Overview
This page gives a description of the input section that controls and modifies how an atomic geometry relaxation is done. The code does a full geometry relaxation via force-elimination, enforcing all user-specified symmetry, and respecting all constraints specified by the user in this section. The geometry relaxation input section is optional; all the user need do is add a "do relax" instruction in the command options and the code will automatically attempt to relax the atomic positions to give zero forces.
The default units for the geometry relaxation section are Ry for energies and bohr for distances.
Current constraints, important version notes:
- "end geometry" keyword required to end this section, new to version 2.53
- "frame " keyword new, added in version 2.54
- "vgfixM" keywords new, added in version 2.62
Input options
The geometry relaxation input section can appear anywhere in the "run phase" section of the input file (as of v2.53), begins with the "geometry relaxation" keyword and ends with the "end geometry" keyword.
The following are the common input keywords in this section. All the keywords are optional, and can appear in any order within this section. The keywords must be left-justified, but all data input is free-format.
- geometry relaxation - begin relaxation input section
- ...
- gsteps - maximum number of geometry steps
- n_step_g
- ghistory - max number of geometry steps to use in blend history
- n_hist_g
- no ges - (*)turns off dynamic SCF guessing, reverts to overlapping spherical atom guess
(*) Moved to run phase data section (2.61) - gconv - force convergence criterion (Ry/bohr)
- force_conv
- gblend - initial update factor: R(2)=R(1)+gblend*F(1)
- gblend
- gmethod - specify relaxation method
- gmethod (BROYDEN, ASD, DMDAT, DAMPED, STEEPEST, ...)
- timestep - time step for dynamical relaxation schemes
- time_step <!-- deprecated, as of 2.67, using label-based selection is better
- only relax - about constraints
- n_atom (first n_atom if n_atom>0, last |n_atom| else) -->
- gfixed - fix positions for sequence of atoms about constraints
- start_atom end_atom (sequence to fix, inclusive)
- grelax - relax positions for sequence of atoms about constraints
- start_atom end_atom (sequence to relax, inclusive)
- frame - constraint: fix atom I, vector I-J, plane I-J-K about constraints
- fixed_atom vec_atom plane_atom
- vgfixM - axis-related constraints for atoms (new in 2.62) about constraints
- num_fixed_atoms x_axis y_axis z_axis (number of atoms bound by this constraint, and Cartesian axis)
fixed_atom(1) fixed_atom(2) ... fixed_atom(num_fixed_atoms) (list of atoms bound by this constraint)
- M = "a": free motion along axis, no motion normal to axis
- M = "d": fix axis-projected distance between listed atoms, free in-plane
- M = "p": no motion along axis vector, free in-plane
- bond - fix bond distance between two atoms (new in 2.68) about constraints
- atom_1 atom_2
- layers | slinky - relax interlayer spacings or atoms (new in 2.68) about constraints
- num_layer_atoms x_axis y_axis z_axis (number of atoms in layer stack, and Cartesian layer direction
atom_top atom_next ... atom_base (list of layer atoms, in order, from top (surface) atom to base (deep) atom
<!-- hold back RMV/SLIC until published - slic1 - local internal coordinate (sphere) projection for 1-fold atom (2.68) about constraints
- atom_1fold atom_bond atom_atom_anchor_
- slic2 - local internal coordinate (cylinder) projection for 2-fold atom (2.68) about constraints
- atom_2fold atom_bond1 atom_atom_bond2_
- slpar1 (expert) adjust slic/sphere scalephi (g-update factor) about constraints
- scalephi_ (default=
- slpar2 (expert) adjust slic/sphere scaletheta (g-update factor) about constraints
- scaletheta_ (default=
- fank0 | fnakf | fankb (expert) force transformation scheme in slic/sphere
- M = "0": use raw force on sphere anchor atom
- M = "f": combine full 1-fold force into sphere anchor atom
- M = "b" (default) : combine bond-project 1-fold force into sphere anchor
- ...
- dynamics - invokes MD input section
- ... MD data input
- end dynamics - end of MD input section
- ...
- end geometry relaxation - end of relaxation input section
Applying constraints
During relaxation, the code will automatically enforce the requirement that the atoms respect the symmetry specified in the setup phase. Other constraints currently available in the code include:
– fix selected atoms in place and allow the rest to totally relax
– fix the frame of reference defined by three atoms
– fix selected atoms using a single type of axis-based constraint It is not avaulable to apply more sophisticated constraints on the system. For example, it is not possible to fix the C-O bond distance for a CO molecule otherwise free to relax when adsorbed on a metal slab.
To select sequences of atoms to be relaxed/fixed-in-place, one can use:
- repeated invocations of gfixed
- repeated invocations of grelax
The gfixed and grelax constraints cumulatively select out sequences of atoms to fix or relax, respectively. These keywords take two numbers as their argument for each invocation. To select one atom, input as a sequence of length one, i.e., enter the atom number twice (e.g. "2 2" to select the second atom). The keywords grelax and gfixed cannot be mixed in one input file!
Atom specification in these various constraints can either be by index within the list (e.g., "1 4" will select the first through fourth atom in a geometry list), or (as of 2.67) by the atom label within the list (e.g., "AT0001 AT0029" will select the atom sequence from the atom labeled "AT0001" through the atom labeled "AT0029", wherever those atom labels appear first in the list).) This allows for more flexible selections of atom sequences that do not required counting. Special names allow further flexibility in atom selection: "FIRST" and "LAST" will specify, naturally, the first atom, and last atom in the geometry list (e.g. "AT0002 LAST" will select the atoms from AT0002 through the last atom inclusive), provided that you have not already named an atom with "FIRST" or "LAST" (in which case the code will use the labeled atoms instead).
Different sets of contraints modify how the relaxations are done: either imposing more sophisticated constraints on the atomic positions, or transforming the coordinate relaxations from simple Cartesian to more effective framces.
The frame keyword can be useful for removing the artificial translational and rotational degrees of freedom in a molecular calculation when doing a NEB calculation (the artificial band forces of the NEB can couple to these spurious degrees of freedom and cause problems). The keyword requires three arguments: fixed_atom vec_atom k_plane_atom. This fixes the position of the first atom listed, fixed_atom, constrains the position of atom vec_atom such that the direction of the vector (R[J]-R[I]) does not change, and constrains the position of atom k_plane_atom such that the plane defined by I-J-K does not change (if I-J-K is desired to be linear, enter K as a negative number).
The vgfixM family of constraints allow for some interesting manipulations of the atom positions in a relaxation. These can be particularly useful in trapping semiconductor defect calculations in reduced symmetry configurations.
The bond constraint does a very simplistic approach to fixing bond length between two atoms
The slinky | layer constraint transforms from a relaxation of atoms using Cartesian positions and forces, into a relaxation of interlayer spacings between atoms. This can be very useful for slab surface relaxations, eliminating the "caterpillar" effect that slows convergence. The layers are indicated by the order of the atoms, be sure to start from top slab layer, to deep layer.
All these constraints must occur after all gfixed|grelax invocations are completed. Multiple constraints can be chosen. The code does check that the sets of atoms affected in multiple invocations of different constraints are completely disjoint, so as to avoid conflicts.
Important : The code does not check whether constraints ( gfixed, grelax, frame, vgfixM, etc ) are compatible with the symmetry specified in the input. If your constraints are incompatible with the symmetry (or with each other), then when the code completes the first force calculation and attempts a geometry update, it will stop the calculation when it discovers the (constrained) update violates symmetry.
Relaxation methods
The default relaxation method is BROYDEN. This is the modified Broyden scheme due to D.D. Johnson, [Phys. Rev. B 38,12807 (1988)]. This works very well when one is within a basin that has well-behaved second derivatives, but fails (badly) outside of its radius of convergence. The damped dynamics scheme originated by Hannes Jonsson is implemented in DAMPED, but is not recommended for use in Quest calculations. The force-coordinate damping leads to orientation-dependent trajectories and trajectories that can (and will) violate symmetry. DMDAT modifies this to use the atomic force-vector, and makes the trajectory independent of system orientation, and thereby conserves symmetry. However, the recommended method for relaxation when BROYDEN fails is ASD. The Accelerated Steepest Descent method (P.A. Schultz, unpublished) is a very robust and remarkably effective geometry minimizer.
Troubleshooting
- The geometry relaxation violated symmetry on the first update step.
- Check that constraints imposed on atomic relaxation are consistent with the symmetry.
- The geometry "blew up" on the first relaxation update step.
- Reduce: gblend, if doing Broyden; timestep, if doing a dynamical relax. The gblend/timestep factors for minimization usually work best when between 0.5 and 6.0, and the smaller these factors, the shorter the first jump.
- During Broyden, my relaxation "blew up" or violated symmetry.
- If you are not in a potential well, the second derivative matrix might be ill-behaved and Broyden will fail. Switch to the ASD method until in a basin (forces getting smaller and energy getting lower).
- Using ASD (or damped dynamics), the geometry update violated symmetry.
- Numerical noise – one atom of symmetrically related set of atoms got damped but others did not. Start geometry relaxation again with the last good geometry.
- Geometry relax will not converge (with Broyden, esp), and is in a well.
- Numerical noise – the convergence criterion might be smaller than the inaccuracy in the computed forces. Either increase force convergence criterion, or use a denser grid (to get more accurate forces).