Computational Chemistry Basics — What DFT and Molecular Modelling Actually Do
A geometry-optimised structure with bond lengths quoted to three decimal places looks authoritative, and that is exactly why computational chemistry needs to be understood rather than trusted blindly. Every number a calculation prints out rests on a stack of deliberate approximations — some essentially exact, some genuinely uncertain — and knowing which is which is the difference between using these methods well and being misled by them. This article explains, honestly and without overclaiming, what a geometry optimisation and a DFT calculation are actually doing.
The approximation that makes the whole enterprise tractable
Solving the full electronic Schrödinger equation for a molecule exactly is not possible for anything beyond the very smallest systems. The first and most important simplification is the Born-Oppenheimer approximation: because a nucleus is roughly 1,836 times heavier than an electron (even for hydrogen, the lightest), nuclei move far more slowly than electrons. The approximation treats the nuclei as effectively stationary while solving for the electronic structure, then moves the nuclei and re-solves — separating nuclear motion from electronic motion instead of solving both simultaneously. Almost every method described below sits on top of this one approximation.
The potential energy surface — what "optimising a geometry" means
With nuclei fixed at a given set of positions, solving for the electronic energy gives one number. Sweep through every possible set of nuclear positions and that number traces out a potential energy surface (PES) — energy as a function of geometry. A geometry optimisation is a search across this surface for a point where the energy gradient (the first derivative with respect to every nuclear coordinate) is zero: a stationary point, which may be a true minimum (a stable structure) or a saddle point (a transition state).
Worked example 1 — following the gradient downhill for water. A geometry optimisation on H₂O is started from an arbitrarily bent, incorrect guess geometry. Explain what the calculation does at each step, and what result to expect.
At the starting geometry the energy gradient is non-zero, pointing in the direction that lowers the energy fastest. The optimisation algorithm takes a step along (roughly) that direction, recomputes the electronic energy and the new gradient, and repeats — following the surface downhill, step by step, until the gradient becomes essentially zero everywhere. For water, this process converges to a structure close to the experimentally known geometry: an H–O–H bond angle near 104.5° and an O–H bond length near 0.96 Å.
Crucially, a single-point energy calculation — one electronic-energy evaluation at a fixed, unoptimised geometry, with no gradient-following at all — reports a number for whatever geometry it was given, including a badly guessed one. Quoting a single-point energy from a non-equilibrium structure as if it described the stable molecule is one of the most common beginner mistakes in the field.
Two families of method — and why DFT reformulates the whole problem
Wavefunction-based methods start from Hartree-Fock (HF) theory, in which each electron is treated as moving in the average field of all the others (a mean-field approximation). This misses "electron correlation" — the instantaneous, correlated avoidance real electrons show toward each other — and post-HF methods (MP2, CCSD(T) among others) add correlation back in systematically, at rapidly increasing computational cost as the treatment gets more complete.
Density functional theory (DFT) takes a different route entirely. Instead of solving for the many-electron wavefunction — an object that depends on 3N coordinates for N electrons and becomes extraordinarily expensive to handle as N grows — DFT reformulates the problem in terms of the electron density ρ(r), a much simpler function of just three spatial coordinates, however many electrons are present. This rests on the Hohenberg-Kohn theorems, which prove that the ground-state energy is, in principle, a unique functional of the electron density alone. The practical implementation, the Kohn-Sham approach, introduces a fictitious system of non-interacting electrons constructed to reproduce the same density as the real interacting system, splitting the total energy into kinetic energy, the classical Coulomb (Hartree) repulsion, nuclear-electron attraction, and one final term — the exchange-correlation functional — that packages together every remaining piece of quantum many-body complexity.
In practice, every real DFT calculation uses an APPROXIMATE exchange-correlation functional (LDA, GGA functionals such as PBE, or hybrid functionals such as B3LYP) — and different functionals can give meaningfully different answers for the same system.
This is the single most important honest caveat about DFT: it is not a black box that produces "the" correct answer. It is a formally exact framework running on a necessarily approximate ingredient, and functional choice is a genuine, consequential decision the calculation's user must make and justify.
Worked example 2 — why two functionals disagree on a hydrogen-bonded dimer. Two calculations on the same water dimer are run with two different functionals: one plain GGA functional with no dispersion correction, and the same functional with an added empirical dispersion correction term. Explain why their computed binding energies are expected to differ, and which is more reliable.
Standard GGA exchange-correlation functionals are built almost entirely from local electron-density information (the density and its gradient at each point). Dispersion (van der Waals) attraction, however, is fundamentally a non-local, long-range correlation effect between electrons on separate, weakly interacting fragments — exactly the kind of interaction a purely local functional systematically under-describes. Adding an empirical dispersion correction restores much of this missing attraction directly, generally bringing the computed binding energy closer to high-level reference values (from methods like CCSD(T)). The dispersion-corrected result is the more reliable one for this specific, genuinely well-documented failure mode of standard functionals.
Basis sets — the other approximation hiding inside every calculation
In practice, molecular orbitals (or, in Kohn-Sham DFT, the auxiliary orbitals) are expressed as a linear combination of a finite set of mathematical basis functions centred on each atom — typically Gaussian-type functions in most quantum chemistry software. A minimal basis set (like STO-3G) supplies too few, too inflexible functions to describe how an atom's electron density genuinely polarises and distorts once that atom is embedded in a real molecule — toward a bond, toward a lone pair. Larger, split-valence and polarised basis sets (6-31G*, triple-zeta sets such as def2-TZVP) supply more flexibility and generally converge toward the theoretical complete-basis-set limit as they grow, at correspondingly higher computational cost.
Worked example 3 — reasoning about basis-set adequacy. A dipole moment computed with a minimal basis set differs noticeably from the same calculation repeated with a larger, polarised basis set. Which is more trustworthy, and why?
A dipole moment is highly sensitive to how well the calculation can describe the polarisation of electron density away from each atom's isolated, spherical shape — exactly the flexibility a minimal basis set lacks by construction (it typically has no polarisation functions at all). The larger, polarised basis set is expected to be more trustworthy for this property specifically, because it can represent the actual distorted electron distribution inside the molecule far more realistically. This is a general pattern, not unique to dipole moments: properties that depend sensitively on how electron density redistributes are the properties most sensitive to basis-set quality.
What computational chemistry is genuinely good at, and where it struggles
Modern DFT is reliably useful for equilibrium geometries, vibrational frequencies, relative energies between similar systems, and mapping out reaction mechanisms and transition states, with due care. It genuinely struggles with dispersion interactions in standard, non-dispersion-corrected functionals; with strongly correlated or multi-reference systems (some transition-metal complexes, diradicals), where a single-determinant Kohn-Sham description can fail qualitatively; and with absolute reaction barrier heights, which routinely disagree by several kilocalories per mole between different functionals applied to the same reaction — a gap that matters, since through the Arrhenius relationship even a few kcal/mol of error in a barrier height can shift a predicted rate by orders of magnitude.
Errors that appear most often
- Treating a computed number as more objective than experiment simply because it has more decimal places. Every result is only as good as the method, functional and basis set chosen, and disagreement with experiment must be diagnosed, not dismissed.
- Assuming a bigger basis set or a "fancier" functional is automatically better for every property. The right method genuinely depends on the property and the system's electronic character; there is no single universally best choice.
- Confusing "the geometry converged" with "the geometry is a minimum". A converged stationary point could be a saddle point instead. Confirming a true minimum needs a separate vibrational-frequency (Hessian) calculation, where all frequencies must come out real; an imaginary frequency signals a saddle point, not a stable structure.
- Thinking DFT computes the many-electron wavefunction directly. It formally works with the electron density via the Kohn-Sham auxiliary orbitals — a related but conceptually distinct object, and a subtlety worth getting right at this level.
Where this appears in postgraduate chemistry
| Context | Typical demand |
|---|---|
| CSIR-NET / GATE physical chemistry | Conceptual questions on Hartree-Fock, DFT, basis sets and the Born-Oppenheimer approximation, generally at the level of "what does the method do" rather than a full derivation |
| Postgraduate research | This is where the topic is actually used day to day — screening reaction mechanisms, rationalising spectroscopic and structural data, and guiding synthetic target selection |
Before reaching for a computational method, make sure the underlying quantum-chemical fundamentals — orbitals, quantum numbers, electron configurations — are solid.
Open the Quantum Numbers Calculator →Preparing for CSIR-NET, GATE, IIT-JAM or CUET-PG? ABC Chemistry runs dedicated competitive-exam batches online and fully online for students across India — details at abcchemistry.in.