Oxide Process

The oxide process is thermal oxidation — the reaction that turns the top of a silicon wafer itself into silicon dioxide by baking it in an oxygen or steam ambient at 800–1200°C. It is the only thin film in the entire fab that consumes the substrate to grow, rather than depositing foreign atoms on top of it, and that single fact is why it still produces the cleanest, most electrically stable insulator silicon technology has:

Wafer clean (RCA)
Remove organics, metals, native oxide
Load furnace / tube
800–1200°C, N2 purge
Oxidation ambientpick dry or wet by target thickness
Dry O2
Slow growth, dense, high-quality
Wet H2O / steam
Fast growth, thicker films
Anneal / densify
N2 soak, stabilize interface
Thickness metrology
Ellipsometry / reflectometry

Growing the oxide also consumes the silicon beneath it. Oxygen or water vapor diffuses through the existing oxide layer and reacts at the silicon interface, so the film doesn't sit on top of the wafer the way a deposited film does — it eats into it. Roughly 44% of the final oxide thickness is grown *into* the original silicon surface, which is why oxidation is tracked by two numbers at once: the oxide thickness grown, and the silicon consumed to produce it. That consumption is exactly what later lets etched recesses (like STI trenches) get backfilled with oxide that bonds seamlessly rather than sitting as a foreign deposited layer.

The growth rate follows the Deal-Grove model, and it isn't linear. Early in the process, oxidant reaches bare silicon quickly and growth is reaction-rate-limited — roughly linear in time. Once the oxide gets thick enough, the oxidant has to diffuse through the film it already built before it can react, and growth becomes diffusion-limited — proportional to the square root of time. That's why doubling a thin gate oxide's growth time doesn't double its thickness, and why furnace recipes for a 2 nm gate oxide look nothing like the recipes for a 1 µm field oxide even though both run the identical chemistry.

Thermal Oxidation: Growth Curve and Consumption dry vs. wet rate, and the silicon that disappears underneath Oxide thickness vs. time time (linear → parabolic) thickness wet (H2O) — fast dry (O2) — slow, dense linear parabolic (√t) What actually grew where Silicon substrate original Si surface SiO2 — above surface (~56%) SiO2 — consumed into Si (~44%) total oxide same chemistry, two regimes: reaction-limited early (linear), diffusion-limited once the film thickens (√time)

Dry and wet oxidation trade speed for density, and fabs use both on purpose. Dry oxidation (O2 only) grows slowly but produces a denser, more defect-free film with a cleaner silicon interface — the reason every modern gate oxide, from the thickest I/O transistor to the thinnest logic device, is grown dry despite the extra furnace time. Wet oxidation (steam, or O2 with H2 added — "pyrogenic" oxidation) grows several times faster because water diffuses through existing oxide more easily than O2 does, which makes it the practical choice for thick field oxide and isolation structures where electrical quality at the atomic level matters far less than throughput.

The same reaction built isolation long before shallow trench isolation existed. LOCOS (local oxidation of silicon) grew thick "bird's beak" oxide laterally under a patterned nitride mask to isolate neighboring transistors, and it was the industry's isolation scheme of choice for roughly two decades before trench-based STI took over below the 0.25 µm node — STI etches a trench and backfills it with deposited oxide instead of growing it, trading LOCOS's gentler topology for a far smaller isolation footprint.

Hitting an exact target thickness means watching the film grow in real time, not trusting a timer. A furnace recipe calibrated on yesterday's test wafer can drift — tube temperature creeps, gas flow varies, a previous run leaves residue — so advanced oxidation, especially for a gate oxide only a few nanometers thick, closes the loop with in-situ optical metrology instead of running blind:

In-Situ Thickness Monitoring During Growth reflected-light interference fringes count off thickness as the film grows Optical probe geometry Silicon wafer, in furnace tube growing SiO2 film probe beam in reflected out reflections from top and bottom of the film interfere Interference signal vs. time fringe 1 fringe 2 target reached oxidation time → each fringe corresponds to a known thickness increment — count fringes, and the recipe knows exactly when to pull the wafer

That optical signal only matters because it's being checked against a physical model, not a lookup table. Real-time control compares what the probe is reporting to what the Deal-Grove equation predicts for the current temperature and ambient, and adjusts before the film ever drifts off target — the same closed-loop discipline that lets other fab steps swap a fixed timer for a live decision:

Closed-Loop Oxidation Control model prediction and live measurement are reconciled every cycle, not just at the end Measure R(t) interference fringe count Compare to model Deal-Grove prediction at T, t Within tolerance? target thickness window Pull wafer, end cycle still short of target — trim temperature/time, keep growing, measure again Why model comparison, not just the raw signal a parabolic-regime film and a linear-regime film can share an optical signature at one instant Deal-Grove tells the controller which regime it's actually in, so it knows how fast the film is really advancing the loop treats the optical reading as evidence to check against physics, not as the answer on its own

And the thing that actually decides whether an oxide is any good isn't its thickness at all — it's the interface it leaves behind. A perfectly dimensioned film with a poor silicon interface is still a bad gate oxide:

What Decides Oxide Quality: the Si/SiO2 Interface thickness is just one number — interface defects are what actually break a transistor Silicon substrate SiO2 — the bulk film amorphous, excellent bulk insulator on its own the interface — where quality is actually won or lost interface trap state (Dit) mobile ion (Na+, K+) drifts under bias, shifts VT fixed oxide charge (Qf) dry oxidation, RCA cleans, and the post-oxidation N2 anneal all exist mainly to minimize these three defects, not to hit a thickness number

Those three interface defects are why the RCA clean and the N2 anneal in the flowchart above aren't optional housekeeping — they're the steps actually being optimized for. Mobile ionic contamination (sodium and potassium, mainly from handling and chemicals) drifts through the oxide under electric field and temperature, slowly shifting a transistor's threshold voltage over its operating life — the entire reason RCA cleaning became a universal pre-oxidation step industry-wide rather than a plant-specific habit. Interface trap states (Dit) sit at dangling, unsatisfied silicon bonds right at the Si/SiO2 boundary and act as unwanted charge-trapping sites that degrade carrier mobility and noise; a post-oxidation anneal in hydrogen-containing ambient (forming gas) terminates many of these dangling bonds and can drop Dit by an order of magnitude. Fixed oxide charge (Qf), trapped within a few nanometers of the interface during growth itself, shifts the threshold voltage uniformly across a wafer in a way that's at least predictable and correctable — unlike mobile ions, which are correctable only by keeping them out in the first place.

Read the oxide process as the one step in the whole flow that doesn't add anything to the wafer — it transforms the wafer. That self-consuming nature is exactly why thermal oxide forms the most defect-free, electrically trustworthy dielectric in the fab, and why every scaling generation that could still afford the furnace time keeps a thermally grown oxide at the single interface — the gate — where electrical quality matters more than anywhere else on the chip. Thickness is what the recipe targets; the interface is what the device actually depends on.

Take oxide process further

Ask the copilot about this term, or have our engineers assess it against your process.