From atoms to bands
An isolated atom has sharp, separate energy levels: its electrons can only sit at certain energies. Bring two atoms close together and each level splits in two. One is a lower "bonding" combination, with the electron shared between the atoms. The other is a higher "antibonding" one. That's the same physics that holds the H₂ molecule together.
Bring N atoms together and each level splits into N levels, packed into a limited range of energy. A real crystal has around 10²³ atoms, so its levels are so close together that they form a continuous band. Between bands there can be ranges of energy with no electron states at all: band gaps.
Below is a chain of atoms, each with two levels (call them "s-like" and "p-like"). Change how many atoms there are and how far apart they sit.
- Levels in each band
- —
- Width of the s-like band
- —
- Width of the p-like band
- —
- Gap between the bands
- —
Things to try
- 1 atom gives two sharp levels. 2 atoms split each one into a bonding and antibonding pair. 60 atoms and the levels crowd into bands.
- Push the atoms closer: their electron clouds overlap more, so the bands widen. Pull them apart and the bands shrink back to the atomic levels on the right.
- Push them very close: the two bands overlap and the gap vanishes. Squeezing really does turn some insulators into metals at extreme pressures.
Each level holds two electrons, one spin-up and one spin-down. So a band made from N atoms holds exactly 2N electrons. Keep that in mind for section 4.
Electrons as waves: what "k" means
In a crystal every atom has identical surroundings, so an electron's wave has to look the same around every atom, apart from a shift in its phase from one atom to the next. This is Bloch's theorem (1928). Each state in a band is labeled by a number k, the wavevector, which says how fast that phase winds as you step from atom to atom.
A band structure is simply a plot of energy against k. Drag the slider to see what the wave looks like on each atom, and where that puts its energy.
- Phase shift between neighbors
- —
- Wavelength
- —
- Energy
- —
- k = 0: every atom in phase. Every overlap is bonding, so this is the bottom of the band.
- k = ±π/a (a is the spacing between atoms): neighbors exactly opposite. Every overlap is antibonding: the top of the band.
- Show the twin: a much shorter wave, k + 2π/a, passes through exactly the same values at every atom. In this picture the state is defined by its value on each atom, so the two are the same state. That means k only needs to run from −π/a to π/a, a range called the Brillouin zone. The grayed parts of the plot just repeat it.
The shape of a band matters. Its slope is how fast an electron in that state travels. Its curvature sets the electron's effective mass: a sharply curved band means light, nimble electrons, and a flat band means heavy, sluggish ones.
The density of states
A band structure is a map: the energy for every k. Often you want an answer to a simpler question: how many states are there at each energy? That's the density of states (DOS), g(E). It's a histogram: sample k evenly, compute the energies, and count how many land in each small energy window. (Darts again.)
Where a band is flat, lots of k values share nearly the same energy, so the DOS piles up into a peak. Where a band is steep, the states are spread thin. The peaks that come from flat spots are called van Hove singularities. Move the pointer over the chart (or use the slider) to probe an energy.
The DOS depends on dimension. In a 1D chain it spikes at the band edges. In 2D it has a sharp peak in the middle, from "saddle points" where the band is flat in one direction. In 3D it rises gently, like a square root, from each band edge. Experiments measure DOS-like quantities directly (photoemission and scanning tunneling spectroscopy, for example). The DOS also controls heat capacity, optical absorption, and how strongly a material responds to doping.
Filling the bands: metals, semiconductors and insulators
Now pour in the electrons. They fill the states from the lowest energy up, two per state, and the energy where filling stops is the Fermi level. Whether a material conducts electricity depends almost entirely on where that level lands.
- Inside a band: metal. Empty states sit just above the full ones, so the slightest push from a voltage moves electrons into them and current flows.
- In a gap: insulator or semiconductor. Every state below is full and the nearest empty one is a whole gap away. A completely full band can't carry current. A semiconductor is just an insulator with a small enough gap (up to roughly 3 eV) that heat, light or a few impurity atoms can get electrons across.
The model below is a chain of two alternating kinds of atom, which opens a gap between two bands. Choose how many electrons there are per repeating unit, the size of the gap, and the temperature.
- Verdict
- —
- Where the Fermi level sits
- —
- Electrons heat has kicked across the gap
- —
Things to try
- Silicon-like (1.1 eV gap, room temperature): a tiny fraction of electrons get thermally kicked across. That's why pure silicon conducts only a little, and conducts better as it heats up, the opposite of a metal.
- Diamond-like (5.5 eV gap): essentially nothing gets across at room temperature. An excellent insulator.
- Doped silicon (2.02 electrons per unit): the extra electrons have to go into the upper band, where they're free to move. Adding a sprinkling of phosphorus atoms to silicon does exactly this, and it's the basis of every transistor.
- Metal (1 electron per unit): the lower band is only half full, so the Fermi level sits in the middle of a band. Materials with an odd number of outer electrons per repeating unit, like sodium and copper, are often metals for this reason.
Why we need DFT
The toy models above used made-up numbers for how strongly neighboring atoms interact. To predict a real material's bands, you have to solve quantum mechanics (the Schrödinger equation) for all of its electrons at once. That runs straight into what Walter Kohn called the exponential wall.
The full quantum state of N electrons, the wavefunction Ψ(r₁, r₂, …, rN), depends on all their positions at once: 3N coordinates. Even on a crude grid of 10 points per coordinate, writing it down takes 103N numbers.
It's the curse of dimensionality again, the same one from Bishop's book. The way around it came from two remarkable results:
- Hohenberg and Kohn (1964) proved that the electron density n(r), meaning how much electron there is at each point in space, determines everything about the lowest-energy state, including its energy. The density lives in ordinary 3D space however many electrons there are.
- Kohn and Sham (1965) turned that into a practical recipe. Replace the real, interacting electrons with imaginary non-interacting ones that move in an effective potential, chosen so that they have exactly the same density. Non-interacting electrons are easy: each one obeys its own one-electron equation.
n(r) = Σi |ψi(r)|² (summed over the occupied orbitals)
The effective potential has three parts: the pull of the nuclei, the average repulsion from all the other electrons (the Hartree term), and the exchange–correlation term Vxc, which bundles up every remaining quantum effect. DFT would be exact if we knew Vxc exactly. We don't, so it has to be approximated, and the choice of approximation is called the functional. John Perdew calls the sequence of increasingly accurate (and expensive) functionals "Jacob's ladder":
- LDA (local density approximation): treats each point as if it sat in a uniform sea of electrons of the same density. Kohn and Sham's original suggestion.
- GGA, e.g. PBE: also uses how quickly the density changes from point to point. The everyday workhorse of materials science.
- Meta-GGA, e.g. SCAN: adds information about the orbitals' kinetic energy.
- Hybrids, e.g. HSE or B3LYP: mix in some "exact exchange" calculated from the orbitals. Better, especially for band gaps, and several times more expensive.
Higher up the ladder: more accurate, more expensive.
The self-consistent loop
The Kohn–Sham recipe has a chicken-and-egg problem. The potential depends on the density, but the density comes from the orbitals, which depend on the potential. The way out is to iterate: guess a density, build the potential, solve for the orbitals, compute a new density, and repeat until the density coming out matches the density going in. That's called self-consistency, and it's the "iterate until self-consistent" box from the ab initio MD explainer.
Below is a real Kohn–Sham calculation, in one dimension, for a chain of hydrogen-like atoms. It starts from the crudest possible guess, with no repulsion between electrons at all. Step through the loop one round at a time, or press Run.
The effective potential the electrons feel, and the occupied orbitals drawn at their energies
Electron density: what went in and what came out
How much the density changed each round (log scale)
- Iteration
- 0
- Change in density
- —
- Total energy
- —
- Gap between highest filled and lowest empty orbital
- —
Things to try
- Step through a few rounds: at first the electrons crowd around the nuclei, since nothing pushes them apart. Their repulsion then lifts the effective potential, the orbitals spread out, and the change from one round to the next shrinks steadily.
- Mixing: real codes don't jump straight to the new density. They blend a fraction β of it into the old one. Set β to 1 with 8 atoms and the density sloshes back and forth forever without settling. This "charge sloshing" is a real headache in big calculations, and codes use clever mixing schemes to tame it.
- More atoms: the occupied levels start to bunch together. In a crystal with 10²³ atoms they would merge into a band, and these Kohn–Sham energies εi are exactly what a DFT band structure plots.
- Turn off exchange–correlation: the answer shifts, which is why the choice of functional matters.
The 1D atoms use a "softened" Coulomb attraction (a standard trick for 1D models), and the exchange–correlation term is a simple LDA-style stand-in, so the numbers aren't real hydrogen's. The loop itself is exactly what full DFT codes do in 3D.
A real material: silicon
Here is the band structure of silicon, the material in nearly every computer chip, calculated live in your browser. A real crystal is 3D, so k becomes a 3D vector, and the plot follows a path between special points of the Brillouin zone, which by convention are labeled Γ (the center, k = 0), X, L, K and U. The DOS on the right is being built by sampling random k-points: darts again.
Move the pointer over the chart to read off an energy.
- Gap in this model
- computing…
- Measured gap
- 1.12 eV
- Plain DFT (PBE functional)
- about 0.6 eV
- Width of the valence bands
- —
How to read it
- Blue bands below zero are the valence bands. They're completely full, holding the 8 outer electrons of the two silicon atoms in each repeating unit (4 per atom). Zero is set at the top of the valence bands.
- Orange bands above are the conduction bands. They're empty at low temperature.
- The gap is indirect. The top of the valence bands is at Γ, but the bottom of the conduction bands sits most of the way to X. Because they're at different k, an electron can't simply drop straight across the gap and emit light; it needs a vibration of the crystal to make up the difference in momentum. That's why silicon is poor at emitting light, and why LEDs and lasers use direct-gap materials such as gallium arsenide and gallium nitride instead.
To run instantly, this plot uses the empirical pseudopotential method of Cohen and Bergstresser (1966): three numbers fitted to optical experiments stand in for the full DFT potential. It gets the shape right. Its gap comes out at about 0.8 eV, against 1.12 eV measured. A standard DFT calculation with the PBE functional gives a very similar picture but a gap of only about 0.6 eV. More on why in the next section.
Doing it for real
A DFT band structure calculation on a real material goes roughly like this:
- Describe the crystalThe repeating unit (its lattice vectors) and the atoms inside it. For silicon: the diamond structure, 5.43 Å cube edge, 2 atoms per primitive cell.
- Choose the approximationsA functional (often PBE), pseudopotentials (which replace each nucleus and its tightly bound inner electrons with something smoother), a plane-wave cutoff (how fine-grained the orbitals can be), and a k-point grid such as 8×8×8.
- Run the self-consistent loopIterate on that k-grid until the density stops changing, exactly as in section 6.
- Compute the bandsFreeze the density and solve the Kohn–Sham equations at k-points along a path through the Brillouin zone. Plot ε against k.
- Compute the DOSSolve on a much denser k-grid and make a histogram. Splitting it by atom or orbital gives the projected DOS, which shows where each band comes from.
- Check convergenceRaise the cutoff and refine the k-grid until the answers stop changing.
Popular codes include VASP, Quantum ESPRESSO, ABINIT, CASTEP and GPAW. The same self-consistent calculation also gives the forces on the atoms, which is how ab initio MD gets its forces at every step.
Where DFT struggles
- Band gaps come out too small. PBE gives silicon about 0.6 eV instead of 1.12 eV, and some small-gap semiconductors, such as germanium, come out as metals. Strictly, the Kohn–Sham energies were never meant to be exact electron energies. Fixes include hybrid functionals (such as HSE) and many-body "GW" calculations, at a much higher cost.
- Strongly correlated materials. In some transition-metal oxides and rare-earth compounds, electrons interact so strongly that standard functionals get even the basic picture wrong, sometimes predicting a metal where there's really an insulator. Common patches are DFT+U and dynamical mean-field theory (DMFT).
- Van der Waals attraction (the weak pull from the Lennard-Jones curve in the molecular dynamics explainer) is missing from LDA and GGA, and has to be added as a correction.
- Excited states. Ground-state DFT describes the lowest-energy state. Light absorption and other excitations need extensions such as time-dependent DFT.
Even so, DFT is one of the most widely used methods in all of science, from batteries and catalysts to drugs and planetary interiors. A handful of DFT papers, including the one introducing the PBE functional, are among the most-cited scientific papers ever written.
Glossary: words you'll hear
| Term | What it means |
|---|---|
| Band | A continuous range of allowed electron energies, formed when atomic levels merge in a solid. |
| Band gap | A range of energy with no electron states, between the top of the valence band and the bottom of the conduction band. |
| Valence / conduction band | The highest filled bands / the lowest empty bands. |
| Fermi level | The energy up to which states are filled. |
| k, wavevector | The label of a band state: how fast its phase winds from one repeating unit to the next. Also called crystal momentum. |
| Brillouin zone | The range of k values that are actually different (−π/a to π/a in 1D). In 3D it's a polyhedron. |
| Γ, X, L, K, U… | Conventional names for special points in the Brillouin zone. Γ is always k = 0. |
| Direct / indirect gap | Whether the top of the valence band and the bottom of the conduction band are at the same k. |
| Effective mass | How heavy an electron behaves in a band, set by the band's curvature. |
| DOS | Density of states: the number of electron states per unit of energy. |
| Projected DOS (PDOS) | The DOS split into contributions from each atom or orbital. |
| Kohn–Sham equations | The one-electron equations at the heart of practical DFT. |
| Functional (LDA, PBE, HSE…) | The approximation used for the exchange–correlation energy. |
| SCF | Self-consistent field: the loop that iterates until the density stops changing. |
| Pseudopotential | A smoothed-out stand-in for a nucleus plus its tightly bound inner electrons. |
| Plane-wave cutoff | How many waves are used to build each orbital. Higher is more accurate and slower. |
| k-point grid | The set of k values sampled in the Brillouin zone during a calculation. |
| Smearing | Slightly blurring occupations near the Fermi level, or the DOS, so metals converge and plots look smooth. |