01Why etch exists at all
Lithography decides where. Etch decides the actual three-dimensional geometry. Everything downstream inherits whatever shape the etch produced.
Early ICs transferred patterns with liquid chemistry. Wet etching is isotropic: it eats sideways as fast as down, so it undercuts the mask. Tolerable at 10 µm. Fatal once vertical dimensions approached horizontal ones.
Two independent lines converged on the modern tool. One physical — sputtering, to pattern metal stacks no wet chemistry could handle selectively. One chemical — oxygen plasmas for stripping resist (ashing), motivated by cutting liquid waste. They merged when reactive halogen gases went into a sputtering apparatus, giving reactive ion etching.
The payoff is anisotropic etching: near-vertical sidewalls that preserve the mask dimension. That property, not speed or cleanliness, is why plasma won.
02How the reactor got here
Worth knowing because reactor architecture is exactly what a surrogate model would be asked to help design.
Barrel reactors — wafers on a quartz boat in a tube with external electrodes. Fine for stripping; no temperature control, poor uniformity. Then the Reinberg radial-flow reactor made patterning production-viable.
Parallel plate — two configurations differing only in which electrode holds the wafer. Wafer on the grounded plate at high pressure was "plasma mode." Wafer on the powered plate at low pressure was RIE mode. Make the powered electrode smaller and insert a blocking capacitor, and a negative DC self-bias develops so ions arrive with far more energy. That asymmetry is the whole trick.
Batch → single wafer — hexode batch RIE ran twenty-plus wafers, but 200 mm uniformity demands killed it. Single-wafer tools needed higher rates, driving three responses:
Response 1 — magnetic enhancement
MERIE makes electrons spiral, lengthening their path to the wall, raising ionization. Effective at low pressure; negligible above roughly 200 mTorr where the mean free path is already short.
Response 2 — multiple frequencies
Dual- and triple-frequency CCP decouples plasma generation from ion energy. A high frequency sustains the discharge; a low frequency on the wafer electrode sets how hard ions land. Lam and TEL both build these; typical sets are 2 / 27 / 60 MHz. Mostly dielectric etch. TEL also ships a variant that DC-biases the top electrode to inject ballistic electrons.
Response 3 — high-density sources
- ICP / TCP — coil outside a dielectric window sustains the plasma; separate RF biases the stage. ~10× the ion density of a CCP. Trenches, gates, high-κ, aluminum.
- ECR — microwaves plus a field tuned so electron cyclotron frequency matches the drive. 2.45 GHz needs about 875 G.
- Helicon — bounded whistler waves from a shaped antenna. Never mainstream.
- Surface wave plasma — microwave, no DC magnetic field. The EEDF in the generation zone is non-Maxwellian, and VUV flux at the wafer is lower, which matters for damage.
- Downstream / remote — plasma generated far away so only neutrals arrive. No ions, no UV, hence isotropic. Mostly resist stripping.
03The hardware nobody puts in the model
These dominate real chamber behaviour and drift, and they're systematically missing from clean physics simulations.
Electrostatic chuck — clamps the wafer and controls its temperature. The wafer–chuck gap is pressurized with helium backside gas, a few to tens of Torr, in a gap at or below the He mean free path so the heat transfer coefficient stops caring about small gap variations. Helium specifically for its heat transfer and high ionization potential — leakage doesn't perturb the plasma. He leak-up rate doubles as a live clamping diagnostic.
Chucks also run hot (above 150 °C, to raise by-product vapour pressure in high-κ etch) or cryogenic — now the enabling technology for the deepest 3D NAND holes.
Focus and shadow rings fix edge uniformity and are consumable. Their erosion is a major source of chamber-to-chamber and time-dependent drift.
Chamber wall condition is the sleeper variable. Radical composition depends heavily on heterogeneous reactions at the walls, so whatever film sits there changes your process. Hence seasoning, between-wafer cleans, and first-wafer effects. Optical windows are heated so polymer doesn't blind the endpoint detector.
04The mechanism, properly
Anisotropy comes from ions crossing the sheath. But the etching itself comes from a synergy that neither ions nor neutrals produce alone.
Volatility gates everything — and then never limits anything
The product has to leave. Its evaporation rate must far exceed the etch rate, or it redeposits. For the common products — SiF₄, SiCl₄, SiBr₄, CO — evaporation is orders of magnitude faster than any etch rate, so volatility is essentially never the limiter. The widely repeated intuition that Si etches slower in HBr because SiBr₄ is less volatile is wrong for exactly this reason.
What does sit on the surface is the chemisorbed layer. Run the thermal desorption numbers and a species needs roughly 0.7 eV binding to survive 0.1 s at room temperature — above physisorption, below most chemical bonds. So chemisorbed radicals live indefinitely, and etching proceeds only when ions remove them.
Ion–neutral synergy
The founding result. Run the experiment yourself:
The mechanism is chemical sputtering — the collision cascade from the impacting ion drives reactions on a picosecond timescale that form desorbable products. Two quantities parameterize all of it: sticking and reaction coefficients for neutrals (usually fitted, rarely measured), and the ion-assisted etch yield. Yield goes linearly with √E above a threshold energy of order tens of eV. It also depends on angle of incidence, and that dependence differs between chemistries — which is why HBr gives straighter sidewalls than Cl₂.
05The sheath, and what actually hits the wafer
Everything the reactor does reaches the wafer only through the ion energy and angular distributions. This is the interface variable, and the cleanest place to put a model boundary.
The sheath is the space-charge layer between quasi-neutral plasma and any surface. Electrons are faster than ions, so surfaces charge negative and a potential drop forms — and nearly all the applied voltage falls here.
The shape of the resulting IEDF is set by one dimensionless ratio: ion transit time across the sheath divided by the RF period. Cross in much less than one cycle and the ion samples a single instantaneous voltage, so the distribution is wide and bimodal. Cross over many cycles and it sees only the time average, so the distribution is narrow. Push the sliders until you can predict the shape before it redraws.
06Three materials, three stories
Etch is not one process. The controlling physics genuinely differs per material — which is why a single surrogate generalizing across materials is a research claim, not a packaging decision.
Silicon in halogens
Spontaneous reactivity ranks F > Cl > Br, tracking Si–halogen bond strength. F etches Si fast with no ions at all, so F-rich chemistries need sidewall protection or cooling to stay anisotropic. Cl and Br barely react spontaneously — so Cl₂/HBr give anisotropy almost for free.
Except: heavily doped n-type silicon etches dramatically faster in Cl and Br. Fermi level shift, Cl⁻ formation, field-assisted transport through the chlorinated layer. A model without dopant as an input will be wrong on real device structures.
Ion bombardment doesn't just remove the layer, it thickens it — Cl coverage rises with ion energy via knock-on implantation and creation of new binding sites. Pure Cl₂ produces sharp microtrenches from specular reflection off smooth sidewalls; adding HBr roughens them enough to scatter ions broadly, flattening the trench out.
Silicon dioxide in fluorocarbons
A completely different control mechanism, and the one most likely to embarrass a naive optimizer. CF₂ and CF₃ don't etch SiO₂ spontaneously; they deposit a fluorocarbon film, and that film's steady-state thickness sets the etch rate. Selectivity to Si falls straight out of it. Find the working window:
Low-κ dielectrics
Carbon-containing silicas (MSQ, SiOC(H)), then porous versions to push κ lower. Porosity usually raises etch rate — but not by the simple density-scaling factor, and with enough fluorocarbon filling large pores it can etch slower than the dense material.
07Diagnostics — where the data comes from
The answer to "what could you actually train or validate on."
OES is the production workhorse. Emission is mostly electron-impact excitation, so intensity depends on both species density and the electron distribution — fundamentally qualitative. Great for endpoint, poor for absolute densities. Actinometry fixes part of that by adding a trace rare gas with a nearby excited state, so the intensity ratio gives relative density independent of the electron distribution.
Its famous failure mode is small open area. Run the trace and try to call endpoint yourself:
Interferometry reads fringes from top and bottom reflections and gives depth and rate directly. Mass spectrometry is broadest in principle but needs differential pumping and beam chopping, and is heavily biased toward wall species. Vacuum-transfer XPS is how the halogenated surface layer was pinned down — a couple of monolayers, ~2 nm, resolved into SiCl/SiCl₂/SiCl₃ by binding energy. LD-LIF and the spinning wall give sub-monolayer sensitivity and the only real handle on the wall-state problem.
08The four numbers a process engineer lives on
If you build an evaluation harness, these are what it reports. Not L2 error on a field.
Uniformity — 3σ/mean, or (Emax−Emin)/2E, in percent. About 3% within-wafer is good. The 3σ form assumes randomness; real nonuniformity is systematic, center-to-edge or azimuthal.
Selectivity — target rate over mask or stop-layer rate. Deceptively hard to measure: post-etch dielectric often reads thicker than it started, because of deposits and because oxygen penetrates and oxidizes the substrate below.
Loading — rate falls as exposed area rises, because reactant is consumed. Strong when neutral-driven (SF₆, NF₃ on lots of Si), weak when ion-driven (Cl₂ on Si), because ion flux doesn't deplete.
ARDE / RIE lag — rate falls as the feature deepens and narrows. Neutral transport, ion transport, charging and shadowing all contribute. Contacts of different diameter on the same wafer etch at different rates. You'll watch this emerge in the next section.
09Profile lab
A working feature-scale Monte Carlo. Ions launch with an angular spread, reflect off sidewalls, and remove material with a yield that depends on energy and incidence angle. Neutrals arrive isotropically, deposit passivation or etch chemically, and get consumed on first contact — which is where ARDE comes from.
Two trenches, wide and narrow, etched under identical conditions. Break them on purpose.
The named failure modes
Two more that don't fit a cross-section. Etch stop in HAR dielectric features is attributed to charging and to polymer accumulating where neutral transport is worst. And twisting — deep holes wandering off axis from statistical variation in local polymer and charge, self-propagating once started because ions neutralize on glancing collisions and continue as fast neutrals.
10Where a surrogate actually fits
Not from the review. This is the bridge to what you'd be building.
Three scales, weakly coupled, each with a different incumbent tool:
- Reactor (~30 cm, ms–s) — HPEM and commercial equivalents. EM fields, electron energy transport, fluid-chemical kinetics. Output: species fluxes to the wafer.
- Sheath (mm, ns–µs) — output is the IEDF and IADF. The clean interface variable.
- Feature (nm, minutes) — Monte Carlo profile evolution, like the lab above. Output: the CD, bow, taper, angle the customer measures.
The honest questions this raises for a foundation-model pitch:
- Which scale is the bottleneck? Reactor-scale runs are workstation-hours, not supercomputer-days. The pain may be parameter-space size or calibration, not per-run cost.
- What's the eval metric? A model at 3% field error can get the sign wrong on how bowing responds to bias frequency, which makes it useless.
- Absolute accuracy or delta accuracy? Being right about the direction and magnitude of a recipe change is far easier and often sufficient.
- Does the pretrained backbone transfer? Fusion plasma is hot, fully ionized, magnetized, few species. Etch is cold, ~99.99% neutral gas, weakly magnetized, dominated by a chemistry network with no analogue in the fusion corpus.
- Where does the training data come from, and do the licence terms permit training on it?