# Fix styles In LAMMPS, a \"fix\" is any operation that is computed during timestepping that alters some property of the system. Essentially everything that happens during a simulation besides force computation, neighbor list construction, and output, is a \"fix\". This includes time integration (update of coordinates and velocities), force constraints or boundary conditions (SHAKE or walls), and diagnostics (compute a diffusion coefficient). New styles can be created to add new options to LAMMPS. Fix_setforce.cpp is a simple example of setting forces on atoms to prescribed values. There are dozens of fix options already in LAMMPS; choose one as a template that is similar to what you want to implement. Here is a brief description of methods you can define in your new derived class. See fix.h for details. ------------------------- ------------------------------------------------------ setmask determines when the fix is called during the timestep (required) init initialization before a run (optional) init_list store pointer to neighbor list; called by neighbor list code (optional) setup_pre_exchange called before atom exchange in setup (optional) setup_pre_force called before force computation in setup (optional) setup called immediately before the first timestep and after forces are computed (optional) min_setup_pre_force like setup_pre_force, but for minimizations instead of MD runs (optional) min_setup like setup, but for minimizations instead of MD runs (optional) initial_integrate called at very beginning of each timestep (optional) pre_exchange called before atom exchange on re-neighboring steps (optional) pre_neighbor called before neighbor list build (optional) pre_force called before pair & molecular forces are computed (optional) post_force called after pair & molecular forces are computed and communicated (optional) final_integrate called at end of each timestep (optional) end_of_step called at very end of timestep (optional) write_restart dumps fix info to restart file (optional) restart uses info from restart file to re-initialize the fix (optional) grow_arrays allocate memory for atom-based arrays used by fix (optional) copy_arrays copy atom info when an atom migrates to a new processor (optional) pack_exchange store atom\'s data in a buffer (optional) unpack_exchange retrieve atom\'s data from a buffer (optional) pack_restart store atom\'s data for writing to restart file (optional) unpack_restart retrieve atom\'s data from a restart file buffer (optional) size_restart size of atom\'s data (optional) maxsize_restart max size of atom\'s data (optional) setup_pre_force_respa same as setup_pre_force, but for rRESPA (optional) initial_integrate_respa same as initial_integrate, but for rRESPA (optional) post_integrate_respa called after the first half integration step is done in rRESPA (optional) pre_force_respa same as pre_force, but for rRESPA (optional) post_force_respa same as post_force, but for rRESPA (optional) final_integrate_respa same as final_integrate, but for rRESPA (optional) min_pre_force called after pair & molecular forces are computed in minimizer (optional) min_post_force called after pair & molecular forces are computed and communicated in minimizer (optional) min_store store extra data for linesearch based minimization on a LIFO stack (optional) min_pushstore push the minimization LIFO stack one element down (optional) min_popstore pop the minimization LIFO stack one element up (optional) min_clearstore clear minimization LIFO stack (optional) min_step reset or move forward on line search minimization (optional) min_dof report number of degrees of freedom *added* by this fix in minimization (optional) max_alpha report maximum allowed step size during linesearch minimization (optional) pack_comm pack a buffer to communicate a per-atom quantity (optional) unpack_comm unpack a buffer to communicate a per-atom quantity (optional) pack_reverse_comm pack a buffer to reverse communicate a per-atom quantity (optional) unpack_reverse_comm unpack a buffer to reverse communicate a per-atom quantity (optional) dof report number of degrees of freedom *removed* by this fix during MD (optional) compute_scalar return a global scalar property that the fix computes (optional) compute_vector return a component of a vector property that the fix computes (optional) compute_array return a component of an array property that the fix computes (optional) deform called when the box size is changed (optional) reset_target called when a change of the target temperature is requested during a run (optional) reset_dt is called when a change of the time step is requested during a run (optional) modify_param called when a fix_modify request is executed (optional) memory_usage report memory used by fix (optional) thermo compute quantities for thermodynamic output (optional) ------------------------- ------------------------------------------------------ Typically, only a small fraction of these methods are defined for a particular fix. Setmask is mandatory, as it determines when the fix will be invoked during the timestep. Fixes that perform time integration (\*nve\*, *nvt*, *npt*) implement initial_integrate() and final_integrate() to perform velocity Verlet updates. Fixes that constrain forces implement post_force(). Fixes that perform diagnostics typically implement end_of_step(). For an end_of_step fix, one of your fix arguments must be the variable \"nevery\" which is used to determine when to call the fix and you must set this variable in the constructor of your fix. By convention, this is the first argument the fix defines (after the ID, group-ID, style). If the fix needs to store information for each atom that persists from timestep to timestep, it can manage that memory and migrate the info with the atoms as they move from processors to processor by implementing the grow_arrays, copy_arrays, pack_exchange, and unpack_exchange methods. Similarly, the pack_restart and unpack_restart methods can be implemented to store information about the fix in restart files. If you wish an integrator or force constraint fix to work with rRESPA (see the [run_style](run_style) command), the initial_integrate, post_force_integrate, and final_integrate_respa methods can be implemented. The thermo method enables a fix to contribute values to thermodynamic output, as printed quantities and/or to be summed to the potential energy of the system.