Quantum Chemistry

Density Functional Theory (DFT)

Hohenberg-Kohn 1964 — ground-state energy is a unique functional of electron density ρ(r); B3LYP, PBE workhorses

Density functional theory (DFT) maps the N-electron Schrödinger problem to a problem of the 3-coordinate electron density ρ(r). Hohenberg and Kohn proved in 1964 that the ground-state energy is a unique functional of ρ. Kohn and Sham then introduced auxiliary one-electron orbitals (1965) that reproduce the exact density via a fictitious non-interacting reference. Modern functionals — LDA, PBE (GGA), B3LYP (hybrid), M06-2X, ωB97X — power roughly 80% of computational chemistry papers. A. D. Becke published B3 in 1988 and the B3LYP hybrid was assembled around 1993; PBE arrived in 1996. Walter Kohn and John Pople shared the 1998 Nobel Prize in Chemistry.

  • FoundationHohenberg-Kohn 1964
  • Practical equationsKohn-Sham 1965
  • Hybrid scalingO(N4); pure DFT O(N3)
  • Most-cited functionalB3LYP (~50% of papers)
  • Solid-state defaultPBE (Perdew-Burke-Ernzerhof 1996)
  • NobelKohn & Pople 1998

Interactive visualization

Press play, or step through manually. The visualization is yours to drive — try it before reading on.

Open visualization fullscreen ↗

Watch the 60-second explainer

A condensed visual walkthrough — narrated, captioned, under a minute.

Why DFT matters

  • ~80% of computational chemistry papers use DFT. The dominant electronic-structure method since the late 1990s. B3LYP alone appears in roughly half of computational papers; PBE dominates solid-state work; M06-2X and ωB97X-D are competitive for non-covalent interactions.
  • Cost scaling makes 1000-atom systems tractable. Pure GGA functionals scale as O(N3) where N is basis size. Hybrid functionals scale as O(N4) because of exact exchange. Compare CCSD(T) at O(N7) — capped at ~30 atoms — and DFT at > 1000 atoms is routine.
  • Hohenberg-Kohn 1964: a half-page proof that changed everything. Pierre Hohenberg and Walter Kohn published two theorems showing the ground-state energy is determined entirely by the electron density. The proofs are short and rigorous; in principle no information is lost relative to wavefunction methods.
  • Kohn-Sham 1965 made it practical. Walter Kohn and Lu Sham introduced a fictitious system of non-interacting electrons that produces the same density as the real one. The resulting Kohn-Sham equations are one-electron Schrödinger-like equations solvable by SCF iteration — the same machinery as Hartree-Fock.
  • Materials Genome Initiative runs on DFT. The 2011 US initiative cataloged the electronic structure of essentially every known stoichiometric crystal in databases like Materials Project, OQMD, and AFLOW — tens of thousands of compounds, all computed with PBE and a fixed pseudopotential library.
  • Drug discovery uses DFT for binding-pose energetics. Mechanism-of-action studies for kinase inhibitors, oxidative metabolism prediction (cytochrome P450 site selectivity), and conformational energy ranking all rely on DFT — usually B3LYP or ωB97X-D — for thousands of poses.
  • The Walter Kohn / John Pople 1998 Nobel. Kohn for DFT itself; Pople for the development of computational methods including the Gaussian program, which made these methods routinely usable. They split the prize evenly.

Common misconceptions

  • DFT is exact in principle. The Hohenberg-Kohn theorem is exact in principle. But the universal functional F[ρ] is not known explicitly, so all practical DFT uses approximate exchange-correlation functionals. The functional is the source of essentially all DFT errors.
  • Kohn-Sham orbitals are real. No. They are the orbitals of a fictitious non-interacting system constructed to give the same density as the real interacting system. KS orbital energies are not strictly ionization energies (only the HOMO energy maps to −IP for an exact functional, by Janak's theorem), and KS orbital shapes only approximately represent the real wavefunction.
  • DFT is variational without caveats. The energy E[ρ] is variational with respect to the exact functional. Approximate functionals can give energies below the true ground state, especially for stretched bonds and transition states.
  • Hybrid functionals are always better. Hybrids improve thermochemistry of small organic molecules. They are often worse for transition-metal complexes, where the Hartree-Fock exchange admixture overlocalizes d-electrons. PBE and SCAN often outperform B3LYP for crystals and metals.
  • DFT handles dispersion forces. Pure local and semilocal functionals miss London dispersion entirely. Grimme's -D3 (2010) and -D4 (2019) corrections add an empirical pairwise dispersion term −C6/r6 to the energy. Without it B3LYP underbinds the benzene dimer by ~10 kJ/mol.
  • Better functional always means better results. Functional rankings depend on the property and the system. M06-2X is excellent for non-covalent interactions; PBE excels for solids; SCAN is balanced; B3LYP is a standard reference for organics. There is no universal best functional.

From wavefunctions to densities and back

The wavefunction ψ(r1,...,rN) for an N-electron system depends on 3N spatial coordinates. Even storing it on a grid for benzene (N = 42) is impossible — 100042 points exceeds the number of atoms in the observable universe. The density ρ(r) = N ∫ |ψ|2dr2...drN depends on just 3 coordinates regardless of N. Hohenberg and Kohn's first theorem says the ground-state ρ uniquely determines the external potential Vext(r) up to a constant, which determines the Hamiltonian, which determines ψ and every other property. Their second theorem says the true ground-state ρ minimizes a universal functional E[ρ] = ∫Vext(r)ρ(r)dr + F[ρ]. So in principle solving for ρ (3D problem) gives all ground-state information without ever computing ψ (3N-D problem).

The Kohn-Sham construction makes this practical. Introduce a fictitious system of N non-interacting electrons with a Slater-determinant wavefunction built from one-electron orbitals φi. The density is ρ(r) = Σ|φi(r)|2. Choose the effective potential veff(r) such that this fictitious density equals the real interacting density. The total energy decomposes into E = Ts[ρ] + ∫Vext(r)ρ(r)dr + J[ρ] + Exc[ρ], where Ts is the non-interacting kinetic energy, J is the classical Hartree (Coulomb) repulsion, and Exc is the exchange-correlation functional that absorbs all the residual quantum effects. The Kohn-Sham one-electron equations resemble Hartree-Fock; they're solved by SCF iteration with cost O(N3) for the dominant matrix operations.

The catch is that the exact Exc[ρ] is unknown. All practical DFT approximates it. Local Density Approximation (LDA) takes Exc[ρ] = ∫εxcHEG(ρ(r))ρ(r)dr from the homogeneous electron gas energy density. Generalized Gradient Approximation (GGA) adds dependence on ∇ρ; PBE (Perdew-Burke-Ernzerhof, 1996) is the standard GGA. Hybrid functionals mix a fraction of exact (Hartree-Fock) exchange; B3LYP uses 20% HF exchange combined with Becke's B88 exchange (1988) and the LYP correlation functional. Range-separated hybrids (ωB97X) use long-range HF and short-range DFT; double-hybrids add a fraction of MP2 correlation. Each rung adds cost and improves typical errors but introduces functional-specific failure modes.

DFT functionals — Jacob's ladder, with cost and accuracy

RungFunctional classExamplesCost scalingTypical thermochemistry errorBest use cases
1LDASVWN, PW92O(N3)~70 kJ/molSolids, metals (overbinding)
2GGAPBE, BLYP, PW91O(N3)~30 kJ/molSolid-state, materials, fast screening
3meta-GGATPSS, M06-L, SCANO(N3)~20 kJ/molVersatile; SCAN balances solids and molecules
4Hybrid (global)B3LYP, PBE0O(N4)~15-20 kJ/molOrganic thermochemistry, ground-state
4'Range-separated hybridωB97X-D, CAM-B3LYP, LC-BLYPO(N4)~10-15 kJ/molCharge-transfer states, large π-systems
5Double-hybridB2PLYP, DSD-PBEP86O(N5)~5-10 kJ/molHigh-accuracy benchmark; expensive

HF vs MP2 vs CCSD(T) vs DFT — when to use which

MethodCost scalingTractable system sizeAccuracy on bond energiesStrengthWeakness
Hartree-FockO(N4)~500 atoms50-100 kJ/mol errorReference for post-HF; reproducibleNo correlation, no dispersion
MP2O(N5)~100 atoms~30 kJ/mol errorCheap correlation; good for noncovalentDiverges for stretched bonds
CCSDO(N6)~50 atoms~10 kJ/mol errorReliable for closed-shellMisses triple excitations
CCSD(T)O(N7)~30 atoms~4 kJ/mol error"Gold standard" benchmarkCostly; multireference fails
DFT (PBE)O(N3)> 1000 atoms~30 kJ/mol errorMaterials/solids; fastUnderestimates band gaps
DFT (B3LYP)O(N4)~500 atoms~15-30 kJ/mol errorOrganic chemistry workhorseMisses dispersion; -D3 patch

Applications and examples

  • ~80% of computational chemistry papers use DFT. Surveys of leading journals (J. Chem. Phys., JCTC, J. Phys. Chem.) consistently show B3LYP, PBE, ωB97X-D, and M06-2X dominating the methods sections. CCSD(T) appears mainly as a benchmark or for reactions where accurate barriers are essential.
  • Materials Project and OQMD. Open databases of DFT-computed properties for ~140,000 inorganic crystals (PBE-level, GGA pseudopotentials). They power high-throughput materials screening for batteries, photocatalysts, thermoelectrics, and topological materials.
  • Catalyst design. Heterogeneous catalysis on metal surfaces — N2 dissociation on Fe (Haber), CO2 reduction on Cu, oxygen reduction on Pt — is computed with PBE plus van der Waals corrections. Activation barriers within 20 kJ/mol of experiment guide alloy and facet choice.
  • Drug discovery. Cytochrome-P450 metabolism site prediction, kinase inhibitor binding poses, organic photocatalyst screening (Doyle, Knowles labs) all use B3LYP or M06-2X with -D3. Throughput of 103-105 compounds per study is routine.
  • The 1998 Nobel Prize in Chemistry. Walter Kohn for DFT itself (Hohenberg-Kohn 1964, Kohn-Sham 1965); John Pople for "the development of computational methods in quantum chemistry," principally the Gaussian program. They split the prize evenly. Becke's 1988 functional and the 1993 B3LYP publication had been transformational by then.

Frequently asked questions

What does density functional theory replace and why is it cheaper?

Wavefunction methods solve for ψ(r1,r2,...,rN), a function of 3N spatial coordinates plus spins. The dimensionality scales exponentially with electron count, which is why CCSD(T) scales as N7 and is only practical below ~30 atoms. DFT replaces ψ with the electron density ρ(r), a function of just 3 coordinates regardless of N. The Hohenberg-Kohn theorem (1964) proved that the ground-state energy is a unique functional of ρ, so in principle nothing is lost. In practice the exchange-correlation functional is unknown and must be approximated — but the cost scaling is now O(N3) for pure functionals, O(N4) for hybrids.

What did Hohenberg and Kohn prove in 1964?

Two theorems. First, the external potential Vext(r) is uniquely determined (up to a constant) by the ground-state electron density ρ(r). Since Vext + N (number of electrons obtained by integrating ρ) determines the Hamiltonian, the entire many-electron problem is determined by ρ. Second, there exists a universal functional F[ρ] such that the ground-state energy is E[ρ] = ∫Vext(r)ρ(r)dr + F[ρ], and the true ground-state density minimizes E over all admissible densities. The proofs are about half a page each. The catch: F[ρ] is unknown, so the practical theory consists of approximations — LDA, GGA, hybrid, etc.

What is the Kohn-Sham trick?

Kohn and Sham (1965) introduced a fictitious system of non-interacting electrons that reproduces the same density as the real interacting system. They split F[ρ] into the kinetic energy of the non-interacting reference Ts[ρ], the classical Coulomb (Hartree) energy J[ρ], and the exchange-correlation functional Exc[ρ] which absorbs everything difficult — the difference between the true kinetic energy and Ts, plus the non-classical electron-electron interaction. The Kohn-Sham equations look like Hartree-Fock equations: one-electron Schrödinger-like equations with an effective potential veff(r) = Vext + VH + Vxc. They are solved iteratively just like SCF, but the orbitals are auxiliary — only the resulting density is physical.

Why is B3LYP so widely used?

B3LYP combines Becke's 1988 exchange (B88), the LYP correlation functional from Lee, Yang, and Parr (1988), and an exact-exchange admixture proposed by Becke (1993). It hits a sweet spot: typical errors of 15-30 kJ/mol for organic thermochemistry, runs at hybrid cost (~O(N4)), and has been validated against thousands of benchmarks since the mid-1990s. Roughly half of all computational chemistry papers in the early 2000s used B3LYP. It has known weaknesses — poor for transition metals, no dispersion (so we add a -D3 correction), problematic for excited states — but as a default it remains a standard reference. PBE0 and ωB97X-D have grown as alternatives but neither has displaced B3LYP from the textbook.

What is the difference between LDA, GGA, hybrid, and double-hybrid functionals?

Jacob's ladder of functionals (Perdew). LDA — Local Density Approximation — uses only ρ(r) at each point; based on the homogeneous electron gas. GGA — Generalized Gradient Approximation — adds the gradient ∇ρ(r); examples PBE, BLYP, PW91. Meta-GGA adds the kinetic-energy density τ or the Laplacian of ρ; examples TPSS, M06-L, SCAN. Hybrid functionals mix in a fraction of exact (Hartree-Fock) exchange; examples B3LYP (20% HF), PBE0 (25%), ωB97X (range-separated HF). Double-hybrids add a fraction of MP2 correlation on top; examples B2PLYP, DSD-PBEP86. Each rung adds cost but also accuracy; the tradeoff and the failure modes are functional-specific.

What are DFT's known failure modes?

Five recurring problems. First, dispersion (London) forces are missing in pure local/semilocal functionals; Grimme's -D3 or -D4 correction is the standard patch. Second, self-interaction error — an electron sees itself in the Coulomb integral; this overstabilizes delocalized states and is severe for charge-transfer excitations. Third, transition-metal multireference problems — DFT struggles when more than one Slater determinant is needed (open-shell singlets, bond-breaking). Fourth, band-gap underestimation in solids — semilocal functionals give gaps about 50% too small. Fifth, no systematic improvement — unlike CCSD(T) you cannot just add more terms; you must change functional. Range-separated hybrids and double-hybrids partially address several of these but at increased cost.