Molly.jl Files

The package is rewritten to use Enzyme.jl rather than Zygote.jl for differentiable simulations. This allows mutating code to be used, leading to better performance on CPU and GPU. In addition, faster CUDA kernels are added and support for other GPU backends is added via KernelAbstractions.jl.

Breaking changes

• CoordinateLogger, VelocityLogger and ForceLogger are renamed to CoordinatesLogger, VelocitiesLogger and ForcesLogger respectively for consistency.
• During system setup the gpu keyword argument is replaced by array_type, with CuArray giving the same behaviour as before. Use of other array types such as ROCArray is supported. The use_cell_list keyword argument is replaced with neighbor_finder_type, with the default depending on the backend.
• The second type parameter to System is changed from a Bool indicating whether the system is on the GPU to the array type of the system.
• Giving the force_units and energy_units is no longer required when constructing pairwise interactions such as LennardJones. In addition, mixing functions and shortcut functions can now be given to determine how to combine atomic parameters and when to skip the interaction respectively.
• The arguments to force and potential_energy for pairwise and specific interactions are changed to allow more flexible interaction definitions, depending for example on the velocities or step number.
• Differentiable simulations are no longer compatible with Zygote.jl, but can now be carried out using Enzyme.jl. This is faster and more memory-efficient.
• The solute field to Atom is removed and the atom_type field is added. weight_solute_solvent is no longer an option for LennardJones, though similar functionality is available with atom_type and mixing functions.
• box_volume is renamed to volume.
• run_loggers! is renamed to apply_loggers! to avoid confusion with the run_loggers function argument.
• The abstract types PairwiseInteraction and SpecificInteraction are removed. Custom interactions no longer need to sub-type anything.
• Extensions are added for code requiring Enzyme.jl, CUDA.jl and KernelDensity.jl, so relevant imports are now required before using those features.

Non-breaking changes

• Support for Julia versions before 1.10 is dropped.

New features

• Backends supported by KernelAbstractions.jl can be used to run simulations by using the appropriate array type or the array_type keyword argument during system setup.
• The Berendsen pressure coupling method is added as BerendsenBarostat.
• The density and dipole_moment functions are added.
• The AshbaughHatch and Yukawa potentials are added.
• VolumeLogger and DensityLogger are added.
• The rng keyword argument is added to place_atoms, place_diatomics and random_coord, allowing reproducible coordinate generation.
• The array_type function is added, giving the array type of a system or array.
• The step number can be given as a third argument with a default of 0 to forces, potential_energy and virial acting on a system. This allows time-dependent forces, with the caveat that the step number resets to 1 every time simulate! is called and can also be 0 to calculate forces before the first step.
• A method is added to random_velocities! which takes in and modifies the velocities as the first argument.

Performance improvements

• Better CUDA force and potential energy kernels significantly improve performance on CUDA compatible devices. The GPUNeighborFinder is added for this path but does not calculate neighbors, which are calculated each step in the kernels.
• Simulations are faster on both CPU and GPU due to mutating code leading to fewer memory allocations.
• System setup from a file is made much faster.
• The package is faster to load as a number of dependencies have been removed.
• The performance of implicit solvent on CPU is improved.

Bug fixes

• A bug when using more than two interactions on the GPU is fixed.
• A type conversion bug in CellListMapNeighborFinder construction is fixed.