Ripple and Noise Suppression of Low-Ripple High-Voltage Power Supplies in High-Voltage Electrochemistry and Precision Metrology
Two applications look similar on a datasheet and behave differently in practice. High-voltage electrochemistry, whether electrophoretic deposition or electrochemical polishing, is sensitive to low-frequency ripple because it modulates the reaction rate and produces a deposit whose density varies with the ripple envelope. Precision metrology is sensitive to wideband noise because the supply is a reference and anything riding on the reference appears in the measurement. A supply specified only by total ripple in percent satisfies neither.
Ripple has to be decomposed before it can be suppressed. The components are the switching frequency and its harmonics, the line frequency and its harmonics, low-frequency modulation from the control loop, and random noise from the reference and the sense chain. Each has a different source and a different cure, and treating them as one number leads to adding filter capacitance until the response time is ruined without the low-frequency component being touched.
Switching ripple is the easiest to handle and the least likely to be the real problem. A multistage output filter with a small inductor between stages attenuates the switching component by sixty decibels while keeping the loop stable. Adding capacitance without the inductor produces a resonance with the cable inductance and can make the ripple worse at a particular load.
Line-frequency ripple comes from the input rectifier and appears at twice the line frequency. Feedback can suppress it only if the loop bandwidth is above the ripple frequency, which at high voltage and high impedance is rarely possible. Feed-forward from the rectified input, applied to the modulator, cancels most of it, and the remainder is handled by a passive filter sized for the specific frequency rather than for a decade band.
Low-frequency modulation is the component that matters in electrochemistry. It arises from thermal cycling, from the interaction of the voltage loop with a load whose impedance changes with the process, and from control loops that are close to instability. It appears as a slow envelope on the output, with a period of seconds to minutes, and it produces visible banding in a deposit.
Suppressing low-frequency modulation requires understanding its source rather than filtering it. If the source is thermal, the cure is thermal design; if the source is the load, the cure is a stiffer output or a separate current loop; if the source is the loop, the cure is more phase margin. A filter large enough to attenuate a slow envelope is large enough to make the supply unusable for any process that needs to change voltage.
Wideband noise in a metrology application is dominated by the reference and the sense chain. A buried Zener reference with a low-pass filter, a sense amplifier with a low noise front end, and a shielded, guarded layout reduce the noise floor by an order of magnitude compared with a general-purpose design. The high-voltage section contributes through the divider, which must be shielded against the electric field of its own high-voltage terminal.
Grounding and guarding decide whether the noise floor is achievable. A supply whose return carries the charge current of the output capacitance produces a voltage drop along the return conductor, and that drop appears in the measurement. A separate sense return, brought back to the reference point, and a guard ring driven at the same potential as the measurement node, remove both the drop and the leakage that would otherwise bypass the signal.
Measurement of the ripple itself is often wrong. A probe with a long ground lead picks up the magnetic field of the switching loop and reports ripple that does not exist at the output. A proper measurement uses a coaxial connection at the output terminals, a bandwidth limit appropriate to the application, and a check with the supply off to establish the noise floor of the measurement setup.
Electrochemistry benefits from current-mode control more than from voltage-mode. The reaction rate depends on charge, so a supply in current mode with low current ripple produces a uniform deposit even when the voltage wanders a little. Voltage-mode control with the same voltage ripple can produce a current ripple that is larger, because the load impedance varies during the process.
Metrology benefits from the opposite. A reference supply must hold voltage constant and let current do what it will, and the current noise then flows through the load rather than appearing on the reference. Choosing the control mode by what the application measures, rather than by what is convenient to specify, avoids a category of problems that no filter can fix.
Thermal design sets the noise floor for both applications over time. A divider that heats by ten degrees changes its ratio, and the change appears as drift that cannot be distinguished from noise in a short measurement. Keeping the divider at a stable temperature, either by thermal mass or by a small heater, is cheaper than any electronic correction and more reliable.
Isolation between the high-voltage domain and the control domain is a noise decision as well as a safety one. Fibre links for the sense signal eliminate the ground loop that would otherwise couple switching noise into the control. A supply that uses optocouplers with inadequate common-mode rejection will show noise that changes when the high-voltage cable is moved, which is a clear diagnostic of a ground loop.
Cost rises steeply as the ripple specification tightens. Reducing total ripple from one percent to a tenth requires a better filter and a faster loop; reducing it from a tenth to a hundredth requires a linear post-regulator or a different topology, and the efficiency and volume penalties become significant. The application's actual sensitivity should decide where on that curve to stop.
Specifications should state the bandwidth and the load condition. A ripple figure quoted at no load, with a twenty-megahertz bandwidth, tells nothing about behaviour at the operating point. Quoting ripple at the process current, with the measurement bandwidth stated, and separating the switching component from the low-frequency component, makes the number usable for a decision.
Limits exist. A supply cannot deliver lower noise than its reference and sense chain allow, and no amount of filtering fixes a reference that is thermally unstable. Similarly, a process that draws pulsed current will see voltage noise determined by the output impedance at the pulse frequency, regardless of the steady-state ripple specification.
A practical acceptance test separates the components. Measure output noise with a spectrum analyser from ten hertz to one megahertz at the process current; measure the low-frequency envelope over an hour with a data logger; measure the line-frequency component with a narrow-band filter. Three measurements, each with a known source, and the specification becomes a description of behaviour rather than a single number that hides everything.
Electrochemical polishing shows how the low-frequency envelope interacts with surface finish. The removal rate follows the charge delivered, and a slow ripple in the current produces bands whose spacing corresponds to the ripple period. A supply with a one-minute thermal envelope produces bands a centimetre apart on a polished surface, which is visible to the eye and impossible to polish out afterwards.
Electrophoretic deposition has the same sensitivity in a different form. A varying voltage produces a varying deposition rate, and the resulting layer has a density gradient through its thickness. Sintering then reveals the gradient as a difference in shrinkage, and the part warps. The supply's low-frequency behaviour is thus a dimensional-stability parameter in the finished component.
Precision metrology adds a requirement that ripple specifications do not capture: the noise must be uncorrelated with the measurement window. A supply whose noise is synchronous with a trigger will accumulate into a systematic offset in an integrating measurement, while random noise averages down. Checking the correlation between the noise and the instrument's timing is part of commissioning a reference supply.
Cable and load impedance shape the noise at the point of use. An output with a low source impedance at high frequency delivers a quiet voltage to a capacitive load and a noisy one to an inductive load, because the load converts voltage noise into current noise and back. Characterising the noise at the actual load, rather than at the terminals, is the only measurement that predicts what the process will see.
Compensation networks should be designed for the load, not the bench. A supply tuned on a resistive load can oscillate with a capacitive load at the same current, and the oscillation appears as a low-frequency envelope that no filter attenuates. Specifying the load range, and verifying stability at the extremes, prevents a class of failures that look like noise but are loop instability.
Records of noise at the operating point, taken monthly, catch degradation before it reaches product. A rise in the low-frequency envelope usually precedes a rise in the switching ripple, because thermal and reference drift happen before capacitor ageing becomes measurable. The trend, not the absolute value, is the useful signal.
Whatever the topology, the noise that reaches the process is the noise at the load, and the only way to know it is to measure there. A supply with an excellent terminal specification and a poor cable installation will deliver a worse result than a modest supply with a good installation, and the cost difference between the two installations is usually a few metres of shielded cable and one ground point chosen with care.

