ABINIT: DFT & DFPT Simulation Platform

• ABINIT is a scientific software package that implements density functional theory and density-functional perturbation theory to simulate solids and nanomaterials.
• It incorporates versatile methodologies such as plane-wave, PAW, and wavelet-based representations to deliver precise total energy and vibrational property calculations.
• Extensive code verification, interoperability with external tools, and GPU-accelerated performance enhancements make ABINIT a key tool in reproducible high-throughput research.

Searching arXiv for recent and foundational ABINIT-related papers to ground the article. ABINIT is a scientific software package for first-principles simulations of solids and nanomaterials, centered on density functional theory and extended to response functions, vibrational properties, electron–phonon coupling, many-body perturbation theory, constrained states, and high-performance execution on heterogeneous architectures. Across the literature considered here, ABINIT appears both as a production code for ground-state and linear-response calculations and as a platform for methodological development, including density-functional perturbation theory, projector augmented-wave formalisms, wavelet-based representations, virtual-crystal pseudopotential mixing, self-consistent linear-response Hubbard parameters, and GPU-accelerated plane-wave solvers. Verification studies further position it as an implementation whose numerical behavior can be cross-compared against independent codes for quantities ranging from total energies to zero-point band-gap renormalization [(Poncé et al., 2013); (Poncé et al., 2024); (Verstraete et al., 11 Jul 2025)].

1. Definition and computational scope

ABINIT is described as a “widely used scientific software package implementing density functional theory and many related functionalities for excited states and response properties” (Verstraete et al., 11 Jul 2025). In the works surveyed here, it is used in plane-wave and DFPT settings, in PAW workflows, in a wavelet-based PAW library, in virtual-crystal calculations, in self-consistent linear-response determinations of Hubbard parameters, and in workflows interfacing with BerkeleyGW, EPW, Yambo, Quantum ESPRESSO, BigDFT, and atomistic Green’s-function transport formalisms (Rangel et al., 2016, Li et al., 2023, MacEnulty et al., 2024).

Its core ground-state formulation in the plane-wave pseudopotential setting is illustrated by the Kohn–Sham total energy functional used in a verification study on diamond (Poncé et al., 2013):

$$E_{\mathrm{tot}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{Hxc}}[\rho] + E_{\mathrm{Ewald}} + E_{\mathrm{psp-core}}.$$

The one-electron contribution is written as

$$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$

ABINIT’s methodological span also includes DFPT phonons, electron–phonon matrix elements, Allen–Heine–Cardona self-energies, Wannier-function perturbation theory, GW, DMFT, coupled-cluster interfaces, constrained charge and spin states, flexoelectricity, polarons, and workflow automation (Poncé et al., 2024, Verstraete et al., 11 Jul 2025). A plausible implication is that the code is not defined by a single basis or single physical approximation, but by a modular ecosystem of first-principles engines and post-processing workflows.

2. Ground-state frameworks and basis representations

In the literature provided, ABINIT supports multiple ground-state representations. The most extensively documented is the plane-wave formulation with norm-conserving pseudopotentials and LDA or GGA exchange–correlation functionals. For diamond, one verification study used LDA in the Perdew–Zunger parameterization, a norm-conserving Troullier–Martins pseudopotential generated with fhi98PP, a plane-wave cutoff of $30$ Hartree, a $6\times 6\times 6$ $\Gamma$-centered Monkhorst–Pack grid, and a relaxed lattice parameter of $6.652$ Bohr (Poncé et al., 2013). Under these settings ABINIT yielded $E_{\mathrm{tot}}=-11.444116277$ Ha per cell, or $-5.7220581385$ Ha/atom, with a discrepancy relative to Quantum ESPRESSO below $1\times 10^{-5}$ Ha/atom (Poncé et al., 2013).

The PAW formalism is also explicitly implemented. In one study of transition-metal-doped AlN, the all-electron Kohn–Sham wavefunction $|\psi_n\rangle$ is reconstructed from a pseudo-wave $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$0 through the transformation operator

$$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$1

so that

$$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$2

The real-space density is decomposed as

$$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$3

with on-site terms built from the projector-overlap matrix $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$4 (Syrotyuk et al., 2013). That work used ABINIT in a PAW–PBE0(+HF) setting for Cr-, Mn-, and Fe-doped AlN and reported spin-resolved gaps and ferromagnetic ground states (Syrotyuk et al., 2013).

A distinct development is the wavelet-based PAW method integrated into ABINIT via a basis-independent Fortran 2003 library, “pawlib,” shared with BigDFT (Rangel et al., 2016). In that implementation, the all-electron wavefunction is obtained from a smooth pseudo-wavefunction through the PAW transformation

$$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$5

or, explicitly,

$$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$6

The basis employs orthogonal Daubechies-16 scaling functions and wavelets on coarse and fine grids, with multiresolution adaptivity near atoms (Rangel et al., 2016). The paper states that with basis = 'bigdft', icoulomb = 1 or 2, nlevel = 2, hcoarse ≃ 0.4 bohr, and hfine ≃ 0.2 bohr, the code can simulate charged and special-boundary-condition systems “with frozen-core all-electron precision” (Rangel et al., 2016).

ABINIT 2025 further reports support for meta-GGA functionals including SCAN, rSCAN, and $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$7SCAN through LibXC, quasi-Fermi-level occupations for photodoping with occopt = 9, high-temperature DFT with an extended-plane-wave model via useextfpmd, and constrained charge or spin states via constraint_kind, ratsph, chrgat, and spinat (Verstraete et al., 11 Jul 2025). These are specific named capabilities in the 2025 overview, though the provided details combine summary and survey language.

3. Linear response, phonons, and force constants

ABINIT’s DFPT implementation is central to several studies. A general Sternheimer form is stated in the 2025 overview as

$$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$8

with $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$9 projecting onto the conduction subspace (Verstraete et al., 11 Jul 2025). This underlies the calculation of phonons, dielectric tensors, Born effective charges, elastic constants, and more.

In the diamond verification work, phonons are obtained from the dynamical equation

$30$0

with the normalizations and eigenvalue relations given in Eqs. 4–5 of that paper (Poncé et al., 2013). Using the same DFT settings as for the ground state, ABINIT reproduced high-symmetry phonon frequencies at $30$1, $30$2, and $30$3 to within $30$4 of QE/Yambo after enforcing the acoustic sum rule at $30$5. The tabulated values are as follows (Poncé et al., 2013).

Point	Mode	ABINIT
$30$6	$30$7 optical	$30$8
$30$9	$6\times 6\times 6$0 transverse	$6\times 6\times 6$1
$6\times 6\times 6$2	$6\times 6\times 6$3 transverse	$6\times 6\times 6$4

ABINIT is also used to compute real-space interatomic force constants from reciprocal-space DFPT data. In the phonon-transport study on ZnO and ZnS, DFPT returns the dynamical matrix $6\times 6\times 6$5, Born effective charge tensors $6\times 6\times 6$6, and the electronic dielectric tensor $6\times 6\times 6$7, after which the interatomic force constants are obtained by inverse Fourier transform (Bachmann et al., 2011):

$6\times 6\times 6$8

with

$6\times 6\times 6$9

That work states that ABINIT internally splits $\Gamma$0 into a long-range dipole–dipole part, handled in reciprocal space to account for the $\Gamma$1 decay, and a short-range part truncated in real space; the final real-space IFCs were truncated at a radius of $\Gamma$2, sufficient to converge the ballistic thermal conductance to $\Gamma$3 (Bachmann et al., 2011). The DFT setup there used LDA, norm-conserving Troullier–Martins pseudopotentials, a $\Gamma$4 Hartree cutoff, and $\Gamma$5 or $\Gamma$6 Monkhorst–Pack meshes depending on polytype (Bachmann et al., 2011).

The 2025 survey extends the DFPT picture to flexoelectricity, natural optical activity, phonon angular momentum, and transport within the EPH module, with named tasks such as eph_task 7, 8, 9, and 13, and Anaddb support for phonon angular momentum via PHANGMOM (Verstraete et al., 11 Jul 2025). This suggests a continued expansion of the linear-response infrastructure beyond standard phonons.

4. Electron–phonon coupling and zero-point effects

ABINIT is repeatedly associated with electron–phonon calculations at both DFT and beyond-DFT levels. In the diamond verification study, the first-order electron–phonon matrix element is defined as (Poncé et al., 2013)

$\Gamma$7

The paper emphasizes that the non-arbitrary observable is $\Gamma$8 summed over degenerate bands and phonon branches, denoted “SGKK2.” With $\Gamma$9 band pairs, identical pseudopotentials, and identical grids, the relative discrepancy

$6.652$0

was found to satisfy $6.652$1 for all tested $6.652$2, i.e. below $6.652$3 (Poncé et al., 2013).

The same study uses the Allen–Heine–Cardona expression in the on-the-mass-shell plus adiabatic approximation for zero-point motion renormalization,

$6.652$4

with imaginary broadening $6.652$5, a homogeneous $6.652$6 $6.652$7-grid with $6.652$8 irreducible points, and additional statistical sampling up to $6.652$9 random $E_{\mathrm{tot}}=-11.444116277$0-points in ABINIT using a Sternheimer formulation and an active space of $E_{\mathrm{tot}}=-11.444116277$1 bands versus $E_{\mathrm{tot}}=-11.444116277$2 explicit bands (Poncé et al., 2013). The converged direct band-gap renormalization in diamond due to electron–phonon coupling is reported as $E_{\mathrm{tot}}=-11.444116277$3, and the ABINIT versus QE+Yambo discrepancy in zero-point renormalization is below $E_{\mathrm{tot}}=-11.444116277$4 (Poncé et al., 2013).

The 2024 verification-and-validation paper expands this picture. It states that ABINIT implements the non-adiabatic Allen–Heine–Cardona self-energy $E_{\mathrm{tot}}=-11.444116277$5 in a plane-wave/DFPT framework, splitting it into a dynamical Fan term and a static Debye–Waller term (Poncé et al., 2024):

$E_{\mathrm{tot}}=-11.444116277$6

In the Rayleigh–Schrödinger linearized form,

$E_{\mathrm{tot}}=-11.444116277$7

That work compares ABINIT, Quantum ESPRESSO, EPW, and ZG, and reports “excellent agreement” between software implementing the same formalism, as well as “good agreement” between DFPT and WFPT methods (Poncé et al., 2024). It further states that the Debye–Waller term is momentum dependent and that neglecting $E_{\mathrm{tot}}=-11.444116277$8 leads to mass-enhancement errors exceeding $E_{\mathrm{tot}}=-11.444116277$9, whereas a Luttinger approximation yields $-5.7220581385$0–$-5.7220581385$1 error in $-5.7220581385$2 for semiconductors (Poncé et al., 2024). In the same study, ABINIT and EPW yield electron spectral functions with differences in $-5.7220581385$3 under $-5.7220581385$4 and in $-5.7220581385$5 under $-5.7220581385$6 across a $-5.7220581385$7 window around the band edges at $-5.7220581385$8, with validation to “< 1%” for both real and imaginary parts of $-5.7220581385$9 (Poncé et al., 2024).

A practical many-body workflow is described in the BerkeleyGW–ABINIT–EPW paper. There, ABINIT supplies the DFT and DFPT starting point: ground-state wavefunctions in netCDF (prtwf = 3) and perturbation potentials in netCDF (prtpot = 2), with SCF on an $1\times 10^{-5}$0 grid, DFPT on a $1\times 10^{-5}$1 $1\times 10^{-5}$2-mesh, and outputs such as wfk_coarse.nc, ddb_coarse.nc, and vscf_q,a_*.nc (Li et al., 2023). The wrapper elph_interface then converts these into BerkeleyGW-style inputs, while sympert.x performs gauge-recovering symmetry unfolding (Li et al., 2023). ABINIT is thus positioned as the source of the DFPT perturbation potentials and dynamical matrices in a GW perturbation theory workflow.

5. Interfaces, workflows, and multicode ecosystems

A recurrent characteristic of ABINIT in the provided literature is interoperability. The diamond verification work relies on “two independent implementations: Quantum Espresso + Yambo and ABINIT” (Poncé et al., 2013). The zero-point validation study compares “Abinit, Quantum ESPRESSO, EPW, ZG” and both DFPT and WFPT methodologies (Poncé et al., 2024). The GWPT workflow explicitly combines BerkeleyGW, ABINIT, and EPW, with ABINIT furnishing the DFT/DFPT stage and BerkeleyGW and EPW consuming its outputs (Li et al., 2023). The wavelet-PAW paper integrates a shared pawlib into both ABINIT and BigDFT (Rangel et al., 2016).

The atomistic Green’s-function study provides another example of ABINIT as an upstream provider of data rather than a standalone end-to-end engine. There, ABINIT computes equilibrium structures, dynamical matrices, Born charges, dielectric tensors, and real-space IFCs, which are then converted into harmonic blocks $1\times 10^{-5}$3 for AGF transport (Bachmann et al., 2011). The downstream quantities include the phonon Green’s function,

$1\times 10^{-5}$4

the transmission

$1\times 10^{-5}$5

and the linear-response thermal conductance per unit area

$1\times 10^{-5}$6

The paper characterizes the overall approach as “parameter free” because no empirical force constants or fitted interatomic potentials enter: all interatomic potentials come from LDA + DFPT in ABINIT (Bachmann et al., 2011).

The 2025 survey extends this workflow orientation to community frameworks including AbiPy, Atomate2, AiiDA, and MLACS (Verstraete et al., 11 Jul 2025). It mentions input generators, automated convergence studies, workflow automation, MongoDB provenance tracking, machine-learning-assisted canonical sampling, and broad high-throughput integration (Verstraete et al., 11 Jul 2025). Since the details provided are survey-style, a cautious reading is that ABINIT has become embedded in broader reproducibility and automation stacks rather than being used only via manually authored standalone inputs.

6. Specialized methodologies and domain-specific applications

Several papers document specialized ABINIT capabilities beyond conventional ground-state DFT and phonons.

One such capability is the “alchemical mixing approximation,” described as the pseudopotential-specific implementation of the virtual crystal approximation offered in ABINIT (Scharoch et al., 2014). In that approach the mixed pseudopotential for the cation site is constructed as

$1\times 10^{-5}$7

with the local part, non-local projector strengths, and core-charge quantities mixed linearly (Scharoch et al., 2014). In the study of $1\times 10^{-5}$8, this functionality was used to compute equilibrium geometries, elastic constants from Hellmann–Feynman stresses, pressure derivatives, and band gaps within MBJLDA (Scharoch et al., 2014). The reported composition-dependent structural relations include

$1\times 10^{-5}$9

$|\psi_n\rangle$0

and

$|\psi_n\rangle$1

while the band-gap bowing in ABINIT–VCA is stated to be $|\psi_n\rangle$2 (Scharoch et al., 2014). The authors interpret the overestimated bowing as a recognized VCA artifact linked to the lack of local bond relaxation (Scharoch et al., 2014).

Another specialized development is the self-consistent-field linear-response determination of Hubbard parameters $|\psi_n\rangle$3 and $|\psi_n\rangle$4, together with the lrUJ utility (MacEnulty et al., 2024). In that framework, the scalar responses are

$|\psi_n\rangle$5

for the charge-like perturbation $|\psi_n\rangle$6, and similarly

$|\psi_n\rangle$7

for the spin-antisymmetric perturbation $|\psi_n\rangle$8 (MacEnulty et al., 2024). The resulting definitions are

$|\psi_n\rangle$9

The lrUJ post-processor fits $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$00 and $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$01 using polynomial regression up to degree $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$02, computes RMS errors, extracts slopes at zero, and propagates uncertainties to $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$03 and $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$04 (MacEnulty et al., 2024). The paper highlights pitfalls linked to PAW projector choices (dmatpuopt), SCF mixing schemes, and non-linearity (MacEnulty et al., 2024). A plausible implication is that ABINIT’s DFT+$$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$05 infrastructure is treated not merely as a parameterized correction but as a response-based, in situ methodology.

ABINIT is also used for hybrid-like PAW treatments of correlated subspaces. In the AlN doping study, the exchange–correlation functional is PBE0, with exact exchange applied only to selected $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$06 orbitals through keywords such as ixc 11, hfcalc 1, hfdiag 2, and orbpawu "d" (Syrotyuk et al., 2013). That work reported direct spin-down band gaps of $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$07, $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$08, and $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$09 for Cr, Mn, and Fe doping, indirect spin-up gaps of $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$10, $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$11, and $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$12, and ferromagnetic ordering with local moments $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$13, $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$14, and $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$15 on Cr, Mn, and Fe, respectively (Syrotyuk et al., 2013).

7. Verification, performance, and evolving hardware

Verification is a prominent theme in the ABINIT literature. The 2013 comparison with QE/Yambo is explicitly framed as a “rigorous and careful study” of the quantities entering zero-point renormalization in diamond, from the total energy through phonon frequencies and electron–phonon matrix elements to the final renormalization (Poncé et al., 2013). The paper reports discrepancies below $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$16 Ha/atom for the total energy, below $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$17 for high-symmetry phonon frequencies, below $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$18 for the squared electron–phonon matrix elements, and below $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$19 for the zero-point renormalization of each eigenenergy (Poncé et al., 2013). It attributes the residual differences to negligible differences in force-constant and self-consistent-potential routines and characterizes ABINIT as highly accurate and robust for electron–phonon and vibrational properties (Poncé et al., 2013).

The 2024 validation study generalizes this cross-code perspective to four codes and three methods, again concluding “excellent agreement” where the same formalism is implemented and emphasizing the physical importance of the momentum dependence of the Debye–Waller term (Poncé et al., 2024). This directly addresses a possible misconception that Debye–Waller effects are merely rigid shifts with no $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$20-dependence.

Performance on modern hardware appears in the 2026 GPU paper and the 2025 survey. The GPU study formulates the plane-wave Kohn–Sham equations as

$$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$21

with the plane-wave expansion

$$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$22

It highlights GPU-efficient implementations of FFTs via cuFFT or rocFFT, dense linear algebra via cuBLAS or rocBLAS, and eigensolvers via cuSOLVER, rocSOLVER, or ELPA, together with two diagonalization strategies: Locally Optimal Block Preconditioned Conjugate Gradient and Chebyshev polynomial filtering (Lygatsika et al., 13 Apr 2026). The study states that Chebyshev filtering uses only two all-to-all communications per SCF inner iteration, versus $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$23 for LOBPCG, and that its arithmetic intensity scales with the polynomial degree $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$24, making it highly compute-bound on GPUs (Lygatsika et al., 13 Apr 2026).

For a 255-atom Ti interface with 4096 bands and 10 SCF steps, the reported speedups are approximately $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$25 for the filtering step, $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$26 for the Rayleigh–Ritz step, and $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$27 overall on NVIDIA hardware, with 2 GPU nodes comparable to 8 CPU nodes in wall time; on AMD hardware the corresponding numbers are approximately $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$28, $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$29, and $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$30 (Lygatsika et al., 13 Apr 2026). The paper also reports that one GPU node uses about $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$31 of the energy of one CPU node on NVIDIA systems and about $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$32–$$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$33 on AMD systems (Lygatsika et al., 13 Apr 2026). The 2025 overview likewise presents GPU offloading via OpenMP 5.0 and, for a 1280-atom $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$34 cell with 6144 bands, reports LOBPCG on 16 CPUs at $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$35 s, LOBPCG on 4 GPUs at $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$36 s, and Chebyshev filtering on 4 GPUs at $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$37 s (Verstraete et al., 11 Jul 2025).

These studies indicate that the modern trajectory of ABINIT includes algorithmic redesign around batched FFTs, dense BLAS kernels, GPU-aware MPI, band parallelization for DFPT, and reductions in Rayleigh–Ritz bottlenecks (Verstraete et al., 11 Jul 2025, Lygatsika et al., 13 Apr 2026). This suggests that ABINIT’s identity as a first-principles code increasingly depends not only on formal physical content but also on scalable numerical linear algebra and distributed-memory layout design.

8. Scientific position and recurring themes

Across the papers considered, ABINIT occupies three distinct but connected roles. First, it is a production engine for first-principles observables, including total energies, structural relaxations, phonons, electron–phonon couplings, transport inputs, band structures, and magnetic properties [(Bachmann et al., 2011); (Syrotyuk et al., 2013); (Poncé et al., 2013)]. Second, it is a testbed for new methodological developments, such as wavelet-PAW integration, linear-response $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$38 and $$E_{1\mathrm{el}} = \frac{1}{N_k}\sum_{k,n\in\mathrm{occ}}\langle nk|\hat T+\hat V_{\mathrm{psp}}|nk\rangle + E_{\mathrm{psp-core}}.$$39, long-range dipole-plus-quadrupole corrections in electron–phonon interpolation, and constrained-charge or finite-temperature schemes (Rangel et al., 2016, MacEnulty et al., 2024, Poncé et al., 2024, Verstraete et al., 11 Jul 2025). Third, it functions as a hub in multicode workflows, passing standardized outputs into transport, GW, GWPT, and Wannier-interpolation pipelines [(Bachmann et al., 2011); (Li et al., 2023)].

A recurring theme is verification by independent implementation. The diamond studies do not treat agreement as incidental; they use it as evidence that “advanced theoretical formalisms” are correctly realized numerically [(Poncé et al., 2013); (Poncé et al., 2024)]. Another recurring theme is controlled approximation: rigid-ion approximations, active-subspace truncations, VCA mixing, polynomial-response fitting, and special-displacement approaches are all explicitly scrutinized rather than merely adopted [(Scharoch et al., 2014); (MacEnulty et al., 2024); (Poncé et al., 2024)]. Finally, the 2025 and 2026 papers place hardware evolution and workflow automation alongside physics developments, indicating that ABINIT’s contemporary scope includes both predictive modeling and the reproducibility infrastructure needed for large campaigns (Verstraete et al., 11 Jul 2025, Lygatsika et al., 13 Apr 2026).

Taken together, these works depict ABINIT as a first-principles platform whose distinguishing features are methodological breadth, DFPT-centered response capabilities, interoperability with external many-body and workflow tools, and an increasingly explicit commitment to code verification, validation, and hardware-aware performance engineering [(Poncé et al., 2013); (Poncé et al., 2024); (Verstraete et al., 11 Jul 2025)].