Download VASP 6.6.0 Vienna Ab initio Simulation 2026

FEATURE

• X-ray absorption spectra (XAS) using the Bethe-Salpeter equation (BSE).
  • Calculate exciton wavefunction for core excitations.
• Constrained random-phase approximation (cRPA).
  • spectral cRPA.
  • multi-centre Coulomb matrix elements.
• Checkpointing of finite-difference calculations (CHECKPOINT_FD when IBRION = 6)
  • Restarting phonon calculations.
  • Splitting phonon calculations.
  • Restarting and splitting calculations of the electron-phonon potential (ELPH_POT_GENERATE).
• Electron-phonon
  • Computing electron-phonon matrix elements using meta-GGAs.
  • Use reciprocal space for contraction when computing electron-phonon matrix elements.
  • Computing the transport properties in CRTA using Wannier interpolation.
• Nuclear magnetic resonance (NMR)
  • Spin-orbit coupling for NMR chemical shieldings (LSOSHIFT = .TRUE.).
  • ZORA scalar-relativistic chemical shieldings (LZORA = .TRUE.).
  • Updated chemical shieldings output in OUTCAR (cf. LNMRLEG = .TRUE.).
  • Print current response (WRT_NMRCUR) and the nucleus-independent chemical shielding (NUCIND).
  • Output to Magres format.
• Bethe-Salpeter equation (BSE)
  • Adds Gaussian smearing option for Lanczos algorithm (IBSE = 3).
• GPU
  • OpenMP offloading for Intel and AMD GPUs: DFT and hybrid functionals (beta).
• Exchange-correlation functionals
  • Short-range EXX within dielectric-dependent range-separated hybrid functionals, enabling RS-DDH.
  • Link to simple DFT-D3 package.
  • Support for LOPTICS = .TRUE. and kinetic-energy-density-dependent meta-GGAs.
• Machine-learning force fields
  • Improved interface for thermodynamic integration (TI).
  • Thermodynamic integration for particle insertion using MLFF and empirical potentials (ML_LEMPPOT, ML_EMPPOT_RCUT, ML_SRPOT_B0, ML_SRPOT_N0, ML_SRPOT_S0,ML_MOPOT_NM, ML_MOPOT_DM, ML_MOPOT_RM, ML_MOPOT_RKM, ML_MOPOT_IJM).
  • Add “delta” mode (ML_MODE = delta): always adding the prediction from a given ML_FF to the ab initio calculation results.
  • Experimental support for GRACE force fields in prediction-only mode (requires VASPml, Tensorflow and cppflow).
• HDF5
  • Write phonon frequencies and eigenvectors to vaspout.h5.
  • HDF5 output for cRPA calculations and GW electron self-energy.
  • Write exciton wavefunction to vaspout.h5.
• Output
  • Write the metaGGA potential μ (WRT_POTENTIAL = xcmu).
  • Write the augmented total (core + valence) pseudo densities (WRT_DENSITY = density gradient laplacian).
  • Write the kinetic energy density (LTAU and TAUCAR file).

IMPROVEMENT

• Electron-phonon
  • List of carrier density ranges in electron-phonon transport calculations (ELPH_SELFEN_CARRIER_DEN_RANGE).
  • Reduced memory consumption in electron-phonon calculations.
• Machine-learning force fields
  • New output in ML_LOGFILE: type-dependent output for various reported errors, species quantities for forces, and normalized errors for energies, forces and stress (NORME lines).
  • Reduced memory footprint for ML_MODE = train with ML_CALGO = 1.
  • Reduced memory footprint at the cost of fitting performance with ML_SAVECMAT = .FALSE. (experimental).
  • Improve default settings: shared memory behavior (ML_NCSHMEM), spilling factor switched always switched on, the calculation will end if the spilling factor becomes critically large (>0.9), and ML_OUTBLOCK = 10.
• Many-body perturbation theory
  • Added support for NKRED in RPA and GW (quartic scaling).
  • Improved RPA forces and EXX-energies in low-scaling GW calculations (LFOCKSTD).
  • Increased number of frequency points for the low-scaling GW (i.e., NOMEGA > 24).
  • Support for fast EXX mode in RPA and GW.
  • Double-counting corrections for RPA forces when using DFT+U (issue 67).
  • Reduced memory usage in BSE calculations
  • Improved support for single precision BSE with Lanczos algorithm.
• Change parameters in semilocal functionals (XCm_Pn).
• Greatly improved stability of noncollinear GGA and metaGGA calculations.
• Allowed the Ewald-cutoff parameter to be changed.
• Implement KSPACING for KPOINTS_OPT file using KSPACING_OPT.
• New CMake build system as an alternative to the traditional makefile.include’s. See Install VASP with CMake for instructions.

BUGFIX

• MD driven by VASPml (inside VASP or LAMMPS) crashed if the number of MPI tasks was larger than the number of atoms.
• The use of the SCPC method was broken for ALGO = A.
• The PLUGINS/LOCAL_POTENTIAL energy contribution was not correctly added to the total energy for ALGO ≠ Normal.
• The eigenvalues of the dynamical matrix written to vaspout.h5 had the wrong sign.
• The use of symmetry was broken in ALGO = TIMEEV.
• The “selfen_carrier_per_cell” and “selfen_carrier_den” variables were swapped in the vaspout.h5 file in the “/results/electron_phonon/electrons/chemical_potential” group.
• For  these INCAR tags, ELPH_SELFEN_CARRIER_DEN, ELPH_SELFEN_CARRIER_PER_CELL, or ELPH_SELFEN_MU in the electron-phonon driver only the first element was written to vasprun.xml or vaspout.h5 file.
• RPA forces for spin-polarized systems were wrong.
• Tamm-Dancoff was not working correctly when VASP was built with ELPA.
• Closing vaspin.h5 in INIT_MPI broke the restart from HDF5 capability.
• NBANDS was reduced if the net magnetic moment is negative.
• Setting the tag ELPH_SELFEN_ENERGY_WINDOW was not updating the values of ELPH_SELFEN_BAND_START and ELPH_SELFEN_BAND_STOP for the computation of the electron-phonon matrix elements.
• Correctly account for external electric field when determining symmetry operations using EFIELD and IDIPOL = 1-3.
• PROCAR_OPT was broken for VASP 6.5.0 and 6.5.1.
• For very dense k-meshes, a crash could occur in the subroutine TETIRR, cf. the forum post (https://www.vasp.at/forum/viewtopic.php?t=19800).
• The ionic CG algorithm (IBRION = 2) for the Brent algorithm determined the bracketing interval improperly.
• Fixed I_CONSTRAINED_M = .TRUE. calculations on GPU, which crashed or produced wrong results.
• Fixed computation of mobilities for linear grids (ELPH_TRANSPORT_DRIVER = 1), the Onsager coefficients from the electron and hole resolved transport function were not computed.
• Fixed incorrect handling of spin channels in electron–phonon matrix element calculations for ISPIN = 2, and optimized potential interpolation to process each spin channel separately.
• The zero-field-splitting (LDMATRIX = .TRUE.) led to randomly wrong results, especially with the GNU compiler.
• Fix integer overflows resulting in errors or NaNs appearing in long MD runs (combinations of large NSW, ML_OUTBLOCK, NBLOCK, KBLOCK values).
• With zero conductivity, NaN appeared for other transport coefficients, now instead they are set to zero.
• Fixed wrong formatting of partial DOS in DOSCAR when f-states are present in noncollinear runs. Previously, f-state contributions were written in a new line instead of using the columns after the d-states.
• Running DFPT for electric fields LEPSILON = .TRUE. or ionic displacements IBRION = 7-8 was inadvertently blocked for deorbitalized metaGGAs.
• Update ionic positions for CONTCAR, XDATCAR, and CHGCAR during interactive mode (IBRION = 11).
• Order of electron eigenvalues if LDIAG = .FALSE. between ionic steps was not conserved (FERDO).
• The exchange and correlation components of the BEEF functional (GGA = BF) were not multiplied by the parameters AGGAX and AGGAC.
• During a geometry relaxation, the new atomic positions and cell parameters were not passed to libMBD, cf. https://www.vasp.at/forum/viewtopic.php?t=20071.
• Fixed deadlock in VCAIMAGES runs caused by inconsistent STOPCAR-check MPI_allreduce calls between MLFF and pure-DFT images.
• If there was an empty PENALTYPOT file or none, the HILLSPOT was not written from scratch.
• MDALGO = 5 was not interacting with the PENALTYPOT or HILLSPOT file.
• `spin = up` and `spin = down` in the corresponding `wannier90.X.win` input files were not specified for spin-polarized calculations.
• Fixed ML_MODE = select when atom types for training structures were switched around.
• The spilling factor was calculated incorrectly in the fast prediction mode if the order of atom species in the current structure differed from that of the force field.
• An issue in VCAIMAGES where MLFF calculations in any of the images failed unless a non-empty ML_FF file was also present in the top directory.
• Sign error in the surface-normal component of the reciprocal-space Ewald force and NaN for large vacuum spacing for 2D truncation (KERNEL_TRUNCATION/IDIMENSIONALITY = 2).
• Fix the output for the electron-phonon renormalization of the gaps for ISPIN = 2.
• Deprecate PLUGINS/MACHINE_LEARNING due to inconsistency between units for VASP and ASE stresses – use PLUGINS/FORCE_AND_STRESS instead.
• Neighbor list for DFT-D3 was determined for a fixed radius (50.2 Å, and 21.167 Å; two-body interaction cutoff, and coordination number cutoff), so VDW_RADIUS and VDW_CNRADIUS did not work correctly for values larger than these.
• ELPH_DRIVER = mels with ISPIN = 2 produced incorrect output due to improper k-point counter handling in the spin-dependent accumulator.
• The POSCAR scaling parameters were not correctly read in from vaspin.h5.
• Fix potential MPI deadlock when running with KPAR > 1 by synchronizing SCF break/abort decisions across all ranks to avoid divergent control flow due to small roundoff differences in energy terms (e.g., Hartree) between KPAR groups.