Metallic Glass: An Alloy Frozen Before It Can Crystallise

What a Metallic Glass Actually Is

Cool almost any molten metal and it crystallises almost instantly. Its atoms are near-spheres held by non-directional metallic bonds, its melt is only about 1-5 mPa·s viscous — the same range as water, and kinematically thinner still because it is so dense — and a viable nucleus needs only a few dozen atoms, so within microseconds the liquid's disorder is erased. A metallic glass is an alloy in which that never happens: the melt is cooled past the point where atoms can still rearrange, freezing the liquid structure in place. The result is a solid with no grains, no grain boundaries and no long-range order.

The frozen liquid is not random. A metallic glass is only about 0.5% less dense than the same alloy after it has crystallised — a remarkably small deficit, where vitreous silica sits some 17% below crystalline quartz. That near-parity is the fingerprint of dense random packing: efficiently packed, solute-centred clusters, many of them icosahedral. Frederick Charles Frank explained in 1952 why this frustrates crystallisation — a 13-atom icosahedron is more tightly bound than the equivalent close-packed cluster, but its fivefold symmetry cannot tile space, so the locally favoured arrangement is incompatible with every periodic lattice.

The diagnostic is blunt: a crystal gives sharp Bragg peaks, while a metallic glass gives a single broad halo a few degrees wide — near 2θ ≈ 37° for Zr-based alloys with Cu Kα radiation — and nothing else. It is still a metal: opaque, lustrous, a free-electron conductor. The word glass describes the atomic arrangement and the glass transition, not transparency.

The Race Against Nucleation

Crystallisation needs nuclei, and classical nucleation theory prices them. A spherical embryo of radius r pays surface energy 4πr²σ and is repaid volume free energy −(4/3)πr³ΔG_v, so the barrier peaks at the critical radius r* = 2σ/ΔG_v:

ΔG* = 16πσ³ / 3ΔG_v²

Work it for Vitreloy 1 at 900 K, the nose of its transformation curve. With a heat of fusion ≈ 8.2 kJ/mol over a molar volume ≈ 9.8 cm³/mol and a liquidus of 993 K, the 93 K undercooling gives ΔG_v ≈ ΔH_fΔT/(T_lV_m) = 8200 × 93 / (993 × 9.8×10⁻⁶) ≈ 7.8×10⁷ J/m³. Taking the crystal-liquid interfacial energy as σ ≈ 0.04 J/m², the barrier is ΔG* ≈ 1.8×10⁻¹⁹ J — about 1.1 eV, against k_BT = 1.24×10⁻²⁰ J — so ΔG*/k_BT ≈ 14 and r* ≈ 1 nm, a couple of hundred atoms. Deeper down ΔG_v grows, r* shrinks to a few dozen atoms, and the barrier all but disappears.

The cube and the square are why glass-forming ability is so temperamental: raise σ by 25% and ΔG* nearly doubles, which — sitting in an exponent — cuts the nucleation rate by roughly six orders of magnitude.

Nucleation also needs atoms to move, and the kinetic prefactor scales as 1/η. Near the liquidus ΔG_v → 0 and the barrier is effectively infinite; near the glass transition the viscosity reaches ≈ 10¹² Pa·s — that value is the operational definition of T_g — and nothing can rearrange. The product of driving force and mobility peaks in between, and plotted as time-to-transform against temperature it traces the C-shaped TTT (time-temperature-transformation) diagram, whose leftmost point is the nose.

Reading the TTT Nose: Why 1 K/s Is Enough

The nose is the only obstacle. Any cooling curve that stays to the left of it reaches the glass transition without ever crossing into crystallisation. For Vitreloy 1 — Zr_41.2Ti_13.8Cu_12.5Ni₁₀Be_22.5, cast by Atakan Peker and William Johnson at Caltech in 1993 — the TTT nose sits at ~900 K with a ~60 s incubation. The rate that just clears it is

R_c ≈ (T_liquidus − T_nose) / t_nose = (993 − 900) K / 60 s ≈ 1.6 K/s

which is why Vitreloy 1 needs only ~1 K/s and casts ~5 cm thick. Its reduced glass transition temperature follows from T_g = 623 K (350 °C) and a liquidus of 993 K (720 °C): T_rg = 623/993 ≈ 0.63, just under David Turnbull's old rule of thumb that alloys with T_rg ≳ 2/3 vitrify easily — a guide, not a law.

Now the other end. Klement, Willens and Duwez reported the first metallic glass in Nature in 1960: Au₇₅Si₂₅ splat-quenched at ~10⁶ K/s, a droplet fired onto a chilled copper substrate by a shock-wave gun at Caltech. Metglas Fe₇₈Si₉B₁₃ still demands that rate, and the rate dictates the geometry. Heat leaves a sheet of half-thickness L in roughly t ≈ L²/α; with α ≈ 2×10⁻⁶ m²/s and a ~25 µm ribbon (L ≈ 12.5 µm), t ≈ (1.25×10⁻⁵)²/(2×10⁻⁶) ≈ 8×10⁻⁵ s. Shedding ~1200 K in that time is of order 10⁷ K/s — comfortably past the requirement, and exactly why the ribbon is that thin and why the copper wheel runs at 20-50 m/s.

Inoue's Rules: Why Three Elements Beat One

Six orders of magnitude separate Au₇₅Si₂₅ from Vitreloy 1. Akihisa Inoue, whose Tohoku University group opened the bulk-glass era with La-Al-Ni in 1989, distilled the gap into three empirical rules through the 1990s:

• Three or more constituent elements — the confusion principle: with five species, every candidate crystal needs a specific stoichiometry the melt does not locally have.
• Atomic size mismatch above 12%. Vitreloy 1 spans beryllium (metallic radius ≈ 1.12 Å) to zirconium (≈ 1.60 Å), a ~40% spread, so small atoms fill the interstices of the large ones and drive free volume down.
• Negative heats of mixing. Zr-Be, Zr-Ni and Zr-Cu all mix exothermically, so the liquid develops chemical short-range order. That lowers its energy, raises σ, and forces any would-be crystal to unmix five elements by long-range diffusion before it can grow.

The combined effect is a melt that is both crowded and slow: Vitreloy 1 is several hundred to a thousand times more viscous at its liquidus than molten iron, of order 2-5 Pa·s against ~5 mPa·s. Crystallising it is not simply nucleating one phase; it is nucleating several phases and chemically sorting five elements into each. This is also why good glass formers cluster near deep eutectics — the compositions where the liquid survives to unusually low temperature and the competing crystals have the least to gain.

The Hardware: Wheels, Copper Moulds and Transformer Cores

Melt spinning is the workhorse for marginal glass formers: the alloy is induction-melted in a quartz crucible, then ejected by argon overpressure through a slot nozzle onto a water-cooled copper wheel spinning at 20-50 m/s. Ribbon peels off continuously at ~25 µm thick and up to about 200 mm wide — the process Allied Chemical commercialised as Metglas in the 1970s.

Its biggest market is unglamorous and enormous. Amorphous Fe-Si-B has no magnetocrystalline anisotropy and very high resistivity, so core loss runs about 0.2-0.3 W/kg at 1.4 T and 60 Hz against roughly 1 W/kg for grain-oriented silicon steel — a ~70% cut in the no-load loss a distribution transformer burns around the clock. The trade-offs are real: saturation induction is ~1.56 T against ~2.03 T, so cores are bulkier, and brittle ribbon cannot be punched into laminations, so cores are wound then annealed in a magnetic field. India and China have installed millions of amorphous-core transformers on that arithmetic.

Bulk casting is different work: arc-melt the charge under high-purity argon on a water-cooled copper hearth, flip and remelt it for homogeneity, then suction- or tilt-cast into a copper mould. Oxygen is the enemy — a few hundred ppm in a Zr-based melt seeds a metastable quasicrystalline phase and can raise the critical cooling rate by orders of magnitude, so sponge purity and crucible cleanliness are specified, not assumed. Pd-based record holders are additionally fluxed in molten B₂O₃ to dissolve oxide particles that would otherwise nucleate crystals. Between T_g and crystallisation the glass becomes a workable liquid at 10⁵-10⁷ Pa·s, so parts can be blow-moulded or hot-embossed like a polymer and quenched back to glass — Liquidmetal casings and SIM-eject tools, unlubricated strain-wave gears developed at NASA JPL for cryogenic mechanisms, surgical blades.

How It Is Tested, Measured and Specified

Diffraction first. A broad halo and no Bragg peaks is the standard amorphicity check, but it is not conclusive: laboratory XRD is blind to crystallinity below roughly 1-2 vol%, so transmission electron microscopy with selected-area diffraction — diffuse rings, no spots — is the confirming test.

Calorimetry defines the material. A DSC scan at 20 K/min on Vitreloy 1 shows an endothermic step at T_g = 623 K, then an exothermic crystallisation peak near T_x ≈ 705 K. The supercooled-liquid window ΔT_x = T_x − T_g ≈ 80 K is the entire thermoplastic processing envelope, and the size of the exotherm says how much of the sample was genuinely glassy. Archimedes weighing before and after deliberate crystallisation resolves the ~0.5% density deficit to ±0.05%, a practical free-volume gauge.

Mechanical testing uses conventional standards: Vickers microhardness per ASTM E384 (Vitreloy 1 measures ~530-560 HV), compression on 2:1 cylinders per ASTM E9, tension on flat dogbones per ASTM E8, notched three-point bend toughness per ASTM E1820. The headline numbers are self-consistent: a yield stress of ~1.9 GPa against a Young's modulus of ~96 GPa gives 1.9/96 ≈ 0.020 — the ~2% elastic strain quoted for the alloy, roughly ten times what structural steel tolerates before yielding (355 MPa over 200 GPa is only ~0.18%). Ferromagnetic ribbon is specified differently: ASTM A883/A883M covers inductive measurement of amorphous strip, and cores are bought on watts per kilogram at a stated induction and frequency. One rule governs real shops — Vitreloy 1 is 22.5 at.% beryllium, so melting, grinding and machining fall under the OSHA beryllium standard and its 0.2 µg/m³ eight-hour exposure limit.

How It Fails, and What It Is Not

Having no dislocations is where the strength comes from and where the trouble starts. A crystal yields by moving dislocations and work-hardens as they tangle; a glass yields through shear transformation zones — clusters of a few dozen atoms shuffling cooperatively — which create free volume locally and leave the sheared material softer than its surroundings. Deformation therefore runs away into shear bands only ~10-20 nm wide. In tension one band does the whole job: the sample fails at its elastic limit with essentially no plastic strain, leaving the vein pattern that marks a momentarily molten layer.

Three other limits matter. Annealing below T_g removes free volume and embrittles the glass — as-cast Vitreloy 1 is reported around 20-60 MPa·m^1/2, but a relaxed sample can fall below 10. Fatigue is worse than the strength suggests, with reported endurance ratios of ~0.05-0.15 against ~0.4-0.5 for steels, which ended the 1990s vogue for metallic-glass golf-club faces. And the material is metastable: hold it above T_x and it devitrifies permanently. NASA's Genesis probe flew Zr-based bulk metallic glass among its solar-wind collectors at L1; in 2004 its return capsule's deceleration sensors proved to have been installed backwards, the parachute never deployed, and it hit the Utah desert at roughly 300 km/h. Much of the science survived — and so, in fragments, did the glass.

Finally, the look-alikes. Quench-hardened steel is the commonest confusion: martensite is made by a fast quench and is hard and brittle, yet it is a crystal — a body-centred tetragonal lattice formed by diffusionless shear with carbon trapped in solution, giving Bragg peaks and no glass transition. Window glass is a glass for the opposite reason: the soda-lime silicate melt it is drawn from is already thousands of times more viscous than any liquid metal and stiffens steeply as it cools, whereas a metal melt at 1-5 mPa·s must be outrun. Nanocrystalline alloys such as Finemet (Fe₇₃.₅Si₁₃.₅B₉Nb₃Cu₁) begin as melt-spun metallic glass and are then deliberately crystallised into 10-15 nm grains. Amorphous silicon and diamond-like carbon are disordered too, but are deposited from vapour and never pass through a liquid.

Amorphous alloys and the quench each one demands

Alloy	Critical cooling rate	Largest amorphous section	Where it is used
Au₇₅Si₂₅ (Klement, Willens & Duwez, 1960)	~10⁶ K/s	splat-quenched foil, ~10 µm	the original proof; laboratory only
Fe₇₈Si₉B₁₃ (Metglas 2605SA1)	~10⁶ K/s	melt-spun ribbon, ~25 µm	distribution-transformer cores, EMI shielding
Zr41.2Ti13.8Cu12.5Ni10Be22.5 (Vitreloy 1, 1993)	~1 K/s	cast rod, ~5 cm	casings, unlubricated gears, sporting goods
Pd₄₀Cu₃₀Ni₁₀P₂₀ (fluxed with B₂O₃, 1997)	~0.1 K/s	cast rod, ~7 cm	laboratory benchmark for glass-forming ability
Pure nickel or copper	~10¹² K/s	nanoscale droplets and thin films only	not achievable as a bulk part