GPUMD-v4.0

GPUMD-v4.0 release

• The nep executable

  • Extended the upper limit of l_max from 4 to 8. #754
  • Enabled outputting per-atom and per-structure descriptors. #832 #837
  • Added the option to write energy_weight in train.xyz. #877
  • Enabled training TNEP models with atomic reference values. #907
  • Published the NEP89 model, a universal NEP model for both inorganic and organic materials across 89 elements. #1000
  • Added the fine_tune keyword in nep.in to allow for finetuning the NEP89 model. #932 #959

• The gpumd executable

  • Added the ILP+SW potential. #761
  • Added the ILP+NEP potential. #835 #860 #951
  • Enabled box optimization for minimize. #841
  • Implemented compute_adf for calculating angular distribution function. #875
  • Added a keyword compute_angular_rdf to calculate angular dependent RDF. #884
  • Added a keyword compute_extrapolation to calculate the extrapolation level of NEP during MD run. #936
  • Added the dump_xyz keyword, which is more general than the dump_exyz keyword. #976
  • Added interface to the DP model. #827 #981
  • Implemented the general EAM/Alloy potential. #994
  • Implemented free energy calculation method for liquids using thermodynamic integration. #999

• Common parts

  • Enabled HIP compiling. #768 #822

• Breaking changes:

  • Changed to always output the 9-component box in thermo.out. #854
  • Changed to output 9 instead of 3 virial components using the compute keyword. #876
  • Changed to calculate the stress tensor using instant velocities instead of target temperature. #946

GPUMD-v3.9.5

• Deprecated:

  • [GPUMD and NEP] NEP2 is not supported any more. Now there are only NEP3 and NEP4. #689 #687

• New or enhanced features:

  • [NEP] Introduced use_typewise_cutoff and use_typewise_cutoff_zbl keywords for NEP training, which could lead to more flexible cutoff schemes and improved performance. #660 #684
  • [GPUMD] Allow for creating multiple xyz files using dump_exyz during a run. #629
  • [GPUMD] Output the total electron stopping power loss during the run with electron_stop. #634
  • [GPUMD] Add electric field (add_efield) based on given fixed charges. #662
  • [GPUMD] Add force (add_force) to a group of atoms. #658
  • [GPUMD] When velocity is missing from run.in, now the code initializes the velocities with a temperature of 300 K, instead of showing the previous "undefined" behavior.
  • [GPUMD] correct_velocity can now be used to correct the velocities for different groups of atoms. #668
  • Better error reporting when reading training/test datasets. #699 #697

• Bug fixes:

  • [GPUMD] For the compute_rdf command, stop calculation when any periodic thickness is smaller than 2.5 times of the RDF cutoff. #649
  • [GPUMD] Fix a bug introduced after adding dipole/polarizability prediction in GPUMD-v3.9.2. The NEP energy was wrongly calculated when multi-GPU is used for MD runs. #661

GPUMD-v3.9.4

• Improved some hyperparameters for NEP4 training, which should make the convergence faster.
• Published the latest version of the UNEP-v1 model for 16 metals and their alloys (https://arxiv.org/abs/2311.04732)

GPUMD-v3.9.3

• Introduced the ensemble ti_rs keyword to enable nonequilibrium free-energy calculations along an isobaric path.
• Introduced the ensemble ti_as keyword to enable nonequilibrium free-energy calculations along an isothermal path.
• Improved automation for computing absolute Gibbs free energy with ensemble ti_spring, eliminating the need for post-processing.
• Improved automation for calculating spring constants with ensemble ti_spring.
• Fixed a memory bug introduced in GPUMD-v3.9.2 for compute_shc.
• Improved the all-group option for compute_shc.

GPUMD-v3.9.2

• New features

  • ensemble ti: This keyword is used to set up an equilibrium thermodynamic integration integrator. It is for testing purpose and its only differece from ti_spring keyword is that the lambda value is fixed instead of changing.
  • ensemble nphug: This keyword sets up a Hugoniot thermostat integrator.
  • ensemble mirror: This keyword is employed to configure a momentum mirror shock wave simulation, where atoms are deflected by a moving momentum mirror to generate a shock wave.
  • ensemble piston: This keyword is employed to configure a piston shock wave simulation, where a fixed wall of atoms is displaced at a specified velocity to generate a shock wave.
  • dump_piston: Piston simulations commonly involve millions of atoms. Dumping all the virial and velocity data for each atom can lead to excessively large output files, making data processing cumbersome. The dump_piston command addresses this by calculating spatial thermo information during the simulation.
  • dump_dipole: Predicts the dipole on the fly using a TNEP model, see https://doi.org/10.1021/acs.jctc.3c01343
  • dump_polarizability: Predicts the polarizability on the fly using a TNEP model, see https://doi.org/10.1021/acs.jctc.3c01343

• Enhancements and changes

  • Changed the default values of basis_size for NEP training from (12, 12) to (8, 8).
  • Improved the default regularization methods.
  • Added stress_train.out and stress_test.out during NEP training
  • Added option for compute_shc to calculate the SHC for all the groups in a grouping method simultaneously.

GPUMD-v3.9.1

• Fixed a few bugs related to MTTK integrator and replicate, see #523 for details
• Improved the documentation

GPUMD-v3.9

• New features:
  • A new keyword electron_stop in run.in to apply electron stopping.
  • A new keyword compte_rdf in run.in to compute the radial distribution function (RDF).
  • A new keyword mc in run.in to perform efficient MCMD simulations in canonical, semi-grand canonical, and variance-constraint semi-grand canonical ensembles with a NEP model (only).
  • A new keyword dftd3 in run.in to add the D3 dispersion correction to a NEP model (only).
  • A new keyword replicate in run.in to replicate the initial model.
  • A set of new ensemble options npt_mttk, nph_mttk, and nvt_mttk.
  • A new keyword ensemble ti_spring for free-energy calculations using the nonequilibrium thermodynamic integration method.
  • A new keyword ensemble msst for multi-scale shock technique (MSST) simulations.
  • A new keyword compute_lsqt to compute the electronic transport properties by coupling MD and linear scaling quantum transport (LSQT).
  • A new option fire for the keyword minimize in run.in.
  • A new option has_potential for the keyword dump_exyz in run.in.

GPUMD-v3.8

• Bugfix:
  • The target temperature in ensemble nvt_lan keeps to be the initial one, not changing linearly from the initial one to the final one.
• New features:
  • Added the model_type keyword in nep.in, which can be set to 0, 1, or 2, corresponding to NEP models for potential, dipole, and polarizability, respectively. Tutorials for training dipole and polarizability NEP models are created.
  • Added path-integral molecular dynamics (PIMD) and related techniques such as ring-polymer molecular dynamics (RPMD) and thermostatted RPMD (TRPMD). The new keywords include ensemble pimd, ensemble rpmd, ensemble trpmd, and dump_beads.
  • Added the homogeneous non-equilibrium molecular dynamics Evans-Cummings (HNEMDEC) mehod for calculating thermal conductivity and related transport coefficients in multicomponent systems. The new keyword is compute_hnemdec and the related output file is onsager.out.
  • Added an option to output velocity in NetCDF trajectory file. See the updated dump_netcdf command.

GPUMD-v3.7

• Bugfix:
  • fixed a bug for the plumed interface #395
  • fixed some bugs for the prediction mode in nep.in. #391 #366
• Enhancement
  • Changed the default of basis_size from 8 8 to 12 12 #402
  • Documented the plumed interface #394
  • Sped up NEP training with fullbatch #388
  • Improved the virial_train.out and virial_test.out file format: they have 12 instead of 2 columns now. #399
  • Warned for too small energy values in the training/test data sets #378
• New features:
  • added an active-learning scheme for NEP training #396
  • Added an option to speed up MD simulations with NEP using tabulated radial functions #392

GPUMD-v3.6

• Bugfix:
  • fixed a bug for the plumed interface when the number of atoms exceeds 1024 #339
  • fixed a filesize mismatch bug #310
• Enhancement
  • Can train NEP models up to 103 elements now #334
  • NEP training now converges much faster than before for many-element systems #361
• New features:
  • added the compute_msd keyword for mean-square displacement calculation #324
  • added the compute_viscosity keyword for viscosity calculation #325
  • added the dump_observer keyword for running MD with the average of an ensemble of NEP models or reporting model uncertainty #332
  • added the lambda_shear keyword in nep.in #320