🧪 Module 2 · Chemistry and Materials Science · Chapter 2.4 · 13 min read

Thin-Film Deposition — ALD, PVD, CVD

Does the atom arrive one at a time, sputtered, or from a gas?

Prerequisites

What you'll learn here

Explain the core difference between ALD, PVD, and CVD
Show why a self-limiting ALD reaction yields atomic-scale control
Describe the 4 phases of an HfO₂ ALD cycle and the HfCl₄ + H₂O chemistry
Map each layer of the SIDRA stack to the right deposition technique
Explain how the BEOL thermal budget (< 400°C) constrains the deposition choice

Hook: How Do You Deposit 5 Nanometres?

A memristor’s active layer is 5 nm HfO₂ — roughly 25 atomic layers. If the film thickness strays by even one atom:

Too thin → leakage current explodes, oxide fails.
Too thick → set voltage shifts, cells no longer uniform.
Rough → local field spikes, early breakdown.

On a 300 mm wafer you have billions of cells. You need ±1 atom consistency per cell. How? The answer is Atomic Layer Deposition (ALD) — a self-limiting chemical reaction. Each cycle grows exactly one atomic layer, no more, no less.

But not every layer is grown by ALD. Copper interconnects use PVD sputtering; thick dielectrics use CVD. This chapter walks through all three, their roles in the SIDRA stack, and why each choice is deliberate.

Intuition: Three Growth Philosophies

Three different “how do I get atoms onto the wafer” strategies:

Technique	Intuition	Analogy
ALD	Alternate chemicals, each pulse grows one layer.	Painting — every brush stroke is exactly one coat.
CVD	Gas mixture reacts on the substrate, continuous growth.	Breath on a window — uniform but no thickness control.
PVD	Atoms physically knocked off a target, fly to the wafer.	Sandblasting — line-of-sight, fast, directional.

Key difference: ALD is chemically self-limiting — once the surface is saturated, the reaction stops. That’s why no chaos, atomic control. CVD and PVD are fast and high-throughput but lack per-atom control.

Role split in SIDRA:

ALD: HfO₂ (memristor dielectric), NbOx (OTS selector), TiN barrier (thin).
PVD: W/TaN/Cu interconnects, thick electrodes.
CVD: SiO₂ fill, Si₃N₄ passivation, thick BEOL dielectrics.

Formalism: The Self-Limiting ALD Reaction

L1 · Başlangıç

One HfO₂ ALD cycle has 4 phases:

Precursor pulse (HfCl₄): Reacts with -OH groups on the wafer surface, binds an Hf-Cl layer. Once no sites remain, it stops (self-limiting).
Purge (N₂): Excess HfCl₄ gas is swept away.
Reactant pulse (H₂O): Swaps Cl bonds for H, leaves -OH groups. Again saturates and stops.
Purge: HCl byproduct and excess H₂O are swept away.

One cycle = 1.1 Å of HfO₂ (about one monolayer). For 5 nm → ≈ 45 cycles. Each cycle takes 2–5 seconds.

Net reaction:

\mathrm{HfCl_4} + 2\,\mathrm{H_2O} \rightarrow \mathrm{HfO_2} + 4\,\mathrm{HCl}

L2 · Tam

Why does self-limiting matter? The reaction depends on surface reactive-site count, not surface area. Once sites saturate, extra precursor cannot react — every point on the wafer gets the same amount. Result: conformal coating even on 3D structures (~100%). Trench, via, high-aspect-ratio pillar — all coated evenly.

Growth rate (GPC = Growth Per Cycle):

HfO₂: 1.1 Å/cycle (HfCl₄ + H₂O, 300°C).
Al₂O₃: 1.0 Å/cycle (TMA + H₂O, 200°C).
TiN: 0.4 Å/cycle (TiCl₄ + NH₃, 350°C).

Temperature window: each reaction has an ideal band:

Too cold → precursor condenses, parasitic growth.
Too hot → precursor decomposes, self-limiting breaks, it drifts to CVD mode.

HfO₂ window ≈ 200–350°C.

Critical for SIDRA: the BEOL (Back-End-of-Line) thermal budget is < 400°C. Underlying copper metallization diffuses at higher T. ALD fits comfortably; that’s why memristors can be integrated in BEOL.

L3 · Derin

PE-ALD (Plasma Enhanced): O₂ plasma instead of H₂O. High-quality oxide at < 200°C. Needed for integration above GST.

Area-selective ALD: Grow only on chosen surfaces — block the others with a self-assembled monolayer (SAM). Can skip a litho step; active research at 3 nm node.

Spatial ALD: Wafer passes through spatially separated gas zones rather than time-separated pulses. 10× throughput, widely used in OLED manufacturing.

ALE (Atomic Layer Etch): The inverse of ALD — each cycle removes one atomic layer. Cl₂ adsorption + Ar ion pulse. Critical for GAA transistor fin shaping.

PVD physics model: Ar⁺ ions strike target atoms, transfer momentum. Knocked-off atoms fly to and deposit on the substrate. Sputtering yield:

Y = \frac{3}{4\pi^2} \alpha \frac{4 M_i M_t}{(M_i + M_t)^2} \frac{E}{U_s}

where $M_i$ is ion mass, $M_t$ target mass, $U_s$ bond energy. For Cu with Ar at 500 eV, Y ≈ 2.5.

CVD variants:

LPCVD (low pressure): Si₃N₄, poly-Si.
PECVD (plasma): low-T SiO₂, amorphous Si.
Epitaxial CVD: single-crystal Si, SiGe (MOCVD / MBE).

Experiment: Watch an ALD Cycle

The animation below shows one HfO₂ ALD cycle in 4 phases. Observe how the precursor saturates the surface, why a purge is needed, and how exactly one monolayer remains at the end.

Cycle: 0

Thickness (Å): 0.0

HfCl₄ + 2 H₂O → HfO₂ + 4 HCl ↑

1. Precursor pulse (HfCl₄)

HfCl₄ molecules react with surface OH groups; **exactly one** molecule binds per surface site (self-limiting). This is ALD's magic: excess gas does no harm.

Note: After each cycle the layer counter increments and thickness grows ≈ 1.1 Å. For 5 nm → ≈ 45 cycles. In a real reactor this is ≈ 3 minutes — slow but perfect.

Quick Quiz

1/5What distinguishes ALD from CVD?

Lab Exercise

A SIDRA chiplet needs a 4 nm HfO₂ + 5 nm NbOx + 20 nm TiN + 200 nm Cu stack.

(a) How many ALD cycles for HfO₂? (GPC = 1.1 Å/cycle) (b) If each cycle takes 3 seconds, how long is just the HfO₂ step? (c) Which technique do you pick per layer? (ALD / CVD / PVD) (d) Why is growing 200 nm Cu by ALD a bad idea?

Answers

(a) 4 nm / 0.11 nm = ≈ 37 cycles.

(b) 37 × 3 s = 111 s ≈ 1.85 min. In practice, with reactor stabilization + purges, ~5 min.

(c) HfO₂ → ALD (atomic control). NbOx → ALD (stoichiometry critical). TiN → ALD (thin barrier, conformal). Cu → PVD seed + ECD electroplating (ALD too slow for 200 nm).

(d) Cu ALD GPC ≈ 0.3 Å/cycle. 200 nm ⇒ ≈ 6700 cycles × 5 s = 9.3 hours/wafer. Not economical. PVD deposits 200 nm Cu in ≈ 30 seconds.

Cheat Sheet

ALD — self-limiting, atomic control, conformal. Slow but perfect. For HfO₂, NbOx, TiN.
PVD — physical sputter. Fast, for metals. Low conformality.
CVD — gas-phase reaction. Medium speed, medium control. For thick dielectrics.
HfO₂ ALD: HfCl₄ + H₂O, 300°C, 1.1 Å/cycle.
BEOL thermal budget < 400°C — the constraint that gates memristor integration.

Vision: The Future of Deposition

Area-selective ALD: skip litho — reduces mask count at the 3 nm node.
Spatial ALD: 10× throughput; widespread in OLED/flexible electronics, reaching silicon.
ALE (Atomic Layer Etch): atomic-precision removal for GAA transistors, 3D NAND.
Supercycle ALD: alternate precursors across cycles for composition gradients (e.g. ~5% Al doping into HfO₂).
MLD (Molecular Layer Deposition): organic-inorganic hybrids; buffer layers on ferroelectric HZO.
PEALD with remote plasma: sub-150°C oxides — critical for upper layers in 3D stacks.
Vapor-phase infiltration: inorganic diffusion into polymers; flexible memristors.
Cold spray / aerosol deposition: room-T ceramic films — heterogeneous integration, chiplet-level packaging.

For post-Y10 SIDRA, area-selective ALD is the biggest lever: fewer masks drop wafer cost by ~30% and boost memristor density 4× in the same footprint.

Prerequisites

What you'll learn here

🪝 Hook: How Do You Deposit 5 Nanometres?

🧭 Intuition: Three Growth Philosophies

📐 Formalism: The Self-Limiting ALD Reaction

🧪 Experiment: Watch an ALD Cycle

📝 Quick Quiz

🛠️ Lab Exercise

🗂️ Cheat Sheet

🔮 Vision: The Future of Deposition

📚 Further Reading