Skip to content
LumenaryAcademy
Processes 75 min

Phase transformations

How a new phase is born and grows — nucleation and the critical radius, the Avrami equation simulated, TTT and CCT diagrams, martensite, and precipitation hardening.

Code tested · 5AI-draftedAwaiting review by a materials scientist

Your guides: Chloe · The Materials Scientist, Margaret · The Tester, Gideon · The Failure Analyst

MotivationH

Same steel, different worlds

Take two identical bars of a steel with 0.8 wt% carbon, heat both until they are austenite, and cool them differently. Let one cool slowly in the furnace: it becomes pearlite, soft enough to machine. Plunge the other into water: it becomes martensite, so hard that a file skids off it, and so brittle that it may crack. The composition is the same and so, below 727 °C, is the equilibrium phase diagram of phase diagrams. What differs is the path, and the path is the business of phase transformations.

Japanese swordsmiths exploited this centuries before anyone could explain it. They coated a blade in clay, thinly along the edge and thickly along the spine, before quenching it. The edge cooled fast and became martensite, hard enough to hold an edge; the spine cooled more slowly and stayed tougher. The boundary between the two shows on the polished blade as the wavy line called the hamon. In 1906, Alfred Wilm in Germany found that an aluminium alloy quenched from a high temperature slowly hardened by itself over several days at room temperature. His alloy, Duralumin, was soon flying in airships and aircraft, but the reason, invisible particles a few nanometres across, was not seen until 1938.

A phase diagram says where a material is heading. This lesson asks how a new phase starts (nucleation), how it spreads (growth and the kinetics of the whole transformation), and how engineers use the answers in steels and aluminium alloys.

IntuitionH

Why change needs a push

Below its melting point, a solid has a lower Gibbs energy than the liquid, so the liquid "wants" to freeze. But the solid cannot appear everywhere at once. It must start as tiny clusters of atoms, and a cluster has a surface, an interface between solid and liquid, which costs energy. For a very small cluster, most of its atoms are at the surface, the surface cost outweighs the gain from the atoms inside, and the cluster is more likely to dissolve than to grow. Only a cluster that by chance grows past a certain critical size gains more than it pays, and then it grows steadily. That chance event is nucleation.

So every transformation needs undercooling (or, for transformations on heating, overheating): the further below the equilibrium temperature, the larger the driving force and the smaller the critical cluster. But most solid-state transformations also need atoms to diffuse, and diffusion slows exponentially as the temperature falls (diffusion). Just below the equilibrium temperature there is plenty of mobility but almost no driving force; far below it there is plenty of driving force but almost no mobility. The transformation is fastest in between. This competition is the key to the C-shaped curves of the TTT diagram later in the lesson.

Analogy, not a formal construction

Starting a new club in a school takes a founding group. One or two enthusiasts give up, because the effort of organising outweighs the fun; once a dozen have joined, the club is self-sustaining and grows. If the school day is very busy, nobody has time to start anything; if the idea is only slightly appealing, nobody bothers. New clubs appear fastest when the idea is attractive and people have time to move around.

DefinitionsH

Nucleation and the critical radius

Definition— homogeneous and heterogeneous nucleation

Homogeneous nucleation happens at random anywhere in the parent phase, with no help. Heterogeneous nucleation happens on something already present, such as a mould wall, an impurity particle, a grain boundary or a dislocation, which supplies part of the interface for free.

Model a nucleus as a sphere of radius rr. Let ΔGv>0\Delta G_v \gt 0 be the drop in Gibbs energy per unit volume when the new phase forms (the driving force), and γ\gamma the energy per unit area of the interface. Then forming the nucleus changes the Gibbs energy by

ΔG(r)=−43πr3 ΔGv+4πr2 γ.\Delta G(r) = -\tfrac43\pi r^3\,\Delta G_v + 4\pi r^2\,\gamma.

Theorem— Critical radius and nucleation barrier

ΔG(r)\Delta G(r) has a maximum at the critical radius and nucleation barrier

r∗=2γΔGv,ΔG∗=16πγ33 ΔGv2.r^* = \frac{2\gamma}{\Delta G_v}, \qquad \Delta G^* = \frac{16\pi\gamma^3}{3\,\Delta G_v^2}.

Clusters smaller than r∗r^* tend to shrink and larger ones tend to grow. For solidification near the melting point TmT_\text{m}, ΔGv≈ΔHv ΔT/Tm\Delta G_v \approx \Delta H_v\,\Delta T/T_\text{m}, where ΔHv\Delta H_v is the latent heat per unit volume and ΔT=Tm−T\Delta T = T_\text{m} - T the undercooling, so r∗∝1/ΔTr^* \propto 1/\Delta T and ΔG∗∝1/ΔT2\Delta G^* \propto 1/\Delta T^2.

Proof

Setting dΔG/dr=−4πr2ΔGv+8πrγ=0d\Delta G/dr = -4\pi r^2\Delta G_v + 8\pi r\gamma = 0 gives r∗=2γ/ΔGvr^* = 2\gamma/\Delta G_v, and the second derivative there, −8πγ<0-8\pi\gamma \lt 0, confirms a maximum. Substituting, ΔG∗=−43π8γ3ΔGv2+4π4γ2ΔGv2γ=16πγ33ΔGv2\Delta G^* = -\tfrac43\pi\tfrac{8\gamma^3}{\Delta G_v^2} + 4\pi\tfrac{4\gamma^2}{\Delta G_v^2}\gamma = \tfrac{16\pi\gamma^3}{3\Delta G_v^2}. For the driving force, ΔGv=ΔHv−TΔSv\Delta G_v = \Delta H_v - T\Delta S_v, and at TmT_\text{m} it vanishes, so ΔSv=ΔHv/Tm\Delta S_v = \Delta H_v/T_\text{m}; if ΔHv\Delta H_v and ΔSv\Delta S_v barely change over the undercooling, ΔGv≈ΔHv(1−T/Tm)=ΔHv ΔT/Tm\Delta G_v \approx \Delta H_v(1 - T/T_\text{m}) = \Delta H_v\,\Delta T/T_\text{m}.

∎

The rate at which nuclei form per unit volume follows a Boltzmann factor, I=I0exp⁡(−ΔG∗/kBT)I = I_0\exp(-\Delta G^*/k_\text{B}T), with a prefactor I0I_0 that counts atoms and their attempts to join a cluster. Because ΔG∗\Delta G^* sits in an exponent and depends on 1/ΔT21/\Delta T^2, the rate switches on abruptly. The block computes this for liquid copper, with a handbook latent heat, the interface energy that David Turnbull estimated from his 1950 undercooling experiments, and an order-of-magnitude prefactor.

python
import math

# Homogeneous nucleation of solid copper in its undercooled liquid.
k_B = 1.380649e-23        # J/K
N_A = 6.02214076e23       # 1/mol
Tm = 1357.8               # K, melting point
dH = 13.26e3              # J/mol, enthalpy of fusion (handbook value, rounded)
Vm = 7.1e-6               # m^3/mol, molar volume of solid copper
gamma = 0.177             # J/m^2, solid-liquid interface energy (Turnbull's 1950 estimate)
I0 = 1e40                 # nuclei per m^3 per s: attempt rate, order of magnitude only

dHv = dH / Vm             # J/m^3
v_atom = Vm / N_A         # m^3 per atom
print(f"latent heat per volume {dHv:.2e} J/m^3, volume per atom {v_atom:.2e} m^3\n")
print(" dT (K)   dGv (J/m^3)   r* (nm)   atoms in nucleus   dG*/kT   log10 I (per m^3 per s)")
for dT in (20, 50, 100, 150, 200, 220, 240, 260):
    dGv = dHv * dT / Tm                      # driving force per unit volume
    r_star = 2 * gamma / dGv
    dG_star = 16 * math.pi * gamma**3 / (3 * dGv**2)
    n_star = 4 / 3 * math.pi * r_star**3 / v_atom
    T = Tm - dT
    barrier = dG_star / (k_B * T)
    log_I = math.log10(I0) - barrier / math.log(10)
    print(f"{dT:6d}   {dGv:11.2e}   {r_star * 1e9:7.2f}   {n_star:16.0f}   {barrier:6.0f}   {log_I:8.1f}")

# Undercooling at which a 1 cm^3 sample would nucleate about once per second (I = 1e6 per m^3 per s).
lo, hi = 50.0, 400.0
for _ in range(60):
    dT = (lo + hi) / 2
    dGv = dHv * dT / Tm
    barrier = 16 * math.pi * gamma**3 / (3 * dGv**2) / (k_B * (Tm - dT))
    lo, hi = (dT, hi) if barrier > math.log(I0 / 1e6) else (lo, dT)
print(f"\nhomogeneous nucleation becomes noticeable near dT = {dT:.0f} K ({dT / Tm:.2f} Tm)")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

At 20 K of undercooling the critical nucleus would hold about 760 000 atoms and the barrier is 6647 kBTk_\text{B}T: the predicted rate, 10−284710^{-2847} nuclei per cubic metre per second, means never. At 200 K the critical nucleus has shrunk to about 760 atoms, with a radius of 1.3 nm, and the rate has climbed through more than 2800 orders of magnitude. Homogeneous nucleation becomes noticeable near 198 K of undercooling, 0.15 TmT_\text{m}; Turnbull's carefully cleaned droplets of many metals froze after undercooling by roughly a fifth of their absolute melting temperatures, the same order.

+30−30123radius r (nm)ΔG (aJ)surface: +4πr²γvolume: −(4/3)πr³ΔGᵥΔG* at r*r* = 1.29 nmtotal ΔG(r)
The Gibbs energy of a spherical copper nucleus at 200 K of undercooling, computed from the values in the block. The surface term (rising) wins for small clusters and the volume term (falling) for large ones. Their sum has a maximum, the nucleation barrier ΔG* ≈ 1.2 aJ (1 aJ = 10⁻¹⁸ J), at the critical radius r* ≈ 1.3 nm.

Real liquids almost never get that far, because they nucleate heterogeneously. A nucleus forming as a cap on a flat surface, meeting it at a contact angle θ\theta, has the same critical radius but needs fewer atoms, and the barrier is reduced by the factor

S(θ)=14 (2+cos⁡θ)(1−cos⁡θ)2,S(\theta) = \tfrac14\,(2 + \cos\theta)(1 - \cos\theta)^2,

which is 1 for θ=180°\theta = 180° (the surface does not help) and falls to 0.156 at θ=60°\theta = 60° and 0.013 at θ=30°\theta = 30°. Because the factor multiplies an exponent of tens to thousands of kBTk_\text{B}T, a good substrate makes nucleation easy at undercoolings of a few kelvin. Foundries exploit this on purpose: adding grain refiners, such as aluminium–titanium–boron master alloys that release tiny boride particles into molten aluminium, gives many nuclei and a fine, uniform grain size. Cloud seeding with silver iodide particles does the same for ice in supercooled clouds, whose pure droplets can stay liquid down to about −38 °C.

ComputationC

Growth and the Avrami equation

Once nuclei exist, they grow. When atoms only need to hop across the interface, as in solidification of a pure metal or recrystallisation, the interface moves at a steady speed vv and a grain's radius grows in proportion to tt. When growth needs solute to diffuse to or from the interface, as for a carbon-poor ferrite grain growing in austenite, the radius grows only as Dt\sqrt{Dt}, the diffusion length of the previous lesson.

The fraction transformed, f(t)f(t), then rises along an S-shaped curve: slowly at first while nuclei are few and small, quickly in the middle, and slowly at the end as growing grains meet one another (impingement) and the untransformed regions shrink. The theory was worked out between 1937 and 1941 by Andrey Kolmogorov, William Johnson and Robert Mehl, and Melvin Avrami, and goes by their initials, JMAK.

Theorem— The Johnson–Mehl–Avrami–Kolmogorov equation

Let nuclei appear at random positions in the untransformed material, at a constant rate II per unit volume, and let each grow as a sphere at constant speed vv until it meets others. Then the fraction transformed is

f(t)=1−exp⁡ ⁣(−π3 Iv3t4),f(t) = 1 - \exp\!\left(-\frac{\pi}{3}\,I v^3 t^4\right),

and more generally f=1−exp⁡(−ktn)f = 1 - \exp(-kt^n), where the Avrami exponent nn depends on how nucleation and growth happen. In two dimensions (a thin film), the same assumptions give f=1−exp⁡(−π3Iv2t3)f = 1 - \exp(-\tfrac{\pi}{3}Iv^2t^3).

Proof

Imagine that grains could grow through each other, and that nuclei could also form inside already transformed material ("phantom" nuclei). The total volume fraction they would occupy, counting overlaps more than once, is the extended fraction fextf_\text{ext}. A nucleus born at time τ\tau has radius v(t−τ)v(t - \tau) at time tt, so in three dimensions fext=∫0tI⋅43πv3(t−τ)3 dτ=π3Iv3t4f_\text{ext} = \int_0^t I\cdot\tfrac43\pi v^3(t - \tau)^3\,d\tau = \tfrac{\pi}{3}Iv^3t^4. Now account for overlap. In a short interval, the extended fraction grows by dfextdf_\text{ext}, and because nuclei are placed at random, only the part that lands in untransformed material, a fraction 1−f1 - f of it, is real: df=(1−f) dfextdf = (1 - f)\,df_\text{ext}. Integrating, −ln⁡(1−f)=fext-\ln(1 - f) = f_\text{ext}, which gives the result. Phantom nuclei cause no error: a phantom born inside a grain growing at the same speed stays inside that grain. In two dimensions the area of a disc, πv2(t−τ)2\pi v^2(t - \tau)^2, replaces the sphere's volume.

∎

The exponent nn is a fingerprint of the mechanism. With interface-controlled growth in three dimensions, n=4n = 4 for a constant nucleation rate and n=3n = 3 for site saturation, when all nuclei are present at the start; with diffusion-controlled growth the corresponding values are 2.5 and 1.5. In two dimensions each is one smaller. Plotting ln⁡(−ln⁡(1−f))\ln(-\ln(1-f)) against ln⁡t\ln t gives a straight line of slope nn, the Avrami plot.

The block tests the theory by simulation. Nuclei appear at random places and times in a periodic unit square, and each grows as a disc at constant speed; a 256 × 256 grid records when each cell is first reached. It compares the simulated fraction with the JMAK formula, fits the Avrami exponent, and repeats the experiment with site saturation.

python
import math
import numpy as np

rng = np.random.default_rng(7)
N = 256                                   # grid of N x N cells on a unit square, periodic edges
xs = (np.arange(N) + 0.5) / N
X, Y = np.meshgrid(xs, xs)
v = 0.01                                  # growth speed: distance per unit time

def arrival_times(nuclei):
    """Time at which each cell is first reached by a growing disc.  nuclei: rows (x, y, t0)."""
    first = np.full((N, N), np.inf)
    for x0, y0, t0 in nuclei:
        dx = np.abs(X - x0); dx = np.minimum(dx, 1 - dx)     # periodic (minimum image)
        dy = np.abs(Y - y0); dy = np.minimum(dy, 1 - dy)
        np.minimum(first, t0 + np.hypot(dx, dy) / v, out=first)
    return first

def avrami_exponent(t, f):
    keep = (f > 0.05) & (f < 0.95)
    slope, _ = np.polyfit(np.log(t[keep]), np.log(-np.log(1 - f[keep])), 1)
    return slope

t = np.linspace(0.25, 10, 40)

# 1. Constant nucleation rate I (per unit area per unit time) for 10 time units.
I = 45.0
count = rng.poisson(I * 1.0 * 10)
nuclei = np.column_stack([rng.random(count), rng.random(count), rng.uniform(0, 10, count)])
first = arrival_times(nuclei)
f_sim = np.array([np.mean(first <= ti) for ti in t])
f_jmak = 1 - np.exp(-math.pi / 3 * I * v**2 * t**3)
print(f"constant nucleation: {count} nuclei (phantoms included)")
print("   t    simulated f   JMAK f")
for i in range(3, 40, 4):
    print(f"{t[i]:5.2f}   {f_sim[i]:11.3f}   {f_jmak[i]:6.3f}")
print(f"fitted Avrami exponent n = {avrami_exponent(t, f_sim):.2f}  (theory: 3 in two dimensions)")
print("\nthe square at t = 5 (# transformed, . not yet):")
for row in range(0, N, 12):
    print("".join("#" if first[row, col] <= 5 else "." for col in range(0, N, 6)))

# 2. Site saturation: all nuclei present at t = 0.
count = 120
nuclei = np.column_stack([rng.random(count), rng.random(count), np.zeros(count)])
first = arrival_times(nuclei)
f_sim = np.array([np.mean(first <= ti) for ti in t])
print(f"\nsite saturation: {count} nuclei at t = 0")
print(f"fitted Avrami exponent n = {avrami_exponent(t, f_sim):.2f}  (theory: 2 in two dimensions)")
print(f"half transformed at t = {np.interp(0.5, f_sim, t):.2f}; "
      f"JMAK predicts {math.sqrt(math.log(2) / (math.pi * count * v**2)):.2f}")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

The simulated fractions follow the JMAK curve to within about 0.02 all the way from 0.4 % to 99.8 % transformed; the small excess late on comes from the random draw, which happened to produce 458 nuclei where 450 were expected. The fitted exponents are 3.08 and 2.08, against 3 and 2 in theory, and with site saturation the half-transformation time, 4.28, matches the formula's 4.29. The snapshot at t=5t = 5, when the square is 45 % transformed, shows what the formula does not: discs of every size, the older ones large and already merging, the newest ones single cells.

Worked examplesH

TTT and CCT diagrams

Steel heat treatment is planned with two kinds of map, both introduced by Edgar Bain and his colleagues at United States Steel from 1930. An isothermal transformation or time–temperature–transformation (TTT) diagram is made by quenching small samples of austenite to a fixed temperature below 727 °C, holding them for different times, and recording when the transformation starts (often taken as 1 %) and finishes (99 %). Plotted against a logarithmic time axis, the start and finish curves are shaped like a C: slow near 727 °C, where the driving force is small, slow at low temperatures, where diffusion is slow, and fastest at the nose in between.

The block builds an illustrative start curve from this reasoning: an Arrhenius factor for diffusion times a driving force that grows as the cube of the undercooling, with the nose placed at 550 °C and 1 s, roughly where it sits for a plain eutectoid steel. In the figure the finish curve lies a factor of 7.7 later at every temperature, which is what the JMAK equation with n=3n = 3 gives between 1 % and 99 % transformed.

1003005007008000.1110100100010⁴time (s, logarithmic scale)temperature (°C)727 °CMsmartensitepearlitebainiteaustenitestartfinish400 K/s10 K/s
An illustrative TTT diagram for a plain eutectoid steel, computed from the model in the next block, with two cooling paths. Above the nose the product is pearlite, finer at lower temperatures; below it, bainite. Below the martensite start temperature Ms, about 220 °C here, austenite that has not transformed becomes martensite. The slow cooling (10 K/s) crosses the C-curves and gives pearlite; the fast quench (400 K/s) misses the nose and gives martensite.

A part being quenched does not jump to one temperature and wait; it cools continuously. A continuous-cooling transformation (CCT) diagram, measured with dilatometers that cool samples at controlled rates, is the more useful map for that, and its curves sit lower and later than the TTT curves. A rough way to get one from the other is Scheil's additivity rule: treat the cooling path as a series of short isothermal holds and say the transformation starts when the fractions of the incubation time used up add to one, ∑Δt/τ(T)=1\sum \Delta t/\tau(T) = 1. The block applies it to find the slowest cooling that still misses the nose, the critical cooling rate for a fully martensitic structure.

python
import math

# An illustrative model of the start of transformation in a plain eutectoid steel:
# rate ~ (diffusion, Arrhenius) x (driving force, ~ undercooling cubed), so
#   tau(T) = tau_n * exp(Q/R (1/T - 1/T_n)) * ((Te - T_n) / (Te - T))**3,
# with Q chosen so that the "nose" (the shortest tau) is at T_n.  Temperatures in kelvin.
R = 8.314
Te = 727 + 273.15             # eutectoid temperature: no transformation above it
T_n, tau_n = 550 + 273.15, 1.0   # nose at 550 C, 1 s (roughly like a plain eutectoid steel)
Q = 3 * R * T_n**2 / (Te - T_n)   # d(ln tau)/dT = 0 at T_n
Ms = 220 + 273.15             # martensite start, roughly, for 0.76 wt% C

def tau(T):
    if T >= Te:
        return math.inf
    return tau_n * math.exp(Q / R * (1 / T - 1 / T_n)) * ((Te - T_n) / (Te - T)) ** 3

print(f"Q = {Q / 1000:.0f} kJ/mol\n")
print("isothermal start times (the TTT 'start' curve):")
for T_c in (700, 650, 600, 550, 500, 450, 400, 300):
    print(f"  {T_c} C: {tau(T_c + 273.15):7.2f} s")

def cooling_start(rate, T0=800 + 273.15, dt=1e-3):
    """Cool at a constant rate (K/s) from T0.  Scheil's additivity rule: the transformation
    starts when the sum of dt / tau(T) reaches 1.  Return the start temperature (K) or None."""
    total, T = 0.0, T0
    while T > Ms:
        total += dt / tau(T)
        if total >= 1:
            return T
        T -= rate * dt
    return None

print("\ncontinuous cooling from 800 C:")
for rate in (1, 5, 20, 50, 100, 150, 200):
    T = cooling_start(rate, dt=min(0.01, 0.2 / rate))
    result = f"starts at {T - 273.15:.0f} C" if T else "no transformation before Ms: martensite"
    print(f"  {rate:4d} K/s: {result}")

lo, hi = 50.0, 400.0                 # bisection for the critical cooling rate
for _ in range(30):
    mid = (lo + hi) / 2
    lo, hi = (mid, hi) if cooling_start(mid, dt=0.2 / mid) else (lo, mid)
print(f"\ncritical cooling rate for martensite: about {mid:.0f} K/s")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

The model's start times, 33 s at 700 °C, 1 s at the nose and 31 s at 300 °C, have the shape of a real plain-carbon steel's diagram. On continuous cooling, the transformation starts lower than any isothermal curve would suggest: at 627 °C for 20 K/s and at 536 °C for 100 K/s, and cooling faster than about 203 K/s from 800 °C avoids it altogether. Real plain carbon steels need critical rates of the same order, which only a vigorous water quench of a thin section can deliver.

That is why alloy steels exist. Adding manganese, chromium, nickel and molybdenum, or a few thousandths of a per cent of boron, slows the diffusional transformations and shifts the C-curves to the right, so slower quenches, in oil or even air, still produce martensite, and thicker sections harden right through. This property is hardenability, and it is measured by the Jominy end-quench test (standardised as ASTM A255 and ISO 642): a bar austenitised in a furnace is sprayed with water on one end only, so each position along it has cooled at a different, known rate, and hardness measured along the bar maps hardness against cooling rate.

ExplanationH

Martensite

If austenite is cooled so fast that carbon has no time to diffuse, it still transforms, but in a completely different way. Below the martensite start temperature MsM_\text{s}, whole regions of the crystal shear into a new structure in a coordinated movement of atoms, each moving less than one interatomic distance, with interfaces sweeping through the crystal at a large fraction of the speed of sound. Because there is no diffusion, the product, martensite, has exactly the composition of the austenite it came from: it is iron supersaturated with carbon. The carbon atoms, trapped in gaps that are too small for them in a body-centred lattice, stretch it along one axis into a body-centred tetragonal cell, whose axial ratio c/ac/a grows with carbon content (approximately 1+0.0451 + 0.045 per wt% carbon). The strained lattice resists the motion of dislocations very strongly, which is why martensite is so hard and, in higher-carbon steels, so brittle.

Definition— athermal martensite

In steels, the amount of martensite depends on how far below MsM_\text{s} the steel is cooled, not on how long it is held there. Each new degree of undercooling produces new plates almost instantly; holding does almost nothing. Transformation is complete only at the martensite finish temperature MfM_\text{f}, and austenite that has not transformed when cooling stops is retained austenite.

Two empirical formulas from the steel industry put numbers on this. K. W. Andrews' 1965 equation estimates the start temperature from composition (wt%): Ms (°C)≈539−423 C−30.4 Mn−17.7 Ni−12.1 Cr−7.5 MoM_\text{s}\,(°\text{C}) \approx 539 - 423\,\text{C} - 30.4\,\text{Mn} - 17.7\,\text{Ni} - 12.1\,\text{Cr} - 7.5\,\text{Mo}. Carbon lowers MsM_\text{s} dramatically, from about 500 °C in a very low-carbon steel to roughly 220 °C in a eutectoid steel and lower still at higher carbon contents, which is why high-carbon tool steels can retain austenite and are sometimes cooled below 0 °C to finish the transformation. And Koistinen and Marburger's 1959 equation gives the fraction transformed at a temperature TT below MsM_\text{s}: fM=1−exp⁡[−0.011 (Ms−T)]f_\text{M} = 1 - \exp[-0.011\,(M_\text{s} - T)], with temperatures in kelvin or Celsius.

Martensitic transformations are not only for steels. In nickel–titanium alloys (developed at the US Naval Ordnance Laboratory around 1960) the martensite plates can move back and forth: deformed when martensitic, the alloy returns to its original shape when warmed (the shape-memory effect), and just above its transformation temperature it can be stretched by several per cent and spring back (superelasticity), which is used in self-expanding stents, orthodontic wires and actuators.

Worked examplesH

Precipitation hardening

Wilm's Duralumin, and the 2xxx, 6xxx and 7xxx aluminium alloys that build aircraft, bicycle frames and car parts today, are strengthened by precipitation hardening (also called age hardening). The phase diagram makes it possible: in the aluminium–copper system, aluminium dissolves up to 5.65 wt% copper at 548 °C, but almost none at room temperature. The treatment has three steps.

  1. Solution treatment. Heat an alloy with, say, 4 wt% copper into the single-phase field, around 500 to 530 °C, until all the copper is dissolved.
  2. Quench. Cool it quickly to room temperature, too fast for the equilibrium phase to form. The result is a supersaturated solid solution, with far more copper dissolved than equilibrium allows, and many quenched-in vacancies that will help it diffuse.
  3. Ageing. Hold it at room temperature (natural ageing) or at about 150 to 190 °C (artificial ageing). The copper comes out of solution as a dense population of tiny particles.

The particles do not go straight to the equilibrium phase θ (Al₂Cu). Its interface with aluminium would cost too much energy, so the nucleation barrier is high. Instead the alloy passes through a sequence of metastable structures, each with an easier interface: first Guinier–Preston zones, discs of copper atoms one or two atomic layers thick and a few nanometres across, discovered independently by André Guinier and George Preston in 1938 from faint streaks in X-ray patterns; then the coherent θ″ and semi-coherent θ′; and finally θ. This route, where the first product is not the most stable phase but the easiest to nucleate, is common enough to have a name: Ostwald's step rule.

0.1110100100010⁴ageing time (hours, logarithmic scale)hardnesspeak agedunderagedoveraged190 °C150 °C
Schematic ageing curves for an aluminium–copper alloy at two temperatures. Hardness rises as fine precipitates form, peaks, and then falls as they coarsen (overageing). At the higher temperature diffusion is faster, so the peak comes sooner but is lower, because the driving force for precipitation is smaller and the particles are coarser.

Why do the particles strengthen the alloy, and why does too much ageing soften it again? Strength comes from obstacles to dislocations (defects and dislocations). While the particles are tiny and coherent with the matrix, dislocations shear through them, and the resistance grows as the particles grow. Once the particles are larger and further apart, dislocations find it easier to bow out between them and pass by, leaving loops around them (the Orowan mechanism), and the resistance now falls as the spacing grows. The peak strength sits at the crossover. With longer ageing, the particles coarsen: large ones grow at the expense of small ones, which dissolve because their high surface-to-volume ratio raises their solubility, and Lifshitz, Slyozov and Wagner showed in 1961 that the mean radius grows as t1/3t^{1/3}. Fewer, larger, more widely spaced particles mean an overaged, softer alloy. Elasticity, plasticity and strengthening puts numbers on both mechanisms.

The same idea strengthens nickel superalloys in jet-engine turbine blades, packed with cube-shaped precipitates of the ordered phase Ni₃(Al,Ti); maraging steels; and some copper alloys for springs and connectors.

ExplanationH

Transformations in research and industry

Simulation. The phase-field method describes a microstructure by smooth fields that change across interfaces a few nanometres wide and evolves them from the free energy, predicting dendrites, precipitate shapes and martensite plates without tracking each interface by hand (see Long-Qing Chen's review); open-source frameworks include PRISMS-PF and MOOSE. For precipitation, numerical models combining classical nucleation, growth and coarsening, such as the Kampmann–Wagner approach in MatCalc (Austria) and Thermo-Calc's precipitation module (Sweden), predict particle sizes and ageing curves from a heat-treatment schedule.

Seeing transformations. Synchrotron X-ray and neutron diffraction follow phase fractions inside bulk samples in real time during heating and cooling. Atom-probe tomography maps individual atoms in three dimensions and can resolve GP zones and the first clusters of solute atoms, the stages that classical theory treats most crudely. Transmission electron microscopy (microscopy) sees precipitates and martensite plates directly.

Engineering. Steelmakers design controlled rolling and cooling around CCT diagrams, producing advanced high-strength steels for car bodies with tailored mixtures of ferrite, bainite, martensite and retained austenite. As of 2026, integrated computational materials engineering increasingly links process simulation, transformation models and property prediction into one chain, with machine learning used to speed up the individual models (see multiscale modelling).

Common misconceptionsH

Common misconceptions

Misconception

"The more you undercool, the faster the transformation." Only for transformations that need no long-range diffusion. For diffusional transformations, speed rises with undercooling at first and then falls, giving the nose of the C-curve. Martensite is the exception that proves the point: it needs no diffusion, so it can form at low temperatures where diffusional transformations are frozen out.

Misconception

"TTT diagrams tell you what happens when you quench a part." They describe isothermal holds. A part cooling continuously needs a CCT diagram, whose curves are lower and later, and different positions in a thick part cool at different rates.

Misconception

"Martensite forms because carbon diffuses into the iron fast." Martensite forms precisely because carbon has no time to diffuse: it is the product of a diffusionless shear that traps carbon in the lattice.

Misconception

"Quenching hardens an aluminium alloy, as it hardens steel." Straight after the quench, a precipitation-hardening aluminium alloy is at its softest: the supersaturated solution is soft and easily formed. The hardness comes later, during ageing. Aircraft rivets made of some of these alloys used to be kept in freezers after quenching to delay natural ageing until they were driven.

Exercises

Exercises

Take R=8.314R = 8.314 J mol⁻¹ K⁻¹. NumPy's polyfit fits straight lines.

1. Phase transformations

Warm-upHuman

A precipitate forming in a solid has an interface energy γ=0.20\gamma = 0.20 J m⁻² and a driving force ΔGv=1.0×108\Delta G_v = 1.0 \times 10^8 J m⁻³. Treating the nucleus as a sphere and ignoring strain energy, what is the critical radius in nanometres?

2. Phase transformations

Warm-upHuman

Nuclei of a new phase form at a constant rate throughout a bulk (three-dimensional) sample, and each grows as a sphere whose radius increases at a constant speed. What is the Avrami exponent nn?

3. Phase transformations

CoreHuman

Why does the start curve of a TTT diagram for the pearlite transformation have a nose, a temperature at which the transformation begins soonest?

Choose one

4. Phase transformations

CoreHuman

A recrystallisation follows the Avrami equation f=1−exp⁡(−ktn)f = 1 - \exp(-kt^n) with n=2n = 2, and is half complete after 40 s. After how many seconds is it 90 % complete? Give the nearest whole number.

5. Phase transformations

CoreHuman

A steel with Ms=220M_\text{s} = 220 °C is quenched to 20 °C. Using the Koistinen–Marburger equation fM=1−exp⁡[−0.011 (Ms−T)]f_\text{M} = 1 - \exp[-0.011\,(M_\text{s} - T)], what percentage of retained austenite remains? Give the nearest whole number.

6. Phase transformations

CoreHuman

Estimate the martensite start temperature, in °C to the nearest degree, of a steel with 0.50 wt% carbon and 1.00 wt% manganese (and no other alloying elements), using Andrews' formula Ms=539−423 C−30.4 Mn−17.7 Ni−12.1 Cr−7.5 MoM_\text{s} = 539 - 423\,\text{C} - 30.4\,\text{Mn} - 17.7\,\text{Ni} - 12.1\,\text{Cr} - 7.5\,\text{Mo}.

7. Phase transformations

CoreComputational

Write avrami_fit(t, f), which fits the Avrami equation f=1−exp⁡(−ktn)f = 1 - \exp(-kt^n) to measured fractions transformed and returns (n, k). Use the Avrami plot: a straight line of ln⁡(−ln⁡(1−f))\ln(-\ln(1-f)) against ln⁡t\ln t, using only the points with 0.01<f<0.990.01 \lt f \lt 0.99 (readings of exactly 0 or 1, or very close to them, carry no information about the curve and would break the logarithms).

Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

8. Phase transformations

StretchComputational

Write start_on_cooling(rate, tau, T0, Ms, dt=0.01), which applies Scheil's additivity rule to a steel cooled at a constant rate (K/s) from T0 (K). tau is a function giving the isothermal incubation time (s) at temperature T (it may return math.inf above the equilibrium temperature). Step forward in time by dt, adding dt / tau(T) at each step; return the temperature at which the sum first reaches 1, or None if the steel cools to Ms first. The tests use a constant incubation time and the lesson's C-curve model.

Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

Further reading and resources