From atoms to parts
An aluminium engine block is cast, solution-treated, quenched and aged. Its fatigue life depends on the pores and dendrites left by solidification, on the size of the precipitates formed during ageing, on the residual stresses from quenching, and on how all of those vary from a thin wall to a thick boss. Each of those features is set at a different scale: atoms jump to form precipitates nanometres across, dendrites grow over micrometres to millimetres, stresses build up over the whole casting. To design the alloy, the process and the part together, an engineer needs models that connect the scales.
No single simulation can do it. A kilogram of aluminium holds about atoms, and a heat treatment of eight hours is about molecular dynamics steps. The answer is a ladder of models, each built for one range of sizes and times, each taking its parameters from the rung below and handing its results to the rung above. The two previous lessons, atomistic simulation and density functional theory, are the bottom rungs. This lesson climbs the rest: phase-field models for microstructure, crystal plasticity and finite elements for mechanical behaviour, the engineering practice that links them, and the discipline of checking whether any of it can be trusted.
A ladder of models
Each model on the ladder throws away detail that does not matter at its scale and keeps a few quantities that do. Atomistic simulations keep atoms and lose electrons; phase-field models keep smooth fields such as composition and lose atoms; crystal plasticity keeps slip on crystal planes and loses individual dislocations; finite-element models of a part keep stress and strain and lose individual grains.
There are two ways to connect the rungs. In hierarchical (or sequential) multiscale modelling, a lower-scale model is run once to compute parameters that a higher-scale model then uses: DFT gives the energy barrier for a vacancy jump, kinetic Monte Carlo turns barriers into a diffusion coefficient, a phase-field model uses the diffusion coefficient to grow precipitates, a strength model turns precipitate sizes into a yield stress, and a finite-element model of the part uses the yield stress. Most industrial work is hierarchical. In concurrent multiscale modelling, models at two scales run at the same time in different regions of one simulation: quantum mechanics around a crack tip or a reacting molecule embedded in a classical force field (QM/MM, recognised by the 2013 Nobel Prize in Chemistry), or full atomistic detail near a dislocation core embedded in a continuum, as in the quasicontinuum method. Concurrent methods are powerful but delicate, because the seams between the regions must not reflect waves or create spurious forces.
Phase fields and diffuse interfaces
Microstructures are made of regions, grains, phases, precipitates and domains, separated by interfaces, and they evolve by moving those interfaces. Tracking each interface as a sharp moving boundary is hard when interfaces merge, split and change shape. The phase-field method replaces each sharp interface by a thin layer across which a smooth field changes from one value to another, and lets that field evolve by a partial differential equation. Topology changes then take care of themselves.
Definition— order parameter
A phase field or order parameter is a field over space that labels the local state of the material: for example in the parent phase and 1 in the new one, or the local composition . It is conserved if its integral over the system cannot change (composition, because atoms are neither created nor destroyed) and non-conserved otherwise (which grain or which ordered variant a region belongs to).
Definition— Cahn–Hilliard free-energy functional
The free energy of a non-uniform system is written, following John Cahn and John Hilliard (1958), as
where is the free energy per unit volume of uniform material, with a minimum for each phase (a double well such as ), and the gradient-energy coefficient makes sharp changes costly. Interfaces cost energy because they force the material through the hump of and because they have gradients.
Definition— Allen–Cahn and Cahn–Hilliard equations
A non-conserved order parameter relaxes down the gradient of (the Allen–Cahn equation); a conserved one moves by a diffusion flux driven by gradients in , which plays the role of a chemical potential (the Cahn–Hilliard equation):
and are mobilities. Both equations always lower ; Cahn–Hilliard also conserves .
Theorem— The equilibrium interface
For , a flat interface between and is in equilibrium with the profile , of width , and its energy per unit area is .
Proof
In equilibrium , so . Multiplying by and integrating from far away, where and , gives , so . This logistic equation has the solution with . The excess energy is , because the two terms are equal. Changing variable from to , .
∎
The two material parameters and therefore fix two observables: the interface energy, which is physical and measurable, and the interface width, which in most applications is made far wider than the real interface (micrometres instead of nanometres) so that the grid can resolve it. Doing that without changing the physics needs care, the subject of "thin-interface" analyses by Alain Karma and others.
Verifying a phase-field code
The block solves the Allen–Cahn equation twice by explicit finite differences. First, a sharp step in one dimension relaxes into an interface, which is compared with the exact profile and interface energy from the theorem. Second, a circular grain in two dimensions shrinks because its curved boundary has extra energy. In the limit of a thin interface, Allen–Cahn boundaries move with a speed proportional to their curvature, , so the grain's area should fall at the constant rate , whatever its size: the law of curvature-driven grain growth.
import math
import numpy as np
# Allen-Cahn: d(phi)/dt = -L (f'(phi) - kappa lap(phi)), with f = W phi^2 (1 - phi)^2.
# Free energy F = integral of f(phi) + (kappa/2) |grad phi|^2. Units are arbitrary but consistent.
L_mob, kappa, W = 1.0, 1.0, 2.0
def fprime(p):
return 2 * W * p * (1 - p) * (1 - 2 * p)
# 1. A flat interface in 1-D: relax a sharp step and compare with the exact profile.
dx, dt = 0.1, 0.002
x = np.arange(-15, 15 + dx / 2, dx)
phi = (x > 0).astype(float)
for _ in range(10000): # to t = 20
lap = np.zeros_like(phi)
lap[1:-1] = (phi[2:] - 2 * phi[1:-1] + phi[:-2]) / dx**2
phi += dt * L_mob * (kappa * lap - fprime(phi))
phi[0], phi[-1] = 0.0, 1.0
delta = math.sqrt(2 * kappa / W)
x0 = np.interp(0.5, phi, x) # where the interface ended up
exact = 0.5 * (1 + np.tanh((x - x0) / delta))
grad = np.gradient(phi, dx)
gamma = np.sum(W * phi**2 * (1 - phi) ** 2 + 0.5 * kappa * grad**2) * dx
print(f"1-D interface centred at x = {x0:.3f}; largest deviation from the exact tanh profile {np.max(np.abs(phi - exact)):.1e}")
print(f"interface energy: simulated {gamma:.4f}, exact sqrt(2 kappa W)/6 = {math.sqrt(2 * kappa * W) / 6:.4f}")
# 2. A circular grain in 2-D shrinks under its own curvature: dA/dt = -2 pi L kappa.
n, h, dt = 128, 0.5, 0.05
X, Y = np.meshgrid((np.arange(n) - n / 2) * h, (np.arange(n) - n / 2) * h)
R0 = 20.0
phi = 0.5 * (1 - np.tanh((np.sqrt(X**2 + Y**2) - R0) / delta))
times, areas = [], []
for step in range(3001):
if step % 500 == 0:
times.append(step * dt)
areas.append(phi.sum() * h * h)
lap = (np.roll(phi, 1, 0) + np.roll(phi, -1, 0) + np.roll(phi, 1, 1) + np.roll(phi, -1, 1) - 4 * phi) / h**2
phi += dt * L_mob * (kappa * lap - fprime(phi))
print("\n t area radius")
for t, A in zip(times, areas):
print(f"{t:5.0f} {A:7.1f} {math.sqrt(A / math.pi):6.2f}")
slope = np.polyfit(times[1:], areas[1:], 1)[0]
print(f"dA/dt: simulated {slope:.3f}, sharp-interface theory -2 pi L kappa = {-2 * math.pi * L_mob * kappa:.3f}")The one-dimensional interface matches the exact profile to everywhere, and its energy, 0.3329, is within 0.1 per cent of the exact 0.3333. The circular grain shrinks from a radius of 20 to 10.4 while its area falls in a straight line, at per unit time against the theoretical , a 2.5 per cent difference that comes from the finite width of the interface and the grid, and shrinks when both are refined.
Spinodal decomposition with Cahn–Hilliard
When an alloy whose free-energy curve bends downwards () is quenched into the two-phase region, it is unstable to any small composition fluctuation and separates without nucleation, by spinodal decomposition (phase transformations). Cahn's linear analysis of 1961 showed that fluctuations of wavenumber grow at the rate : long waves grow slowly because atoms must diffuse far, short waves are suppressed by the gradient energy, and in between there is a fastest-growing wavelength that sets the scale of the early microstructure.
The block quenches a 50:50 alloy with , whose two phases are , on a 128 × 128 periodic grid. It uses a semi-implicit spectral scheme: in Fourier space the stiff fourth-derivative term is treated implicitly and the nonlinear term explicitly, which allows time steps hundreds of times longer than an explicit scheme would. It tracks the characteristic length of the pattern from the structure factor, the same quantity a small-angle scattering experiment measures.
import math
import numpy as np
rng = np.random.default_rng(1958)
# Cahn-Hilliard: dc/dt = M lap( f'(c) - kappa lap(c) ), with f(c) = (c^2 - 1)^2 / 4, so the two
# phases are c = -1 and c = +1. A 50:50 alloy (mean c = 0) quenched into the spinodal region.
n, M, kappa, dt = 128, 1.0, 1.0, 0.5
k = 2 * np.pi * np.fft.fftfreq(n, d=1.0)
k2 = k[:, None] ** 2 + k[None, :] ** 2
c = 0.05 * rng.standard_normal((n, n)) # small random fluctuations about c = 0
c -= c.mean()
def characteristic_length(c):
"""2 pi / (mean wavenumber), weighting each wavenumber by the structure factor |c_k|^2."""
S = np.abs(np.fft.fft2(c - c.mean())) ** 2
S[0, 0] = 0
return 2 * np.pi * S.sum() / (np.sqrt(k2) * S).sum()
print(f"linear theory: fastest-growing wavelength 2 pi sqrt(2 kappa / |f''(0)|) = {2 * np.pi * math.sqrt(2 * kappa):.2f}\n")
print(" time length mean c min c max c")
t, records = 0.0, []
for target in (10, 25, 50, 100, 200, 400, 800):
while t < target - 1e-9: # semi-implicit spectral time step
c_hat = np.fft.fft2(c)
mu_hat = np.fft.fft2(c**3 - c)
c = np.real(np.fft.ifft2((c_hat - dt * M * k2 * mu_hat) / (1 + dt * M * kappa * k2**2)))
t += dt
Lc = characteristic_length(c)
records.append((t, Lc))
print(f"{t:7.0f} {Lc:6.2f} {c.mean():+.1e} {c.min():+.3f} {c.max():+.3f}")
late = [(tt, ll) for tt, ll in records if tt >= 100]
p = np.polyfit(np.log([r[0] for r in late]), np.log([r[1] for r in late]), 1)[0]
print(f"\ncoarsening: length grows as t^{p:.2f} between t = 100 and 800 (Lifshitz-Slyozov: t^(1/3) at late times)")
print("\nthe microstructure at t = 800 (the top half of the box, '#' where c > 0):")
for row in c[:64:4]:
print("".join("#" if v > 0 else "." for v in row[::2]))Early on, at , the pattern's characteristic length is 9.4, close to the 8.9 predicted by linear theory, and the composition has separated only to about ±0.25. By the two phases have reached their equilibrium compositions near ±1, and from then on the pattern coarsens: small domains dissolve and large ones grow to reduce the total interface energy, with the characteristic length growing as about , approaching the law of Lifshitz, Slyozov and Wagner for diffusion-controlled coarsening. The mean composition stays at zero to within throughout: the scheme conserves atoms exactly, as Cahn–Hilliard must. The interconnected, sponge-like pattern is characteristic of spinodal decomposition at equal fractions; it is seen in iron–chromium ferrite aged near 475 °C, where it embrittles duplex stainless steels, in copper–nickel–tin spring alloys strengthened on purpose by it, and in the phase-separated borosilicate glass from which porous glass is made.
Phase-field models are now used for solidification (the dendrites of castings and welds, and of the melt pools in additive manufacturing), precipitation in aluminium alloys and nickel superalloys, grain growth and recrystallisation, ferroelectric and magnetic domains, crack growth (phase-field fracture), and the growth of lithium dendrites in batteries.
Crystal plasticity and finite elements, in outline
Phase-field models predict structure; the next rungs predict properties from it. A single crystal yields when the resolved shear stress on one of its slip systems reaches a critical value , Schmid's law from elasticity, plasticity and strengthening: under a tensile stress , the resolved shear stress is , where is the Schmid factor. A polycrystal is harder to predict, because each grain is oriented differently and must deform compatibly with its neighbours. The block computes Schmid factors for the 12 slip systems of a face-centred cubic crystal, for single crystals and for a random polycrystal.
import numpy as np
rng = np.random.default_rng(1938)
# The 12 slip systems of an FCC crystal: four {111} planes, each with three <110> directions.
planes = [np.array(p) / np.sqrt(3) for p in ([1, 1, 1], [-1, 1, 1], [1, -1, 1], [1, 1, -1])]
directions = [np.array(d) / np.sqrt(2) for d in ([1, 1, 0], [1, -1, 0], [1, 0, 1], [1, 0, -1], [0, 1, 1], [0, 1, -1])]
systems = [(p, d) for p in planes for d in directions if abs(p @ d) < 1e-12]
def max_schmid(axis):
"""Largest Schmid factor |cos(phi) cos(lambda)| over the slip systems, for tension along `axis`."""
axis = np.asarray(axis, dtype=float) / np.linalg.norm(axis)
return max(abs((p @ axis) * (d @ axis)) for p, d in systems)
print(f"{len(systems)} slip systems")
for name, axis in [("[001]", [0, 0, 1]), ("[011]", [0, 1, 1]), ("[111]", [1, 1, 1]), ("[123]", [1, 2, 3])]:
m = max_schmid(axis)
print(f"single crystal pulled along {name}: largest Schmid factor {m:.3f}, yields at {1 / m:.2f} tau_c")
# A random polycrystal: in each grain's own crystal axes the tension axis points in a random direction.
A = rng.normal(size=(20000, 3))
A /= np.linalg.norm(A, axis=1, keepdims=True)
P = np.array([p for p, d in systems])
D = np.array([d for p, d in systems])
m = np.max(np.abs((A @ P.T) * (A @ D.T)), axis=1)
print(f"\nrandom grains: mean largest Schmid factor {m.mean():.3f}; spread {m.min():.3f} to {m.max():.3f}")
print(f"Sachs estimate (each grain slips on its own best system): sigma_y = {1 / m.mean():.2f} tau_c")
print("Taylor estimate (each grain must follow the polycrystal's strain): sigma_y = 3.06 tau_c")A single crystal pulled along a cube direction yields at , along a body diagonal at , and along a low-symmetry direction such as [123] at only . In a random polycrystal the largest Schmid factor ranges from 0.274 to 0.500 between grains, with a mean of 0.453. If each grain simply slipped on its own best system (Sachs's assumption) the polycrystal would yield at about . But grains cannot deform independently without opening gaps; G. I. Taylor showed in 1938 that if every grain must take up the same strain as the whole, each needs five independent slip systems, and the yield stress rises to for random face-centred cubic metals. Measured values lie between the two bounds, close to Taylor's for most metals, which is why the Taylor factor of about 3 converts single-crystal into polycrystal strength.
Crystal plasticity finite-element (CPFE) models do better than either bound. They mesh a representative volume of real grains, often taken from an EBSD map (microscopy) or a phase-field simulation, give every integration point the orientation of its grain and a constitutive law for slip on each system with hardening, and solve for equilibrium and compatibility with the finite-element method of Lumenary ME's finite element method. They predict how texture evolves during rolling, why some grains concentrate strain and nucleate fatigue cracks, and how stress partitions between phases in dual-phase steels. Homogenisation then turns the response of the representative volume into effective properties, a stress–strain law or an anisotropic yield surface, that a finite-element model of the whole part can use with one element spanning thousands of grains.
Integrated computational materials engineering
Integrated computational materials engineering (ICME), named in a 2008 report of the United States National Research Council, is the practice of linking these models into a chain that runs from processing to structure to properties to performance, so that a material and the process that makes it can be designed with the component, rather than chosen afterwards from a handbook. The chain idea was set out by Gregory Olson at Northwestern University, whose group designed steels by computation that, developed commercially, were later qualified for aircraft landing gear.
The thermodynamic backbone of most ICME chains is CALPHAD: assessed databases of the Gibbs energies of phases as functions of composition and temperature, fitted to experiments and increasingly to DFT, from which phase diagrams, phase fractions and driving forces are computed for alloys of ten or more elements (phase diagrams). Commercial packages include Thermo-Calc from Sweden and FactSage from Canada and Germany, and the open-source pycalphad and OpenCalphad are used in research. Mobility databases add diffusion, so precipitation and homogenisation can be simulated.
The National Research Council's report described industrial programmes such as Ford's Virtual Aluminum Castings, which linked casting, heat-treatment, microstructure and fatigue models to design engine components, and the 2011 United States Materials Genome Initiative aimed to halve the time and cost of bringing new materials to market by combining computation, data and experiment; Europe, China, Japan and others run comparable programmes. As of 2026, the most active frontier is additive manufacturing, where every part has its own thermal history: models chain the melt pool, its solidification, the resulting grains and defects, residual stress and distortion, and fatigue life, and machine-learned surrogates of the slower models make process optimisation fast enough to use (Lumenary ME's additive manufacturing and simulation-driven design).
Verification, validation and uncertainty
A model chain is only as good as the confidence that can be placed in its answer, and engineering practice separates three questions.
- Verification: are the equations solved correctly? Code is checked against exact solutions, as in the Allen–Cahn block, and results against refinement of the grid and time step.
- Validation: are they the right equations? Predictions are compared with experiments that were not used to calibrate the model, over the range of conditions where it will be used.
- Uncertainty quantification (UQ): how uncertain is the answer, given uncertain inputs, parameters and model form?
The ASME V&V standards for computational solid mechanics and fluid dynamics, and similar guidance from NASA and others, set out how to document all three. The simplest UQ method is Monte Carlo: sample the uncertain inputs from their distributions, run the model for each sample, and look at the spread of the outputs. The block does this for the last link of a simple chain, the Hall–Petch relation from elasticity, plasticity and strengthening, with illustrative inputs of the right order for a low-carbon ferritic steel: a grain size measured by microscopy as 20 µm with a standard uncertainty of 1.5 µm, and the two coefficients known to about ±10 and ±8 per cent.
import numpy as np
rng = np.random.default_rng(2011)
# Yield strength of a ferritic steel from the Hall-Petch relation, sigma_y = sigma_0 + k / sqrt(d),
# with uncertain inputs. Illustrative values of the right order for a low-carbon steel.
N = 200_000
sigma_0 = rng.normal(70.0, 7.0, N) # MPa: friction stress, fitted to tests (+-10 %)
k_hp = rng.normal(0.74, 0.06, N) # MPa m^0.5: Hall-Petch coefficient
d = rng.normal(20e-6, 1.5e-6, N) # m: grain size measured by microscopy (+-7.5 %)
sigma_y = sigma_0 + k_hp / np.sqrt(d)
print(f"nominal prediction: {70 + 0.74 / np.sqrt(20e-6):.0f} MPa")
print(f"Monte Carlo: mean {sigma_y.mean():.1f} MPa, standard deviation {sigma_y.std():.1f} MPa")
print(f" 1st percentile {np.percentile(sigma_y, 1):.0f} MPa, 99th {np.percentile(sigma_y, 99):.0f} MPa")
# First-order (linearised) propagation, for comparison
d_k, d_d = 1 / np.sqrt(20e-6), -0.5 * 0.74 * (20e-6) ** -1.5
var = 7.0**2 + (d_k * 0.06) ** 2 + (d_d * 1.5e-6) ** 2
print(f"linearised: standard deviation {np.sqrt(var):.1f} MPa")
# Which input matters most? Freeze each one at its nominal value in turn.
print("\nfreeze one input at its nominal value -> remaining standard deviation")
for name, s0, k, dd in [("friction stress", 70.0, k_hp, d), ("Hall-Petch k", sigma_0, 0.74, d),
("grain size", sigma_0, k_hp, 20e-6)]:
print(f" {name:16s} {np.std(s0 + k / np.sqrt(dd)):6.1f} MPa")The nominal prediction is 235 MPa, but the honest answer is 236 ± 16 MPa, with 1 per cent of cases below 198 MPa. (The mean is slightly above the nominal value because is curved, so errors in grain size do not cancel.) Here linearised propagation, using derivatives, gives the same 16.4 MPa, because the model is nearly linear over the range; for strongly nonlinear chains it does not, which is why Monte Carlo and its more efficient relatives are used. Freezing one input at a time shows where the uncertainty comes from: removing the uncertainty in cuts the spread from 16.4 to 9.4 MPa, while perfect knowledge of the friction stress or the grain size helps much less. The cheapest way to a better prediction is therefore better calibration of , not more micrographs. Real ICME chains do the same analysis across many links, and the answer often surprises.
Multiscale modelling in practice
The tools are a mixture of open-source research codes and commercial packages. For phase-field modelling, open frameworks include MOOSE from Idaho National Laboratory, PRISMS-PF from the University of Michigan, NIST's FiPy and OpenPhase from Ruhr University Bochum, while MICRESS, from Aachen, is widely used in industry for solidification. For crystal plasticity, DAMASK from the Max Planck Institute in Düsseldorf and many in-house codes work alongside general finite-element packages that accept user-defined material laws. Process simulation software for casting, forging and welding embeds microstructure models, and CALPHAD packages provide the thermodynamics.
Three trends dominate in 2026. Machine-learned surrogates trained on many runs of a slow model replace it inside optimisation and UQ loops, and machine-learned interatomic potentials (AI-assisted materials discovery) make atomistic inputs cheaper. Digital twins of components, models kept up to date with data from the part in service, are being built for turbine blades, pressure vessels and additively manufactured parts, with the materials chain inside them. And data infrastructure, shared formats, provenance and FAIR (findable, accessible, interoperable, reusable) data, increasingly decides whether a model chain can be reused by anyone except the group that built it.
Good practice runs through all of it: know which rung each parameter came from, verify each code, validate each link and the chain against independent experiments, propagate uncertainties, and state the range of conditions outside which the model has not been tested.
Common misconceptions
Misconception
"Multiscale modelling means one simulation from electrons to parts." Almost all practical multiscale modelling is hierarchical: separate models at separate scales, joined by passing parameters. Concurrent coupling is used only where a small region needs finer detail.
Misconception
"The phase-field interface width is the real interface width." Usually it is a numerical parameter, made hundreds of times wider than real interfaces so that a grid can resolve it. The model must be shown to give results that do not depend on it, or corrected by a thin-interface analysis.
Misconception
"A validated model can be used anywhere." Validation holds for the conditions tested. Outside them, for a new alloy, a higher strain rate or a different microstructure, the model is an extrapolation.
Misconception
"Refining the mesh removes the uncertainty." Refinement removes discretisation error only. Uncertainty in parameters and in the form of the model remains and is usually larger.
Exercises
1. Multiscale modelling
Warm-upHumanWhich of these processes should be modelled with a conserved order parameter, and so with the Cahn–Hilliard equation rather than Allen–Cahn? Select all that apply.
2. Multiscale modelling
CoreHumanA phase-field model uses the double well with J m⁻³ and a gradient-energy coefficient J m⁻¹. Using the lesson's theorem, what interface energy does it give, in mJ m⁻²?
3. Multiscale modelling
CoreHumanIn a Cahn–Hilliard simulation with gradient-energy coefficient and an alloy whose free-energy curvature at its composition is (in the same units, lengths in grid spacings), what is the fastest-growing wavelength of the early composition fluctuations? Give the nearest whole number of grid spacings.
4. Multiscale modelling
CoreHumanA team compares the precipitate sizes predicted by its phase-field model of an aluminium alloy with transmission electron micrographs of samples aged at a temperature that was not used to fit the model's parameters. What kind of activity is this?
5. Multiscale modelling
CoreHumanSingle crystals of a face-centred cubic alloy have a critical resolved shear stress of 10 MPa. Using Taylor's factor for a random polycrystal, , estimate the tensile yield stress of a polycrystal of the same alloy with a random texture, in MPa, to the nearest whole number. (Ignore grain-size strengthening.)
6. Multiscale modelling
CoreComputationalWrite allen_cahn_step(phi, dt, dx, L=1.0, kappa=1.0, W=1.0), one explicit time step of the 1-D Allen–Cahn equation
on a periodic grid of spacing dx, using the three-point Laplacian . Return the new array and leave the input unchanged.
7. Multiscale modelling
StretchComputationalWrite propagate(model, means, sds, n=100_000, seed=0) for Monte Carlo uncertainty propagation. Each input is independent and normally distributed with mean means[i] and standard deviation sds[i]. Draw n samples of every input with np.random.default_rng(seed), evaluate model(*inputs) on the arrays of samples (the model works element-wise), and return (mean, standard_deviation) of the outputs as plain floats.
Further reading and resources
- Free energy of a nonuniform system. I. Interfacial free energy · J. W. Cahn and J. E. Hilliard · 1958The Journal of Chemical Physics. The gradient-energy free-energy functional on which phase-field models rest.
- On spinodal decomposition · J. W. Cahn · 1961Acta Metallurgica. The theory of spinodal decomposition and its fastest-growing wavelength.
- A microscopic theory for antiphase boundary motion and its application to antiphase domain coarsening · S. M. Allen and J. W. Cahn · 1979Acta Metallurgica. The Allen–Cahn equation for a non-conserved order parameter.
- Phase-field models for microstructure evolution · L.-Q. Chen · 2002Annual Review of Materials Research. A widely read review of the method and its applications.
- Overview of constitutive laws, kinematics, homogenization and multiscale methods in crystal plasticity finite-element modeling · F. Roters, P. Eisenlohr, L. Hantcherli, D. D. Tjahjanto, T. R. Bieler and D. Raabe · 2010Acta Materialia. A thorough review of crystal plasticity and its coupling to finite elements.
- Integrated Computational Materials Engineering: A Transformational Discipline for Improved Competitiveness and National Security · National Research Council · 2008The National Academies Press. The report that named and defined ICME, with industrial case studies.
- Materials Genome InitiativeThe United States initiative, launched in 2011, to halve the time and cost of developing new materials with computation, data and experiment.
- DAMASK · Max Planck Institute, DüsseldorfOpen-source crystal plasticity and multiphysics simulation.
- FiPy · NISTAn open-source Python solver for partial differential equations, with phase-field examples.