Materials from the Schrödinger equation
Suppose you want a new cathode for a lithium-ion battery: a compound that holds lithium at a high voltage, does not fall apart when the lithium leaves, and is made of elements that are not scarce. There are tens of thousands of candidate compositions. Making and testing each would take lifetimes. Since the 2000s a large part of that search has been done first on computers, with density functional theory (DFT): from nothing but the atomic numbers and an arrangement of atoms, it computes the energy of a crystal, and from energies come lattice parameters, elastic constants, voltages, stabilities, defect energies and band structures.
DFT is now the most widely used method of computational materials science, and its results fill open databases of well over a hundred thousand inorganic compounds. Lumenary Chemistry's computational chemistry introduced it for molecules, alongside Hartree–Fock and the hierarchy of functionals chemists use. This lesson takes the solid-state view: periodic crystals, plane waves, pseudopotentials and k-points, what DFT gets right and wrong about solids, and how high-throughput calculations are used to decide which compounds can exist. It builds on the bands of electronic structure and bands, which DFT computes for real materials.
The many-electron problem
Within the Born–Oppenheimer approximation from atomistic simulation, the nuclei are fixed and the electrons obey the Schrödinger equation with, in Hartree atomic units,
kinetic energy, attraction to the nuclei, and repulsion between every pair of electrons. The last term couples all the electrons together, so is a function of coordinates that cannot be broken into independent pieces.
The way out is to notice that the energy depends on the electrons only through a few simple quantities. The electron density , the number of electrons per unit volume at each point, is a function of only three coordinates, whatever the number of electrons. DFT is the theory that makes the density, rather than the wavefunction, the thing to compute.
Hohenberg–Kohn and Kohn–Sham, in outline
Theorem— Hohenberg–Kohn (1964)
(1) The ground-state density of a system of interacting electrons determines the external potential up to a constant, and therefore every property of the ground state. (2) There is a functional , with universal (the same for every material), whose minimum over all reasonable densities with electrons is the exact ground-state energy, reached at the exact ground-state density.
The theorem proves that the functional exists but not what is. Walter Kohn and Lu Jeu Sham made it practical in 1965 by inventing a system of non-interacting electrons constructed to have the same density as the real one. Their kinetic energy can be computed exactly from one-electron orbitals, and everything that is difficult is gathered into one small term.
Definition— Kohn–Sham equations
The orbitals of the non-interacting system satisfy
where is the classical electrostatic (Hartree) potential of the density, are occupations, and the exchange–correlation potential contains every quantum many-body effect: the Pauli exclusion between electrons of the same spin, their correlated avoidance of each other, and the correction to the kinetic energy. If were known exactly, the equations would give the exact ground-state energy and density.
The potential depends on the density, and the density on the orbitals that the potential produces, so the equations are solved by iteration until input and output agree: the self-consistent field (SCF) cycle.
The self-consistent cycle in miniature
The block runs a real SCF cycle on a toy: a one-dimensional chain of atoms, each a smooth potential well, with one electron per atom, which makes it a metal. The orbitals are expanded in plane waves at a set of k-points, as in production codes, and occupied with a little Fermi smearing. The electrons repel each other through a model interaction with the same long-range form as the Coulomb repulsion in a solid ( in reciprocal space); to keep the code short there is no exchange–correlation term, which in real DFT would be added to the potential in the same line. The chain is computed twice: with one atom in the periodic cell, and with eight, which must give exactly the same answer but turns out to be much harder to converge.
import math
import numpy as np
# The Kohn-Sham cycle in miniature (Hartree atomic units, one dimension).
# A chain of atoms 4 bohr apart, each a Gaussian well, with one electron per atom: a metal.
# The electrons repel through a model interaction with the long-range 1/G^2 form of the
# Coulomb repulsion; real DFT adds an exchange-correlation potential in the same place.
a, depth, width, kT = 4.0, 1.0, 0.8, 0.02 # kT: Fermi smearing, as in calculations on metals
def scf(M, mixing, kerker_q0=None, max_iter=80, tol=1e-6):
"""Self-consistent density for a supercell of M atoms; returns the residual history and density."""
Ls = M * a
m = np.arange(-6 * M, 6 * M + 1) # plane waves exp(i(k + G)x), G = 2 pi m / Ls
G = 2 * math.pi / Ls * m
nk = 32 // M # the same k-point density for every cell
ks = -math.pi / Ls + 2 * math.pi / Ls * (np.arange(nk) + 0.5) / nk
nfft = 48 * M
Gf = 2 * math.pi * np.fft.fftfreq(nfft, d=Ls / nfft)
Vext = np.zeros(nfft, dtype=complex) # Fourier coefficients of the wells
lattice = np.isclose((Gf * a / (2 * math.pi)) % 1, 0) | np.isclose((Gf * a / (2 * math.pi)) % 1, 1)
Vext[lattice] = -depth * width * math.sqrt(2 * math.pi) / a * np.exp(-0.5 * (Gf[lattice] * width) ** 2)
v = np.zeros(nfft)
v[1:] = 4 * math.pi / Gf[1:] ** 2 # model repulsion; G = 0 is cancelled by the ions
n_in = np.full(nfft, M / Ls) # first guess: a uniform density
history = []
for _ in range(max_iter):
Veff = Vext + v * np.fft.fft(n_in) / nfft # V_ext + V_Hartree[n_in], in reciprocal space
Vmat = Veff[(m[:, None] - m[None, :]) % nfft]
eps, vecs = [], []
for k in ks: # solve the one-electron equations at each k
e, c = np.linalg.eigh(np.diag(0.5 * (k + G) ** 2) + Vmat)
eps.append(e[: M + 4])
vecs.append(c[:, : M + 4])
eps = np.array(eps)
lo, hi = eps.min() - 1, eps.max() + 1 # Fermi level, by bisection on the electron count
for _ in range(60):
mu = 0.5 * (lo + hi)
f = 1 / (np.exp(np.clip((eps - mu) / kT, -60, 60)) + 1)
lo, hi = (mu, hi) if 2 * f.sum() / nk < M else (lo, mu)
n_out = np.zeros(nfft) # the density these orbitals produce
for ik in range(nk):
c = np.zeros((nfft, M + 4), dtype=complex)
c[m % nfft] = vecs[ik]
psi = np.fft.ifft(c, axis=0) * nfft / math.sqrt(Ls)
n_out += (2 / nk) * np.sum(f[ik] * np.abs(psi) ** 2, axis=1)
residual = n_out - n_in
history.append(math.sqrt(np.mean(residual**2)) * Ls / M)
if history[-1] < tol:
break
if kerker_q0 is None: # simple (linear) mixing
n_in = n_in + mixing * residual
else: # Kerker: damp the long-wavelength part
R = np.fft.fft(residual) * Gf**2 / (Gf**2 + kerker_q0**2)
n_in = n_in + mixing * np.real(np.fft.ifft(R))
return history, n_out
cases = [("1-atom cell, simple mixing 0.5", 1, 0.5, None),
("8-atom cell, simple mixing 0.5", 8, 0.5, None),
("8-atom cell, simple mixing 0.002", 8, 0.002, None),
("8-atom cell, Kerker mixing 0.5", 8, 0.5, 3.0)]
print("density residual after iteration: 1 5 10 20 40 80")
densities = {}
for name, M, mix, q0 in cases:
h, n = scf(M, mix, q0)
cols = " ".join(f"{h[i - 1]:.0e}" if i <= len(h) else " - " for i in (1, 5, 10, 20, 40, 80))
verdict = f"converged in {len(h)}" if h[-1] < 1e-6 else "not converged"
print(f"{name:34s} {cols} {verdict}")
densities[(M, q0)] = n
n1, n8 = densities[(1, None)], densities[(8, 3.0)]
print(f"\nlargest difference between the 8-atom and 1-atom densities: {np.max(np.abs(n8 - np.tile(n1, 8))):.1e} electrons/bohr")With one atom per cell, simple mixing (taking half of the new density each time) converges steadily in 31 iterations. With eight atoms per cell the same recipe starts well, then blows up between the fifth and tenth iteration and never recovers. Cutting the mixing to 0.002 stops the explosion but leaves the cycle crawling, with the residual still 0.3 after 80 iterations. The fix is Kerker mixing, which damps the long-wavelength components of each density update: it converges in 37 iterations, and the eight-atom density agrees with the one-atom density to electrons per bohr, as it must.
Exchange–correlation: the one approximation
Everything in Kohn–Sham DFT is exact except , and the quality of a calculation is mostly the quality of its exchange–correlation functional. The main families, as used for solids:
| Approximation | Uses | Examples | In solids |
|---|---|---|---|
| LDA | the local density | Perdew–Zunger, PW92 fits to quantum Monte Carlo for the electron gas | overbinds: lattice parameters about 1 % short, cohesive energies too large |
| GGA | + the density gradient | PBE, PBEsol | PBE underbinds slightly (about 1 % long); PBEsol is tuned for solids' lattice parameters |
| meta-GGA | + the kinetic-energy density | SCAN, r²SCAN | better formation energies and structures at modest extra cost |
| hybrid | + a fraction of exact exchange | HSE06 (screened), PBE0 | much better band gaps; several times to tens of times more costly with plane waves |
| DFT+U | + an on-site penalty for localised d or f electrons | PBE+U | corrects some transition-metal oxides, with an empirical |
Two further corrections are routine. Semilocal functionals miss most of the van der Waals attraction, so layered materials such as graphite come out nearly unbound unless a dispersion correction (Grimme's D3 or D4, or a nonlocal vdW functional) is added. And strongly correlated materials, where electrons on transition-metal or rare-earth ions are localised, need DFT+U or methods beyond DFT such as dynamical mean-field theory.
The band-gap problem. The Kohn–Sham eigenvalues are energies of fictitious non-interacting electrons, and the gap between them is not the fundamental gap (the energy to remove an electron minus the energy to add one). Even the exact functional would miss the difference, a jump in called the derivative discontinuity; LDA and GGA also suffer from self-interaction, which makes electrons too delocalised. The result, as Gideon warned in electronic structure and bands, is that LDA and GGA band gaps are typically 30 to 50 per cent too small, silicon's coming out near 0.6 eV against the measured 1.12 eV and germanium's sometimes closing altogether. Mott insulators such as NiO can even come out metallic. Hybrid functionals such as HSE06 bring most semiconductor gaps within about 10 to 20 per cent; the GW approximation, a many-body perturbation theory built on top of DFT, is the standard for accurate quasiparticle gaps, at much higher cost.
Plane waves, pseudopotentials and k-points
In a crystal, Bloch's theorem (electronic structure and bands) lets each orbital be written as , a sum of plane waves over reciprocal lattice vectors . Plane waves are the natural basis for solids: they are automatically periodic, they treat all of space equally, and their quality is controlled by a single number, the cut-off energy : only plane waves with are kept. The number of plane waves grows as the cell volume times , and the cost of solving the Kohn–Sham equations roughly as the cube of the number of electrons.
Two numerical parameters must be converged in every calculation. The block tests both in one dimension, with the same Gaussian wells as before but without self-consistency. First the cut-off, for a smooth, shallow well and for a deep, narrow one; then the number of k-points used to integrate over the Brillouin zone, for a filled band (an insulator) and a half-filled one (a metal).
import math
import numpy as np
# Hartree atomic units (hbar = m_e = 1, energies in hartree, 1 Ha = 27.211 eV), one dimension.
a = 4.0 # bohr: one atom per cell
def V_G(G, depth, width):
"""Fourier coefficient of a row of Gaussian wells -depth exp(-x^2 / 2 width^2), one per cell."""
return -depth * width * math.sqrt(2 * math.pi) / a * np.exp(-0.5 * (G * width) ** 2)
def bands(k, ecut, depth, width, n_bands=2):
"""Lowest eigenvalues at wavevector k, keeping plane waves with (k+G)^2 / 2 <= ecut."""
m_max = int(math.sqrt(2 * ecut) * a / (2 * math.pi)) + 2
G = 2 * math.pi / a * np.arange(-m_max, m_max + 1)
G = G[0.5 * (k + G) ** 2 <= ecut]
H = np.diag(0.5 * (k + G) ** 2) + V_G(G[:, None] - G[None, :], depth, width)
return np.linalg.eigvalsh(H)[:n_bands], len(G)
# 1. Cut-off convergence: a smooth, shallow well (like a pseudopotential) against a deep,
# narrow one (like the bare potential near a nucleus).
smooth, sharp = (1.0, 0.8), (30.0, 0.1)
ref_s = bands(0.0, 2000, *smooth)[0][0]
ref_h = bands(0.0, 2000, *sharp)[0][0]
print("E_cut (Ha) plane waves error, smooth well (meV) error, sharp well (meV)")
for ecut in (2, 5, 20, 50, 100, 200, 500):
e_s, n = bands(0.0, ecut, *smooth)
e_h, _ = bands(0.0, ecut, *sharp)
print(f"{ecut:10d} {n:11d} {abs(1000 * 27.211 * (e_s[0] - ref_s)):24.3f} {abs(1000 * 27.211 * (e_h[0] - ref_h)):23.1f}")
print(f"lowest level: smooth well {ref_s * 27.211:.3f} eV, sharp well {ref_h * 27.211:.2f} eV")
# 2. k-point convergence of the band energy per atom: two electrons per atom fill band 1
# (an insulator); one electron per atom half-fills it (a metal).
def band_energy(nk, electrons):
ks = -math.pi / a + 2 * math.pi / a * (np.arange(nk) + 0.5) / nk # a uniform grid
e1 = np.sort([bands(k, 30, *smooth, n_bands=1)[0][0] for k in ks])
if electrons == 2:
return 2 * e1.mean()
return 2 * e1[: nk // 2].sum() / nk # lowest half, two spins each
ref_ins, ref_met = band_energy(4000, 2), band_energy(4000, 1)
print("\n k-points error, insulator (meV) error, metal (meV)")
for nk in (2, 4, 8, 16, 32, 64):
print(f"{nk:9d} {abs(1000 * 27.211 * (band_energy(nk, 2) - ref_ins)):22.1e} {abs(1000 * 27.211 * (band_energy(nk, 1) - ref_met)):18.1e}")The smooth well's lowest level is converged to 0.2 meV with just 5 plane waves. The deep, narrow well, whose lowest level lies at −315 eV like a core state, is still 656 meV wrong with 25 plane waves and needs about 41 to come within a few meV. In three dimensions the count of plane waves is roughly the cube of the one-dimensional count, so this is the difference between hundreds and many thousands of plane waves per atom. For the k-points, the insulator's band energy converges exponentially, from about 12 meV with 4 points to meV with 16, while the metal's error falls only as , still 0.45 meV with 64 points: the sharp Fermi surface of a metal, where occupation jumps from full to empty, is hard to integrate. That is why calculations on metals use denser k-point grids and smear the occupations, as the SCF block did.
The deep well is the reason for pseudopotentials. Core electrons are tightly bound, sit deep in energy, and hardly change when atoms bond; valence wavefunctions must wiggle near the nucleus to stay orthogonal to them, and those wiggles need enormous cut-offs. A pseudopotential replaces the nucleus and its core electrons by a smooth, weaker potential that gives the same valence energies and the same wavefunctions outside a core radius. Norm-conserving pseudopotentials (Hamann, Schlüter and Chiang, 1979) were the first practical kind; David Vanderbilt's ultrasoft pseudopotentials (1990) relaxed a constraint to allow lower cut-offs; Peter Blöchl's projector augmented-wave (PAW) method (1994) reconstructs the full all-electron wavefunction from the smooth one when needed and is used by many plane-wave codes. Curated, tested libraries, such as SSSP on Materials Cloud and PseudoDojo, record which cut-off each pseudopotential needs.
k-points. Integrals over the Brillouin zone, for the density and the energy, are replaced by sums over a grid of k-points, usually the evenly spaced grids of Hendrik Monkhorst and James Pack (1976), reduced by the crystal's symmetry. Bigger cells have smaller Brillouin zones and need fewer k-points; metals need more, with smearing of the occupations near the Fermi level.
What DFT predicts well, and badly
With a standard GGA functional such as PBE, well-converged calculations give, for typical inorganic solids (as of 2026, from many benchmark studies):
| Property | Typical performance |
|---|---|
| lattice parameters | within 1–2 % (PBE slightly long, LDA slightly short, PBEsol and r²SCAN closer) |
| bulk and elastic moduli | within about 10–15 %; PBE tends to be soft |
| phonon frequencies | within a few per cent where the lattice parameter is right |
| relative energies of structures, defect and surface energies | trends reliable; errors of tenths of an eV for defects in some oxides |
| formation energies from the elements | errors of order 0.1 eV per atom, reduced by fitted corrections for gases and transition-metal oxides |
| magnetic moments of metals | good for iron, cobalt, nickel |
| band gaps | systematically too small (GGA), much better with hybrids or GW |
| van der Waals bonding, strongly correlated oxides | poor without dispersion corrections or DFT+U and beyond |
| finite temperature | DFT is a ground-state, 0 K theory; heat capacities and free energies need phonons or molecular dynamics added |
Two kinds of error must be kept apart. The numerical error (cut-off, k-points, pseudopotential, convergence of the SCF) can be made as small as you like, and a 2016 comparison of forty codes and methods by Kurt Lejaeghere and colleagues found that careful modern codes agree with each other on the equations of state of 71 elemental solids to within a fraction of the differences between functionals. The physical error of the functional cannot be reduced by convergence, only by a better functional, and it is what limits agreement with experiment.
High-throughput DFT and the convex hull
Once DFT could be run automatically and reliably, it became possible to compute every known inorganic crystal, and many hypothetical ones, in the same way. The Materials Project, begun in 2011 at Lawrence Berkeley National Laboratory and MIT, holds standardised DFT results for well over a hundred thousand compounds, with structures, formation energies, band structures, elastic tensors and battery voltages, freely available through a website and an API. Other open databases include AFLOW (Duke University), the OQMD (Northwestern University), JARVIS-DFT (NIST), the European NOMAD archive, which stores raw calculations from many groups, and Materials Cloud, based in Switzerland.
The most-used number in these databases is a compound's energy above the convex hull. Plot the formation energy per atom of every computed phase in a chemical system against composition. The lower convex hull of those points connects the phases that are stable at 0 K; any phase above it can lower its energy by decomposing into the hull phases on either side, in proportions given by the lever rule of phase diagrams. The block builds the hull for an imaginary A–B system with made-up but typical numbers.
import numpy as np
# Formation energies (eV per atom) of compounds in an IMAGINARY A-B system, made up for this
# example but of a typical size. x is the fraction of B atoms.
phases = {"A": (0.0, 0.0), "A3B": (0.25, -0.160), "A2B": (1 / 3, -0.170), "AB": (0.5, -0.300),
"AB2 (form 1)": (2 / 3, -0.220), "AB2 (form 2)": (2 / 3, -0.205), "AB3": (0.75, -0.120),
"B": (1.0, 0.0)}
def lower_hull(points):
"""Lower convex hull of (x, E) points, by Andrew's monotone-chain algorithm."""
hull = []
for p in sorted(points):
while len(hull) >= 2:
(x1, y1), (x2, y2) = hull[-2], hull[-1]
if (x2 - x1) * (p[1] - y1) - (y2 - y1) * (p[0] - x1) <= 0: # not a left turn
hull.pop()
else:
break
hull.append(p)
return hull
hull = lower_hull(list(phases.values()))
hx, hy = np.array(hull).T
names = {v: k for k, v in phases.items()}
print("phase x_B E_form (eV/atom) above hull (meV/atom) decomposes into")
for name, (x, E) in phases.items():
above = E - np.interp(x, hx, hy)
if above < 1e-9:
products = "nothing: on the hull"
elif np.isclose(hx, x).any():
products = names[hull[int(np.argmin(np.abs(hx - x)))]] + " (same composition)"
else:
j = int(np.searchsorted(hx, x))
f = (x - hx[j - 1]) / (hx[j] - hx[j - 1]) # lever rule, fraction of atoms
products = f"{1 - f:.2f} {names[hull[j - 1]]} + {f:.2f} {names[hull[j]]}"
print(f"{name:13s} {x:5.3f} {E:16.3f} {1000 * above:21.0f} {products}")Five phases lie on the hull: the two elements, A3B, AB and the first form of AB2. A2B lies 37 meV per atom above it and would decompose into two-thirds A3B and one-third AB (by atoms); AB3 lies 45 meV above and would decompose into AB2 and B; the second form of AB2 is a polymorph 15 meV above the first.
Being above the hull does not make a compound impossible. Diamond is above graphite; many useful materials, from martensite to most glasses, are metastable. A 2016 analysis of almost 30 000 known, synthesised inorganic crystals in the Materials Project by Wenhao Sun and colleagues found that a large fraction of them lie above the computed hull, most by a few tens of meV per atom or less, with a long tail whose size depends strongly on the chemistry. High-throughput screens therefore use a threshold of a few tens of meV per atom, then check the survivors more carefully.
Querying the database from Python looks like this (not run here: it needs the mp-api package, the network and a free API key; see the documentation, since the interface changes between versions):
# Not run here: needs the mp-api package, network access and an API key.
from mp_api.client import MPRester
with MPRester("YOUR_API_KEY") as mpr:
docs = mpr.materials.summary.search(
chemsys="Li-Fe-P-O",
energy_above_hull=(0, 0.05), # eV per atom
fields=["material_id", "formula_pretty", "energy_above_hull", "band_gap"],
)
for d in docs:
print(d.material_id, d.formula_pretty, d.energy_above_hull, d.band_gap)As of 2026, machine-learned models trained on these databases screen millions of hypothetical structures before any DFT is run, and DFT then checks the most promising candidates; that partnership, and the debates about how many genuinely new materials it has found, are the subject of AI-assisted materials discovery.
Codes and good practice
Plane-wave codes dominate solid-state work. Quantum ESPRESSO, developed by an international consortium led from Italy, and ABINIT, from Belgium and partners, are open source; VASP, from the University of Vienna, is licensed and widely used in industry and in the Materials Project; CASTEP was developed in the United Kingdom. CP2K combines Gaussian and plane-wave bases for large systems and molecular dynamics; GPAW, from Denmark, uses real-space grids; all-electron codes such as WIEN2k and FHI-aims provide reference accuracy; ONETEP, CONQUEST and BigDFT use linear-scaling methods for thousands of atoms. Python tools such as pymatgen and ASE build inputs and analyse outputs, and workflow systems such as AiiDA record the full provenance of every calculation so that it can be reproduced.
The silicon input in electronic structure and bands computed a band structure at fixed geometry. A second typical job relaxes both atoms and cell, here for magnesium oxide in the rock-salt structure (not run here: Quantum ESPRESSO is not available in the browser, and the pseudopotential file names are placeholders for ones from a tested library):
&CONTROL
calculation = 'vc-relax' ! relax atoms and cell together
prefix = 'mgo'
pseudo_dir = './pseudo'
/
&SYSTEM
ibrav = 2 ! face-centred cubic
celldm(1) = 7.96 ! starting lattice parameter in bohr (0.421 nm)
nat = 2, ntyp = 2
ecutwfc = 50, ecutrho = 400 ! cut-offs in rydberg: check convergence!
/
&ELECTRONS
conv_thr = 1.0d-9
/
&IONS
/
&CELL
press_conv_thr = 0.1 ! kbar
/
ATOMIC_SPECIES
Mg 24.305 Mg.pbe.UPF
O 15.999 O.pbe.UPF
ATOMIC_POSITIONS alat
Mg 0.0 0.0 0.0
O 0.5 0.5 0.5
K_POINTS automatic
6 6 6 0 0 0Good practice is the same in every code: converge the cut-off and k-points for the property you need (energy differences converge faster than total energies, stresses more slowly); use the same settings for every structure you compare; check that the SCF and the geometry have truly converged; test the sensitivity of your conclusion to the functional; and report the code, version, pseudopotentials and settings, so that someone else can reproduce the result.
Common misconceptions
Misconception
"DFT is exact, so its predictions are exact." The Hohenberg–Kohn and Kohn–Sham theory is exact; every functional in use is an approximation. Converging the numerics brings you closer to the functional's answer, not necessarily to nature's.
Misconception
"Kohn–Sham eigenvalues are the energies of electrons, so their differences are excitation energies." They are the energies of fictitious non-interacting electrons. Apart from the highest occupied level, which in exact DFT equals minus the ionisation energy, they have no strict physical meaning, and their gap underestimates the true band gap.
Misconception
"On the hull means it can be made; above the hull means it cannot." The hull is a 0 K, perfect-crystal construction with the errors of the functional. Metastable compounds above it are made every day, and some compounds on it resist every attempt at synthesis.
Misconception
"More plane waves and more k-points always improve agreement with experiment." They reduce numerical error only. Once converged, extra effort changes nothing, and the remaining disagreement belongs to the functional, the neglect of temperature, or the structure model.
Exercises
Hartree atomic units: 1 Ha = 27.211 eV, 1 bohr = 0.052918 nm.
1. Density functional theory
Warm-upHumanA plane-wave DFT calculation on a cubic cell takes one hour. You double the cell's edge length (keeping the same material, cut-off and k-point density) to model a defect. If the cost grows as the cube of the number of electrons, roughly how many hours will the new calculation take?
2. Density functional theory
CoreHumanA PBE calculation gives silicon a band gap of about 0.6 eV; experiment gives 1.12 eV. Which explanation is correct?
3. Density functional theory
CoreHumanWhich of these errors in a DFT calculation can be reduced as far as you like by spending more computer time with the same functional? Select all that apply.
4. Density functional theory
CoreHumanAn SCF calculation on a 200-atom aluminium slab converges in the first few iterations and then oscillates wildly, with the density swinging from one side of the cell to the other. What is the most appropriate first remedy?
5. Density functional theory
CoreHumanIn a binary A–B system, a database lists formation energies of 0 for A and for B, −0.40 eV/atom for AB (50 % B) and −0.15 eV/atom for A₃B (25 % B). No other phases are known. How far above the convex hull is A₃B, in meV per atom?
6. Density functional theory
CoreComputationalWrite count_plane_waves(a, ecut) returning the number of plane waves at the point () for a simple cubic cell of edge a bohr with cut-off ecut hartree. The reciprocal lattice vectors are for integers , and a plane wave is kept if (atomic units). Count exactly, by checking every integer triple that could qualify.
7. Density functional theory
StretchComputationalWrite energy_above_hull(phases, name) for a binary system. phases maps each phase name to (x, E): its fraction of the second element, from 0 to 1, and its formation energy in eV/atom. The two elements are included, at (0.0, 0.0) and (1.0, 0.0). Return the energy of phase name above the lower convex hull of all the points, in eV/atom (0 for a phase on the hull).
Further reading and resources
- Inhomogeneous electron gas · P. Hohenberg and W. Kohn · 1964Physical Review. The theorems that make the electron density the central quantity.
- Self-consistent equations including exchange and correlation effects · W. Kohn and L. J. Sham · 1965Physical Review. The Kohn–Sham equations, solved in essentially every DFT code.
- Generalized gradient approximation made simple · J. P. Perdew, K. Burke and M. Ernzerhof · 1996Physical Review Letters. The PBE functional, the workhorse of solid-state DFT.
- Special points for Brillouin-zone integrations · H. J. Monkhorst and J. D. Pack · 1976Physical Review B. The k-point grids used in almost every calculation on crystals.
- Projector augmented-wave method · P. E. Blöchl · 1994Physical Review B. The PAW method, which combines the efficiency of pseudopotentials with all-electron accuracy.
- Efficient iteration scheme for self-consistent pseudopotential calculations · G. P. Kerker · 1981Physical Review B. The preconditioner that tames charge sloshing in metals.
- Commentary: The Materials Project: A materials genome approach to accelerating materials innovation · A. Jain, S. P. Ong, G. Hautier, W. Chen and colleagues · 2013APL Materials. How the Materials Project computes and shares its data.
- Reproducibility in density functional theory calculations of solids · K. Lejaeghere and many colleagues · 2016Science. Forty codes and methods compared on the same 71 elemental crystals; the precise ones agree closely.
- The thermodynamic scale of inorganic crystalline metastability · W. Sun, S. T. Dacek, S. P. Ong, G. Hautier and colleagues · 2016Science Advances. How far above the convex hull known, synthesised compounds lie.
- Nobel Prize in Chemistry 1998: Walter Kohn and John A. Pople · Nobel Prize Outreach
- Materials Project documentationHow the database's calculations are set up and corrected, and how to use its API.