rf generator

An RF generator is the solid-state power amplifier that drives a plasma tool, locked to a single legal frequency — almost always 13.56 MHz, held to within ±0.05%, a window of just ±6.78 kHz — and its job is to push a clean sinewave into a load that refuses to hold still. It is not the matching network and it is not the plasma; it is the source, and the number on its front panel reads forward power in watts. What the wafer actually receives is set by how well that fixed-frequency energy couples into a plasma whose impedance drifts every time pressure, gas chemistry, or density changes. The generator's real problem has nothing to do with making plasma "stronger": hold the frequency dead steady, and keep delivering into a mismatch that is moving underneath it. **The generator's defining product is a fixed frequency, not adjustable power.** The 13.56 MHz standard exists because it is an ISM allocation under FCC Part 18 — a frequency the regulators reserved for industrial heating so a plasma tool does not jam radio traffic — and a generator that wanders outside its ±6.78 kHz window is not merely imprecise, it is out of compliance. Every other frequency in the fab is a relative of that anchor: 27.12 MHz and 40.68 MHz are its second and third harmonics, 2 MHz and 400 kHz sit below it for bias duty, and the 60 MHz and 100 MHz VHF sources sit above it for density. A modern generator synthesizes this reference from a crystal or direct-digital-synthesis core stable to parts per million, then amplifies it; the amplifier can be told to deliver anywhere from tens of watts to 15 kW, but the one thing it may not do is let the frequency slip, because the whole downstream chain — cable length, match component values, sheath dynamics — is tuned to that single number. **Modern generators are solid-state LDMOS amplifiers, and that architecture set what the box can do.** The vacuum-tube generators that ENI and others shipped into the first plasma tools have been replaced by laterally-diffused MOSFET (LDMOS) power stages running in switch-mode Class-D or Class-E, and the reason is efficiency: a linear Class-A stage runs at roughly 30% drain efficiency and a Class-AB at about 50%, while a switch-mode Class-D reaches 85% and Class-E 90%. To put 2,000 W into the cable, a Class-A design must draw 6,667 W from the rail and burn 4,667 W as heat, whereas a Class-E stage draws 2,222 W and dissipates only 222 W — a twenty-fold reduction in the heat the chassis must remove. Production generators from Advanced Energy, MKS/ENI, Comet, Kyosan, and Daihen reach multi-kilowatt ratings by combining pallets of 250 W to 1 kW LDMOS devices through hybrid combiners, so a 10 kW generator is really a dozen small amplifiers summed in phase, and losing one pallet degrades the output rather than killing it. **A plasma is a nonlinear load that manufactures its own harmonics, and the generator has to survive them.** The sheath in front of an electrode rectifies — it passes electron current in bursts and ion current as a trickle — so even a perfect 13.56 MHz drive comes back distorted, with energy at every harmonic. Modeling the sheath as a half-wave rectifier puts 21.2% of the peak current at the 27.12 MHz second harmonic and 4.2% at the 54.24 MHz fourth, and that reflected harmonic power flows back into the amplifier's finals. This is why a generator carries an output harmonic filter and why its transistors are rated to endure a reflection coefficient $|\Gamma|$ approaching 1: at ignition the chamber is nearly a pure reactance, essentially an open circuit, and for a few milliseconds every watt the generator makes bounces straight back. The protection strategy is foldback — the generator senses reflected power and throttles its own output to keep the finals inside their safe operating area — and a generator is specified to deliver into any load phase at high standing-wave ratio without damage precisely because ignition guarantees this on every strike. **Frequency-agile tuning replaced the moving match for the transitions that matter.** A traditional match network re-tunes by driving two vacuum-variable capacitors with stepper motors, and that mechanical stroke takes 0.1 to 2 s; a frequency-agile generator instead nudges its own drive frequency by a few percent to null the reflection electronically, in under a millisecond. The physics is simple: a matched network of loaded quality factor $Q$ = 5 converts a small fractional frequency shift into a canceling reactance of about ${2 Z_0 Q}$ per unit detuning, so a ±5% tuning window erases 25 $\Omega$ of drift-induced reactance and a ±10% window erases 50 $\Omega$. The diagram makes the consequence concrete. A fixed 13.56 MHz source facing 50 $\Omega$ of uncompensated reactance delivers only 80.0% of its forward power; the agile generator, having absorbed that same 50 $\Omega$ inside its ±5% window, still delivers 94.1%, and across the whole 25 $\Omega$ band it holds a flat 100% while the fixed source is already sliding down its curve. The generator has, in effect, swallowed the fast part of the matching job that the mechanical network is too slow to do. **Pulsing is a timing specification of the generator, not a plasma recipe.** Once processes began switching the RF on and off at kilohertz rates to control radical chemistry and surface charging, the burden landed on the generator to raise and collapse its envelope in microseconds and to hold the setpoint flat within each burst. A motor-driven match cannot follow this — its 1 s stroke tracks pulse rates no faster than about 1 Hz — and even a fast servo match tops out near 10 Hz, while a switched solid-state match reaches roughly 1 kHz and a frequency-agile generator re-nulls every pulse up to about 10 kHz, a 10,000-fold span shown as the log bars at right. The generator also keeps two amplifiers in step: a high-frequency source that sets plasma density and a low-frequency bias that sets ion energy are phase-locked and pulsed synchronously, sometimes in multi-level patterns that cycle among several power states within one period, and only a source whose matching is electronic can honor that timing. **The generator regulates forward power, but the wafer feels delivered power.** A dual-directional coupler at the output continuously measures forward and reflected power, and how the control loop uses those two numbers decides what "1,000 W" even means. In forward-power mode the servo holds the outgoing wave constant and lets the reflected fraction eat into what the plasma receives; against 50 $\Omega$ of residual reactance that is a 20.0% shortfall, and against 100 $\Omega$ a full 50.0% — the plasma gets half of what the panel claims. In load-power (delivered-power) mode the loop instead servos on forward minus reflected, pushing harder to make up the loss so the plasma sees the true setpoint. The choice matters for repeatability: two chambers with slightly different matches run identically in load-power mode and diverge in forward-power mode, which is why advanced etch and deposition recipes almost always specify delivered power and lean on the coupler, not the amplifier's output meter, as the source of truth. | Uncompensated reactance | Fixed source $|\Gamma|$ | Delivered, fixed | Delivered, agile ±5% | Delivered, agile ±10% | |---|---|---|---|---| | 0 $\Omega$ | 0.000 | 100.0% | 100.0% | 100.0% | | 15 $\Omega$ | 0.148 | 97.8% | 100.0% | 100.0% | | 30 $\Omega$ | 0.287 | 91.7% | 99.8% | 100.0% | | 50 $\Omega$ | 0.447 | 80.0% | 94.1% | 100.0% | | 75 $\Omega$ | 0.600 | 64.0% | 80.0% | 94.1% | | 100 $\Omega$ | 0.707 | 50.0% | 64.0% | 80.0% | ```flowchart Reference (13.56 MHz crystal / DDS core, +-6.78 kHz) -> LDMOS power stage (Class D/E, 85-90% efficient) -> Pallet combiner (sum of 250 W - 1 kW devices) -> Harmonic filter + dual-directional coupler -> measure forward / reflected -> control loop: hold forward power OR delivered power -> frequency-agile trim (+-5%, nulls reflection in <1 ms) -> match network -> plasma load ``` Read the RF generator through a *signal-source* lens rather than a *plasma-load* lens: its central problems are the ones every precision power source shares — synthesize a spectrally clean, frequency-stable carrier, amplify it efficiently without melting, survive a load that reflects everything you send during ignition, and close a control loop fast enough to track a target that moves at the pulse rate. The matching network worries about the plasma's impedance; the generator worries about frequency, ruggedness, and loop speed, and the modern design's decisive move was to fold the fast half of the matching job into the source itself by making frequency the tuning knob. Get the source right — a stable reference, an efficient rugged amplifier, and a coupler-based loop that regulates the power the wafer actually feels — and the plasma is driven cleanly; get it wrong and no matching network downstream can rebuild a carrier that was never clean or steady to begin with.