Ceramic capacitors lose capacitance when you apply voltage

Why voltage reduces capacitance

X5R and X7R multilayer ceramic capacitors (MLCCs) use a barium-titanate-based ferroelectric as the dielectric. Under an electric field, ferroelectric polarization approaches saturation, so raising the field further adds less and less polarization — in other words, the permittivity drops. The permittivity under a DC bias is lower than at zero volts, and that is the DC bias characteristic.

How much it drops depends on the field strength across each dielectric layer. For the same capacitance, a smaller case size or a lower rated voltage means a thinner dielectric, so the same applied voltage produces a stronger field and a larger drop. A larger case or higher rated voltage means a thicker dielectric and a smaller drop.

C0G (NP0) uses a paraelectric material, so its DC bias shift is essentially zero. Its permittivity is low, though, so for the same case size it only reaches 1/10 to 1/100 the capacitance of X7R.

How much it actually drops

This varies by manufacturer and part, so always check the specific characteristic curve. As a rough guide, here is the capacitance a 10µF part retains at 3.3 V.

A nominal 10µF part can differ threefold in effective capacitance at 3.3 V depending on whether it's a 1005/6.3V or a 3216/25V. If the circuit needs 10µF, do not pick the smallest part rated 10µF — pick the part that still measures 10µF at your operating voltage. That may end up being a 22µF or 47µF part number.

Keeping the applied voltage at a third or less of the rating keeps the loss under 20% for most parts. Choosing 10V or higher for a 3.3V rail, 16V or higher for 5V, and 35-50V for 12V, keeps the DC bias effect small.

Temperature, aging and tolerance stack on top

DC bias isn't the only factor. Three more multiply into the effective capacitance.

Temperature coefficient. X7R holds ±15% over -55 to +125°C; X5R holds ±15% over -55 to +85°C. Y5V runs +22/-82% over -30 to +85°C, leaving only 20% of capacitance at high temperature. Y5V should not be used for power decoupling.

Aging. The ferroelectric domains realign over time and capacitance falls logarithmically. For X7R that's 1-2.5% per decade of time (-3 to -7% at 1000 hours). Soldering resets it once the part exceeds the Curie temperature (about 125°C), and the decline starts again.

Tolerance. K = ±10%, M = ±20%.

Stack these worst-case for a 1608/6.3V/10µF X5R, grade M part at 3.3V, 85°C, one year in: 10µF × 0.8 (tolerance) × 0.45 (DC bias) × 0.85 (temperature) × 0.95 (aging) ≈ 2.9µF. Thirty percent of the catalog value.

Everything except tolerance is on the datasheet's characteristic curves. Most manufacturers' web-based part search tools plot the DC bias curve per part number — always check it when selecting a part.

Where the lost capacitance matters

LDO stability. An LDO's output capacitor is part of its phase compensation. When the datasheet says "1µF ceramic minimum at the output," that's the effective value. Use a 0402/6.3V/1µF part on a 3.3V rail and you get an effective 0.5µF, which can fall short of spec and oscillate.

Switching regulator output ripple and the control loop. Halve the output capacitance and ripple voltage doubles, while the control loop's crossover frequency rises and phase margin shrinks. Design without accounting for effective capacitance and the real board tends toward marginal oscillation.

Time constants and filters. RC time constants and LC cutoffs shift in proportion to the capacitance. Use C0G for timing capacitors, analog filters, and oscillator circuits.

Acoustic noise. Ferroelectrics are also piezoelectric, so an AC voltage makes them vibrate mechanically and ring the board ("singing capacitors"). This shows up with X7R on a supply carrying audible ripple, or on an audio coupling capacitor. The fix is C0G, a film capacitor, or an anti-singing part (e.g. with metal terminals).

Selection procedure

1. Determine the effective capacitance the circuit actually needs, at the operating voltage and maximum temperature.
2. Read the retained fraction at the operating voltage off the candidate part's DC bias curve, then multiply by the temperature coefficient (±15%) and tolerance.
3. If the effective capacitance falls short, look for a part with a higher rated voltage, a larger case, or a higher nominal capacitance.
4. Use C0G for timing, filtering, oscillators, and audio. C0G above 1µF gets large and expensive, so consider a film capacitor there too.
5. For verification, measure effective capacitance under DC bias with an LCR meter, or confirm it through circuit behavior (LDO transient response, ripple).