the bead has a peak of about 12.4dB (about 4.2 times) at resonance.

So noise near 48kHz is larger beyond the capacitor than without the bead. A bead works at high frequency (MHz to 100MHz), but at low frequency (tens of kHz to a few MHz) it is just a small inductor and forms unintended peaking with the downstream C. If noise from the supply or digital circuits overlaps this resonant frequency, the result is worse on the bench.

How to add damping

The fix is to add resistance to the resonant circuit and lower Q. Adding a series resistor Rd = 0.5Ω to the LC filter above lowers Q to about 0.57 and the 12.4dB peak disappears. There are three ways.

1. Put a resistor of a few hundred mΩ to 1Ω in series between the bead and the capacitor. It is reliable, but the voltage drop and heat increase, so it suits analog circuits with small load current.

2. Use a capacitor with higher ESR, or add one. Multilayer ceramics have a low ESR of a few mΩ to tens of mΩ and give no damping. Adding a tantalum or conductive polymer capacitor (ESR tens to hundreds of mΩ) in parallel makes that ESR act as the damping resistance.

3. Choose a bead with a higher Rdc (a few hundred mΩ). Where no large current is needed and the bead is only for EMI, it doubles as damping. It also adds voltage drop, so it trades off against the IR drop in the next section.

Voltage drop and ringing on analog supplies with sudden load changes

Putting a bead in series with a load that draws a few mA to tens of mA briefly at each conversion, such as an ADC's AVDD, causes two problems.

One is the IR drop across Rdc. A 20mA current step through a part with Rdc = 0.08Ω drops 0.08Ω × 20mA = 1.6mV. For a 12-bit, 3.3V full-scale ADC, 1LSB is about 0.8mV, so this alone is a 2LSB error.

The other is ringing from the LC resonance above. When the load current changes in a step, the second-order system of the bead's L and the decoupling C is excited. For Q = 4.15 (undamped), the overshoot is theoretically about 68% of the step (from exp(−πζ/√(1−ζ²)) with ζ = 1/(2Q)). The characteristic impedance √(L/C) ≈ 0.33Ω times a 20mA step gives an amplitude of about 6.6mV, much larger than the IR drop, at about 48kHz and ringing for tens of µs after each step.

If this ringing lands right after the ADC samples, the result shifts by a few LSB only when it coincides with the conversion, a hard-to-reproduce fault. Departing from the decoupling values the ADC datasheet specifies (often "no bead, C = 0.1µF + 1µF close by") and combining a large bead with a large C on your own judgment makes this resonance likely.

Where the noise current goes

Figure 5 contrasts where noise current goes through the bead in each frequency band. At low frequency the current passes through the bead almost unattenuated and reaches the downstream capacitor (the resonance in the previous section occurs here). Around 100MHz most of the current turns into heat inside the bead.

Where to use a bead and where not to

A bead helps only when you want to turn noise in the band where R is high into heat along the current path.

Where it bites on real hardware

• Noise got worse after adding a bead to the supply line. Suspect LC peaking first. Compare the voltage before and after the bead on a scope; if the downstream side is larger between tens of kHz and a few MHz, it is resonance, and a temporary series damping resistor that settles it confirms it.
• A large bead and a large C were added to an ADC or DAC analog supply. Swapping the datasheet's parts for larger ones lowers the resonant frequency, and the larger L also raises Q, so the ringing can get worse. If you change the values, calculate the resonant frequency and Q first.

Evaluate the bead as a whole system, including the downstream C, wiring inductance and load impedance. As with real RC filter attenuation, a component's own specs do not determine the behavior. Also see passive versus active filters, MLCC DC bias and via inductance.