Compendium

Computational Chemistry Deep Dive

21 min read

Computational Chemistry Deep Dive

A Tier 2 deep-dive into the methods, software, and workflows of modern computational chemistry — wavefunction theory (HF through CCSD(T), MRCI, CASSCF/CASPT2/NEVPT2), density-functional theory (LDA, GGA, hybrid, range-separated, double-hybrid, meta-GGA, dispersion corrections), basis sets and pseudopotentials, implicit and explicit solvation, excited-state methods, wavefunction analysis (NBO, AIM, ELF, NCI), the major production software (Gaussian, ORCA, NWChem, Psi4, Q-Chem, GAMESS, MOLPRO, MRCC, Turbomole, ADF, Dalton, CFOUR, CP2K, Quantum ESPRESSO, VASP, CASTEP, ABINIT, GPAW, PySCF, OpenMolcas), classical and reactive molecular dynamics (OPLS-AA, AMBER, CHARMM, MMFF, GAFF, ReaxFF), machine-learning interatomic potentials (ANI, AIMNet2, MACE, Allegro, NequIP), QM/MM, enhanced sampling, and Python workflow tooling (cclib, ASE, MolModa, pwtools). Sub-disciplines glue together; choosing the right level of theory and software for a question matters as much as running the calculation.

See also

The hierarchy of electronic-structure methods

Two complementary trajectories climb the accuracy ladder. Wavefunction theory (WFT) — Hartree-Fock + correlation corrections — is systematically improvable but expensive (CCSD(T) scales O(N⁷), full CI factorially). Density-functional theory (DFT) reframes the problem as a functional of electron density — affordable (O(N³)-O(N⁴) typical) but with built-in approximation in the exchange-correlation functional that is not systematically reducible.

Hartree-Fock (HF, mean-field)

Single-determinant wavefunction; each electron sees average field of others. Scales O(N⁴) formally, O(N²·⁵-N³) with screening. Missing: dynamic correlation (Coulomb hole). Captures static-field bonding well; covalent bond lengths typically too short by ~0.01-0.02 Å, dipole moments too large.

Restricted HF (RHF, closed shells), unrestricted HF (UHF, radicals — but spin contamination), restricted open-shell HF (ROHF, no contamination but no Brillouin in cross-shell).

Møller-Plesset perturbation (MP2, MP3, MP4)

Treat correlation as perturbation. MP2 scales O(N⁵); recovers ~80-90% of valence correlation. Workhorse for ground-state thermochemistry on medium molecules; weakness with multireference systems (transition metals, diradicals, bond-breaking).

SCS-MP2 (spin-component scaled, Grimme 2003) and SOS-MP2 (Head-Gordon scaled-opposite-spin) reweight singlet/triplet pair correlation; small accuracy bumps. dRPA (direct random-phase approximation) and dRPA+SOSEX competitive for dispersion.

Coupled cluster (CC)

Exponential ansatz e^T |HF⟩ — preserves size consistency and includes infinite-order corrections.

  • CCSD — singles + doubles. O(N⁶). Good for benchmarking medium systems.
  • CCSD(T) — perturbative triples added. O(N⁷). “Gold standard” for single-reference closed-shell organic thermochemistry; sub-kcal/mol against experiment in atomization energies, reaction enthalpies, barrier heights, ionization potentials, electron affinities.
  • CCSDT, CCSDTQ — full triples, quadruples. Hugely expensive (O(N⁸), O(N¹⁰)); only for small molecules / W4-style composite benchmarks.

Variants: ROCCSD(T), UCCSD(T) for open shells; LR-CCSD, EOM-CCSD for excited states. Domain-based local pair natural orbital (DLPNO) CCSD(T) — Neese-Riplinger 2009 onward — reduces CCSD(T) to O(N) effective for large molecules; ORCA implementation routine for systems up to 200+ atoms. PNO-CCSD(T), LNO-CCSD(T) (Kállay, MRCC) competitors.

Multireference methods

Single-determinant ansatz fails for: bond-breaking, diradicals, transition-metal d-d states, near-degenerate electronic configurations, conical intersections, excited states with multiple character.

  • CASSCF (complete active space self-consistent field). Choose “active space” (n electrons in m orbitals — CAS(n,m)); full CI in that space, SCF orbital optimization outside. Scales factorially in active space — practical limit CAS(16,16). Mature implementations: OpenMolcas, ORCA, Gaussian, MOLPRO, GAMESS.
  • RASSCF (restricted active space). Partition active space into RAS1/RAS2/RAS3 to enlarge effective active space.
  • DMRG (density matrix renormalization group). Block, BAGEL, CheMPS2 implementations. Enables active spaces up to CAS(50,50) tractable. Useful for polyenes, mixed-valence systems, FeMoco.
  • MRCI (multireference CI). CAS reference + external CI. Davidson size-extensivity correction MRCI+Q.
  • CASPT2 / NEVPT2. Second-order perturbation theory on CAS reference. CASPT2 (Roos-Andersson) needs IPEA shift and imaginary-shift level-shifting against intruder states; NEVPT2 (Angeli-Cimiraglia-Malrieu) intruder-free by construction — increasingly preferred. CASPT2 still standard for transition-metal excited states (Roos, Lindh, Pierloot, González).
  • MRCC (multireference coupled cluster). Mk-MRCC, Brillouin-Wigner; specialized package MRCC (Kállay).
  • FCIQMC, AS-FCIQMC. Stochastic full-CI on active space (Alavi-Thom). NECI code.

Composite thermochemistry

Combine cheap large-basis with expensive small-basis to approximate large-basis CCSD(T)/CBS.

  • Gaussian-n (G2, G3, G4, G4(MP2)). Pople-Curtiss schemes.
  • CBS-Q, CBS-QB3, CBS-APNO. Petersson Wesleyan.
  • Wn theories (W1, W2, W3, W4). Martin-Karton; sub-kJ/mol on small molecules.
  • HEAT, ATcT. “High-accuracy extrapolation of thermochemistry.”
  • DLPNO-CCSD(T)/CBS. Practical large-molecule pseudo-benchmark.

Density-functional theory (DFT)

Hohenberg-Kohn (1964) and Kohn-Sham (1965) put the electron density (3D) at center stage. Kohn-Sham DFT reintroduces orbitals via a fictitious non-interacting system → solve self-consistent one-electron equations + add exchange-correlation potential v_xc(r).

Functional accuracy traces “Jacob’s Ladder” (Perdew):

Rung 1 — LDA

Local density approximation; v_xc depends only on ρ(r). VWN, PW92 parameterizations. Severe overbinding (~30 kcal/mol on small molecules); modern use restricted to solids and certain materials (cohesive energies of simple metals OK).

Rung 2 — GGA

Generalized gradient approximation; v_xc(ρ, ∇ρ). PBE (Perdew-Burke-Ernzerhof 1996) — universal standard in solid-state DFT, plane-wave codes. BLYP (Becke 1988 + Lee-Yang-Parr) — popular early molecular GGA. PW91 — predecessor to PBE. RPBE — improved chemisorption energies (Hammer-Nørskov 1999). revPBE, PBEsol, AM05.

Rung 3 — Meta-GGA

Add kinetic-energy density τ(r). TPSS (Tao-Perdew-Staroverov-Scuseria 2003), revTPSS, M06-L (Truhlar), SCAN (Sun-Ruzsinszky-Perdew 2015 — strongly constrained and appropriately normed). r²SCAN (Furness-Sun 2020) — numerically stable SCAN. Meta-GGAs increasingly common in materials science as best non-hybrid option.

Rung 4 — Hybrid (exact-exchange admixture)

Mix a fraction of HF exact exchange.

  • B3LYP (Becke 1993) — 20% HF, dominant in organic chemistry. Workhorse but flawed for π-stacking, large molecules (delocalization error), transition metals.
  • PBE0 (PBE 1996/Adamo-Barone 1999) — 25% HF; more “honest” than B3LYP; common in inorganic and materials.
  • M06-2X (Truhlar 2008) — 54% HF; excellent for non-covalent, kinetics, thermochemistry of main-group; weak for transition metals (use M06 — 27% HF — for TM).
  • B3PW91, B97-D, HSE06 (range-separated screened-exchange for solids; Heyd-Scuseria-Ernzerhof 2003).

Rung 5 — Range-separated and double-hybrid

  • CAM-B3LYP (Yanai-Tew-Handy 2004) — long-range corrected B3LYP; charge-transfer excited states.
  • ωB97X-D (Chai-Head-Gordon 2008) — Becke 97 family with range-separation + empirical dispersion. Excellent general-purpose hybrid for medium molecules; routinely used in pharma CADD.
  • ωB97M-V (Mardirossian-Head-Gordon 2016) — meta-hybrid range-separated with VV10 nonlocal correlation. Among top general-purpose functionals on GMTKN benchmarks.
  • LRC-ωPBE, LRC-ωPBEh — Rohrdanz-Herbert.
  • Double hybrids. Add MP2-like correlation: B2PLYP (Grimme 2006), DSD-PBEP86 (Kozuch-Martin 2011), revDSD-PBEP86-D4 (Santra-Sylvetsky-Martin 2019). O(N⁵) but better than hybrids for thermochemistry and kinetics.

Dispersion corrections

Standard DFT (through hybrid) misses long-range dispersion (no R⁻⁶ asymptote in exchange-correlation hole). Add empirical or non-local correction:

  • D3 (Grimme 2010). C6/C8 atom-pair coefficients with damping. D3(BJ) — Becke-Johnson damping, most common. Per-functional fit parameters.
  • D4 (Caldeweyher-Bannwarth-Grimme 2017, 2019). Charge-dependent dispersion; better for polar systems.
  • MBD (many-body dispersion; Tkatchenko-Scheffler). Beyond-pairwise correction; default in MaterialsStudio CASTEP.
  • VV10, rVV10 (Vydrov-Van Voorhis 2010). Non-local functional; included in ωB97M-V, B97M-rV.
  • vdW-DF, vdW-DF2 (Langreth-Lundqvist). Non-local; common in plane-wave solids.

Time-dependent DFT (TDDFT)

Standard tool for excited states; vertical excitation energies. LR-TDDFT (linear response) for absorption spectra. Tamm-Dancoff approximation (TDA) speeds + stabilizes triplets. Limitations:

  • Charge-transfer (CT) states underestimated by global hybrids — use range-separated (CAM-B3LYP, ωB97X-D) or LR-corrected.
  • Double-excitations missed entirely (single-particle response only).
  • Conical intersections — TDDFT topologically wrong near degeneracy with ground state (use spin-flip TDDFT — Krylov; or MR methods).
  • Rydberg states — diffuse basis required.

Alternatives: CIS (cheap; qualitative), CIS(D), CC2 (Christiansen-Koch-Jørgensen iterative O(N⁵); good general-purpose excited-state method), ADC(2) (algebraic-diagrammatic construction; equivalent to CC2 quality; Schirmer-Dreuw), ADC(3), EOM-CCSD (Koch-Jørgensen-Bartlett; benchmark quality but O(N⁶)), SAC-CI (Nakatsuji).

Basis sets

Discretize molecular orbitals as LCAO. Choice trades accuracy vs cost.

Pople basis sets

  • STO-3G. Minimal; only didactic / very crude.
  • 3-21G, 6-31G. Split-valence; small.
  • 6-31G(d), 6-31G(d,p), 6-311G(d,p), 6-311+G(2d,p), 6-311++G(2d,p). Polarization (d on heavy / p on H) + diffuse (+ on heavy / ++ also on H) extensions. Decent for organics.

Dunning correlation-consistent

cc-pVnZ (n = D, T, Q, 5, 6): each shell adds correlating polarization functions in a balanced manner. aug-cc-pVnZ adds diffuse. cc-pVTZ is the “default” small-but-trustworthy basis for organics; aug-cc-pVTZ for charge-transfer/anion/Rydberg work; cc-pVQZ/cc-pV5Z for CBS extrapolation. Weak point: more $/CPU than Karlsruhe def2 family for hybrid DFT.

Karlsruhe (Ahlrichs)

def2-SVP (split-valence + 1 polarization), def2-TZVP (triple-zeta), def2-TZVPP (TZ + 2 polarization), def2-QZVP, def2-QZVPP. def2-TZVP is the modern workhorse for DFT — balanced, well-tested, ECPs for heavy elements built-in (Stuttgart-Köln-Dresden ECP series).

Plane-wave basis (for solids)

Used in VASP, Quantum ESPRESSO, CP2K, CASTEP, ABINIT, GPAW. Plane-wave cutoff energy E_cut (typically 400-600 eV for PAW potentials). PAW (projector-augmented wave; Blöchl 1994) or pseudopotentials (Vanderbilt USPP, Hartwigsen-Goedecker-Hutter HGH, Trouillier-Martins, Optimized Norm-Conserving Vanderbilt ONCV). All-electron alternative: FLAPW (WIEN2k, Elk).

Mixed Gaussian + plane-wave (GPW)

CP2K Quickstep — Gaussians for orbitals + plane-waves for density. Excellent for liquids, biomolecules in periodic boundary conditions.

Effective core potentials (ECPs)

Replace core electrons with potential. Stuttgart-Köln (def2 series), LANL2DZ/LANL2TZ (Hay-Wadt), CRENBL, SBKJC. Essential for heavy elements (relativistic; computational savings). Scalar relativistic ECPs default. Spin-orbit coupling adds explicit treatment.

Relativistic all-electron

  • DKH2/DKH3 Douglas-Kroll-Hess scalar relativistic Hamiltonian.
  • ZORA zeroth-order regular approximation — used in ADF, ORCA.
  • X2C exact two-component — modern preferred for 5d, actinides; available in ORCA, Psi4, MOLPRO, Turbomole, OpenMolcas.

Solvation models

Implicit solvation replaces explicit solvent with dielectric continuum; ~95% of ground-state computational thermochemistry on solvated species uses one of these.

  • PCM (polarizable continuum model). Miertus-Scrocco-Tomasi 1981. Cavity from atom-centered spheres; electrostatic interaction iteratively. Gaussian default.
  • CPCM (conductor-like PCM). Barone-Cossi 1998. Simpler than PCM; widely available.
  • COSMO (Klamt-Schüürmann 1993). Foundation of ADF/Turbomole solvation. COSMO-RS (real solvents — beyond dielectric) extends to phase equilibria.
  • SMD (Marenich-Cramer-Truhlar 2009). Adds non-electrostatic terms (cavity, dispersion, structural). Best-calibrated for solvation free energies (~0.5 kcal/mol vs experimental).
  • SS(V)PE, IEFPCM — integral equation formulations.

For solvent-mediated transition states, explicit + implicit (microsolvation) often needed: 1-5 explicit solvent molecules + bulk PCM.

Explicit solvation

QM/MM (below) or full classical MD with TIP3P, TIP4P/2005, SPC/E water and OPLS/AMBER organic solvents. Free-energy methods (thermodynamic integration, FEP, replica-exchange) recover absolute solvation energies.

Wavefunction analysis

Numbers from the calculation; interpretation tools convert them to chemical narrative.

Population analysis

  • Mulliken — basis-set dependent, often unreliable but ubiquitous default.
  • Löwdin — symmetric orthogonalization.
  • Natural Population Analysis (NPA; Weinhold). Within NBO framework; more basis-set robust.
  • CM5, Hirshfeld, CHELPG, RESP — ESP-fitted partial charges (RESP standard for force-field parametrization in AMBER/GAFF).
  • DDEC6 (Manz) — modern atom-in-molecule charges for solids and molecules.

Bonding analysis

  • NBO (Natural Bond Orbital; Weinhold). Lewis-like localization → bond orbitals + lone pairs + Rydberg. NBO 7 current; integrated into Gaussian, ORCA, ADF, NWChem, Q-Chem. Donor-acceptor (E²) perturbation analysis quantifies hyperconjugation, anomeric, n→σ* effects.
  • AIM (Bader Atoms-in-Molecules). Topological analysis of ρ(r) — bond critical points (BCPs), ring (RCPs), cage (CCPs). Quantitative bond orders from ρ(BCP), Laplacian ∇²ρ, ellipticity. AIMAll, Multiwfn.
  • ELF (Electron Localization Function; Becke-Edgecombe). Domains of electron pair localization — bonding pairs, lone pairs visualizable.
  • EDA (Energy Decomposition Analysis). Decompose interaction energy into electrostatic, exchange-repulsion (Pauli), orbital, dispersion. ETS-NOCV (Mitoraj-Michalak-Ziegler), ALMO-EDA (Head-Gordon), LMO-EDA (Su-Li), SAPT (Symmetry-Adapted Perturbation Theory; Jeziorski-Moszynski-Szalewicz — distinguishes electrostatic / induction / dispersion / exchange).
  • NCI (Non-Covalent Interaction plots). Yang-Contreras-García 2010 — visualize weak interactions in reduced density gradient space (RDG = s(ρ) vs sign(λ₂)ρ). NCIPLOT software.
  • LOL (Localized Orbital Locator). Like ELF, simpler.

Multiwfn (Tian Lu, BNU)

Free wavefunction analysis suite. NBO, AIM, NCI, ELF, electrostatic potential, Fukui functions, conceptual DFT descriptors, ESI, charge decomposition, orbital composition. Workflow standard.

Frontier molecular orbital and conceptual DFT

HOMO/LUMO energies + gap → reactivity proxies (Pearson hardness η = (LUMO-HOMO)/2, electronegativity χ, electrophilicity ω = χ²/2η — Parr 1999). Fukui functions f(r) = (∂ρ/∂N) for nucleophilic/electrophilic susceptibility.

Production software — the landscape

Gaussian (Frisch et al., Gaussian Inc.)

Most-cited quantum chemistry code (>100,000 citations). Current: Gaussian 16 (Revision C.01-C.02), Gaussian 09 still common in legacy. Strengths: huge functional library, NBO, IRC, Freq + thermochemistry, polished GUI (GaussView). Weak in: multireference (no DMRG, modest CASSCF), no DLPNO-CCSD(T) (use ORCA), restrictive licensing (per-CPU, no academic group-share).

ORCA (Frank Neese, MPI Mülheim → MPI für Kohlenforschung)

Free for academic use; current ORCA 6 (2024 release; 5.x widely deployed). Strengths: DLPNO-CCSD(T) (best implementation), excellent multireference (CASSCF, NEVPT2, MRCI, DMRG via Block interface), TDDFT, EPR/NMR property modules (ORCA was originally an EPR code), excited-state methods, broad functional support including modern (ωB97M-V, r²SCAN), DKH/ZORA/X2C, COSMO/CPCM/SMD. Workhorse for inorganic, spectroscopy, large-system CCSD(T).

NWChem (PNNL)

Open-source; massively parallel scaling. Strengths: HPC-friendly, plane-wave + Gaussian hybrid (NWPW), QM/MM, TCE for high-order CC (CCSDT, CCSDTQ on huge clusters). Weak in: GUI; learning curve.

Psi4 (Sherrill, Crawford — multi-university; psicode.org)

Open-source Python-driven. Strengths: SAPT (canonical), DLPNO-CCSD(T) emerging, deep functional library, very clean APIs for method development. Mature but smaller user base than Gaussian/ORCA.

Q-Chem (Krylov, Head-Gordon, multi-PI commercial)

Strengths: ALMO-EDA, EOM-CC family (Krylov), spin-flip methods, ωB97X-D family (Head-Gordon), TDDFT excited-state forces.

GAMESS (Iowa State, Gordon group)

Free; venerable (1980s heritage). Effective fragment potential (EFP), MCSCF, MP2 gradients, RHF/UHF/ROHF. Less actively developed for modern functionals.

MOLPRO (Werner, Knowles)

Commercial; gold standard for high-accuracy benchmark CCSD(T), MRCI, CASPT2 — internally-contracted. Often best CCSD(T) numerics; used for W4 benchmarks.

MRCC (Mihály Kállay, BME Budapest)

Free for academic. Specializes in arbitrary-order CC (CCSDT, CCSDTQ, CCSDTQP). LNO-CCSD(T) — competitor to DLPNO.

Turbomole (Karlsruhe-based; commercial)

Strengths: RI-MP2 with tight scaling, COSMO heritage, fast DFT, easy-to-use ridft. Used in industry pharma R&D.

ADF/AMS (SCM Amsterdam)

Slater-type orbital (STO) basis — natural for relativistic; ZORA, X2C heritage. AMS modular suite includes ADF (molecular DFT), BAND (periodic DFT), DFTB (tight-binding), MD, ReaxFF, COSMO-RS. Strong on heavy-element + spectroscopy.

Dalton (Århus, Helgaker)

Free for academic. Strengths: response theory (linear, quadratic), magnetic properties, NMR shieldings, vibrational corrections.

CFOUR (Stanton-Gauss-Watts)

Free; high-accuracy CC + analytic gradients + relativistic. Used for spectroscopy benchmarks.

OpenMolcas (formerly MOLCAS; Roos heritage)

Free, open. Best CASSCF/CASPT2/RASPT2 implementation; DMRG via interface. Standard for photochemistry, transition-metal multireference, lanthanide/actinide. Lund-Stockholm-Vienna-Uppsala consortium.

Plane-wave codes (solids, surfaces, periodic)

  • VASP (Vienna ab initio Simulation Package; Hafner-Kresse-Furthmüller). Commercial; dominant in solid-state DFT. PAW potentials. GGA + meta-GGA + hybrid (HSE06 standard), GW, BSE, DFT+U, AIMD. Cross-link electronic-structure-and-computational-materials.
  • Quantum ESPRESSO (Giannozzi et al.). Open-source. PWSCF, CP, PHonon, EPW, TDDFT, NEB, DFT+U+V. Workhorse academic solid-state.
  • CASTEP (Cambridge Serial Total Energy Package; Payne, Probert, Refson, Pickard). Commercial; Materials Studio interface. UK academic-friendly.
  • ABINIT (Belgian-led collaboration). Open-source. GW, BSE, DFPT, anharmonicity (TDEP).
  • GPAW (Mortensen-Hammer-Hansen). Open-source. Real-space grid + PAW; LCAO mode; Atomic Simulation Environment (ASE) integration.
  • CP2K (Hutter-VandeVondele). Open-source. GPW basis; AIMD on huge systems; QM/MM; meta-GGA + hybrid via auxiliary density matrix method (ADMM).
  • WIEN2k (Blaha-Schwarz). Commercial; full-potential linearized augmented plane wave (FLAPW) — all-electron benchmark for solids.
  • Elk, Exciting — open FLAPW.
  • FHI-aims (Blum, FHI Berlin). Numerical atomic orbitals; molecules + solids; many-body methods.

Python ecosystems

  • PySCF (Sun et al.). Pure Python + C kernel. HF, DFT, MP2, CCSD(T), CASSCF, DMRG (via Block), AIMD, embedding. Hackable framework for method development; growing fast.
  • Psi4 (with Python API via psi4 module).
  • Pyscf-forge, Pyscf-shciscf plugins.
  • ASE (Atomic Simulation Environment; Dulak-Hammer-Mortensen). Python toolkit for atomistic simulations; drivers for VASP, Quantum ESPRESSO, ORCA, NWChem, GPAW, CP2K, LAMMPS, GAMESS, NEB, vibrations, surface science workflows.

Visualization

  • GaussView (Gaussian Inc.). Standard GUI for Gaussian.
  • Avogadro (open-source; cross-platform).
  • Chemcraft (commercial; Russian developer).
  • VESTA — crystals + densities.
  • VMD (Theoretical and Computational Biophysics Group, UIUC) — MD trajectories + biomolecules.
  • PyMOL — proteins; structural biology heritage.
  • Multiwfn — embeds visualization (cube files).
  • IQmol — Q-Chem-affiliated.

Molecular dynamics — force fields

Atomistic MD treats nuclei classically with empirical potential energy U({r_i}).

Biomolecular force fields

  • OPLS (Jorgensen Yale, 1980s onward). OPLS-AA, OPLS-3, OPLS4 (Schrödinger commercial). Refined for organic/drug-like molecules in water (TIP4P).
  • AMBER (Kollman heritage; Case, Cornell, Walker, Cheatham). ff99SB → ff14SB → ff19SB → ff19SB-OPC for proteins; bsc1 for DNA; OL15 for RNA; GAFF2 (general AMBER force field) for small molecules.
  • CHARMM (Karplus-Brooks-MacKerell Maryland). CHARMM36 protein, CHARMM36m updated protein, CHARMM CGenFF small molecule. Strong in lipid membranes (CHARMM36 lipids).
  • GROMOS (Zürich-Karplus; van Gunsteren). United-atom heritage; 53A6, 54A8.
  • MMFF (Halgren, Merck). MMFF94, MMFF94s; ubiquitous in cheminformatics (RDKit, Schrödinger).

General-purpose / drug-like

  • GAFF, GAFF2. General AMBER force field; auto-parameterized by Antechamber + AM1-BCC charges.
  • CGenFF. CHARMM general; auto-paramertized via CGenFF server.
  • OPLS4. Schrödinger commercial; best for drug-receptor binding in pharma CADD.
  • SMIRNOFF / Open Force Field Initiative. Chodera-Mobley-Wang community-driven SMILES-keyed FF; OpenFF 2.x.

Water models

TIP3P (AMBER default), TIP4P, TIP4P/2005, TIP4P-Ew, TIP5P, SPC, SPC/E, OPC (3-site optimal-point-charge; Izadi-Anandakrishnan-Onufriev), OPC3. Each tuned to different observables; TIP4P/2005 most accurate for bulk water properties; OPC best for protein simulations recently.

Reactive force fields

  • ReaxFF (van Duin-Goddard 2001). Bond-order-based; allows bond breaking/forming. Parameterized for combustion, catalysis surfaces (Pt-H, Cu-O), polymers, oxides. LAMMPS, ADF AMS implementations.
  • AIREBO, REBO — Brenner reactive bond-order; carbon-hydrogen surfaces.
  • CHARMM-FF EVB (empirical valence bond; Warshel) — multi-state reactive; transferable to enzymatic reactions in QM/MM context.

Coarse-grained

MARTINI (Marrink Groningen) — 4-heavy-atom-per-bead mapping; CG lipids, proteins, sugars. MARTINI 3 (2021) refit. Captures lipid bilayer phenomena, membrane proteins, vesicle fusion.

UNRES, CABS (protein-only CG), DPD (dissipative particle dynamics — broader soft matter).

MD engines

  • GROMACS (Lindahl, KTH). Free; ultra-fast on CPU and GPU; biomolecular standard.
  • AMBER pmemd / pmemd.cuda (Walker, San Diego SBGrid). GPU-optimized; commercial license but free for academics.
  • NAMD (UIUC; Schulten heritage). Free; large-system parallel scaling.
  • LAMMPS (Plimpton, Sandia). Free; materials-MD; supports ReaxFF, ML potentials, DPD, granular.
  • OpenMM (Pande heritage; Eastman, Chodera). Free; Python API; GPU; foundation of many ML/QM-MM workflows; common backend for OpenFF and modern free-energy codes.
  • Desmond (D.E. Shaw / Schrödinger). Commercial; original Anton-platform code; ultra-long-trajectory analysis.

Machine-learning interatomic potentials (MLIPs)

Goal: DFT (or higher) accuracy at force-field speed. Train neural network on reference QM data; deploy as classical-MD potential.

Generation 1 — descriptor-based

  • Behler-Parrinello NN (2007). Symmetry function descriptors + per-atom NN.
  • GAP (Gaussian approximation potential; Bartók-Csányi 2010). SOAP descriptors + Gaussian process. Excellent uncertainty quantification.
  • SchNet (Schütt-Müller 2017). Continuous-filter convolutional NN on atomic embeddings.
  • ANI (Smith-Isayev-Roitberg). ANI-1 → ANI-1x → ANI-2x → ANI-1ccx; covers H, C, N, O, S, F, Cl; trained on ω B97X/6-31G*; near-CCSD(T) at force-field speed. Open source torchani.

Generation 2 — equivariant / message-passing

  • NequIP (Smidt-Batzner-Musaelian 2022). E(3)-equivariant; ~100× more data-efficient than non-equivariant.
  • Allegro (Musaelian-Batzner-Smidt 2023). Local equivariant; massive parallel scale.
  • MACE (Batatia-Kovács-Csányi-Ortner 2022). Higher-order equivariant tensor messages. State-of-the-art accuracy/speed on most materials benchmarks; MACE-MP-0 (Batatia 2023) — universal foundation potential covering periodic table.
  • AIMNet, AIMNet2 (Zubatyuk-Roitberg-Isayev 2019, 2024). Atom-in-molecule NN with explicit charge equilibration; CCSD(T) accuracy on drug-like molecules.
  • PaiNN, PainNN+ — polarizable atom interaction neural network.

Universal / foundation MLIPs

  • MACE-MP-0 (Batatia et al. 2023). Trained on MPTraj (Materials Project); covers periodic table.
  • CHGNet (Deng-Zhong-Schmidt-Persson 2023). Materials Project-trained; magnetic moments included.
  • M3GNet (Chen-Ong UCSD 2022).
  • GNoME (Google DeepMind 2023) — vast training set, used for crystal-structure stability prediction at scale.
  • Orb (Orbital Materials 2024) — commercial universal MLIP.

Workflow

  1. Generate reference data with DFT or CCSD(T) (active learning loops typical).
  2. Train model on energies + forces (+ stresses for solids).
  3. Validate on holdout + physical test (RDF, vibrational frequencies, melting point).
  4. Deploy in LAMMPS, ASE, OpenMM via integration interfaces.

Production examples: ANI-2x for medchem conformer scanning; MACE-MP-0 for catalyst screening; NequIP for catalysis surfaces; AIMNet2 for protein-ligand sampling.

QM/MM and embedding

Hybrid scheme: QM region (active site) treated quantum mechanically; MM region (solvent, scaffold) treated classically. Boundary handled by link-atom or boundary-orbital schemes (Singh-Kollman, Maseras-Morokuma IMOMM, Truhlar).

QM/MM implementations

  • ChemShell (Sherwood-Brooks-Catlow). Pure QM/MM driver; couples Turbomole/ORCA/Gaussian/NWChem with DL_POLY/GULP.
  • CHARMM-GAMESS, CHARMM-Q-Chem — native interfaces.
  • NWChem QM/MM. Native.
  • AMBER + Q-Chem/Gaussian/Terachem.
  • GROMACS + CP2K / GROMACS + ORCA (recent integrations).
  • CP2K QM/MM — native; popular for enzyme AIMD.

Subsystem DFT, embedded MF

Frozen-density embedding (FDE; Wesolowski-Warshel 1993). DMET (Knizia-Chan 2012) — density matrix embedding theory; for strongly correlated active spaces inside larger lattice/molecule.

Multiscale enzyme catalysis

QM/MM is the de-facto method for enzyme reaction mechanisms. Notable: Warshel (Nobel 2013), Karplus (Nobel 2013), Levitt (Nobel 2013). Recent work increasingly couples QM/MM with enhanced sampling (metadynamics, US) to compute free-energy barriers, not just enthalpic.

Enhanced sampling

Vanilla MD samples Boltzmann ensemble; rare events (barrier crossings, conformational transitions, binding/unbinding) need biasing.

  • Umbrella sampling (Torrie-Valleau 1977). Harmonic restraints along reaction coordinate; PMF via WHAM (weighted histogram analysis method; Kumar-Bouzida-Swendsen).
  • Metadynamics (Laio-Parrinello 2002). Deposit Gaussian hills on collective variables (CVs); flatten free-energy surface. Well-tempered metadynamics (Barducci-Bussi-Parrinello 2008) — convergent. PLUMED library implements metadynamics, ABF, US, FES reweighting; integrates with GROMACS, NAMD, LAMMPS, OpenMM, Q-Chem.
  • Adaptive Biasing Force (ABF; Darve-Pohorille 2001). Estimate mean force and bias against it; flat FES on CV.
  • Replica-exchange MD (REMD; Sugita-Okamoto 1999). Multiple replicas at different T (T-REMD) or with different Hamiltonians (H-REMD); swap periodically.
  • Accelerated MD (aMD; Hamelberg-Mongan-McCammon 2004). Boost low-energy regions to escape barriers.
  • Steered MD (SMD). Pull with constant force/velocity along reaction coordinate.
  • String method (E-Vanden-Eijnden). Find minimum free-energy path between two states.
  • Forward-flux sampling (Allen-Frenkel). For very rare events (nucleation, protein folding).
  • WESTPA, OpenPathSampling — weighted-ensemble libraries.

CV design: distance, angle, dihedral, RMSD, coordination number, contact map, principal components from tICA/MSM, deep-learned CVs (Tiwary, Bonati-Parrinello DeepLDA, DeepMD-kit collective).

Workflows and orchestration

cclib

Python parser for output files of 20+ codes (Gaussian, ORCA, NWChem, Psi4, Q-Chem, GAMESS, MOLPRO, Turbomole, ADF). Single API to extract energies, geometries, frequencies, atomic charges, MO coefficients. Hutchison-O’Boyle Pittsburgh.

ASE — Atomic Simulation Environment

Python; calculator interface to all major QM and MD codes. Drives NEB (nudged-elastic band transition-state search), vibrations, EOS, surface generation, slab construction. De-facto standard for materials science Python workflows; integrates with GPAW, VASP, Quantum ESPRESSO, ORCA, LAMMPS, MACE, NequIP, OpenMM.

MolModa

Web-based platform (Durrant lab Pitt) for cross-code molecular modeling workflows.

pwtools

Python toolkit for plane-wave electronic-structure post-processing (Schmerler).

AiiDA, ASR, fireworks, jobflow, atomate2

Workflow engines for high-throughput materials/chemistry computing. Persist provenance, automate restart, cluster scheduling. Materials Project uses atomate2/jobflow for ~150,000 DFT calculations indexed.

Materials Cloud, Materials Project, NOMAD, OQMD

Public repositories of DFT-calculated materials data. Materials Project (LBNL; Persson) ~150k structures + properties. OQMD (Wolverton, Northwestern) ~1M. NOMAD (Berlin) ~12M with raw I/O. AFLOW (Curtarolo Duke). JARVIS (NIST).

Selected validation sets and benchmarks

  • GMTKN55 (Goerigk-Hansen-Bauer-Ehlert-Najibi-Grimme 2017). 1,505 benchmark reactions across 55 subsets. Defining benchmark for general-purpose DFT functional ranking.
  • W4-11, W4-17 (Karton-Martin). Sub-kJ/mol atomization energies on small molecules.
  • S22, S66, S66x8 (Hobza). Non-covalent interaction benchmarks.
  • ACCDB (Domínguez-Soria-Castaña-Truhlar 2017). Accuracy compilation.
  • BH9, BH76, BH76RC — barrier heights.
  • MOR41, MOBH35 — metal-organic reactions.
  • 3BC, NCIBLIND10 — non-covalent.
  • WCCR10 — transition-metal complexes.
  • HEAVY28 — heavy-element thermochemistry.
  • QMugs, PubChemQC (>3M-molecule DFT databases for ML training).

Pharma and industry workflows

Lead optimization computational stack

  1. 2D filtering. Lipinski, PAINS, REOS, Veber filters.
  2. 3D conformer generation. OMEGA (OpenEye), RDKit ETKDG, BCL conformer.
  3. Docking. Glide (Schrödinger), GOLD, AutoDock Vina, ICM, OEDocking.
  4. Free-energy perturbation (FEP+). Schrödinger commercial — relative binding free energy ~1 kcal/mol against experiment for congeneric series. PMX, FEP-toolkit, perses (open-source alternatives).
  5. QM site refinement. DFT (B3LYP-D3, ωB97X-D) or DLPNO-CCSD(T) on representative cluster around binding site to validate poses, predict pKa shifts.
  6. MD stability. Desmond, AMBER, OpenMM; check trajectory stability of bound poses.
  7. ADMET prediction. Mostly ML now (SwissADME, ADMET-AI, in-house QSAR models).

Catalysis / reaction discovery

  1. Mechanism hypothesis. Generate from chemical intuition or via automated reaction-graph tools (CRN — Chemical Reaction Network; Reaction Mechanism Generator RMG-Py for combustion).
  2. Intermediate optimization. DFT (ωB97X-D, M06, B3LYP-D3 with def2-TZVP).
  3. Transition state location. Berny QST2/QST3 (Gaussian), eigenvector-following (ORCA), NEB (ASE-VASP/QE).
  4. IRC verification. Confirm TS connects intended minima.
  5. Higher-level energy. DLPNO-CCSD(T)/CBS or canonical CCSD(T) for benchmark; single-point at DFT geometry.
  6. Solvation + thermal corrections. PCM/SMD + Gibbs free energy at T.
  7. Microkinetic modeling. Build rate-law model from elementary steps; CatMAP (Stanford), Cantera, AMSpy.

Open challenges and frontier topics

  • Multireference at scale. DMRG, FCIQMC, AS-FCIQMC, ASCI extend active spaces but still bottlenecks for FeMoco, photosystem II, plutonium speciation.
  • Excited states beyond TDDFT. ADC(3), EOM-CCSDT, multireference RIXS. Conical intersection accuracy.
  • Quantum embedding. DMET, projection-based embedding (Manby-Miller), bootstrap embedding bridging single-reference DFT with multireference active region.
  • Quantum computing for chemistry. VQE (variational quantum eigensolver), QPE (quantum phase estimation) — sub-quantum-supremacy demonstrations on small systems (H₂, LiH); error correction barriers remain. Cross-link quantum-computing-engineering.
  • ML-driven exchange-correlation functional development. DM21 (DeepMind 2021) — neural-network XC functional matched best hybrid on common benchmarks.
  • Reactive ML potentials at chemical accuracy. Ongoing — TS energetics under 1 kcal/mol still hard.
  • Coupling MLIPs with enhanced sampling for free-energy surfaces. DeepMD-kit, MACE + PLUMED; active research.

Adjacent

Further reading

  • Szabo, A., Ostlund, N.S. — Modern Quantum Chemistry: Introduction to Advanced Electronic Structure Theory. Dover 1996 — the canonical wavefunction-theory textbook (HF through MP / CI / CC).
  • Helgaker, T., Jørgensen, P., Olsen, J. — Molecular Electronic Structure Theory. Wiley 2000 — definitive monograph on WFT.
  • Parr, R.G., Yang, W. — Density-Functional Theory of Atoms and Molecules. Oxford 1989 — foundational DFT.
  • Jensen, F. — Introduction to Computational Chemistry, 3rd ed. Wiley 2017 — broad pedagogical overview.
  • Cramer, C.J. — Essentials of Computational Chemistry: Theories and Models, 2nd ed. Wiley 2004 — practical and concise.
  • Frenkel, D., Smit, B. — Understanding Molecular Simulation. Academic 2002 — MD foundations.
  • Allen, M.P., Tildesley, D.J. — Computer Simulation of Liquids, 2nd ed. Oxford 2017 — classical reference.
  • Crabtree, R.H. — The Organometallic Chemistry of the Transition Metals, 7th ed. Wiley 2019 — for TM electronic structure and ligand-field interpretation context.
  • Anslyn, E.V., Dougherty, D.A. — Modern Physical Organic Chemistry. University Science Books 2006 — MO/computational context for organic mechanism.
  • Smith, M.B. — March’s Advanced Organic Chemistry, 8th ed. Wiley 2020 — reactions whose mechanisms benchmark computational methods.
  • Frenking, G., Shaik, S., eds. — The Chemical Bond: Fundamental Aspects of Chemical Bonding + The Chemical Bond: Chemical Bonding Across the Periodic Table. Wiley-VCH 2014 — modern bonding-analysis perspectives (NBO, EDA, AIM, ELF).

Graph View

Backlinks