Lecture 31 · Part IX: From energies and forces to materials properties
Phonons and dynamical stability
How do second derivatives of the energy become vibrational spectra, and what does an imaginary mode mean?
You will be able to explain
- how second derivatives of the energy (force constants) become phonon frequencies through the dynamical matrix
- what acoustic and optical branches are, how to read a phonon band structure and DOS, and why a supercell computes exact phonons only at the wavevectors that fit in it
- how to set up and converge a finite-displacement or DFPT phonon calculation in VASP and phonopy, including polar crystals
- what an imaginary mode means, how to tell a real instability from an artefact, and how to follow it to the stable structure
- how phonons give free energies, thermal expansion (QHA) and phase stability at high temperature, including phases stabilized by anharmonicity
Before this
The Born–Oppenheimer approximation, Bloch's theorem and band structure, Forces, stress and structural relaxation
Time
About 115 minutes
By Osman Goni Ridwan · Last reviewed
DFT & VASP Course · Lecture 311 / 33
Every structure you relax in VASP ends with forces close to zero. This lecture asks what happens next to second order: how the energy changes when the atoms move a little, what vibrations that implies, and whether the structure you found is a true minimum or only a balanced saddle point that a small push would destroy.
Those second derivatives are the force constants. Turned into a dynamical matrix they give the phonons, the vibrational waves of a crystal. Phonons answer three practical questions. Is the structure dynamically stable? What are its vibrational spectra (infrared, Raman, neutron scattering)? And what are its free energy, entropy and heat capacity, which decide which phase is stable when the crystal is hot? The last part of the lecture is about that third question, and in particular about phases that are unstable in a harmonic calculation at 0 K but stable at high temperature.
Energies are in eV and lengths in Å, so force constants come in eV/Ų. Masses are in unified atomic mass units (u). Frequencies are quoted as ordinary frequencies ν in THz, the unit VASP prints first; ω = 2πν is the angular frequency.
Press P to present one slide at a time, S for speaker notes.
Is a structure with zero forces stable?
Not necessarily. Zero forces mark a stationary point; only the second derivatives tell a minimum from a saddle.
DFT & VASP Course · Lecture 312 / 33
A relaxation in VASP stops when the forces on all atoms are small (EDIFFG, Lecture 21). A vanishing force means that the energy surface is flat at that point, to first order. It does not say whether the surface curves upward in every direction, like the bottom of a bowl, or curves downward in some, like the top of a mountain pass. Both are stationary points of the potential-energy surface. Only the first is a local minimum.
High symmetry is the usual way to land on a saddle point. If the starting structure is cubic, every force that would lower the symmetry cancels by symmetry, so the relaxation can never leave the cubic structure, however unfavourable it is. VASP keeps the symmetry of the starting structure during a relaxation, so the run ends with zero forces and an apparently good answer. The textbook example is a cubic perovskite such as SrTiO₃. Relaxed in the cubic cell it has zero forces, but rotating the oxygen octahedra lowers its energy, and below about 105 K the real crystal is indeed tetragonal with rotated octahedra.
The test is the curvature of the energy in every direction the atoms can move: the second derivatives of the energy. In a crystal, the natural directions are the collective vibrations, the phonons, each with a squared frequency ω². A positive ω² means the energy rises along that pattern, so the atoms swing back and vibrate. A negative ω² means the energy falls; the frequency ω is then imaginary, and the structure is unstable along that pattern. A structure is dynamically stable when ω² > 0 for every phonon at every wavevector, apart from the three acoustic modes at Γ, which are rigid translations with ω² = 0.
The rest of the lecture builds this test from scratch, shows how VASP and phonopy compute it, explains what to do when it fails, and then asks the harder question: what happens when the atoms are hot, and vibrate far from the point where the second derivatives were taken.
What does the energy look like near equilibrium?
A bowl: to second order in the displacements, the energy is a quadratic form set by the force constants.
- Energy at rest: the Born–Oppenheimer energy with every nucleus on its site.
- Forces on the sites: zero at a relaxed structure, so this term drops out.
- Force constants ΦIα,Jβ=∂2E/∂uIα∂uJβ, in eV/Ų.
- Anharmonic terms: neglected in the harmonic approximation, essential when hot.
DFT & VASP Course · Lecture 313 / 33
Label the nuclei by I and the Cartesian directions by α, β ∈ {x, y, z}. Let uIα=RIα−RIα0 be the displacement of nucleus I from its reference position RI0. The Born–Oppenheimer energy E({R}) of Lecture 1, the energy VASP computes for clamped nuclei, can be expanded as a Taylor series in these displacements:
E=E0+Iα∑∂uIα∂E0uIα+21Iα,Jβ∑∂uIα∂uJβ∂2E0uIαuJβ+3!1∑∂u∂u∂u∂3E0uuu+…The first derivatives are minus the forces, FIα=−∂E/∂uIα. At a relaxed structure they vanish, and the first nonzero term is the quadratic one. Its coefficients are the second-order force constants,
ΦIα,Jβ=∂uIα∂uJβ∂2E0=−∂uJβ∂FIα0.For N atoms they form a 3N × 3N matrix, the Hessian of the energy. Because mixed second derivatives do not depend on the order of differentiation, the matrix is symmetric: ΦIα,Jβ=ΦJβ,Iα.
The harmonic approximation keeps the expansion up to second order. It works when the displacements are small compared with the distance over which the curvature changes. At room temperature thermal displacements in a typical solid are a few hundredths to about a tenth of an ångström, so the approximation is usually good, and it is the basis of almost every phonon calculation. It fails when the atoms move far from the expansion point: close to melting, in soft or "rattling" structures, and, most importantly for this lecture, when the expansion point is itself a saddle point. The cubic and higher terms are the anharmonic terms; they come back in the second half of the lecture.
Why the displacements, not the positions, are the natural variables
Writing the expansion in u rather than R moves the origin to the reference structure, so the constant and linear terms carry all the information about that structure, and the quadratic term carries everything about small motions around it. If the reference structure is not exactly relaxed, the linear term does not vanish. A harmonic calculation that ignores it then describes vibrations around the wrong point; residual forces show up as errors in the force constants and, for soft modes, as spurious imaginary frequencies. This is why phonon calculations start from a tightly relaxed structure.
What does one force constant measure?
Push one atom: column J of Φ lists minus the forces this creates on every atom, per unit displacement.
DFT & VASP Course · Lecture 314 / 33
Read the definition ΦIα,Jβ=−∂FIα/∂uJβ as an experiment. Move atom J by a small amount u in direction β and keep every other atom fixed. To first order, the force that appears on atom I in direction α is
FIα=−ΦIα,Jβu.So column (J, β) of the matrix is the force pattern created by pushing atom J along β, divided by −u. This is also exactly how a finite-displacement calculation measures it, as later slides show.
In the model, a nearest-neighbour spring of stiffness C contributes −C to the off-diagonal element between the two atoms it joins and +C to each of their diagonal elements. Push atom 3 by +0.01 Å: its neighbours 2 and 4 each feel +10 × 0.01 = +0.10 eV/Å, pulled along, and atom 3 itself feels −20 × 0.01 = −0.20 eV/Å, pulled back. With the second-neighbour springs switched on, atoms 1 and 5 also feel +0.02 eV/Å and the self term grows to 24 eV/Ų.
The acoustic sum rule. Every column, and by symmetry every row, of Φ sums to zero. The reason is translation invariance: if all atoms move by the same vector, nothing about the crystal changes, so no forces appear. For a uniform displacement uJβ=dβ the force on atom I is −∑JβΦIα,Jβdβ=0 for any d, so
J∑ΦIα,Jβ=0⟹ΦIα,Iβ=−J=I∑ΦIα,Jβ.The self term is fixed by all the others: an atom's restoring stiffness is the sum of its couplings to the rest. In the figure the forces always add up to zero, which is Newton's third law. In a numerical calculation the sum rule holds only approximately, because forces carry small errors; phonon codes therefore usually symmetrize the force constants to restore it. A violated sum rule shows up as small nonzero, often imaginary, acoustic frequencies at Γ.
How do force constants become vibrational frequencies?
Newton's law plus a trial vibration turns Φ into an eigenvalue problem; its eigenvalues are ω².
- Mass-weighted force constants: the dynamical matrix D, real and symmetric.
- Eigenvector: the pattern of a normal mode, how far and which way each atom moves.
- Eigenvalue ω²: positive for a vibration, negative for an instability. Frequency ν = ω/2π.
DFT & VASP Course · Lecture 315 / 33
In the harmonic approximation the force on atom I is −∑JβΦIα,JβuJβ, so Newton's second law gives the equation on the left. Look for solutions in which every atom oscillates at the same angular frequency ω: uIα(t)=eIαe−iωt/MI. Dividing the amplitude by MI is a convenience: substitute, divide both sides by MI, and the equation becomes a symmetric eigenvalue problem for the matrix DIα,Jβ=ΦIα,Jβ/MIMJ.
Because D is real and symmetric (Lecture 2), its 3N eigenvalues are real and its eigenvectors can be chosen orthonormal. Each eigenvector is a normal mode: a collective motion in which all atoms vibrate in step at one frequency. Any small motion of the atoms is a superposition of normal modes, which vibrate independently of each other in the harmonic approximation.
The eigenvalues are ω², not ω. A positive ω² gives a real frequency and an oscillation. A negative ω² gives ω=±i∣ω∣, and the "oscillation" e−iωt becomes e±∣ω∣t: a displacement along that mode grows exponentially instead of swinging back. That is the instability of the opening slide, now in algebraic form. A mode with ω² < 0 is called an imaginary mode; by convention its frequency is printed with an "i" or plotted as a negative number.
Units. With Φ in eV/Ų and M in u, Φ/M is an angular frequency; one unit of it corresponds to ν = 15.633 THz. Spectroscopists also use wavenumbers and energies: 1 THz = 33.356 cm⁻¹ = 4.1357 meV. VASP prints all of these on each line of its phonon output.
A worked example: carbon monoxide. For a diatomic molecule the only vibration is the bond stretch, with frequency ω=k/μ, where k is the bond's force constant and μ=M1M2/(M1+M2) the reduced mass. For ¹²C¹⁶O, μ = 6.856 u and the measured harmonic wavenumber is 2169.8 cm⁻¹, that is ν = 65.05 THz. Inverting, k = μω² = 1902 N/m = 118.7 eV/Ų. A free molecule has 3N degrees of freedom but no restoring force for its 3 translations and 3 rotations (2 for a linear molecule), so its Hessian has 6 (or 5) zero eigenvalues; CO has 3 × 2 − 5 = 1 vibration.
How can a crystal of 10²³ atoms have only a few phonon branches?
Periodicity: each mode is a wave with wavevector q, so only the n atoms of one cell enter D(q).
- Cell pattern: how the n atoms of one cell move; the same in every cell up to a phase.
- Phonon wavevector in the first Brillouin zone: the phase advances by q·R from cell to cell.
- Force constants between atom i in the home cell and atom j in cell l, summed over cells.
DFT & VASP Course · Lecture 316 / 33
In a crystal, label each atom by its cell l (at lattice vector Rl) and its position i within the cell, i = 1 … n. The force constants depend only on the relative position of the two cells, Φliα,l′jβ=Φ0iα,(l′−l)jβ, because every cell is equivalent. This is the same translational symmetry that gave Bloch's theorem for electrons (Lecture 14), and it has the same consequence: the normal modes can be chosen as waves that look the same in every cell up to a phase eiq⋅Rl.
Insert the wave on the slide into Newton's equations. The sum over all atoms J becomes a sum over cells l and atoms j, the phase factors collect into D(q), and each q gives its own small eigenvalue problem,
jβ∑Diα,jβ(q)ejβ(q)=ω2(q)eiα(q),with a 3n × 3n Hermitian matrix. For every q there are 3n solutions ων(q), ν = 1 … 3n. As q moves through the Brillouin zone each one traces a phonon branch; plotted along a path of high-symmetry points they form the phonon dispersion, the analogue of an electronic band structure.
Counting. A crystal of N cells has 3nN vibrational degrees of freedom. With periodic boundary conditions q takes N distinct values in the first Brillouin zone, each with 3n branches: 3nN modes in all, as it must. Three branches are acoustic: their frequency goes to zero at q = 0 (Γ), where they become rigid translations of the whole crystal. The other 3n − 3 are optical: atoms within a cell move against each other, and the frequency stays finite at Γ. Silicon, with n = 2, has 3 acoustic and 3 optical branches.
The phonon wavevector q is not the electron wavevector k; both live in the same Brillouin zone because both come from the same lattice. A phonon of wavevector q is a lattice wave of wavelength 2π/|q|, and q and q + G (a reciprocal-lattice vector) describe the same motion of the atoms, which is why q can be restricted to the first zone. Some codes put the atomic positions ri into the phase, eiq⋅(Rl+rj−ri); this changes the eigenvectors by phases but not the frequencies.
What do acoustic and optical branches look like?
Acoustic: neighbours move together and ω → 0 at Γ. Optical: they move against each other.
DFT & VASP Course · Lecture 317 / 33
The simplest crystal with both kinds of branch is a chain of alternating masses M1 and M2 joined by identical springs C, with two atoms in each cell of length a. Each atom 1 is bonded to the atom 2 of its own cell and to the atom 2 of the cell to its left. Following the recipe of the previous slide, the 2 × 2 dynamical matrix is
D(q)=(2C/M1−C(1+eiqa)/M1M2−C(1+e−iqa)/M1M22C/M2),and its two eigenvalues are
ω±2(q)=C(M11+M21)±C(M11+M21)2−M1M24sin2(qa/2).The panel shows the matrix as it changes with q: the diagonal is fixed, and the off-diagonal element has a fixed size pattern and a phase that winds with q.
The limits. At Γ (q = 0) the minus sign gives ω = 0: the acoustic mode is a rigid translation, both atoms moving together (click near Γ on the lower branch). The plus sign gives the optical mode, ω2=2C(1/M1+1/M2), with the two atoms moving in opposite directions and the centre of mass of each cell at rest. At the zone boundary X (q = π/a) the two branches are ω2=2C/Mheavy and 2C/Mlight: in each, only one kind of atom moves. Between them is a gap with no vibrations at all; it closes when the masses are equal, because then the chain is really monatomic with cell a/2, and the "optical" branch is just the acoustic branch folded back into the smaller zone.
Sound. Near Γ the acoustic branch is linear, ω≈vsq, with vs=aC/[2(M1+M2)]: long acoustic waves are sound waves, and their slope is the speed of sound. In three dimensions there are two transverse acoustic (TA) branches and one longitudinal (LA), and the optical branches are likewise transverse (TO) or longitudinal (LO).
The chain shows the mode with its displacements drawn sideways so that you can see them; in the model they are along the chain. Note what the mass does: heavy atoms vibrate slowly. Replace hydrogen by deuterium and the hydrogen modes drop by close to √2, the basis of isotope effects in spectroscopy and diffusion.
How do you read a phonon band structure and DOS?
Each line is one branch traced along a path through the zone; the DOS counts the modes at each frequency over the whole zone.
DFT & VASP Course · Lecture 318 / 33
Real crystals are three-dimensional, so the chain's two curves become 3n branches, and q becomes a vector. The figure is the smallest 3D crystal with both kinds of branch: a simple cubic lattice with two atoms per cell, A at the corner and B at the centre (the CsCl structure), held together by springs. Its 6 × 6 dynamical matrix is computed and diagonalized live at every point. It is a model, but every feature you will meet in a real phonon plot is in it.
How the band plot is made. q cannot be drawn in three dimensions, so the plot follows a path of straight lines between high-symmetry points of the Brillouin zone: for a simple cubic lattice, Γ (0, 0, 0), X (½, 0, 0), M (½, ½, 0) and R (½, ½, ½), in units of 2π/a. The horizontal axis is the distance travelled along the path; vertical lines mark the special points. At each q the 3n frequencies are computed and joined in order of size. This is the same picture as an electronic band structure (Lecture 29), and phonopy (--band auto) and VASP (LPHON_DISPERSION with a QPOINTS file) produce it in the same way.
Read the model. Three branches start at zero at Γ: the acoustic branches, sound waves in the long-wavelength limit. Along Γ–X two of them coincide: the two transverse acoustic (TA) modes, with atoms moving perpendicular to q, are degenerate because the fourfold axis along x makes y and z equivalent. The longitudinal acoustic (LA) mode, with atoms moving along q, is stiffer and higher. Their slopes at Γ are the speeds of sound. The other three branches are optical. At Γ they are degenerate, 9.03 THz with the default masses, from ω2=38k1(1/mA+1/mB); this model has no Born charges, so there is no LO–TO splitting. Away from Γ they split into two TO and one LO branch.
Zone-boundary points. At X and R in this model the two kinds of atom decouple completely: the lower modes move only the heavy A atoms, the upper modes only the light B atoms (pick "at R", or click near X). At X the A-only modes are at 4.04 THz (T) and 4.94 THz (L), and the B-only modes at 8.07 and 9.89 THz, exactly (8k1/3)/m and (8k1/3+4k2)/m in the THz conversion. This is the three-dimensional version of the chain's zone boundary, and it is why a large mass difference opens a gap between acoustic and optical branches. Set m_A = m_B and watch the gap close.
The density of states. The band plot shows only the lines of the path. The phonon DOS counts all modes in the whole zone:
g(ν)=Nq1q∑j∑δ(ν−νj(q)),∫g(ν)dν=3n.The figure evaluates it on a uniform mesh of q-points with a small Gaussian broadening, as phonopy does with --mesh. At low frequency g grows as ν², the Debye behaviour of three acoustic branches. Peaks appear where branches are flat, because a flat branch puts many q-points at nearly the same frequency (van Hove singularities). The projected DOS weights each mode by how much of its eigenvector sits on each atom, ∣eA∣2 or ∣eB∣2: here the heavy A atoms carry the low frequencies and the light B atoms the high ones. Each projection integrates to 3, the three degrees of freedom of one atom.
The 3D view. The atom at position r moves by Re[eei(q⋅r−ωt)]/M, the formula of the crystal-modes slide, with the amplitude exaggerated and the time slowed to one cycle every two seconds. Pick TA and LA: the same q, with atoms moving across or along the arrow. Pick TO and LO: now A and B move against each other within each cell. "Optical at Γ" moves every cell identically, the two sublattices against each other, which is the pattern Raman and infrared spectroscopy see. Drag the crystal to look at it from any side.
Two cautions. A path shows only part of the zone: an instability at a general q, off the path, does not appear in the band plot but does appear in the DOS (as weight below zero) and in a mesh calculation, so check both. And the ordering by size makes crossing branches look as if they bounce off each other; phonopy can follow branches by their eigenvectors instead (BAND_CONNECTION).
What does a phonon band structure tell you?
Whether the crystal is stable, how stiff it is, which atoms vibrate where, and how easily heat moves.
| Feature | What it tells you |
|---|---|
| Branches below zero | An instability at that q, or an artefact to test |
| Acoustic slopes at Γ | Speeds of sound; they must agree with the elastic constants |
| Flat optical branches | Localized, weakly coupled vibrations; slow phonons that carry little heat |
| Gap between acoustic and optical | A large mass contrast; fewer ways for phonons to scatter |
| LO above TO at Γ | A polar crystal: Born charges and ε∞ matter |
| Peaks in the DOS | Many modes at one frequency: heat capacity, neutron and infrared spectra |
| Projected DOS | Which atoms vibrate at which frequencies: heavy low, light high |
DFT & VASP Course · Lecture 319 / 33
A phonon band structure is often shown in a paper only to prove "no imaginary frequencies". It carries much more.
Stability. Any branch below zero, anywhere in the zone, means the structure is a saddle point at that q, unless the tests of the artefact slide explain it away. Look at the zone boundary and along the lines as well as at Γ.
Stiffness and sound. Near Γ the acoustic frequencies are linear in q, and the slope v=2πdν/dq is a speed of sound. For a cubic crystal along [100], vLA=C11/ρ and vTA=C44/ρ, with ρ the mass density. The phonon slopes and the elastic constants of Lecture 22 are two measurements of the same thing, and checking one against the other is a good test of both calculations. A TA branch that is almost flat near Γ signals a crystal that is soft against shear, often close to a structural instability.
Heat transport. The slope of a branch anywhere is its group velocity, the speed at which a phonon carries energy. The lattice thermal conductivity is, in kinetic theory, κ=31∑Cqjvqj2τqj, a sum over modes of heat capacity times squared group velocity times lifetime. Flat branches carry little heat. Very flat low-lying optical branches mark weakly bound "rattling" atoms in cages, as in clathrates and skutterudites, which are poor heat conductors. A wide gap between acoustic and optical branches removes many of the three-phonon scattering processes that limit the lifetimes, which helps explain the very high thermal conductivity of boron arsenide. Lifetimes need third-order force constants (phono3py, ShengBTE, TDEP).
Polarity. A jump of the LO branch at Γ, different when Γ is approached from different directions, is the LO–TO splitting of a polar crystal (next slides on polar materials). If your polar crystal shows no such jump, the non-analytic correction is missing.
The DOS and the atoms. The DOS is what enters the free energy, entropy and heat capacity, and what inelastic neutron scattering measures on a powder. The projected DOS tells you which atom does what: heavy atoms dominate low frequencies, light atoms high ones, and hydrogen-related modes, O–H stretches near 3600 cm⁻¹ (about 108 THz), sit far above everything else. In a proton conductor, for example, the projected DOS separates the O–H stretch and bend modes from the vibrations of the oxide lattice.
Comparing with experiment. Inelastic neutron and x-ray scattering measure the dispersion itself, branch by branch. Raman and infrared spectroscopy see only optical modes at Γ, and only those allowed by symmetry selection rules. Measured frequencies include temperature and anharmonic shifts, so compare like with like: harmonic results with low-temperature data, or include the finite-temperature effects of the last part of this lecture.
How does a DFT code compute force constants?
Displace one atom by ±δ in a supercell, compute every force, and take the difference.
- One displaced structure, one full SCF run, gives a whole column: the forces on all atoms.
- Without symmetry: 2 × 3N runs. Symmetry cuts this to a handful for simple crystals.
DFT & VASP Course · Lecture 3110 / 33
DFT gives forces directly: VASP computes them from the electron density by the Hellmann–Feynman theorem, plus Pulay corrections for the basis (Lecture 21). The finite-displacement (or "frozen-phonon") method uses nothing else. Take the relaxed structure, move atom J by +δ along β, run a self-consistent calculation and record the forces on all atoms. Repeat with −δ. The central difference on the slide gives a full column of the force-constant matrix from two calculations.
Two features make this practical. First, one calculation gives the forces on every atom at once, so the number of calculations scales with the number of displacements, 2 × 3N at most, not with the number of matrix elements. Second, symmetry relates columns to each other: if a rotation of the crystal maps atom J onto atom K, the column of K follows from that of J by rotating the force pattern. Codes such as phonopy and VASP with IBRION = 6 find the minimal set of symmetry-distinct displacements and generate the rest. For a high-symmetry crystal such as silicon this is one or two displacements instead of dozens.
The calculation is done in a supercell, several primitive cells repeated in each direction. The reason: VASP's cell is periodic, so displacing one atom also displaces all its periodic images. The forces on atom I then contain the response to atom J and to every image of J. If the supercell is large enough that the force constants have died out before the next image, the images do not matter and the force constants are those of the infinite crystal. A later slide shows what goes wrong when it is too small.
Each displaced structure must be a single-point calculation with the atoms frozen: no relaxation (NSW = 0 or IBRION = −1 in phonopy runs). Relaxing would move the atoms back and destroy the displacement you are measuring.
How large should the displacement be?
Large enough that force noise is small next to the signal, small enough to stay harmonic: about 0.01–0.015 Å.
DFT & VASP Course · Lecture 3111 / 33
A finite difference has two competing errors, and the figure computes both for a model bond whose exact force constant is known.
The anharmonic (truncation) error grows with δ. Expand the force around the equilibrium position, F(u)=−Φu−21E′′′u2−61E′′′′u3−…. In the central difference [F(δ)−F(−δ)]/2δ the even terms cancel: the cubic anharmonicity drops out, and the error is 61E′′′′δ2, second order in δ. For the stiff Morse bond the relative error in the frequency is about 127a2δ2 (half that of the force constant): 0.04% at δ = 0.015 Å. A one-sided difference, −F(δ)/δ, keeps the cubic term, and its error is first order: about 2% at the same δ for this bond, fifty times worse. That is why VASP calls NFREE = 1 strongly discouraged and uses central differences (NFREE = 2) by default. NFREE = 4 adds displacements of ±2δ and cancels the δ² term as well.
The noise error grows as δ shrinks. Every computed force has a small error: from the incompletely converged SCF loop, from the finite FFT grid (the "egg-box" effect, an energy that changes slightly when an atom moves relative to the grid), and from k-point sampling. If each force has a random error σ, the force constant has an error of about σ/(2δ). With σ = 10⁻⁴ eV/Å and δ = 0.015 Å that is 0.005 eV/Ų, irrelevant for a 13 eV/Ų bond but 3% of the soft mode's 0.17 eV/Ų. Switch the figure to the soft mode and raise σ: the noise exceeds the force constant itself, and the computed ω² can come out negative. This is the most common origin of small spurious imaginary modes.
Defaults. VASP uses δ = POTIM = 0.015 Å for IBRION = 5 and 6 (it resets unreasonably large values to this). phonopy's default displacement is 0.01 Å. Both sit near the minimum of the total error for typical bonds, provided the forces are accurate: hence EDIFF = 10⁻⁸ eV and PREC = Accurate for phonon runs. For very soft modes, increase the accuracy of the forces before increasing δ, and check that the frequencies do not change when δ is varied.
Which wavevectors does a supercell compute exactly?
Only those whose waves fit in it; everywhere else the dispersion is interpolated.
DFT & VASP Course · Lecture 3112 / 33
A supercell made of N1×N2×N3 primitive cells is itself a crystal, with a smaller Brillouin zone. A phonon of wavevector q looks the same in every supercell only if its phase returns to the same value after one supercell length, eiq⋅L=1 for every supercell lattice vector L. Those wavevectors, q=∑i(mi/Ni)bi with integers mi, fold onto the Γ point of the supercell, and they are the only ones a supercell calculation contains. A 2 × 2 × 2 supercell of a cubic perovskite, for example, contains Γ, X, M and R exactly, and nothing in between.
Why the supercell is exact at those points. When atom J is displaced in the supercell, so are all its periodic images, and the measured force constant is the sum over images, ΦSC(0,j)=∑mΦ(0,j+mN). At a commensurate q every image carries the same phase, so this sum is exactly what D(q) needs. The dots in the figure mark these q-points; at each of them the supercell result lies on the exact curve, for any model.
Everywhere else it is Fourier interpolation. To get ω(q) at other q, codes assign each supercell force constant to the nearest image of the atom pair (the minimal-image convention) and evaluate D(q) from the formula of the crystal-modes slide. This is exact only if every force constant beyond half the supercell length is zero, because then each sum over images has a single nonzero term. In the short-range model the springs reach two neighbours, and from N = 4 on the interpolated curve is exact everywhere. In the long-range model the springs never stop; the error shrinks with N but never vanishes.
In real materials. Force constants in insulators decay quickly, except for the dipole–dipole part in polar crystals (next slides). In metals, screening by the conduction electrons gives long-range oscillating force constants and sharp dips in the dispersion (Kohn anomalies), which need large supercells. The only reliable check is convergence: compute the dispersion with two supercell sizes and compare. Two practical rules follow. Scale the k-point mesh inversely with the supercell (a primitive-cell 8 × 8 × 8 mesh becomes 4 × 4 × 4 for a 2 × 2 × 2 supercell). And never trust a feature of the dispersion between commensurate points until a larger supercell confirms it, in particular a small imaginary dip.
Can force constants be computed without moving any atom?
Yes: density-functional perturbation theory computes the orbitals' response to a displacement directly.
- First-order change of each occupied orbital when one atom moves.
- Perturbation: how the Hamiltonian (and PAW overlap) changes with that displacement; it depends on the density response, so it is solved self-consistently.
| Finite differences | DFPT in VASP | |
|---|---|---|
| VASP tag | IBRION = 6 | IBRION = 8 |
| Wavevectors | Γ of the cell given | Γ of the cell given |
| Functionals | any | LDA, GGA |
| Step size | needs POTIM | none |
| Elastic tensor | yes, with ISIF ≥ 3 | no |
DFT & VASP Course · Lecture 3113 / 33
Density-functional perturbation theory (DFPT) differentiates the Kohn–Sham equations instead of the forces. Differentiating H^∣ψn⟩=εnS^∣ψn⟩ with respect to a displacement u gives the linear Sternheimer equation on the slide for the first-order orbitals. Its right-hand side contains the change of the self-consistent potential, which depends on the change of the density, which depends on the first-order orbitals: so it is solved by the same kind of self-consistent loop as the ground state. A useful property is that only occupied orbitals are needed. The force constants then follow from the first-order orbitals.
In codes built around DFPT, such as Quantum ESPRESSO and ABINIT, the perturbation can be a wave of any wavevector q, and phonons at any q are computed in the primitive cell without a supercell. Phonon dispersions are then obtained on a coarse q-grid and interpolated.
VASP's implementation is more limited. IBRION = 7 (no symmetry, deprecated) and 8 (with symmetry) compute only Γ-point phonons of the cell you give it, so dispersions still need a supercell. The routines support only LDA and GGA functionals (not meta-GGAs such as r²SCAN, nor hybrids), they do not give the elastic tensor, and they do not work with the empirical dispersion corrections of Lecture 28. Internally VASP even uses small finite differences, for the second derivatives of the exchange–correlation energy and to turn first-order orbitals into force constants. The main advantage is that there is no displacement to choose; VASP's documentation suggests using DFPT once to find the largest displacement whose finite-difference result agrees with it. With r²SCAN or a hybrid functional, use finite differences.
What must be converged in a phonon calculation?
The forces must be clean and the supercell big: phonons are more demanding than energies.
| Setting | Why it matters | Typical choice |
|---|---|---|
| Relaxation | Residual forces shift the expansion point; soft modes turn imaginary | ISIF = 3, EDIFFG ≈ −0.001 eV/Å, same settings as the phonon run |
| SCF accuracy | Force noise ∝ σ/δ | EDIFF = 10⁻⁸ eV, PREC = Accurate |
| Supercell | Exact only at q that fit; images couple | Compare two sizes; Γ alone is not enough |
| k-points | Forces in the supercell | Primitive mesh ÷ supercell multiple |
| Cutoff and grid | Basis error, egg-box noise | Same ENCUT as the relaxation; check optical modes |
| Smearing (metals) | Occupations change as atoms move | ISMEAR = 1 or 2 with small SIGMA; never ISMEAR > 0 for insulators |
DFT & VASP Course · Lecture 3114 / 33
Energies converge faster than forces, and forces faster than force constants, which are differences of forces divided by a small number. Each row of the table is a separate source of error, and the earlier slides showed how each one appears.
Relaxation. The harmonic expansion assumes zero forces at the reference structure, and a cell that is not relaxed is under stress. Relax both the positions and the cell (ISIF = 3) with the same cutoff, k-points and functional you will use for the phonons, and with tight forces. The VASP documentation names residual forces, from a relaxation that stopped at NSW instead of EDIFFG, as a frequent source of spurious imaginary modes. For cell relaxations, the Pulay stress requires a cutoff higher than the default (Lecture 21).
Forces. Use EDIFF = 10⁻⁸ eV and PREC = Accurate for the phonon run itself. VASP's documentation advises against switching on ADDGRID without testing. A useful check is the Γ-point optical frequency against ENCUT, PREC and the k-mesh: it should change by much less than the accuracy you need.
Supercell and k-points. The previous slide is the rule: converge the supercell size, and scale the k-mesh down as the supercell grows so that the k-point density in reciprocal space stays the same. For Si, the VASP documentation gives 12 × 12 × 12 for the primitive cell, 6 × 6 × 6 for a 2 × 2 × 2 supercell and 3 × 3 × 3 for a 4 × 4 × 4 one.
Smearing. In metals, use Methfessel–Paxton smearing (ISMEAR = 1 or 2) or Gaussian smearing with a small width, and check the frequencies against SIGMA. In insulators, ISMEAR > 0 can produce phonon frequencies off by more than 20%.
Why do polar crystals need extra input near Γ?
Long-wavelength LO vibrations create a macroscopic electric field that stiffens them, and a finite supercell cannot see it.
- Born effective charge tensor: the polarization created by moving atom i.
- Electronic dielectric tensor, ions clamped: how the electrons screen the field.
- Lyddane–Sachs–Teller: one polar mode, cubic crystal; ε₀ includes the ions.
| GaAs at Γ | Value |
|---|---|
| TO, measured | 269 cm⁻¹ (8.06 THz) |
| LO, measured | 292 cm⁻¹ (8.76 THz) |
| ε₀ / ε∞ | 12.9 / 10.89 |
| LO from TO and ε | 292 cm⁻¹ |
DFT & VASP Course · Lecture 3115 / 33
In an ionic or polar crystal, moving the atoms of a cell against each other creates an electric dipole. The Born effective charge Zi,γα∗ measures it: the change in the macroscopic polarization along γ per unit displacement of atom i along α (times the cell volume over e), or equivalently the force on atom i per unit applied electric field. It is a tensor, and it can differ a lot from the formal ionic charge.
For an optical mode at very long wavelength, the dipoles of all cells add up to a uniform polarization. If the atoms move along q (longitudinal, LO), this polarization has a nonzero divergence and produces a macroscopic electric field that opposes the motion, adding stiffness. If they move perpendicular to q (transverse, TO), it does not. The result is LO–TO splitting: at Γ the LO frequency lies above the TO frequency. The extra term, shown on the slide in Gaussian units (Ω is the cell volume), is called non-analytic because it depends on the direction from which q approaches zero, not only on |q|.
A finite-displacement supercell is periodic and neutral, with no macroscopic field, so it misses this term: VASP's Γ-point output has LO = TO. The correction needs the Born charges and the electronic dielectric tensor ε∞ (ions clamped), which VASP computes with LEPSILON (DFPT, LDA and GGA) or LCALCEPS (finite electric fields, any functional, gapped systems only) in a separate run on the primitive cell, with a denser k-mesh. phonopy reads them from a BORN file and applies the correction automatically; VASP's own dispersion does it with LPHON_POLAR = .TRUE. plus PHON_DIELECTRIC and PHON_BORN_CHARGES. The long-range dipole–dipole force constants also decay slowly in real space, and without the correction the dispersion near Γ wiggles. Nonpolar crystals (Si, diamond: zero Born charges) and metals (infinite dielectric constant) need no correction.
A check with numbers. For a cubic crystal with one polar optical mode, the Lyddane–Sachs–Teller relation links the splitting to the static and high-frequency dielectric constants. For GaAs, the measured TO frequency of 268.7 cm⁻¹ and ε₀/ε∞ = 12.9/10.89 give ωLO=268.712.9/10.89=292.5 cm⁻¹, against 292.2 cm⁻¹ measured by Raman scattering.
What does an imaginary frequency mean?
A negative ω²: moving along that mode lowers the energy, so the structure sits on a saddle point.
DFT & VASP Course · Lecture 3116 / 33
Take one normal mode and write the energy as a function of the amplitude u along its eigenvector. The harmonic part is 21κu2, where κ is the mode's eigenvalue times its mass, and ω2=κ/m. If κ is positive, u = 0 is a minimum along this mode and the frequency is real. Drag κ below zero: the curvature at u = 0 becomes negative, ω² < 0, and the frequency is imaginary. VASP prints such a mode with the label f/i instead of f, and phonon plots show it as a negative frequency by convention.
An imaginary frequency is therefore not a numerical failure of the diagonalization; it is a statement about the energy surface. The structure is at a stationary point that is a maximum along this mode, and a displacement along the eigenvector lowers the energy. The quartic term, which the harmonic analysis does not see, eventually turns the energy back up, producing a double well. Its minima are at u0=±−κ/β, its depth is κ2/4β, and the curvature at the bottom is −2κ, so the new structure has a real frequency √2 times the magnitude of the old imaginary one, in this model.
The soft mode that becomes unstable defines the distortion: its eigenvector says which atoms move and which way, and its wavevector q says how the distortion repeats from cell to cell. A Γ-point instability is the same in every cell (for example, a ferroelectric shift of the ions in BaTiO₃). A zone-boundary instability alternates from cell to cell and doubles the cell: at the R point of a cubic perovskite, the oxygen octahedra rotate in opposite senses in neighbouring cells.
The frozen-phonon scan. To map the double well, displace the atoms along the eigenvector by several amplitudes, run a static calculation for each (the open circles), and plot E(u). The depth tells you how much energy the distortion gains; if it is only a few meV per formula unit, thermal or zero-point motion may be enough to keep the average structure symmetric, the subject of the last part of the lecture.
What do the soft modes of a perovskite look like?
Patterns that repeat with the soft mode's q: a polar shift at Γ, octahedral tilts at M and R.
DFT & VASP Course · Lecture 3117 / 33
The eigenvector of an imaginary mode is the distortion the structure wants to make, and its wavevector says how that distortion repeats from cell to cell. Cubic perovskites ABO₃ show the three most common patterns, animated here. Each is a single mode of the dynamical matrix: an animation of the eigenvector as it would oscillate if ω² were positive.
Polar mode at Γ. In every cell the B cation moves one way and its oxygen cage the other. Because q = 0, all cells do the same thing, and the crystal acquires a net electric dipole: a ferroelectric distortion. In cubic BaTiO₃ this mode is imaginary in harmonic calculations; freezing it in gives the tetragonal polar phase stable below about 120 °C. No supercell is needed.
Tilt at R, q = (½, ½, ½). Each octahedron rotates about z, and its neighbours in every direction rotate the opposite way: the phase eiq⋅R is ±1 and changes sign from one B site to the next along x, y and z. Because the octahedra share corners, rotating one forces its neighbours in the plane to turn the other way; the R pattern also alternates along z. In Glazer's notation this is a⁰a⁰c⁻ (no tilt about a and b, out-of-phase tilt about c). It doubles the cell in all three directions and gives the tetragonal space group I4/mcm. This is the mode that drives SrTiO₃'s transition near 105 K, and the soft R-point mode measured in BaZrO₃.
Tilt at M, q = (½, ½, 0). The same rotation alternates within each layer, but all layers along z turn the same way: in-phase tilting, a⁰a⁰c⁺, space group P4/mbm when frozen in. Because q has no z component, the cell doubles only in the plane.
The A cations and the apical oxygens do not move in the tilt modes, and the B cations stay at the centres of their octahedra. Many perovskites freeze in several tilts at once: the common orthorhombic Pnma structure of CaTiO₃ and many others combines in-phase and out-of-phase tilts, a⁻a⁻c⁺. Which combination wins is decided by the energy, which is why the follow-the-mode procedure ends with a full relaxation and a new phonon calculation.
The patterns here are drawn from symmetry, not computed: the eigenvectors from a real calculation contain the same pattern, plus small admixtures of other displacements allowed by the symmetry of q. In a VASP OUTCAR or a phonopy output, the eigenvector of each mode lists exactly these per-atom displacement vectors; phonopy can also write animations of a mode for viewers such as V_Sim or Jmol.
Is an imaginary mode real or an artefact?
Test it: artefacts move or vanish when the settings tighten; a real instability persists at a definite q.
| What you see | Likely cause | What to do |
|---|---|---|
| Tiny f/i on the three acoustic modes at Γ | Force noise; sum rule slightly violated | Ignore if < about 0.5 meV (0.1 THz); symmetrize Φ |
| Imaginary modes scattered everywhere | Structure not relaxed; residual stress | Relax again with ISIF = 3 and tight EDIFFG |
| Small dip between commensurate q | Supercell too small | Repeat with a larger supercell |
| Changes with EDIFF, ENCUT, k-mesh, δ | Numerical noise in the forces | Tighten until the frequency stops moving |
| Robust, at a high-symmetry q, clear pattern | A real instability | Follow the mode (next slide) |
| Real, but the phase is seen in experiments when hot | Stabilized by anharmonicity | Finite-temperature phonons (end of lecture) |
DFT & VASP Course · Lecture 3118 / 33
Every imaginary mode deserves an explanation before you draw a conclusion. The first four rows are artefacts of the calculation; the last two are physics.
Acoustic noise at Γ. The three acoustic modes at Γ are exact translations and should have zero frequency. Numerically they come out as small real or imaginary numbers because the forces are not perfectly accurate and the acoustic sum rule is not perfectly satisfied. VASP sorts the modes by ω², from the largest down, so the acoustic modes come last unless there are genuine imaginary modes, which follow them with the most unstable one at the very end. Its documentation treats imaginary parts below about 0.5 meV as noise. phonopy and other codes can impose the sum rule on the force constants. A small imaginary acoustic branch near Γ that persists away from Γ is a different matter: in two-dimensional materials it often comes from the flexural (out-of-plane) mode, whose frequency rises as q², and needs tight settings and the sum rules for rotations as well.
Unrelaxed structures. If many modes at many q-points are imaginary, the structure is usually not at a stationary point at all. Relax the positions and the cell with the settings of the phonon run, with forces below about 0.001 eV/Å, and recompute.
Supercell and noise artefacts. These change when the supercell, the SCF tolerance, the cutoff, the k-mesh or the displacement changes. A real instability does not.
Real instabilities are robust to all of these, sit at a definite wavevector, usually a high-symmetry point (Γ, X, M, R in a cubic perovskite), and have a clear displacement pattern. They mean the structure is a saddle point at 0 K in the harmonic approximation, and a lower-energy, lower-symmetry structure exists nearby. Whether the high-symmetry phase can still exist at some temperature is a separate question, answered by the methods of the last part of the lecture. An example relevant to proton-conducting perovskites: several harmonic DFT calculations find the R-point octahedral-tilt mode of cubic BaZrO₃ imaginary, yet experiments find BaZrO₃ cubic down to 2 K, with that mode real and soft, at 5.6 meV at 2 K and 9.4 meV at room temperature. Reproducing those numbers takes anharmonic force constants and the quantum zero-point motion of the atoms.
How do you find the stable structure behind an imaginary mode?
Distort along the eigenvector in a cell that holds its q, relax, and compute the phonons again.
- Find the softest mode: its q and eigenvector
- Build a cell that folds q to Γ
- Displace along the eigenvector, ± in alternate cells
- Scan E(u), or follow the IRC
- Relax positions and cell, ISIF = 3
- Recompute the phonons: all real?
| Soft mode at q | Diagonal supercell (smallest cell) |
|---|---|
| Γ (0, 0, 0) | primitive cell |
| X (½, 0, 0) | 2 × 1 × 1 |
| M (½, ½, 0) | 2 × 2 × 1 (√2 × √2 × 1) |
| R (½, ½, ½) | 2 × 2 × 2 (a 2-cell fcc cell) |
DFT & VASP Course · Lecture 3119 / 33
A real imaginary mode is a recipe for a better structure. Start with the softest one, the mode with the largest imaginary frequency, which usually leads to the largest energy gain.
Choose the cell. VASP displaces atoms only at Γ of the cell you give it, so the instability must fold to Γ in the cell you use. For a mode at q = (h, k, l) in fractional coordinates of the primitive reciprocal lattice, a supercell of 1/h × 1/k × 1/l primitive cells (1 for a zero component) does that: the X point needs a 2 × 1 × 1 cell, not 2 × 2 × 2. Smaller non-diagonal cells often exist, shown in brackets: for M, the √2 × √2 × 1 cell; for R, the face-centred cell spanned by (0, 1, 1), (1, 0, 1) and (1, 1, 0), which holds two perovskite units (10 atoms), because q·L is an integer for each of its vectors. The tetragonal tilted structure is usually set up in the conventional √2 × √2 × 2 cell of 20 atoms.
Displace. Move the atoms along the eigenvector by a small amplitude, about 0.05 Å for a start. For a zone-boundary mode the phase eiq⋅Rl is ±1, so the displacement changes sign from one primitive cell to the next. The displacement breaks the symmetry, so the subsequent relaxation can leave the saddle point.
Follow the energy down. Either scan the amplitude with static runs and pick the minimum of E(u), as in the double-well figure, or let VASP follow the steepest-descent path from the saddle point: IBRION = 40 runs an intrinsic-reaction-coordinate (IRC) calculation, with the soft-mode eigenvector appended to the POSCAR. Then relax everything, positions and cell (ISIF = 3).
Check again. Compute the phonons of the new structure. The symmetry lowering usually removes the original instability, but another one may appear, especially if the first dispersion showed imaginary modes at several q-points. Repeat until all frequencies are real. The new structure's space group is a subgroup of the old one: for SrTiO₃, the R-point tilt takes the cubic perovskite, space group Pm3̄m (221), to the tetragonal I4/mcm (140) observed below about 105 K.
What do phonons tell you about energy at finite temperature?
Each mode is a quantum oscillator; summing over all modes gives zero-point energy, free energy, entropy and heat capacity.
- Zero-point energy: a quantum oscillator never rests; present even at 0 K.
- Thermal part: negative, from the entropy of the excited vibrations.
- Phonon density of states: how many modes per frequency interval; needs a dense q-mesh.
DFT & VASP Course · Lecture 3120 / 33
In the harmonic approximation the normal modes are independent oscillators, and quantum mechanics gives each one the energy levels ℏω(n+21), n = 0, 1, 2, … (Lecture 2). The partition function of one oscillator is a geometric series, Z=∑ne−ℏω(n+1/2)/kBT=e−ℏω/2kBT/(1−e−ℏω/kBT), and its free energy −kBTlnZ is the bracket on the slide. Summing over all modes gives the vibrational free energy of the crystal. The other quantities follow by differentiation:
S=−∂T∂F,U=F+TS=qν∑ℏω(21+eℏω/kBT−11),CV=∂T∂U.The second form on the slide replaces the sum by an integral over the phonon density of states g(ω), normalized to 3n modes per cell, with f(ω,T) the bracket for one mode. In practice phonopy evaluates the sum on a uniform q-mesh (for example 20 × 20 × 20), and the mesh must be converged.
The zero-point energy ∑21ℏω is present at 0 K. It is typically tens of meV per atom, and much more for light atoms: an O–H stretch near 3600–3700 cm⁻¹ alone carries 0.22–0.23 eV. When two structures with different vibrational spectra are compared, the difference in zero-point energy can change their order, especially for hydrogen-containing materials.
Imaginary modes break this formula. The logarithm ln(1−e−ℏω/kBT) has no meaning for imaginary ω. Codes either skip such modes or fail. A free energy computed with imaginary modes left out is not a free energy of anything; it means the harmonic description of that structure does not apply.
How large are the vibrational contributions?
Zero-point energy is tens of meV per atom; near room temperature, −TS is of the same size and grows quickly.
DFT & VASP Course · Lecture 3121 / 33
To see magnitudes without a full calculation, the figure uses the Debye model: all three branches are treated as sound waves, g(ω)∝ω2, cut off at the Debye frequency ωD=kBΘD/ℏ so that there are exactly three modes per atom. The Debye temperature ΘD measures how stiff the crystal is; it is a few hundred kelvin for most solids, higher for diamond-like light, stiff solids, lower for heavy, soft ones. Per atom, the model gives
EZP=89kBΘD,F=EZP+3kBTln(1−e−ΘD/T)−kBTD3(ΘD/T),D3(x)=x33∫0xet−1t3dt.Read the figure. With ΘD = 400 K the zero-point energy is 38.8 meV per atom. U rises with T as modes are excited. F falls, because −TS grows faster than U; it crosses zero near room temperature for this ΘD and keeps falling. The heat capacity rises as T³ at low temperature (the Debye law, CV=512π4kB(T/ΘD)3) and approaches 3kB per atom, the classical Dulong–Petit value, once T exceeds ΘD.
Why this matters for stability. Energy differences between competing polymorphs are often 10–50 meV per atom. Vibrational free energies differ between structures by amounts of that size at a few hundred kelvin, so a ranking of structures by their 0 K DFT energy can reverse at high temperature. The next slides make this quantitative. A real calculation uses the computed phonon density of states instead of the Debye form; the Debye model is only for intuition.
How does a crystal expand on heating, in DFT?
Quasi-harmonically: phonons soften as the volume grows, so minimizing E(V) + Fvib(V, T) gives V(T).
DFT & VASP Course · Lecture 3122 / 33
A purely harmonic crystal does not expand when heated: its frequencies do not depend on temperature, so its free energy has no temperature-dependent pressure. Real crystals expand because the frequencies depend on the volume. The quasi-harmonic approximation (QHA) keeps the harmonic formulas but lets the phonons change with volume:
F(V,T)=E(V)+Fvib(V,T),G(T,p)=Vmin[F(V,T)+pV].E(V) is the static DFT energy at each volume (Lecture 22); Fvib(V, T) comes from a phonon calculation at each volume. At each temperature the minimum over V gives the equilibrium volume V(T) and the Gibbs free energy G(T). In the figure, each grey curve is F(V) at one temperature; the dashed line connects their minima.
How strongly a mode softens with volume is measured by its Grüneisen parameter, γqν=−∂lnωqν/∂lnV, typically 1–3. The figure gives every mode the same γ. Set γ = 0 and nothing expands. With a positive γ the vibrational free energy is lowered by expanding, and the balance against E(V) gives the volumetric expansion coefficient
αV=V1dTdV=BTVγCV,where BT is the bulk modulus: soft crystals with large γ expand most. Notice also that even at 0 K the minimum lies above the static volume: zero-point motion expands the lattice, here by 0.3% in the lattice constant with γ = 2.
Doing it. Compute E(V) and the phonons at several volumes (phonopy's equation-of-state fit needs at least five) around the static minimum, relaxing the positions at each volume, then fit an equation of state at each temperature; phonopy's QHA tools do this and return V(T), α(T), BT(T), Cp(T) and G(T). For metals, add the electronic free energy. QHA works while the thermal displacements stay small enough for each volume's harmonic description. It fails when anharmonicity at fixed volume matters, as it does at high temperature in soft or strongly anharmonic crystals, and it cannot be applied at all to a phase whose phonons are imaginary at any of the volumes.
Which polymorph is stable at high temperature?
The one with the lower G. A softer phase gains more vibrational entropy and can overtake the ground state.
DFT & VASP Course · Lecture 3123 / 33
Two structures A and B of the same composition are compared by their Gibbs free energies. At zero pressure and with only vibrations,
ΔG(T)=GB−GA=ΔE0+ΔFvib(T),where ΔE₀ is the static DFT energy difference. Phase B is stable where ΔG < 0. At 0 K only the zero-point energies add to ΔE₀. At high temperature each mode contributes kBTln(ℏω/kBT) to F, so
ΔFvib≈kBTqν∑lnωqνAωqνB=3kBTlnΘAΘB per atom (Debye).A phase with lower frequencies has more vibrational entropy: its atoms have more room to move. If B is softer (ΘB < Θ_A), ΔF_vib falls linearly with T, and B overtakes A near T≈ΔE0/[3kBln(ΘA/ΘB)]. With the default settings, 30 meV/atom and Θ_B/Θ_A = 0.8, the crossing is at 504 K, and the high-temperature estimate gives 520 K.
Consequences. First, the transition temperature is very sensitive to ΔE₀: a few meV per atom of error in the DFT energy difference, typical between functionals, moves it by tens or hundreds of kelvin (drag ΔE in the figure). Second, open, low-coordinated structures are often the high-temperature phases, because they are vibrationally soft. Third, other entropies can matter as much as the vibrational one: configurational entropy in alloys and doped oxides, magnetic entropy in magnets, electronic entropy in metals.
The classic case is titanium and zirconium, close-packed (hcp) at room temperature and body-centred cubic (bcc) above 882 °C and 863 °C. The bcc phase is the soft, high-entropy one. But a harmonic calculation of bcc Ti or Zr gives imaginary frequencies, so the formula above cannot even be evaluated for it. How a phase with imaginary harmonic modes can be the stable one is the subject of the next slide.
Can a structure with imaginary modes be stable when hot?
Yes. Hot atoms explore the whole double well, and their average position keeps the high symmetry above Tc.
DFT & VASP Course · Lecture 3124 / 33
The harmonic test of dynamical stability asks about infinitesimal displacements from the high-symmetry structure at 0 K. At high temperature the atoms are not near that point: their thermal displacements can be larger than the distance to the bottom of the double well. What matters then is not the curvature at the top of the barrier but the shape of the whole energy surface that the atoms sample. This is anharmonicity doing real work.
The model. Each cell has one local displacement u (think of the tilt of one octahedron) with an on-site energy 21Au2+41Bu4 and a coupling −Juiuj to its z neighbours that favours equal displacements. For a uniform distortion, the frozen-phonon energy per cell is 21(A−zJ)u2+41Bu4: with zJ > A it is the double well drawn as a line on the left, and the harmonic phonon of the symmetric structure is imaginary, −3.91 THz (dashed). Each cell then feels its neighbours through their average displacement ⟨u⟩, the mean-field approximation, and the thermal distribution of u follows from Boltzmann statistics.
Read the figure. At low T the distribution sits in one well: the average structure is distorted, ⟨u⟩ ≠ 0, and its own soft mode (green) is real. As T rises, the atoms spread over both wells. Above the transition temperature Tc (473 K in this model) the distribution is centred at u = 0: the average structure has the full high symmetry, and the effective soft-mode frequency of the symmetric structure (blue) is real. It rises from zero as ω2∝(T−Tc), the soft-mode behaviour observed in many structural phase transitions. The symmetric phase is then dynamically stable at that temperature, although its 0 K harmonic phonons are imaginary.
What "effective frequency" means. The blue curve uses κeff=kBT/⟨u2⟩−zJ: the stiffness of the harmonic oscillator that has the same thermal fluctuations as the real, anharmonic one. Methods that compute finite-temperature phonons from first principles find such effective, temperature-dependent force constants for all modes at once (next slide).
Real examples. bcc Ti, Zr and Hf, whose harmonic phonons are imaginary but whose high-temperature phonons computed this way are real and agree with neutron scattering. Cubic SrTiO₃ above about 105 K. Halide perovskites such as CsPbBr₃, cubic only when hot. And BaZrO₃, where the R-point mode is close to the boundary and the zero-point motion of the atoms, which this classical model leaves out, helps keep it cubic down to 2 K.
How are phonons and free energies computed at high temperature?
By sampling the atoms' real thermal motion and fitting effective force constants, or by integrating the free energy.
| Method | What it adds | How |
|---|---|---|
| Quasi-harmonic | Volume dependence of ω | Phonons at several volumes; needs real modes |
| Self-consistent phonons (SSCHA, SCAILD; SCPH) | Temperature-dependent ω, quantum or classical | Sample displacements with the trial phonons' thermal amplitudes and refit (SCPH: solve analytically from quartic force constants) |
| TDEP | Effective force constants at T | Fit Φ to forces and displacements from MD at T |
| MD power spectrum | Full anharmonic spectrum | Fourier transform of the velocity autocorrelation |
| Thermodynamic integration | Exact anharmonic free energy | Integrate ⟨E − Eref⟩ from a harmonic reference |
| Machine-learned force fields | Affordable long, large MD | Train on DFT, then run the sampling |
DFT & VASP Course · Lecture 3125 / 33
All of these methods replace "the curvature at one point" by "the energy surface where the atoms actually are at temperature T". They differ in how they sample that surface and whether they treat the nuclei classically or quantum mechanically.
Self-consistent phonons. Guess a set of harmonic force constants. Generate atomic configurations with the displacements those phonons would have at temperature T (for a quantum oscillator, the Gaussian width includes the zero-point motion: ⟨u2⟩=2Mωℏcoth2kBTℏω). Compute DFT forces for these configurations, fit new force constants, and repeat until input and output agree. Imaginary modes in the starting guess are replaced by small real ones to start the loop. The stochastic self-consistent harmonic approximation (SSCHA) does this variationally, minimizing an upper bound to the free energy, and gives free energies as well; SCAILD is a related sampling scheme, while the SCPH method of ALAMODE solves the self-consistent equations analytically from precomputed third- and fourth-order force constants; phonopy can generate the thermally displaced supercells (RANDOM_DISPLACEMENT_TEMPERATURE). Because they include zero-point motion, these methods work at low temperatures too, which classical MD does not.
TDEP (temperature-dependent effective potential) runs ab initio MD at temperature T and finds the force constants that best reproduce the forces seen along the trajectory (next slide). It captures, on average, the anharmonicity the classical MD samples: the shifts of the frequencies with temperature, though not their broadening.
MD power spectrum. The vibrational density of states is the Fourier transform of the velocity autocorrelation function of an MD run. VASP's documentation describes the procedure: thermalize with a Langevin thermostat, then run several constant-energy trajectories with the velocities written out (VELOCITY = .TRUE.), and average the spectra. It needs no harmonic model at all, but gives no dispersion and no eigenvectors directly.
Thermodynamic integration computes the anharmonic free energy exactly (for classical nuclei): switch the energy from a reference Eref (for example the TDEP harmonic model) to the DFT energy along Eλ=(1−λ)Eref+λEDFT, run MD at each λ, and use F=Fref+∫01⟨EDFT−Eref⟩λdλ. It is the most accurate route to phase diagrams of strongly anharmonic solids, and the most expensive.
Cost. All of these need many DFT force evaluations in large supercells, hundreds to tens of thousands. Machine-learned force fields, trained on a modest set of DFT calculations, make long MD and dense sampling affordable (Lecture 34; in VASP, ML_LMLFF). The final numbers should still be checked against DFT on a subset of configurations.
How does an MD run become temperature-dependent phonons?
Fit the harmonic model whose forces best match the forces the atoms actually feel at temperature T.
- DFT forces on every atom in snapshot t of an MD run at temperature T (or of thermally displaced supercells).
- Displacements of the atoms from the average (symmetric) positions in that snapshot.
Then: D(q) from Φeff, frequencies at T, and dynamical stability at T if all ω² > 0.
DFT & VASP Course · Lecture 3126 / 33
In a finite-displacement calculation each snapshot has one atom displaced by 0.01 Å. In TDEP every atom is displaced at once, by the thermal motion at temperature T, which can be a tenth of an ångström or more. Each snapshot gives 3N force components, so a few dozen uncorrelated snapshots of a 100-atom supercell give thousands of equations for the symmetry-independent force constants, which are solved by least squares. Symmetry of the average structure is imposed on Φ, which keeps the number of unknowns small.
The fitted Φeff is not the second derivative at the symmetric structure. It is the best harmonic description of the energy surface as sampled at T: it includes, on average, the effect of the cubic and quartic terms. For the double-well mode it is positive above Tc even though the curvature at the top of the barrier is negative, which is exactly the classical effective stiffness of the previous figure. Third-order force constants can be fitted in the same way, for phonon lifetimes and thermal conductivity.
Practical points.
- Volume. Run the MD at the volume of the crystal at T: from an NpT run, from QHA, or from experiment. Effective phonons at the wrong volume give the wrong answer, whatever the sampling.
- Reference positions. The displacements are measured from the average positions, which in the high-symmetry phase are the symmetric sites. If the time-averaged structure in the MD drifts away from the symmetric one, the symmetric phase is not stable at that T in the simulation, which is itself the answer.
- Convergence. Check the frequencies against the number of snapshots, the supercell size and the length of the run. Snapshots should be separated by more than the correlation time of the slowest modes.
- Classical nuclei. Ordinary MD treats the nuclei classically, which is accurate well above the Debye temperature and wrong at low T, where zero-point motion dominates. At low temperature, sample with quantum amplitudes (self-consistent phonons) or path-integral MD.
An approximate free energy follows as well: the harmonic free energy of Φeff, plus a correction from the difference between the DFT energy and the fitted model averaged along the run. With it, the ΔG(T) of the polymorph slide can be evaluated for phases that QHA cannot handle.
What does a high-temperature stability study look like?
Harmonic phonons first, QHA where they are real, anharmonic sampling where they are not, then compare G(T).
- Relax each candidate; harmonic phonons over the whole zone
- All real: QHA gives V(T) and G(T)
- Imaginary: MD or MLFF at T, at the right volume
- Effective phonons at T (TDEP, SSCHA): all real?
- Anharmonic free energies; ΔG(T) = 0 gives Tc
- Compare with experiment: Tc, a(T), spectra
INCAR: NVT sampling at 1200 K
IBRION = 0 # molecular dynamics MDALGO = 3 # Langevin thermostat LANGEVIN_GAMMA = 10 10 10 # ps⁻¹, one per species TEBEG = 1200 # K ISIF = 2 # fixed cell: set V(T) first POTIM = 2 # fs; 0.5–1 with hydrogen NSW = 10000 ISYM = 0 ML_LMLFF = .TRUE. # optional: learn a force ML_MODE = train # field on the fly
DFT & VASP Course · Lecture 3127 / 33
The steps on the left combine everything in this lecture into one study of which phase is stable at which temperature.
- Harmonic phonons of every candidate, in converged supercells and over the whole Brillouin zone, with the artefact tests of the earlier slide. This sorts the candidates into dynamically stable ones and saddle points.
- Stable candidates: QHA. Phonons at several volumes give V(T), α(T) and G(T), as long as anharmonicity at fixed volume stays small; check this at the highest temperatures with one of the methods below.
- Candidates with imaginary modes that are known, or suspected, to exist when hot. Run MD at the target temperature in a supercell large enough to hold the soft mode's wavevector, at the volume appropriate to that temperature. Check the time-averaged structure: does it stay on the symmetric sites? The INCAR on the right is a minimal NVT run with the Langevin thermostat (MDALGO = 3, LANGEVIN_GAMMA per species in ps⁻¹) at fixed cell (ISIF = 2) and without symmetry (ISYM = 0); the MD lecture (Lecture 33) explains each choice. Switching on a machine-learned force field (ML_LMLFF) lets VASP learn from the first steps and run most of the trajectory at a fraction of the cost.
- Effective phonons at T from the MD snapshots (TDEP) or from self-consistent sampling (SSCHA). If they are all real at T, the phase is dynamically stable at T.
- Free energies. For each phase, at each T: harmonic or QHA free energy where that is valid, TDEP or SSCHA free energy where it is not, thermodynamic integration where the highest accuracy is needed. Find the temperature where ΔG changes sign.
- Validate against measured transition temperatures, thermal expansion and phonon spectra (Raman, infrared, inelastic neutron or x-ray scattering) wherever they exist.
Two cautions. Dynamical stability at T is necessary but not sufficient for thermodynamic stability: a dynamically stable phase can still have a higher G than its competitor. And all of this inherits the error of the functional in ΔE₀, which often dominates the uncertainty in Tc.
Where do phonons appear in VASP?
IBRION = 6 or 8 for Γ-point phonons of the cell given; a supercell plus LPHON_DISPERSION or phonopy for the rest.
| Task | In VASP |
|---|---|
| Γ phonons, finite differences | IBRION = 6 (5: no symmetry, deprecated), NFREE = 2, POTIM = 0.015 Å |
| Γ phonons, DFPT | IBRION = 8 (7 deprecated); LDA and GGA only |
| Accurate forces | EDIFF = 1E-8, PREC = Accurate, NSW = 1 |
| Born charges, ε∞ | LEPSILON or LCALCEPS, unit cell |
| Dispersion and DOS | Supercell + LPHON_DISPERSION + QPOINTS; LPHON_POLAR for polar crystals |
| Output | OUTCAR: “Eigenvectors and eigenvalues of the dynamical matrix”; f real, f/i imaginary; grep THz OUTCAR |
DFT & VASP Course · Lecture 3128 / 33
Γ-point phonons. IBRION = 6 displaces only the symmetry-distinct atoms and directions and fills in the rest by symmetry; IBRION = 5 displaces every atom in every direction and is deprecated, except that it is the only finite-difference mode that honours selective dynamics. That makes it the mode for vibrations of an adsorbate on a frozen surface, for example to get its zero-point energy. NFREE = 2 (the default for IBRION = 5 and 6) is the central difference of the step-size slide; POTIM is the displacement in Å, 0.015 by default. The documentation's recommended run also sets NSW = 1, EDIFF = 1E-8 and PREC = Accurate, and does not need the wavefunction or charge files (LWAVE = LCHARG = .FALSE.).
Reading the output. After the line “Eigenvectors and eigenvalues of the dynamical matrix”, the OUTCAR lists 3N modes from the highest frequency down. Each mode line gives the frequency in THz, 2πTHz, cm⁻¹ and meV, labelled f for a real frequency and f/i for an imaginary one, followed by the displacement eigenvector of every atom. grep THz OUTCAR lists them all. The modes are sorted by ω², largest first: the three acoustic translations, near zero, come last, followed only by any genuine imaginary modes, the most unstable one at the very end. From VASP 6.6.0, frequencies and eigenvectors are also written to vaspout.h5 and can be read with py4vasp.
Dispersions in VASP itself (6.3.2 and later). Compute the force constants in a supercell with IBRION = 6 or 8 and set LPHON_DISPERSION = .TRUE. with a QPOINTS file, written like a line-mode KPOINTS file, along a high-symmetry path of the primitive cell. A uniform q-mesh in QPOINTS with PHON_DOS gives the phonon DOS. For polar materials add LPHON_POLAR = .TRUE. with the ion-clamped dielectric tensor (PHON_DIELECTRIC) and the Born charges (PHON_BORN_CHARGES) from a LEPSILON or LCALCEPS run on the unit cell. The force constants can be reused by renaming vaspout.h5 to vaspin.h5 and setting LPHON_READ_FORCE_CONSTANTS.
Functionals. DFPT (IBRION = 7, 8) and LEPSILON work with LDA and GGA only. With r²SCAN or a hybrid, use IBRION = 6 and LCALCEPS.
What does a phonopy workflow look like?
phonopy writes the displaced supercells, VASP computes their forces, and phonopy builds Φ, D(q) and everything after.
Shell
# 1. displaced supercells from the relaxed POSCAR phonopy-init -d --dim 2 2 2 --pa auto # → SPOSCAR, POSCAR-001 …, phonopy_disp.yaml # 2. one static VASP run per POSCAR-XXX, in disp-XXX/ # 3. collect the forces → FORCE_SETS phonopy-init -f disp-*/vasprun.xml # 4. dispersion, then DOS and thermal properties phonopy --band auto -p phonopy --mesh 20 20 20 -t -p
INCAR for each displaced supercell
IBRION = -1 ; NSW = 0 # never relax PREC = Accurate EDIFF = 1E-8 ENCUT = 520 # as in the relaxation ISMEAR = 0 ; SIGMA = 0.05 LREAL = .FALSE. LWAVE = .FALSE. ; LCHARG = .FALSE.
Polar crystals: add a BORN file from a LEPSILON run on the unit cell.
DFT & VASP Course · Lecture 3129 / 33
phonopy is the most widely used finite-displacement code. It does the symmetry analysis and the bookkeeping; VASP only computes forces. The commands are those of phonopy version 4, which split the old single command in two: phonopy-init prepares and collects calculations, phonopy analyses them.
- Displacements. From the relaxed unit cell in POSCAR,
phonopy-init -d --dim 2 2 2builds a 2 × 2 × 2 supercell (SPOSCAR) and the symmetry-distinct displaced supercells (POSCAR-001, …), each with one atom moved by 0.01 Å.--pa autotells phonopy to find the primitive cell, so the dispersion is plotted in the primitive Brillouin zone; it is the default in version 4. - Forces. Run each displaced supercell as a single-point calculation with the INCAR shown, in its own directory. The k-mesh is the relaxation's mesh divided by the supercell multiple; ENCUT, functional, PAW datasets and smearing must match the relaxation. phonopy's documentation warns explicitly: do not relax the displaced structures.
- Collect.
phonopy-init -f disp-*/vasprun.xmlreads the forces and writes FORCE_SETS. - Analyse.
--band autochooses a standard high-symmetry path (it needs the seekpath Python package);-pplots it.--mesh 20 20 20 -tcomputes the DOS and thermal properties (F, S, CV) on a 20 × 20 × 20 q-mesh; check that the numbers do not change on a denser mesh. If a BORN file with the Born charges and ε∞ is present, the LO–TO correction is applied.
Then repeat with a larger supercell (3 × 3 × 3, or 4 × 4 × 4 for small cells) and compare the dispersions. For the quasi-harmonic approximation, run the whole workflow at several volumes and pass the results with the E(V) data to phonopy's QHA tools.
Lab 31: which crystal is dynamically stable?
Silicon and cubic SrTiO₃. Predict first: which modes at Γ, any imaginary branches, and where an instability would sit.
Part A: silicon, Γ only (VASP alone)
- Relaxed 2-atom primitive cell; IBRION = 6, NFREE = 2, POTIM = 0.015, EDIFF = 1E-8, PREC = Accurate, 12 × 12 × 12 k-points.
- Predict: 6 modes; 3 near zero, 3 degenerate optical modes; no f/i beyond about 0.5 meV. Compare with the Raman line, 520 cm⁻¹ (15.6 THz).
Part B: silicon dispersion (phonopy)
- 2 × 2 × 2 and 3 × 3 × 3 supercells; k-mesh scaled down.
- Predict: all branches real; LO = TO at Γ; differences only between commensurate q.
Part C: cubic SrTiO₃ (optional)
- Relax the 5-atom cubic cell; phonons in a 2 × 2 × 2 supercell.
- Predict: imaginary modes at R (octahedral tilts), possibly at M; the polar Γ mode is very soft: check its sign.
- Freeze the R-mode in a √2 × √2 × 2 cell, relax, recompute: all real?
DFT & VASP Course · Lecture 3130 / 33
Part A. Silicon has two atoms per primitive cell, so VASP prints 6 modes. Three are the acoustic translations, near zero and possibly tiny f/i values from noise. The other three are the optical modes at Γ, degenerate by the cubic symmetry, with no LO–TO splitting because Si is not polar. Compare their frequency with the 520 cm⁻¹ Raman line measured at room temperature, and explain any difference: the functional (the lattice constant it predicts), the temperature of the experiment, and anharmonicity all enter. Also count the displacements VASP makes with IBRION = 6, and compare with the 2 × 3 × 2 = 12 that IBRION = 5 would need.
Part B. Follow the phonopy slide. Before you look at the result, predict where the 2 × 2 × 2 dispersion is exact (the commensurate q-points) and where the 3 × 3 × 3 result might differ. Plot both on one graph. All branches should be real; if a small imaginary pocket appears, use the artefact table to find out why.
Part C. Cubic SrTiO₃ is stable only above about 105 K. A harmonic calculation at 0 K should therefore find it dynamically unstable, and the place to look is the zone-boundary R point, where the octahedral tilts live. Before running, decide which supercell contains R, and therefore which supercell you need. SrTiO₃ is polar, so compute Born charges and ε∞ for the Γ-point optical modes; the ferroelectric-like Γ mode of SrTiO₃ is itself very soft and sensitive to the volume and the functional, which makes it a good test of how carefully you converge. Then follow the mode as on the slide, and check whether the frozen distortion leads to the tetragonal I4/mcm structure.
No outputs are shown here: these are runs for you to make. Record your settings, the frequencies, and every check you did.
Which misconceptions should you drop now?
Six beliefs about phonons and stability. Click a card to turn it over.
“Zero forces mean the structure is stable.”
They mean a stationary point. Only ω² > 0 for every mode at every q rules out a saddle.
“Real Γ-point frequencies prove stability.”
Many instabilities sit at the zone boundary (X, M, R). You need the whole dispersion.
“An imaginary frequency is a numerical bug.”
It can be noise, but it can also be the most important result: a lower-energy structure nearby.
“A smaller displacement is always more accurate.”
Below about 0.01 Å the force noise dominates. Central differences already remove the cubic error.
“A phase with imaginary harmonic modes can never be observed.”
bcc Zr and cubic SrTiO₃ exist when hot: anharmonicity stabilizes them above Tc.
“The quasi-harmonic approximation includes anharmonicity.”
Only through the volume. It misses fixed-volume anharmonicity and fails with imaginary modes.
DFT & VASP Course · Lecture 3131 / 33
Check yourself
One question per outcome. Answer aloud, then click to compare.
How do force constants become phonon frequencies?
Mass-weight Φ and Fourier-sum it over cells into D(q); its eigenvalues are ω²(q), and ν = ω/2π.
Why is a 2 × 2 × 2 supercell exact only at some q?
Only q with eiq·L = 1 fold to its Γ. Elsewhere it interpolates, exact only if Φ vanishes beyond half the cell.
Name four settings of a reliable finite-displacement run.
Tight relaxation; EDIFF 1E-8 and PREC Accurate; δ ≈ 0.01–0.015 Å, central differences; converged supercell, scaled k-mesh.
Cubic perovskite, f/i = 2 THz at R, robust. Next?
Freeze the R mode with alternating signs in a 2 × 2 × 2 or √2 × √2 × 2 cell, relax, recompute: expect tilted octahedra.
How would you show that phase is stable at 1200 K?
MD or MLFF at 1200 K and the right volume; TDEP or SSCHA phonons all real; then compare G(T) with competitors.
DFT & VASP Course · Lecture 3132 / 33
What should you remember?
Second derivatives of the DFT energy decide stability, spectra and, with care, the phases seen when hot.
- Force constants are second derivatives of the energy; mass-weighted and Fourier-summed they form D(q), whose eigenvalues are ω².
- Acoustic branches go to zero at Γ, optical ones do not. Band slopes give sound speeds; the projected DOS shows which atoms vibrate where. A supercell is exact only at q that fit in it.
- Finite differences need tight forces and δ ≈ 0.01–0.015 Å; DFPT avoids δ but in VASP is limited to LDA, GGA and Γ. Polar crystals need Born charges.
- An imaginary mode is a saddle point, or an artefact: test it, then follow it to the lower-symmetry structure.
- Fvib gives zero-point energy, entropy and, through QHA, V(T) and G(T). Phases with imaginary harmonic modes need effective phonons at T (TDEP, SSCHA, MD).
Next: Lecture 32 follows the energy surface between two minima, over the saddle point, to find the barrier for an atom to hop.
DFT & VASP Course · Lecture 3133 / 33
Key terms
- Interatomic force constants
- Second derivatives of the energy with respect to the displacements of two atoms, ΦIα,Jβ=∂2E/∂uIα∂uJβ=−∂FIα/∂uJβ, in eV/Ų. Pushing atom J by u creates the force −ΦIα,Jβu on atom I.
- Dynamical matrix
- The mass-weighted Fourier sum of the force constants, Diα,jβ(q)=∑lΦ0iα,ljβeiq⋅Rl/MiMj, a 3n × 3n Hermitian matrix for each phonon wavevector q. Its eigenvalues are the squared frequencies ω2, its eigenvectors the mode patterns.
- Phonon
- A quantized normal mode of vibration of a crystal: a lattice wave with wavevector q in the Brillouin zone, frequency ων(q) and energy quanta ℏω.
- Dynamical stability
- A structure is dynamically stable when every phonon at every wavevector has ω2>0 (apart from the three acoustic modes at Γ, rigid translations with ω2=0): the energy rises for any small displacement of the atoms. A negative ω2 means a saddle point.
- Potential-energy surface
- The energy E({RI}) as a function of the positions of all nuclei. Its minima are stable or metastable structures; its first-order saddle points are the transition states, whose energies set the barriers between them.
- Local minimum
- A point where the energy rises in every direction (apart from moving all atoms together), although a lower minimum may exist elsewhere. A relaxation usually ends in a local minimum, not necessarily the global one.
- Harmonic approximation
- Keeping the Taylor expansion of the energy in the atomic displacements only up to second order. The motion then separates into independent normal modes with fixed frequencies.
- Anharmonicity
- The third- and higher-order terms of the energy in the atomic displacements. They make phonons interact and change with temperature, cause thermal expansion at fixed pressure and thermal resistance, and can stabilize structures that are harmonically unstable.
- Acoustic sum rule
- Translation invariance: moving all atoms by the same vector creates no force, so ∑JΦIα,Jβ=0. It makes the three acoustic frequencies vanish at Γ. Numerical errors violate it slightly; codes symmetrize the force constants to restore it.
- Normal mode
- A collective motion in which all atoms vibrate at one frequency with fixed relative amplitudes: an eigenvector of the dynamical matrix. Small motions are superpositions of independent normal modes.
- Imaginary mode
- A mode with ω2<0, so that ω is imaginary: a displacement along it lowers the energy and grows instead of oscillating. VASP labels it f/i; plots show it as a negative frequency. It is either a real instability or a numerical artefact.
- Phonon branch
- One of the 3n frequency curves ων(q) of a crystal with n atoms per cell. Three are acoustic (ω→0 at Γ, sound waves); 3n − 3 are optical (atoms in a cell move against each other).
- Finite-displacement method
- Computing force constants from the forces in supercells in which one atom is displaced by a small ±δ (frozen phonons): Φ≈−[F(+δ)−F(−δ)]/2δ. Used by VASP (IBRION = 6) and phonopy.
- Supercell
- A cell made of several primitive cells, used so that a local perturbation (a displaced atom, a defect) is far from its periodic images. A supercell computes phonons exactly only at wavevectors that fold to its Γ point.
- Density-functional perturbation theory (DFPT)
- Linear-response DFT: the first-order change of the Kohn–Sham orbitals under a perturbation (an atomic displacement, an electric field, a strain) is found from the self-consistent Sternheimer equation, giving force constants, Born charges and dielectric tensors without finite displacements.
- Born effective charge
- The tensor Zi,γα∗: the macroscopic polarization along γ created by moving atom i along α (in units of e per cell volume), equivalently the force on atom i per unit electric field. It differs from the formal ionic charge; it vanishes by symmetry in Si and diamond.
- LO–TO splitting
- In polar insulators, longitudinal optical modes near Γ create a macroscopic electric field that raises their frequency above that of the transverse modes. It is added to supercell phonons with the Born charges and ε∞ (the non-analytic correction).
- Soft mode
- A phonon whose frequency is low and falls toward zero, or becomes imaginary, as a structure approaches an instability. Its eigenvector and wavevector describe the distortion that the structure undergoes.
- Vibrational free energy
- Fvib(T)=∑qν[21ℏω+kBTln(1−e−ℏω/kBT)], the harmonic free energy of all phonons. It contains the zero-point energy and the vibrational entropy.
- Zero-point energy
- ∑21ℏω, the vibrational energy a quantum crystal or molecule keeps at 0 K. Typically tens of meV per atom, more for light atoms (about 0.22 eV for one O–H stretch).
- Quasi-harmonic approximation (QHA)
- Harmonic phonons computed at several volumes, so that F(V,T)=E(V)+Fvib(V,T) can be minimized over V at each temperature. Gives thermal expansion and G(T); needs real frequencies at every volume.
- Grüneisen parameter
- γ=−∂lnω/∂lnV, how much a mode softens when the crystal expands; typically 1–3. The volumetric thermal expansion is αV=γCV/(BTV).
- Self-consistent phonons
- Methods (SSCHA, SCAILD, SCPH) that find temperature-dependent effective force constants by sampling atomic displacements with the thermal amplitudes of trial phonons, computing forces, and refitting until input and output agree. Can include zero-point motion.
- TDEP
- Temperature-dependent effective potential: effective harmonic (and higher-order) force constants fitted by least squares to the forces and displacements of a molecular-dynamics run at temperature T.
- Thermodynamic integration
- Computing a free-energy difference by switching the energy gradually from a reference to the target, F=Fref+∫01⟨E−Eref⟩λdλ, with a simulation at each λ. Gives the full anharmonic free energy.
VASP tags in this lecture
EDIFFG- Stopping criterion of an ionic relaxation. A positive value stops when the total energy changes by less than EDIFFG between ionic steps; a negative value stops when the norms of all forces are below |EDIFFG|. It is not used in molecular dynamics.Default: EDIFF × 10 · VASP Wiki: EDIFFG
LPHON_DISPERSION- Phonon dispersion along the QPOINTS path, for the primitive cell, from the supercell force constants of the same run (6.3.2 and later).Default: .FALSE. · VASP Wiki: LPHON_DISPERSION
IBRION- How the nuclei are moved between ionic steps: −1 not at all, 0 molecular dynamics, 1 relaxation with RMM-DIIS, 2 relaxation with conjugate gradients, 3 relaxation with damped molecular dynamics.Default: −1 if NSW = −1 or 0, otherwise 0 (molecular dynamics) · VASP Wiki: IBRION
NSW- Maximum number of ionic steps. With NSW = 0 the nuclei are not moved: a single-point (static) calculation.Default: 0 · VASP Wiki: NSW
NFREE- IBRION = 1: how many previous steps the quasi-Newton optimizer remembers; IBRION = 5, 6: displacements per direction.Default: 2 for IBRION = 5 and 6, otherwise 0 · VASP Wiki: NFREE
POTIM- In ionic relaxations (IBRION = 1, 2, 3), a scaling constant for the step widths; in molecular dynamics, the time step in fs.Default: 0.5 for IBRION = 1, 2 and 3; 0.015 Å for IBRION = 5 and 6; none for molecular dynamics (must be set) · VASP Wiki: POTIM
EDIFF- Stopping criterion of the electronic self-consistency loop, in eV. The loop stops when both the change of the free energy and the change of the band-structure energy between two steps fall below EDIFF.Default: 10−4 eV · VASP Wiki: EDIFF
PREC- Precision mode: sets the default FFT grids (and ENCUT, if you do not).Default: Normal (Medium in VASP 4) · VASP Wiki: PREC
ENCUT- Cutoff energy of the plane-wave basis set, in eV. The VASP Wiki strongly recommends setting it explicitly in every INCAR, and checking the convergence of the quantity you report.Default: the largest ENMAX among the POTCAR files · VASP Wiki: ENCUT
ISMEAR- How partial occupancies are assigned to the electronic states. ISMEAR = 0 is Gaussian smearing; ISMEAR = N > 0 is Methfessel–Paxton smearing of order N, which can give erroneous results for insulators.Default: 1 · VASP Wiki: ISMEAR
ISIF- Which degrees of freedom may change and whether the stress tensor is calculated. ISIF = 2: the atoms move, the cell shape and volume stay fixed, and the stress is calculated; ISIF = 3: atoms, cell shape and volume all change.Default: 0 for molecular dynamics (IBRION = 0) or hybrid functionals (LHFCALC = .TRUE.), otherwise 2 · VASP Wiki: ISIF
ADDGRID- An extra, finer grid for the augmentation charges.Default: .FALSE. · VASP Wiki: ADDGRID
LEPSILON- Static dielectric tensor, Born effective charges and piezoelectric tensor by density-functional perturbation theory (LDA and GGA only).Default: .FALSE. · VASP Wiki: LEPSILON
LCALCEPS- The same quantities as LEPSILON from finite electric fields; works for hybrids and meta-GGA, gapped systems only.Default: .FALSE. · VASP Wiki: LCALCEPS
LPHON_POLAR- Include the long-range dipole–dipole (non-analytic) correction in VASP's phonon dispersion of a polar insulator; needs PHON_DIELECTRIC (ion-clamped dielectric tensor) and PHON_BORN_CHARGES from a LEPSILON or LCALCEPS run (6.3.2 and later).Default: .FALSE. · VASP Wiki: LPHON_POLAR
VELOCITY- Write the velocities to vaspout.h5 (VASP 6.4 and later).Default: .FALSE. · VASP Wiki: VELOCITY
ML_LMLFF- Switch on VASP's machine-learned force field (VASP 6.3.0 and later).Default: .FALSE. · VASP Wiki: ML_LMLFF
MDALGO- Thermostat: 1 Andersen, 2 Nosé–Hoover, 3 Langevin, 4 Nosé–Hoover chain (6.4.0), 5 CSVR (6.4.3).Default: 0 (legacy Nosé) · VASP Wiki: MDALGO
LANGEVIN_GAMMA- Langevin friction per species in ps⁻¹ (MDALGO = 3).Default: 0 for each species · VASP Wiki: LANGEVIN_GAMMA
ISYM- How symmetry is used.Default: 2 for PAW; 3 with hybrids (LHFCALC); 1 with ultrasoft potentials · VASP Wiki: ISYM
Further reading
- A. Togo and I. Tanaka, “First principles phonon calculations in materials science”, Scr. Mater. 108, 1–5 (2015). · doi:10.1016/j.scriptamat.2015.07.021 · arXiv:1506.08498
- S. Baroni, S. de Gironcoli, A. Dal Corso and P. Giannozzi, “Phonons and related crystal properties from density-functional perturbation theory”, Rev. Mod. Phys. 73, 515–562 (2001). · doi:10.1103/RevModPhys.73.515 · arXiv:cond-mat/0012092
- O. Hellman, P. Steneteg, I. A. Abrikosov and S. I. Simak, “Temperature dependent effective potential method for accurate free energy calculations of solids”, Phys. Rev. B 87, 104111 (2013). · doi:10.1103/PhysRevB.87.104111 · arXiv:1303.1145
- L. Monacelli, R. Bianco, M. Cherubini, M. Calandra, I. Errea and F. Mauri, “The stochastic self-consistent harmonic approximation: calculating vibrational properties of materials with full quantum and anharmonic effects”, J. Phys.: Condens. Matter 33, 363001 (2021). · doi:10.1088/1361-648X/ac066b · arXiv:2103.03973
- VASP Wiki, “Phonons from finite differences”. · www.vasp.at
- VASP Wiki, “How to handle imaginary phonon modes”. · www.vasp.at