⚛️ Module 1 · Physics Foundation · Chapter 1.8 · 13 min read

Electrochemistry and Ion Motion

Atoms moving at the heart of the memristor.

Prerequisites

What you'll learn here

  • Explain ion migration and the role of oxygen vacancies
  • Describe filament formation in a SET/RESET cycle
  • Link voltage-time behavior to Arrhenius activation and the Mott-Gurney law
  • Evaluate SIDRA's 10⁶ endurance target in terms of ionic drift

Hook: What If Atoms Moved?

Classic transistor design has a rule: atoms don’t move. Only electrons flow. Chemical bonds, crystal structure, dopants — all fixed. You can charge-discharge a billion times and the chip stays the same.

The memristor breaks this rule. Oxygen atoms inside HfO₂ do move. A positive voltage on one electrode drags them upward, leaving behind an oxygen vacancy that drifts to the other electrode. Vacancies coalesce into a conductive filament. A negative voltage ruptures that filament. Memory here is the position of atoms, not of electrons.

This chapter is the last key to SIDRA’s operation: electrochemistry, ionic diffusion, drift — and why 10⁶ program-erase cycles is a hard ceiling.

Intuition: Oxygen Vacancies Migrate

HfO₂ is not stoichiometrically perfect. During deposition some oxygen atoms go missing — leaving behind an oxygen vacancy (V_O). Effectively these act as +2 charged defects because the missing atom took electrons with it; local net charge is positive.

Apply a voltage: these vacancies drift along the electric field. If bottom is ground and top is +2 V, vacancies drift upward. They pile up near the top electrode, form a “vacancy cluster”. Clusters align → filament. The filament acts like a conducting wire linking top to bottom electrode.

Reverse it: top = −2 V, vacancies drift downward, filament disperses. No path for electrons → high resistance.

This memory is not binary, it’s continuous. The longer voltage is applied, the more vacancies move → in-between conductance levels are analog. That’s why memristors are excellent for analog AI.

Formalism: Drift + Diffusion + Arrhenius

L1 · Intro

Two processes:

  • Drift: ions move along the electric field. Velocity ∝ field.
  • Diffusion: thermal agitation spreads ions randomly. Rate ∝ temperature.

Voltage builds a filament; temperature + time erases it (retention).

L2 · Full

Mott-Gurney ionic current:

Iion=I0sinh⁡(qaE2kT)I_{ion} = I_0 \sinh\left(\frac{qaE}{2kT}\right)
  • aa: atomic hop distance (~0.3 nm)
  • EE: electric field (V/m)
  • kTkT: thermal energy (0.026 eV at room T)

Low field → sinh ≈ linear (ohmic ionic). High field (qaE>kTqaE > kT) → exponential. At SIDRA program voltage (V∼2V \sim 2 V across d∼5d \sim 5 nm → E=4×108E = 4 \times 10^8 V/m), qaE/2kT≈2.3qaE/2kT \approx 2.3, sinh⁡≈5\sinh \approx 5. Doubling the field 10× the ionic current.

Arrhenius activation:

kdrift=k0exp⁡(−EakT)k_{drift} = k_0 \exp\left(-\frac{E_a}{kT}\right)

HfO₂ vacancy migration barrier Ea≈0.8E_a \approx 0.8-1.21.2 eV. Room-T kT=0.026kT = 0.026 eV. Ratio: exp⁡(−40)≈10−17\exp(-40) \approx 10^{-17} → ion motion essentially zero on nanosecond scales without field. Under voltage, EaE_a effectively drops (field-assisted); e.g. at 2 V down to ~0.3 eV → exp⁡(−12)≈10−6\exp(-12) \approx 10^{-6}. Times 101310^{13} attempts/s → 10710^7 Hz → filament forms in milliseconds.

Retention:

tretention=t0exp⁡(EakT)t_{retention} = t_0 \exp\left(\frac{E_a}{kT}\right)

At 85°C (kT=0.031kT = 0.031 eV), t∼10t \sim 10 years for Ea=1E_a = 1 eV. This is SIDRA’s target.

L3 · Deep

Filament geometry: 1-10 nm diameter, 5 nm length. 10-100 vacancies suffice. Conductance scales with filament cross-section. Dopants (Al, Y) in HfO₂ steer directional growth.

Endurance (10610^6 cycles): each SET/RESET permanently displaces a few atoms. This is electromigration fatigue. SIDRA targets 10610^6, above NAND Flash’s 10510^5, but well below DRAM’s 101510^{15}.

Redox (ECM vs VCM): two mechanisms. ECM — Ag or Cu metal ions diffuse from an electrode into the oxide, forming a metallic filament. VCM — oxygen vacancies in HfO₂. SIDRA uses VCM (more controllable, CMOS-compatible).

Stochastic variation: vacancy positions are random → each SET filament differs. Cell-to-cell and cycle-to-cycle σ/μ≈5\sigma/\mu \approx 5-15%15\%. That noise caps analog MVM accuracy (detail in 5.10).

Experiment: Build and Break the Filament

Steps:

  1. SET (+2 V): vacancies drift to the top electrode. A filament appears. Badge turns “LRS”.
  2. Return voltage to 0 — filament persists (the memory effect!). “LRS” stays.
  3. RESET (−2 V): vacancies drift down, filament dissolves. “HRS”.
  4. READ (+0.1 V): probes state without disturbing. LRS → high read current, HRS → low.
  5. Cycle SET/RESET many times — filaments vary slightly each time (stochastic).

Quiz

1/5What forms the conductive filament in a HfO₂ memristor?

Lab Task

HfO₂ active layer d=5d = 5 nm. Program voltage V=2V = 2 V.

(a) Compute electric field EE (V/m). (b) Find qaEqaE in eV (a=0.3a = 0.3 nm). (c) At room T (kT=0.026kT = 0.026 eV), what is qaE/kTqaE/kT? Linear or exponential sinh regime? (d) If temperature rises to 85°C, how does retention change? (Use Ea=1.0E_a = 1.0 eV.)

Answers

(a) E=V/d=2/(5×10−9)=4×108E = V/d = 2 / (5 \times 10^{-9}) = 4 \times 10^{8} V/m.

(b) qaE=1.6×10−19⋅0.3×10−9⋅4×108=1.92×10−20qaE = 1.6 \times 10^{-19} \cdot 0.3 \times 10^{-9} \cdot 4 \times 10^{8} = 1.92 \times 10^{-20} J =0.12= 0.12 eV.

(c) qaE/kT=0.12/0.026≈4.6qaE/kT = 0.12/0.026 \approx 4.6. sinh(4.6) ≈ e4.6/2≈50e^{4.6}/2 \approx 50. Exponential regime — ionic current is large, filament forms quickly.

(d) t85/t25=exp⁡(1.0/0.031−1.0/0.026)=exp⁡(−6.2)≈0.002t_{85} / t_{25} = \exp(1.0/0.031 - 1.0/0.026) = \exp(-6.2) \approx 0.002. Retention shrinks 500× — from 10 years to about a week.

Cheat Sheet

  • Memristor memory: position of ions (oxygen vacancies), not electrons.
  • SET → LRS: + voltage drifts vacancies into a filament.
  • RESET → HRS: − voltage disperses the filament.
  • Mott-Gurney: Iion∝sinh⁡(qaE/2kT)I_{ion} \propto \sinh(qaE/2kT) — exponential at high field.
  • Arrhenius: retention ∝exp⁡(Ea/kT)\propto \exp(E_a/kT); heat kills it fast.
  • ECM vs VCM: metal-atom filament vs oxygen-vacancy filament. SIDRA is VCM.
  • Endurance: ∼106\sim 10^6 cycles — better than Flash, worse than DRAM.
  • Stochastic variation: σ/μ≈5\sigma/\mu \approx 5-15%15\% → noise floor for analog MVM.

Vision: Beyond Memristor Physics

Alternative memory mechanisms:

  • ECM (Ag, Cu filament): faster SET, sharper threshold — but harder CMOS integration (metal contamination).
  • Phase-change (PCM, GST): amorphous↔crystalline, 10 ns speed, 10⁸ endurance. Intel Optane (2017-22), now Samsung.
  • Ferroelectric (HZO, HfZrO₂): polarization memory. 10⁹+ endurance, 10 ns write. Micron + UMC 2024 demo.
  • Protonic memristor: H⁺ migration (not oxygen). Harvard 2024 — 10 ns program, 10⁹ endurance, low V.
  • Organic polymer memristors: bendable electronics, biocompatible neural interfaces.
  • DNA storage: not a base-scale contender but lifetime data (1000+ year retention) — Microsoft + Illumina.
  • Li-ion neuromorphic cell: real ion migration for analog weights; very linear, very symmetric but slow (µs).
  • MIEC (mixed ionic-electronic conductor): carries both ions and electrons; three-terminal analog synapse.
  • Skyrmion memristor: magnetic topological solitons; target 10 fJ/bit and unlimited endurance.

Biggest lever for post-Y10 SIDRA: the protonic memristor — H⁺ is 10× smaller and faster than oxygen, dropping SET time from 100 ns to 10 ns and pushing endurance to 10⁹. Linearity is critical for trainable analog. 2027–2029 horizon.

Further Reading