New and improved features¶

Added support for AWH biasing¶

The AWH (Accelerated Weight Histogram) method is an adaptive biasing method used for overcoming free energy barriers and calculating free energies (see https://doi.org/10.1063/1.4890371). Although AWH can in general bias any system parameter, this change only implements biasing of reaction coordinates. The actual force distribution and coordinate handling is taken care of by the pull code. AWH interacts with the pull code by registering itself as the external potential module for the coordinate that should be AWH biased. The AWH code sets the potential and force for those coordinates. See the reference manual for full documentation.

It includes a feature to compute the time-integrated force correlation, also known as the friction tensor (see e.g. https://doi.org/10.1103/PhysRevLett.108.190602). The friction tensor defines a metric on the coordinate space and the local volume element of this metric is a useful measure for determining which regions need more or less sampling.

Dual pair-list buffer with dynamic pruning¶

The GROMACS simulation engine uses a new dual pair-list algorithm with dynamic pruning in cases where the Verlet buffer is determined automatically (which is the default). This allows further reducing the frequency of pair search (and domain decomposition) while avoiding large Verlet buffers and the previously inherent increased computational cost in the short-ranged nonbonded kernels. This is achieved by constructing an “outer” pair-list built infrequently, which includes many pairs in the list that are outside the cut-off range for most of the lifetime of the list. Such pairs can be pruned out very efficiently every few steps and with that building a smaller, “inner” pair-list with a shorter life-time, and importantly a correspondingly shorter Verlet buffer (still adhering to the specified tolerance), which is then used in the nonbonded kernels. Thanks to this, simulations runs are significantly less sensitive to tuning the search frequency parameter (“nstlist”). When short-ranged interactions are running on the GPU, the dynamic pruning is overlapped with the integration on the CPU, so is usually free. This feature improves all of simulation rate, hardware utilization, and power consumption.

Added physical validation suite¶

These tests run series of short simulations and verify the expected statistical properties, e.g. of energy distributions between the simulations, as a sensitive test that the code correctly samples the expected ensemble.

To run everything locally (which can take a few hours!) use

cmake -DGMX_PHYSICAL_VALIATION=ON ..
make
make check-phys-run

Currently, the script is only running a few systems, checking convergence of energy conservation in NVE system with decaying timestep, and the ensembles generated by a few thermostating and barostating algorithms. Other systems and ensembles covering a broader combination of settings will be added over time.

Added reporting of conserved quantities for coupling algorithms¶

The work that some more coupling algorithms (Berendsen pressure, Berendsen temperature, and Parrinello-Rahman pressure) applies on the system is calculated and integrated. Formulae are in the reference manual.

Added acceleration correction VCM mode¶

New mdp option to remove the center of mass translational velocity and correct the center of mass position, assuming linear acceleration. This is useful e.g. for pulling on a group using an absolute reference.

Changed handling of gmx mdrun -gpu_id¶

As more code is able to be offloaded to the GPU, task assignment has become more complex, and is likely to get more complex still. The -gpu_id command-line option now merely enables the user to restrict which of the detected GPUs are available to the automated task assignment scheme, somewhat like the CUDA_VISIBLE_DEVICES environment variable. For the rare cases where full control is needed, gmx mdrun -gputasks is available and documented in the user guide.

Added log output for equivalent 1x1 pair-list setup¶

The GROMACS NxM setup can use a shorter pair-list buffer than other codes’ simpler 1x1 scheme, so our log files now report the equivalent setup, to help people doing performance and correctness comparisons.

New mdp input for electric fields¶

EW3DC for non-neutral systems¶

Added the corrections to force and energy according to Ballenegger, Arnold, and Cerda, J. Chem. Phys. 131, 094107 2009 (https://doi.org/10.1063/1.3216473). Hinted that people read https://doi.org/10.1021/ct400626b to help make good choices. Added a warning in grompp for charged systems with Ewald.

Reduce rounding errors in SETTLE¶

The parameters for SETTLE are now computed in double precision, which lowers the systematic error.

Made `gmx mdrun -pforce` terminate with non-finite forces¶

The gmx mdrun option -pforce used to print non-finite forces, but would not terminate the run if any were found. Now a fatal error is issued.