Giant Magnetoresistance: The Spin Valve That Made Hard Drives Huge

Mott's Two Currents: Spin as a Resistor Label

In copper an electron's spin is a passenger, irrelevant to how the electron scatters. Inside a ferromagnet it becomes an address label. Nevill Mott made the argument in 1936. Well below the Curie temperature — 1043 K for iron, 1388 K for cobalt, 627 K for nickel — the processes that flip a conduction electron's spin (magnon emission and absorption, spin–orbit scattering off impurities) are far rarer than the ordinary momentum-randomising collisions with phonons and defects. An electron keeps its spin through many scattering events, so the current separates into two nearly independent streams flowing in parallel: majority spins aligned with the local magnetisation and minority spins against it, each with its own resistivity, ρ↑ and ρ↓.

Those resistivities are unequal, and the reason is band structure. Exchange interaction shifts the majority and minority 3d bands rigidly apart — by roughly 1.5–2 eV in Fe and Co, only about 0.3 eV in Ni. In Co and Ni, the so-called strong ferromagnets, the majority d band is completely filled and lies below the Fermi level, so majority carriers at E_F are light, free-electron-like 4s states with almost no empty d states to fall into. Minority carriers sit instead in a dense, flat d band right at E_F. Fermi's golden rule does the rest: a scattering rate is proportional to the density of available final states, so s→d scattering is fast for minority electrons and slow for majority ones.

The upshot is a spin asymmetry α = ρ↓/ρ↑ that is a material property: in permalloy (Ni₈₀Fe₂₀) the bulk asymmetry β = (α−1)/(α+1) is about 0.7, so α ≈ 5–6; in cobalt β ≈ 0.45–0.5, giving α ≈ 3. No field bends any trajectory. It is pure spin-dependent scattering.

The Sandwich: Why Parallel Alignment Is a Short Circuit

Now stack two ferromagnetic films with a non-magnetic metal spacer between them — canonically Fe/Cr/Fe or Co/Cu/Co, the spacer 1–3 nm thick. The spacer must be thin compared with the electron mean free path, roughly 40 nm in copper at 300 K, so a single electron can cross it and sample both magnetic layers before it forgets where it came from. That requirement is why GMR waited for nanometre-controlled film growth in the 1980s.

Parallel magnetisations. A majority-spin electron in the first layer is still a majority-spin electron in the second. It runs the entire stack in its low-resistivity channel, while every minority electron is throttled in both layers. In a parallel resistor network the low branch dominates: the majority spins act as a short circuit and the total resistance is low.

Antiparallel magnetisations. Flip the second layer. An electron that is majority in layer one is minority in layer two, and vice versa, so every electron gets one easy pass and one hard pass. Neither channel is a short circuit and the resistance rises — by tens of percent, against the small fraction of a percent an ordinary metal manages in the same field.

The field's only job is to rotate a magnetisation. It exerts no useful Lorentz force: at 1 mT an electron at copper's Fermi velocity (~1.6 × 10⁶ m s⁻¹) has a cyclotron radius of order a centimetre, a million times the stack's thickness. GMR is magnetically switched, not deflected — so the signal saturates the instant the layers go collinear, then stays flat however hard you push.

The Resistor Algebra and What It Gets Right

Take the crudest model: two identical magnetic layers, resistance proportional to resistivity, spacer ignored, no spin flips. In the parallel state each channel runs the same resistivity twice, so 2ρ↑ and 2ρ↓ sit in parallel and R_P = 2ρ↑ρ↓/(ρ↑+ρ↓). In the antiparallel state each channel sees one of each in series, ρ↑+ρ↓, and the two identical channels in parallel give R_AP = (ρ↑+ρ↓)/2. Subtract and divide:

ΔR/R_P = (ρ↓ − ρ↑)² / (4ρ↑ρ↓) = (α − 1)² / (4α)

α = 2 gives 12.5%; α = 3 gives 33%; α = 5 gives 80%; α = 6 gives 104%. It vanishes quadratically as α → 1 and diverges as α → ∞ — which is why cobalt and permalloy stacks work, why a weakly polarised ferromagnet is useless, and why half-metals, with one spin species at E_F, are the standing goal.

Three caveats. The algebra counts only bulk scattering, whereas spin-dependent reflection at the ferromagnet/spacer interfaces is often the larger term — the Co/Cu interface is famously asymmetric because Co majority states match Cu's Fermi surface well and minority states do not. Geometry matters too: current-in-plane (CIP) GMR, the read-head case, is set by the mean free path, while current-perpendicular-to-plane (CPP) GMR, described by Valet–Fert diffusion theory in 1993, is governed by the far longer spin-diffusion length — ~350 nm in Cu at room temperature, only ~5 nm in permalloy — and runs two to three times larger for the same stack. And real GMR decays roughly as 1/t_spacer once the spacer shunts current.

From Curiosity to Read Head: 1988 to 1997

Two groups arrived independently in 1988. At the Laboratoire de Physique des Solides in Orsay, Albert Fert's group studied MBE-grown (001)Fe/(001)Cr superlattices from Alain Friederich's team at Thomson-CSF — 60 repeats of ~3 nm Fe and ~0.9 nm Cr. As a field near 2 T forced the antiferromagnetically coupled iron layers parallel, the resistance at 4.2 K fell by almost a factor of two (Baibich et al., Phys. Rev. Lett. 61, 2472). They named the effect. At Forschungszentrum Jülich, Peter Grünberg's group measured a far simpler Fe/Cr/Fe trilayer and found about 1.5% — but at room temperature (Binasch, Grünberg, Saurenbach and Zinn, Phys. Rev. B 39, 4828). Grünberg had supplied the missing ingredient in 1986, discovering antiferromagnetic interlayer coupling across a Cr spacer with Brillouin light scattering. The two shared the 2007 Nobel Prize in Physics.

Three engineering steps made it a product. Stuart Parkin at IBM Almaden showed in 1990 that plain sputtering reproduced everything without epitaxy, and that the interlayer coupling oscillates between ferromagnetic and antiferromagnetic as the spacer thickens, period near 1 nm — an RKKY-like oscillation set by extremal calipers across the spacer's Fermi surface, and so a design parameter rather than an accident. By 1991 his sputtered Co/Cu multilayers reached about 65% at room temperature.

Those multilayers still needed teslas to switch, which is hopeless for a sensor. Bernard Dieny and colleagues at IBM Almaden solved it in 1991 with the spin valve: decouple the two magnetic layers, pin one by exchange bias against an antiferromagnet (FeMn then, IrMn or PtMn later — the Meiklejohn–Bean effect of 1956), and leave the other free. A representative stack is Ta/NiFe/Cu(~2.5 nm)/Co/FeMn/Ta: the free layer rotates in roughly 1 mT (10 Oe) while the pinned layer holds still, giving 4–8% over a few oersteds. A sensor sells its slope, not its magnitude.

Then IBM shipped it. The Deskstar 16GP, codenamed Titan and announced in December 1997, was the first drive with GMR read heads: 16.8 GB on five platters at roughly 2.6 Gbit/in². Areal density then compounded at close to 100% per year. GMR heads ruled until tunnel-magnetoresistance heads displaced them from about 2005 — alumina barriers first, MgO within a few years.

How It Is Actually Measured

The transport measurement is unglamorous: four-point or van der Pauw resistance on a patterned film while an in-plane field is swept and R(H) recorded. The signature is a butterfly-shaped curve, symmetric in H, peaking near the coercive field of whichever layer is reversing, negative in sign, and dead flat above saturation. That flatness is diagnostic — an ordinary metal's magnetoresistance keeps climbing as B², while GMR simply stops once the magnetisations are collinear.

Transport alone cannot prove the antiparallel state exists, so it is paired with a probe that sees the magnetisation: MOKE or SQUID/VSM magnetometry for the hysteresis loop, Brillouin light scattering for the coupling strength (Grünberg's original tool), and polarised neutron reflectometry — at instruments such as POLREF at ISIS or NIST's polarised beam reflectometer — which yields a half-order Bragg peak when the magnetic period is twice the structural period, the direct fingerprint of antiferromagnetic alignment. X-ray reflectometry fixes thicknesses and roughness to the ångström. CPP measurement is harder, since a 20 nm stack contributes nanoohms: W. P. Pratt and Jack Bass at Michigan State sandwiched the multilayer between crossed superconducting niobium strips in 1991 and read it with a SQUID.

One warning. The GMR ratio has two conventions: ΔR/R_P, normalised to the parallel state, unbounded and optimistic, and ΔR/R_AP, normalised to the high state and capped at 100%. A stack quoted at 100% under one is 50% under the other, and Fert's halving of the resistance appears in the literature as both. Check the denominator before comparing papers.

The Look-alikes: AMR, TMR, CMR and Ordinary Magnetoresistance

AMR — anisotropic magnetoresistance — is the effect GMR displaced and the one most often confused with it. William Thomson (Lord Kelvin) reported it in iron and nickel in 1857. It needs only a single ferromagnetic film, and its origin is spin–orbit coupling: the scattering cross-section depends on the angle θ between magnetisation and current, ρ(θ) = ρ_⊥ + Δρ cos²θ. In permalloy it is about 2% at room temperature. IBM's 0663 Corsair introduced AMR heads in 1991; GMR beat them with several times the signal.

TMR keeps the sandwich but replaces the metal spacer with a ~1 nm insulator, so conduction is tunnelling rather than diffusion. Amorphous Al₂O₃ barriers (Moodera and Miyazaki, 1995) gave 10–20% at room temperature; crystalline MgO, which coherently filters the slowly decaying Δ₁ Bloch state that is fully spin-polarised in bcc Fe and CoFe, took the groups of Parkin and Yuasa to roughly 180–220% in 2004, and CoFeB/MgO/CoFeB stacks have since reached roughly 600%. That is why TMR, not GMR, sits in today's read heads and in MRAM.

CMR — colossal magnetoresistance — occurs in doped perovskite manganites such as La₀.₆₇Ca₀.₃₃MnO₃, with neither layers nor spacers. A field tips the bulk sample through a coupled ferromagnetic and metal–insulator transition via Zener double exchange; ratios of 10⁴–10⁵% have been reported (Jin et al., 1994), but only near T_C, usually below room temperature, and only in several tesla. Spectacular in a laboratory, useless in a device.

Ordinary magnetoresistance really is the Lorentz force curving trajectories between collisions, scaling as (ω_cτ)². For copper at room temperature in 1 mT, ω_cτ ≈ 4 × 10⁻⁶, so the effect is one part in 10¹¹ — exactly what people wrongly picture on first hearing the word magnetoresistance.

Where It Breaks Down, and What Is Still Open

The two-current model has a temperature-dependent expiry date. As T rises toward T_C, magnon scattering mixes the two channels — an electron that emits a magnon flips its spin — so they stop being independent and the ratio collapses. That spin mixing, together with the growing spin-independent phonon scattering that dilutes the asymmetry — rather than any change in band structure — is why a stack showing 50% at 4.2 K may show only a few percent at 300 K, and why Grünberg's modest 1.5% at room temperature was arguably the more consequential of the two 1988 results.

The remaining failure modes are practical. Interdiffusion at the Co/Cu interface during annealing destroys the asymmetry. Pinholes through a 2 nm spacer, or correlated roughness producing Néel orange-peel coupling, link the layers ferromagnetically so the free layer never reaches a clean antiparallel state. Exchange bias vanishes above the antiferromagnet's blocking temperature, only ~150 °C for FeMn. And as sensors shrink, thermally excited magnon modes in a free layer of order 10³–10⁴ nm³ produce magnetic noise well above the Johnson–Nyquist floor.

Open ground remains. All-metal CPP-GMR with Heusler electrodes such as Co₂MnSi, predicted to be half-metallic, is still pushed for future heads because a metallic stack keeps the resistance–area product low where a tunnel barrier would be prohibitive. How much of a measured asymmetry is bulk and how much interfacial is argued stack by stack. And the same spin-polarised current that reads a bit can write one: the spin-transfer torque predicted independently by John Slonczewski and Luc Berger in 1996 turned GMR's physics into today's STT-MRAM.

Magnetoresistance effects compared. GMR and TMR pair a large signal with a millitesla switching field — which is why they took the read head over from AMR, and why CMR, needing several tesla, never got near one.

Effect	Typical ΔR/R at 300 K	Physical origin	Field for full signal
GMR spin valve (NiFe/Cu/Co/FeMn)	4–8%	Spin-dependent scattering across a metallic spacer	~1 mT
GMR multilayer (Co/Cu, AF-coupled)	up to ~65%	Same mechanism, but the stack must be forced parallel	~1 T
AMR (permalloy; Thomson, 1857)	~2%	Spin–orbit scattering anisotropy in a single film	~1 mT
TMR (CoFeB/MgO/CoFeB)	200–600%	Coherent Δ₁ spin-filtered tunnelling through an insulator	~1 mT
CMR (La₀.₆₇Ca₀.₃₃MnO₃)	10⁴–10⁵% but only near T_C, below 300 K	Double-exchange metal–insulator transition in the bulk	several tesla
Ordinary MR (Cu film)	~10⁻²% at several tesla (~10⁻⁹% at 1 mT)	Lorentz-force orbit bending, scaling as (ω_c τ)²	never saturates; grows as B²